xref: /openbmc/qemu/tcg/i386/tcg-target.c.inc (revision 3a8c4e9e53c6f4aa7c590971950000b174e74fa1)
1/*
2 * Tiny Code Generator for QEMU
3 *
4 * Copyright (c) 2008 Fabrice Bellard
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to deal
8 * in the Software without restriction, including without limitation the rights
9 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
10 * copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
22 * THE SOFTWARE.
23 */
24
25/* Used for function call generation. */
26#define TCG_TARGET_STACK_ALIGN 16
27#if defined(_WIN64)
28#define TCG_TARGET_CALL_STACK_OFFSET 32
29#else
30#define TCG_TARGET_CALL_STACK_OFFSET 0
31#endif
32#define TCG_TARGET_CALL_ARG_I32      TCG_CALL_ARG_NORMAL
33#define TCG_TARGET_CALL_ARG_I64      TCG_CALL_ARG_NORMAL
34#if defined(_WIN64)
35# define TCG_TARGET_CALL_ARG_I128    TCG_CALL_ARG_BY_REF
36# define TCG_TARGET_CALL_RET_I128    TCG_CALL_RET_BY_VEC
37#elif TCG_TARGET_REG_BITS == 64
38# define TCG_TARGET_CALL_ARG_I128    TCG_CALL_ARG_NORMAL
39# define TCG_TARGET_CALL_RET_I128    TCG_CALL_RET_NORMAL
40#else
41# define TCG_TARGET_CALL_ARG_I128    TCG_CALL_ARG_NORMAL
42# define TCG_TARGET_CALL_RET_I128    TCG_CALL_RET_BY_REF
43#endif
44
45#ifdef CONFIG_DEBUG_TCG
46static const char * const tcg_target_reg_names[TCG_TARGET_NB_REGS] = {
47#if TCG_TARGET_REG_BITS == 64
48    "%rax", "%rcx", "%rdx", "%rbx", "%rsp", "%rbp", "%rsi", "%rdi",
49#else
50    "%eax", "%ecx", "%edx", "%ebx", "%esp", "%ebp", "%esi", "%edi",
51#endif
52    "%r8",  "%r9",  "%r10", "%r11", "%r12", "%r13", "%r14", "%r15",
53    "%xmm0", "%xmm1", "%xmm2", "%xmm3", "%xmm4", "%xmm5", "%xmm6", "%xmm7",
54#if TCG_TARGET_REG_BITS == 64
55    "%xmm8", "%xmm9", "%xmm10", "%xmm11",
56    "%xmm12", "%xmm13", "%xmm14", "%xmm15",
57#endif
58};
59#endif
60
61static const int tcg_target_reg_alloc_order[] = {
62#if TCG_TARGET_REG_BITS == 64
63    TCG_REG_RBP,
64    TCG_REG_RBX,
65    TCG_REG_R12,
66    TCG_REG_R13,
67    TCG_REG_R14,
68    TCG_REG_R15,
69    TCG_REG_R10,
70    TCG_REG_R11,
71    TCG_REG_R9,
72    TCG_REG_R8,
73    TCG_REG_RCX,
74    TCG_REG_RDX,
75    TCG_REG_RSI,
76    TCG_REG_RDI,
77    TCG_REG_RAX,
78#else
79    TCG_REG_EBX,
80    TCG_REG_ESI,
81    TCG_REG_EDI,
82    TCG_REG_EBP,
83    TCG_REG_ECX,
84    TCG_REG_EDX,
85    TCG_REG_EAX,
86#endif
87    TCG_REG_XMM0,
88    TCG_REG_XMM1,
89    TCG_REG_XMM2,
90    TCG_REG_XMM3,
91    TCG_REG_XMM4,
92    TCG_REG_XMM5,
93#ifndef _WIN64
94    /* The Win64 ABI has xmm6-xmm15 as caller-saves, and we do not save
95       any of them.  Therefore only allow xmm0-xmm5 to be allocated.  */
96    TCG_REG_XMM6,
97    TCG_REG_XMM7,
98#if TCG_TARGET_REG_BITS == 64
99    TCG_REG_XMM8,
100    TCG_REG_XMM9,
101    TCG_REG_XMM10,
102    TCG_REG_XMM11,
103    TCG_REG_XMM12,
104    TCG_REG_XMM13,
105    TCG_REG_XMM14,
106    TCG_REG_XMM15,
107#endif
108#endif
109};
110
111#define TCG_TMP_VEC  TCG_REG_XMM5
112
113static const int tcg_target_call_iarg_regs[] = {
114#if TCG_TARGET_REG_BITS == 64
115#if defined(_WIN64)
116    TCG_REG_RCX,
117    TCG_REG_RDX,
118#else
119    TCG_REG_RDI,
120    TCG_REG_RSI,
121    TCG_REG_RDX,
122    TCG_REG_RCX,
123#endif
124    TCG_REG_R8,
125    TCG_REG_R9,
126#else
127    /* 32 bit mode uses stack based calling convention (GCC default). */
128#endif
129};
130
131static TCGReg tcg_target_call_oarg_reg(TCGCallReturnKind kind, int slot)
132{
133    switch (kind) {
134    case TCG_CALL_RET_NORMAL:
135        tcg_debug_assert(slot >= 0 && slot <= 1);
136        return slot ? TCG_REG_EDX : TCG_REG_EAX;
137#ifdef _WIN64
138    case TCG_CALL_RET_BY_VEC:
139        tcg_debug_assert(slot == 0);
140        return TCG_REG_XMM0;
141#endif
142    default:
143        g_assert_not_reached();
144    }
145}
146
147/* Constants we accept.  */
148#define TCG_CT_CONST_S32 0x100
149#define TCG_CT_CONST_U32 0x200
150#define TCG_CT_CONST_I32 0x400
151#define TCG_CT_CONST_WSZ 0x800
152#define TCG_CT_CONST_TST 0x1000
153#define TCG_CT_CONST_ZERO 0x2000
154
155/* Registers used with L constraint, which are the first argument
156   registers on x86_64, and two random call clobbered registers on
157   i386. */
158#if TCG_TARGET_REG_BITS == 64
159# define TCG_REG_L0 tcg_target_call_iarg_regs[0]
160# define TCG_REG_L1 tcg_target_call_iarg_regs[1]
161#else
162# define TCG_REG_L0 TCG_REG_EAX
163# define TCG_REG_L1 TCG_REG_EDX
164#endif
165
166#if TCG_TARGET_REG_BITS == 64
167# define ALL_GENERAL_REGS      0x0000ffffu
168# define ALL_VECTOR_REGS       0xffff0000u
169# define ALL_BYTEL_REGS        ALL_GENERAL_REGS
170#else
171# define ALL_GENERAL_REGS      0x000000ffu
172# define ALL_VECTOR_REGS       0x00ff0000u
173# define ALL_BYTEL_REGS        0x0000000fu
174#endif
175#define SOFTMMU_RESERVE_REGS \
176    (tcg_use_softmmu ? (1 << TCG_REG_L0) | (1 << TCG_REG_L1) : 0)
177
178#define have_bmi2       (cpuinfo & CPUINFO_BMI2)
179#define have_lzcnt      (cpuinfo & CPUINFO_LZCNT)
180
181static const tcg_insn_unit *tb_ret_addr;
182
183static bool patch_reloc(tcg_insn_unit *code_ptr, int type,
184                        intptr_t value, intptr_t addend)
185{
186    value += addend;
187    switch(type) {
188    case R_386_PC32:
189        value -= (uintptr_t)tcg_splitwx_to_rx(code_ptr);
190        if (value != (int32_t)value) {
191            return false;
192        }
193        /* FALLTHRU */
194    case R_386_32:
195        tcg_patch32(code_ptr, value);
196        break;
197    case R_386_PC8:
198        value -= (uintptr_t)tcg_splitwx_to_rx(code_ptr);
199        if (value != (int8_t)value) {
200            return false;
201        }
202        tcg_patch8(code_ptr, value);
203        break;
204    default:
205        g_assert_not_reached();
206    }
207    return true;
208}
209
210/* test if a constant matches the constraint */
211static bool tcg_target_const_match(int64_t val, int ct,
212                                   TCGType type, TCGCond cond, int vece)
213{
214    if (ct & TCG_CT_CONST) {
215        return 1;
216    }
217    if (type == TCG_TYPE_I32) {
218        if (ct & (TCG_CT_CONST_S32 | TCG_CT_CONST_U32 |
219                  TCG_CT_CONST_I32 | TCG_CT_CONST_TST)) {
220            return 1;
221        }
222    } else {
223        if ((ct & TCG_CT_CONST_S32) && val == (int32_t)val) {
224            return 1;
225        }
226        if ((ct & TCG_CT_CONST_U32) && val == (uint32_t)val) {
227            return 1;
228        }
229        if ((ct & TCG_CT_CONST_I32) && ~val == (int32_t)~val) {
230            return 1;
231        }
232        /*
233         * This will be used in combination with TCG_CT_CONST_S32,
234         * so "normal" TESTQ is already matched.  Also accept:
235         *    TESTQ -> TESTL   (uint32_t)
236         *    TESTQ -> BT      (is_power_of_2)
237         */
238        if ((ct & TCG_CT_CONST_TST)
239            && is_tst_cond(cond)
240            && (val == (uint32_t)val || is_power_of_2(val))) {
241            return 1;
242        }
243    }
244    if ((ct & TCG_CT_CONST_WSZ) && val == (type == TCG_TYPE_I32 ? 32 : 64)) {
245        return 1;
246    }
247    if ((ct & TCG_CT_CONST_ZERO) && val == 0) {
248        return 1;
249    }
250    return 0;
251}
252
253# define LOWREGMASK(x)	((x) & 7)
254
255#define P_EXT		0x100		/* 0x0f opcode prefix */
256#define P_EXT38         0x200           /* 0x0f 0x38 opcode prefix */
257#define P_DATA16        0x400           /* 0x66 opcode prefix */
258#define P_VEXW          0x1000          /* Set VEX.W = 1 */
259#if TCG_TARGET_REG_BITS == 64
260# define P_REXW         P_VEXW          /* Set REX.W = 1; match VEXW */
261# define P_REXB_R       0x2000          /* REG field as byte register */
262# define P_REXB_RM      0x4000          /* R/M field as byte register */
263# define P_GS           0x8000          /* gs segment override */
264#else
265# define P_REXW		0
266# define P_REXB_R	0
267# define P_REXB_RM	0
268# define P_GS           0
269#endif
270#define P_EXT3A         0x10000         /* 0x0f 0x3a opcode prefix */
271#define P_SIMDF3        0x20000         /* 0xf3 opcode prefix */
272#define P_SIMDF2        0x40000         /* 0xf2 opcode prefix */
273#define P_VEXL          0x80000         /* Set VEX.L = 1 */
274#define P_EVEX          0x100000        /* Requires EVEX encoding */
275
276#define OPC_ARITH_EbIb	(0x80)
277#define OPC_ARITH_EvIz	(0x81)
278#define OPC_ARITH_EvIb	(0x83)
279#define OPC_ARITH_GvEv	(0x03)		/* ... plus (ARITH_FOO << 3) */
280#define OPC_ANDN        (0xf2 | P_EXT38)
281#define OPC_ADD_GvEv	(OPC_ARITH_GvEv | (ARITH_ADD << 3))
282#define OPC_AND_GvEv    (OPC_ARITH_GvEv | (ARITH_AND << 3))
283#define OPC_BLENDPS     (0x0c | P_EXT3A | P_DATA16)
284#define OPC_BSF         (0xbc | P_EXT)
285#define OPC_BSR         (0xbd | P_EXT)
286#define OPC_BSWAP	(0xc8 | P_EXT)
287#define OPC_CALL_Jz	(0xe8)
288#define OPC_CMOVCC      (0x40 | P_EXT)  /* ... plus condition code */
289#define OPC_CMP_GvEv	(OPC_ARITH_GvEv | (ARITH_CMP << 3))
290#define OPC_DEC_r32	(0x48)
291#define OPC_IMUL_GvEv	(0xaf | P_EXT)
292#define OPC_IMUL_GvEvIb	(0x6b)
293#define OPC_IMUL_GvEvIz	(0x69)
294#define OPC_INC_r32	(0x40)
295#define OPC_JCC_long	(0x80 | P_EXT)	/* ... plus condition code */
296#define OPC_JCC_short	(0x70)		/* ... plus condition code */
297#define OPC_JMP_long	(0xe9)
298#define OPC_JMP_short	(0xeb)
299#define OPC_LEA         (0x8d)
300#define OPC_LZCNT       (0xbd | P_EXT | P_SIMDF3)
301#define OPC_MOVB_EvGv	(0x88)		/* stores, more or less */
302#define OPC_MOVL_EvGv	(0x89)		/* stores, more or less */
303#define OPC_MOVL_GvEv	(0x8b)		/* loads, more or less */
304#define OPC_MOVB_EvIz   (0xc6)
305#define OPC_MOVL_EvIz	(0xc7)
306#define OPC_MOVB_Ib     (0xb0)
307#define OPC_MOVL_Iv     (0xb8)
308#define OPC_MOVBE_GyMy  (0xf0 | P_EXT38)
309#define OPC_MOVBE_MyGy  (0xf1 | P_EXT38)
310#define OPC_MOVD_VyEy   (0x6e | P_EXT | P_DATA16)
311#define OPC_MOVD_EyVy   (0x7e | P_EXT | P_DATA16)
312#define OPC_MOVDDUP     (0x12 | P_EXT | P_SIMDF2)
313#define OPC_MOVDQA_VxWx (0x6f | P_EXT | P_DATA16)
314#define OPC_MOVDQA_WxVx (0x7f | P_EXT | P_DATA16)
315#define OPC_MOVDQU_VxWx (0x6f | P_EXT | P_SIMDF3)
316#define OPC_MOVDQU_WxVx (0x7f | P_EXT | P_SIMDF3)
317#define OPC_MOVQ_VqWq   (0x7e | P_EXT | P_SIMDF3)
318#define OPC_MOVQ_WqVq   (0xd6 | P_EXT | P_DATA16)
319#define OPC_MOVSBL	(0xbe | P_EXT)
320#define OPC_MOVSWL	(0xbf | P_EXT)
321#define OPC_MOVSLQ	(0x63 | P_REXW)
322#define OPC_MOVZBL	(0xb6 | P_EXT)
323#define OPC_MOVZWL	(0xb7 | P_EXT)
324#define OPC_PABSB       (0x1c | P_EXT38 | P_DATA16)
325#define OPC_PABSW       (0x1d | P_EXT38 | P_DATA16)
326#define OPC_PABSD       (0x1e | P_EXT38 | P_DATA16)
327#define OPC_VPABSQ      (0x1f | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
328#define OPC_PACKSSDW    (0x6b | P_EXT | P_DATA16)
329#define OPC_PACKSSWB    (0x63 | P_EXT | P_DATA16)
330#define OPC_PACKUSDW    (0x2b | P_EXT38 | P_DATA16)
331#define OPC_PACKUSWB    (0x67 | P_EXT | P_DATA16)
332#define OPC_PADDB       (0xfc | P_EXT | P_DATA16)
333#define OPC_PADDW       (0xfd | P_EXT | P_DATA16)
334#define OPC_PADDD       (0xfe | P_EXT | P_DATA16)
335#define OPC_PADDQ       (0xd4 | P_EXT | P_DATA16)
336#define OPC_PADDSB      (0xec | P_EXT | P_DATA16)
337#define OPC_PADDSW      (0xed | P_EXT | P_DATA16)
338#define OPC_PADDUB      (0xdc | P_EXT | P_DATA16)
339#define OPC_PADDUW      (0xdd | P_EXT | P_DATA16)
340#define OPC_PAND        (0xdb | P_EXT | P_DATA16)
341#define OPC_PANDN       (0xdf | P_EXT | P_DATA16)
342#define OPC_PBLENDW     (0x0e | P_EXT3A | P_DATA16)
343#define OPC_PCMPEQB     (0x74 | P_EXT | P_DATA16)
344#define OPC_PCMPEQW     (0x75 | P_EXT | P_DATA16)
345#define OPC_PCMPEQD     (0x76 | P_EXT | P_DATA16)
346#define OPC_PCMPEQQ     (0x29 | P_EXT38 | P_DATA16)
347#define OPC_PCMPGTB     (0x64 | P_EXT | P_DATA16)
348#define OPC_PCMPGTW     (0x65 | P_EXT | P_DATA16)
349#define OPC_PCMPGTD     (0x66 | P_EXT | P_DATA16)
350#define OPC_PCMPGTQ     (0x37 | P_EXT38 | P_DATA16)
351#define OPC_PEXTRD      (0x16 | P_EXT3A | P_DATA16)
352#define OPC_PINSRD      (0x22 | P_EXT3A | P_DATA16)
353#define OPC_PMAXSB      (0x3c | P_EXT38 | P_DATA16)
354#define OPC_PMAXSW      (0xee | P_EXT | P_DATA16)
355#define OPC_PMAXSD      (0x3d | P_EXT38 | P_DATA16)
356#define OPC_VPMAXSQ     (0x3d | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
357#define OPC_PMAXUB      (0xde | P_EXT | P_DATA16)
358#define OPC_PMAXUW      (0x3e | P_EXT38 | P_DATA16)
359#define OPC_PMAXUD      (0x3f | P_EXT38 | P_DATA16)
360#define OPC_VPMAXUQ     (0x3f | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
361#define OPC_PMINSB      (0x38 | P_EXT38 | P_DATA16)
362#define OPC_PMINSW      (0xea | P_EXT | P_DATA16)
363#define OPC_PMINSD      (0x39 | P_EXT38 | P_DATA16)
364#define OPC_VPMINSQ     (0x39 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
365#define OPC_PMINUB      (0xda | P_EXT | P_DATA16)
366#define OPC_PMINUW      (0x3a | P_EXT38 | P_DATA16)
367#define OPC_PMINUD      (0x3b | P_EXT38 | P_DATA16)
368#define OPC_VPMINUQ     (0x3b | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
369#define OPC_PMOVSXBW    (0x20 | P_EXT38 | P_DATA16)
370#define OPC_PMOVSXWD    (0x23 | P_EXT38 | P_DATA16)
371#define OPC_PMOVSXDQ    (0x25 | P_EXT38 | P_DATA16)
372#define OPC_PMOVZXBW    (0x30 | P_EXT38 | P_DATA16)
373#define OPC_PMOVZXWD    (0x33 | P_EXT38 | P_DATA16)
374#define OPC_PMOVZXDQ    (0x35 | P_EXT38 | P_DATA16)
375#define OPC_PMULLW      (0xd5 | P_EXT | P_DATA16)
376#define OPC_PMULLD      (0x40 | P_EXT38 | P_DATA16)
377#define OPC_VPMULLQ     (0x40 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
378#define OPC_POR         (0xeb | P_EXT | P_DATA16)
379#define OPC_PSHUFB      (0x00 | P_EXT38 | P_DATA16)
380#define OPC_PSHUFD      (0x70 | P_EXT | P_DATA16)
381#define OPC_PSHUFLW     (0x70 | P_EXT | P_SIMDF2)
382#define OPC_PSHUFHW     (0x70 | P_EXT | P_SIMDF3)
383#define OPC_PSHIFTW_Ib  (0x71 | P_EXT | P_DATA16) /* /2 /6 /4 */
384#define OPC_PSHIFTD_Ib  (0x72 | P_EXT | P_DATA16) /* /1 /2 /6 /4 */
385#define OPC_PSHIFTQ_Ib  (0x73 | P_EXT | P_DATA16) /* /2 /6 /4 */
386#define OPC_PSLLW       (0xf1 | P_EXT | P_DATA16)
387#define OPC_PSLLD       (0xf2 | P_EXT | P_DATA16)
388#define OPC_PSLLQ       (0xf3 | P_EXT | P_DATA16)
389#define OPC_PSRAW       (0xe1 | P_EXT | P_DATA16)
390#define OPC_PSRAD       (0xe2 | P_EXT | P_DATA16)
391#define OPC_VPSRAQ      (0xe2 | P_EXT | P_DATA16 | P_VEXW | P_EVEX)
392#define OPC_PSRLW       (0xd1 | P_EXT | P_DATA16)
393#define OPC_PSRLD       (0xd2 | P_EXT | P_DATA16)
394#define OPC_PSRLQ       (0xd3 | P_EXT | P_DATA16)
395#define OPC_PSUBB       (0xf8 | P_EXT | P_DATA16)
396#define OPC_PSUBW       (0xf9 | P_EXT | P_DATA16)
397#define OPC_PSUBD       (0xfa | P_EXT | P_DATA16)
398#define OPC_PSUBQ       (0xfb | P_EXT | P_DATA16)
399#define OPC_PSUBSB      (0xe8 | P_EXT | P_DATA16)
400#define OPC_PSUBSW      (0xe9 | P_EXT | P_DATA16)
401#define OPC_PSUBUB      (0xd8 | P_EXT | P_DATA16)
402#define OPC_PSUBUW      (0xd9 | P_EXT | P_DATA16)
403#define OPC_PUNPCKLBW   (0x60 | P_EXT | P_DATA16)
404#define OPC_PUNPCKLWD   (0x61 | P_EXT | P_DATA16)
405#define OPC_PUNPCKLDQ   (0x62 | P_EXT | P_DATA16)
406#define OPC_PUNPCKLQDQ  (0x6c | P_EXT | P_DATA16)
407#define OPC_PUNPCKHBW   (0x68 | P_EXT | P_DATA16)
408#define OPC_PUNPCKHWD   (0x69 | P_EXT | P_DATA16)
409#define OPC_PUNPCKHDQ   (0x6a | P_EXT | P_DATA16)
410#define OPC_PUNPCKHQDQ  (0x6d | P_EXT | P_DATA16)
411#define OPC_PXOR        (0xef | P_EXT | P_DATA16)
412#define OPC_POP_r32	(0x58)
413#define OPC_POPCNT      (0xb8 | P_EXT | P_SIMDF3)
414#define OPC_PUSH_r32	(0x50)
415#define OPC_PUSH_Iv	(0x68)
416#define OPC_PUSH_Ib	(0x6a)
417#define OPC_RET		(0xc3)
418#define OPC_SETCC	(0x90 | P_EXT | P_REXB_RM) /* ... plus cc */
419#define OPC_SHIFT_1	(0xd1)
420#define OPC_SHIFT_Ib	(0xc1)
421#define OPC_SHIFT_cl	(0xd3)
422#define OPC_SARX        (0xf7 | P_EXT38 | P_SIMDF3)
423#define OPC_SHUFPS      (0xc6 | P_EXT)
424#define OPC_SHLX        (0xf7 | P_EXT38 | P_DATA16)
425#define OPC_SHRX        (0xf7 | P_EXT38 | P_SIMDF2)
426#define OPC_SHRD_Ib     (0xac | P_EXT)
427#define OPC_TESTB	(0x84)
428#define OPC_TESTL	(0x85)
429#define OPC_TZCNT       (0xbc | P_EXT | P_SIMDF3)
430#define OPC_UD2         (0x0b | P_EXT)
431#define OPC_VPBLENDD    (0x02 | P_EXT3A | P_DATA16)
432#define OPC_VPBLENDVB   (0x4c | P_EXT3A | P_DATA16)
433#define OPC_VPBLENDMB   (0x66 | P_EXT38 | P_DATA16 | P_EVEX)
434#define OPC_VPBLENDMW   (0x66 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
435#define OPC_VPBLENDMD   (0x64 | P_EXT38 | P_DATA16 | P_EVEX)
436#define OPC_VPBLENDMQ   (0x64 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
437#define OPC_VPCMPB      (0x3f | P_EXT3A | P_DATA16 | P_EVEX)
438#define OPC_VPCMPUB     (0x3e | P_EXT3A | P_DATA16 | P_EVEX)
439#define OPC_VPCMPW      (0x3f | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
440#define OPC_VPCMPUW     (0x3e | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
441#define OPC_VPCMPD      (0x1f | P_EXT3A | P_DATA16 | P_EVEX)
442#define OPC_VPCMPUD     (0x1e | P_EXT3A | P_DATA16 | P_EVEX)
443#define OPC_VPCMPQ      (0x1f | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
444#define OPC_VPCMPUQ     (0x1e | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
445#define OPC_VPINSRB     (0x20 | P_EXT3A | P_DATA16)
446#define OPC_VPINSRW     (0xc4 | P_EXT | P_DATA16)
447#define OPC_VBROADCASTSS (0x18 | P_EXT38 | P_DATA16)
448#define OPC_VBROADCASTSD (0x19 | P_EXT38 | P_DATA16)
449#define OPC_VPBROADCASTB (0x78 | P_EXT38 | P_DATA16)
450#define OPC_VPBROADCASTW (0x79 | P_EXT38 | P_DATA16)
451#define OPC_VPBROADCASTD (0x58 | P_EXT38 | P_DATA16)
452#define OPC_VPBROADCASTQ (0x59 | P_EXT38 | P_DATA16)
453#define OPC_VPMOVM2B    (0x28 | P_EXT38 | P_SIMDF3 | P_EVEX)
454#define OPC_VPMOVM2W    (0x28 | P_EXT38 | P_SIMDF3 | P_VEXW | P_EVEX)
455#define OPC_VPMOVM2D    (0x38 | P_EXT38 | P_SIMDF3 | P_EVEX)
456#define OPC_VPMOVM2Q    (0x38 | P_EXT38 | P_SIMDF3 | P_VEXW | P_EVEX)
457#define OPC_VPERMQ      (0x00 | P_EXT3A | P_DATA16 | P_VEXW)
458#define OPC_VPERM2I128  (0x46 | P_EXT3A | P_DATA16 | P_VEXL)
459#define OPC_VPROLVD     (0x15 | P_EXT38 | P_DATA16 | P_EVEX)
460#define OPC_VPROLVQ     (0x15 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
461#define OPC_VPRORVD     (0x14 | P_EXT38 | P_DATA16 | P_EVEX)
462#define OPC_VPRORVQ     (0x14 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
463#define OPC_VPSHLDW     (0x70 | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
464#define OPC_VPSHLDD     (0x71 | P_EXT3A | P_DATA16 | P_EVEX)
465#define OPC_VPSHLDQ     (0x71 | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
466#define OPC_VPSHLDVW    (0x70 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
467#define OPC_VPSHLDVD    (0x71 | P_EXT38 | P_DATA16 | P_EVEX)
468#define OPC_VPSHLDVQ    (0x71 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
469#define OPC_VPSHRDVW    (0x72 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
470#define OPC_VPSHRDVD    (0x73 | P_EXT38 | P_DATA16 | P_EVEX)
471#define OPC_VPSHRDVQ    (0x73 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
472#define OPC_VPSLLVW     (0x12 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
473#define OPC_VPSLLVD     (0x47 | P_EXT38 | P_DATA16)
474#define OPC_VPSLLVQ     (0x47 | P_EXT38 | P_DATA16 | P_VEXW)
475#define OPC_VPSRAVW     (0x11 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
476#define OPC_VPSRAVD     (0x46 | P_EXT38 | P_DATA16)
477#define OPC_VPSRAVQ     (0x46 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
478#define OPC_VPSRLVW     (0x10 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
479#define OPC_VPSRLVD     (0x45 | P_EXT38 | P_DATA16)
480#define OPC_VPSRLVQ     (0x45 | P_EXT38 | P_DATA16 | P_VEXW)
481#define OPC_VPTERNLOGQ  (0x25 | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
482#define OPC_VPTESTMB    (0x26 | P_EXT38 | P_DATA16 | P_EVEX)
483#define OPC_VPTESTMW    (0x26 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
484#define OPC_VPTESTMD    (0x27 | P_EXT38 | P_DATA16 | P_EVEX)
485#define OPC_VPTESTMQ    (0x27 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
486#define OPC_VPTESTNMB   (0x26 | P_EXT38 | P_SIMDF3 | P_EVEX)
487#define OPC_VPTESTNMW   (0x26 | P_EXT38 | P_SIMDF3 | P_VEXW | P_EVEX)
488#define OPC_VPTESTNMD   (0x27 | P_EXT38 | P_SIMDF3 | P_EVEX)
489#define OPC_VPTESTNMQ   (0x27 | P_EXT38 | P_SIMDF3 | P_VEXW | P_EVEX)
490#define OPC_VZEROUPPER  (0x77 | P_EXT)
491#define OPC_XCHG_ax_r32	(0x90)
492#define OPC_XCHG_EvGv   (0x87)
493
494#define OPC_GRP3_Eb     (0xf6)
495#define OPC_GRP3_Ev     (0xf7)
496#define OPC_GRP5        (0xff)
497#define OPC_GRP14       (0x73 | P_EXT | P_DATA16)
498#define OPC_GRPBT       (0xba | P_EXT)
499
500#define OPC_GRPBT_BT    4
501#define OPC_GRPBT_BTS   5
502#define OPC_GRPBT_BTR   6
503#define OPC_GRPBT_BTC   7
504
505/* Group 1 opcode extensions for 0x80-0x83.
506   These are also used as modifiers for OPC_ARITH.  */
507#define ARITH_ADD 0
508#define ARITH_OR  1
509#define ARITH_ADC 2
510#define ARITH_SBB 3
511#define ARITH_AND 4
512#define ARITH_SUB 5
513#define ARITH_XOR 6
514#define ARITH_CMP 7
515
516/* Group 2 opcode extensions for 0xc0, 0xc1, 0xd0-0xd3.  */
517#define SHIFT_ROL 0
518#define SHIFT_ROR 1
519#define SHIFT_SHL 4
520#define SHIFT_SHR 5
521#define SHIFT_SAR 7
522
523/* Group 3 opcode extensions for 0xf6, 0xf7.  To be used with OPC_GRP3.  */
524#define EXT3_TESTi 0
525#define EXT3_NOT   2
526#define EXT3_NEG   3
527#define EXT3_MUL   4
528#define EXT3_IMUL  5
529#define EXT3_DIV   6
530#define EXT3_IDIV  7
531
532/* Group 5 opcode extensions for 0xff.  To be used with OPC_GRP5.  */
533#define EXT5_INC_Ev	0
534#define EXT5_DEC_Ev	1
535#define EXT5_CALLN_Ev	2
536#define EXT5_JMPN_Ev	4
537
538/* Condition codes to be added to OPC_JCC_{long,short}.  */
539#define JCC_JMP (-1)
540#define JCC_JO  0x0
541#define JCC_JNO 0x1
542#define JCC_JB  0x2
543#define JCC_JAE 0x3
544#define JCC_JE  0x4
545#define JCC_JNE 0x5
546#define JCC_JBE 0x6
547#define JCC_JA  0x7
548#define JCC_JS  0x8
549#define JCC_JNS 0x9
550#define JCC_JP  0xa
551#define JCC_JNP 0xb
552#define JCC_JL  0xc
553#define JCC_JGE 0xd
554#define JCC_JLE 0xe
555#define JCC_JG  0xf
556
557static const uint8_t tcg_cond_to_jcc[] = {
558    [TCG_COND_EQ] = JCC_JE,
559    [TCG_COND_NE] = JCC_JNE,
560    [TCG_COND_LT] = JCC_JL,
561    [TCG_COND_GE] = JCC_JGE,
562    [TCG_COND_LE] = JCC_JLE,
563    [TCG_COND_GT] = JCC_JG,
564    [TCG_COND_LTU] = JCC_JB,
565    [TCG_COND_GEU] = JCC_JAE,
566    [TCG_COND_LEU] = JCC_JBE,
567    [TCG_COND_GTU] = JCC_JA,
568    [TCG_COND_TSTEQ] = JCC_JE,
569    [TCG_COND_TSTNE] = JCC_JNE,
570};
571
572#if TCG_TARGET_REG_BITS == 64
573static void tcg_out_opc(TCGContext *s, int opc, int r, int rm, int x)
574{
575    int rex;
576
577    if (opc & P_GS) {
578        tcg_out8(s, 0x65);
579    }
580    if (opc & P_DATA16) {
581        /* We should never be asking for both 16 and 64-bit operation.  */
582        tcg_debug_assert((opc & P_REXW) == 0);
583        tcg_out8(s, 0x66);
584    }
585    if (opc & P_SIMDF3) {
586        tcg_out8(s, 0xf3);
587    } else if (opc & P_SIMDF2) {
588        tcg_out8(s, 0xf2);
589    }
590
591    rex = 0;
592    rex |= (opc & P_REXW) ? 0x8 : 0x0;  /* REX.W */
593    rex |= (r & 8) >> 1;                /* REX.R */
594    rex |= (x & 8) >> 2;                /* REX.X */
595    rex |= (rm & 8) >> 3;               /* REX.B */
596
597    /* P_REXB_{R,RM} indicates that the given register is the low byte.
598       For %[abcd]l we need no REX prefix, but for %{si,di,bp,sp}l we do,
599       as otherwise the encoding indicates %[abcd]h.  Note that the values
600       that are ORed in merely indicate that the REX byte must be present;
601       those bits get discarded in output.  */
602    rex |= opc & (r >= 4 ? P_REXB_R : 0);
603    rex |= opc & (rm >= 4 ? P_REXB_RM : 0);
604
605    if (rex) {
606        tcg_out8(s, (uint8_t)(rex | 0x40));
607    }
608
609    if (opc & (P_EXT | P_EXT38 | P_EXT3A)) {
610        tcg_out8(s, 0x0f);
611        if (opc & P_EXT38) {
612            tcg_out8(s, 0x38);
613        } else if (opc & P_EXT3A) {
614            tcg_out8(s, 0x3a);
615        }
616    }
617
618    tcg_out8(s, opc);
619}
620#else
621static void tcg_out_opc(TCGContext *s, int opc)
622{
623    if (opc & P_DATA16) {
624        tcg_out8(s, 0x66);
625    }
626    if (opc & P_SIMDF3) {
627        tcg_out8(s, 0xf3);
628    } else if (opc & P_SIMDF2) {
629        tcg_out8(s, 0xf2);
630    }
631    if (opc & (P_EXT | P_EXT38 | P_EXT3A)) {
632        tcg_out8(s, 0x0f);
633        if (opc & P_EXT38) {
634            tcg_out8(s, 0x38);
635        } else if (opc & P_EXT3A) {
636            tcg_out8(s, 0x3a);
637        }
638    }
639    tcg_out8(s, opc);
640}
641/* Discard the register arguments to tcg_out_opc early, so as not to penalize
642   the 32-bit compilation paths.  This method works with all versions of gcc,
643   whereas relying on optimization may not be able to exclude them.  */
644#define tcg_out_opc(s, opc, r, rm, x)  (tcg_out_opc)(s, opc)
645#endif
646
647static void tcg_out_modrm(TCGContext *s, int opc, int r, int rm)
648{
649    tcg_out_opc(s, opc, r, rm, 0);
650    tcg_out8(s, 0xc0 | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
651}
652
653static void tcg_out_vex_opc(TCGContext *s, int opc, int r, int v,
654                            int rm, int index)
655{
656    int tmp;
657
658    if (opc & P_GS) {
659        tcg_out8(s, 0x65);
660    }
661    /* Use the two byte form if possible, which cannot encode
662       VEX.W, VEX.B, VEX.X, or an m-mmmm field other than P_EXT.  */
663    if ((opc & (P_EXT | P_EXT38 | P_EXT3A | P_VEXW)) == P_EXT
664        && ((rm | index) & 8) == 0) {
665        /* Two byte VEX prefix.  */
666        tcg_out8(s, 0xc5);
667
668        tmp = (r & 8 ? 0 : 0x80);              /* VEX.R */
669    } else {
670        /* Three byte VEX prefix.  */
671        tcg_out8(s, 0xc4);
672
673        /* VEX.m-mmmm */
674        if (opc & P_EXT3A) {
675            tmp = 3;
676        } else if (opc & P_EXT38) {
677            tmp = 2;
678        } else if (opc & P_EXT) {
679            tmp = 1;
680        } else {
681            g_assert_not_reached();
682        }
683        tmp |= (r & 8 ? 0 : 0x80);             /* VEX.R */
684        tmp |= (index & 8 ? 0 : 0x40);         /* VEX.X */
685        tmp |= (rm & 8 ? 0 : 0x20);            /* VEX.B */
686        tcg_out8(s, tmp);
687
688        tmp = (opc & P_VEXW ? 0x80 : 0);       /* VEX.W */
689    }
690
691    tmp |= (opc & P_VEXL ? 0x04 : 0);      /* VEX.L */
692    /* VEX.pp */
693    if (opc & P_DATA16) {
694        tmp |= 1;                          /* 0x66 */
695    } else if (opc & P_SIMDF3) {
696        tmp |= 2;                          /* 0xf3 */
697    } else if (opc & P_SIMDF2) {
698        tmp |= 3;                          /* 0xf2 */
699    }
700    tmp |= (~v & 15) << 3;                 /* VEX.vvvv */
701    tcg_out8(s, tmp);
702    tcg_out8(s, opc);
703}
704
705static void tcg_out_evex_opc(TCGContext *s, int opc, int r, int v,
706                             int rm, int index, int aaa, bool z)
707{
708    /* The entire 4-byte evex prefix; with R' and V' set. */
709    uint32_t p = 0x08041062;
710    int mm, pp;
711
712    tcg_debug_assert(have_avx512vl);
713
714    /* EVEX.mm */
715    if (opc & P_EXT3A) {
716        mm = 3;
717    } else if (opc & P_EXT38) {
718        mm = 2;
719    } else if (opc & P_EXT) {
720        mm = 1;
721    } else {
722        g_assert_not_reached();
723    }
724
725    /* EVEX.pp */
726    if (opc & P_DATA16) {
727        pp = 1;                          /* 0x66 */
728    } else if (opc & P_SIMDF3) {
729        pp = 2;                          /* 0xf3 */
730    } else if (opc & P_SIMDF2) {
731        pp = 3;                          /* 0xf2 */
732    } else {
733        pp = 0;
734    }
735
736    p = deposit32(p, 8, 2, mm);
737    p = deposit32(p, 13, 1, (rm & 8) == 0);             /* EVEX.RXB.B */
738    p = deposit32(p, 14, 1, (index & 8) == 0);          /* EVEX.RXB.X */
739    p = deposit32(p, 15, 1, (r & 8) == 0);              /* EVEX.RXB.R */
740    p = deposit32(p, 16, 2, pp);
741    p = deposit32(p, 19, 4, ~v);
742    p = deposit32(p, 23, 1, (opc & P_VEXW) != 0);
743    p = deposit32(p, 24, 3, aaa);
744    p = deposit32(p, 29, 2, (opc & P_VEXL) != 0);
745    p = deposit32(p, 31, 1, z);
746
747    tcg_out32(s, p);
748    tcg_out8(s, opc);
749}
750
751static void tcg_out_vex_modrm(TCGContext *s, int opc, int r, int v, int rm)
752{
753    if (opc & P_EVEX) {
754        tcg_out_evex_opc(s, opc, r, v, rm, 0, 0, false);
755    } else {
756        tcg_out_vex_opc(s, opc, r, v, rm, 0);
757    }
758    tcg_out8(s, 0xc0 | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
759}
760
761static void tcg_out_vex_modrm_type(TCGContext *s, int opc,
762                                   int r, int v, int rm, TCGType type)
763{
764    if (type == TCG_TYPE_V256) {
765        opc |= P_VEXL;
766    }
767    tcg_out_vex_modrm(s, opc, r, v, rm);
768}
769
770static void tcg_out_evex_modrm_type(TCGContext *s, int opc, int r, int v,
771                                    int rm, int aaa, bool z, TCGType type)
772{
773    if (type == TCG_TYPE_V256) {
774        opc |= P_VEXL;
775    }
776    tcg_out_evex_opc(s, opc, r, v, rm, 0, aaa, z);
777    tcg_out8(s, 0xc0 | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
778}
779
780/* Output an opcode with a full "rm + (index<<shift) + offset" address mode.
781   We handle either RM and INDEX missing with a negative value.  In 64-bit
782   mode for absolute addresses, ~RM is the size of the immediate operand
783   that will follow the instruction.  */
784
785static void tcg_out_sib_offset(TCGContext *s, int r, int rm, int index,
786                               int shift, intptr_t offset)
787{
788    int mod, len;
789
790    if (index < 0 && rm < 0) {
791        if (TCG_TARGET_REG_BITS == 64) {
792            /* Try for a rip-relative addressing mode.  This has replaced
793               the 32-bit-mode absolute addressing encoding.  */
794            intptr_t pc = (intptr_t)s->code_ptr + 5 + ~rm;
795            intptr_t disp = offset - pc;
796            if (disp == (int32_t)disp) {
797                tcg_out8(s, (LOWREGMASK(r) << 3) | 5);
798                tcg_out32(s, disp);
799                return;
800            }
801
802            /* Try for an absolute address encoding.  This requires the
803               use of the MODRM+SIB encoding and is therefore larger than
804               rip-relative addressing.  */
805            if (offset == (int32_t)offset) {
806                tcg_out8(s, (LOWREGMASK(r) << 3) | 4);
807                tcg_out8(s, (4 << 3) | 5);
808                tcg_out32(s, offset);
809                return;
810            }
811
812            /* ??? The memory isn't directly addressable.  */
813            g_assert_not_reached();
814        } else {
815            /* Absolute address.  */
816            tcg_out8(s, (r << 3) | 5);
817            tcg_out32(s, offset);
818            return;
819        }
820    }
821
822    /* Find the length of the immediate addend.  Note that the encoding
823       that would be used for (%ebp) indicates absolute addressing.  */
824    if (rm < 0) {
825        mod = 0, len = 4, rm = 5;
826    } else if (offset == 0 && LOWREGMASK(rm) != TCG_REG_EBP) {
827        mod = 0, len = 0;
828    } else if (offset == (int8_t)offset) {
829        mod = 0x40, len = 1;
830    } else {
831        mod = 0x80, len = 4;
832    }
833
834    /* Use a single byte MODRM format if possible.  Note that the encoding
835       that would be used for %esp is the escape to the two byte form.  */
836    if (index < 0 && LOWREGMASK(rm) != TCG_REG_ESP) {
837        /* Single byte MODRM format.  */
838        tcg_out8(s, mod | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
839    } else {
840        /* Two byte MODRM+SIB format.  */
841
842        /* Note that the encoding that would place %esp into the index
843           field indicates no index register.  In 64-bit mode, the REX.X
844           bit counts, so %r12 can be used as the index.  */
845        if (index < 0) {
846            index = 4;
847        } else {
848            tcg_debug_assert(index != TCG_REG_ESP);
849        }
850
851        tcg_out8(s, mod | (LOWREGMASK(r) << 3) | 4);
852        tcg_out8(s, (shift << 6) | (LOWREGMASK(index) << 3) | LOWREGMASK(rm));
853    }
854
855    if (len == 1) {
856        tcg_out8(s, offset);
857    } else if (len == 4) {
858        tcg_out32(s, offset);
859    }
860}
861
862static void tcg_out_modrm_sib_offset(TCGContext *s, int opc, int r, int rm,
863                                     int index, int shift, intptr_t offset)
864{
865    tcg_out_opc(s, opc, r, rm < 0 ? 0 : rm, index < 0 ? 0 : index);
866    tcg_out_sib_offset(s, r, rm, index, shift, offset);
867}
868
869static void tcg_out_vex_modrm_sib_offset(TCGContext *s, int opc, int r, int v,
870                                         int rm, int index, int shift,
871                                         intptr_t offset)
872{
873    tcg_out_vex_opc(s, opc, r, v, rm < 0 ? 0 : rm, index < 0 ? 0 : index);
874    tcg_out_sib_offset(s, r, rm, index, shift, offset);
875}
876
877/* A simplification of the above with no index or shift.  */
878static inline void tcg_out_modrm_offset(TCGContext *s, int opc, int r,
879                                        int rm, intptr_t offset)
880{
881    tcg_out_modrm_sib_offset(s, opc, r, rm, -1, 0, offset);
882}
883
884static inline void tcg_out_vex_modrm_offset(TCGContext *s, int opc, int r,
885                                            int v, int rm, intptr_t offset)
886{
887    tcg_out_vex_modrm_sib_offset(s, opc, r, v, rm, -1, 0, offset);
888}
889
890/* Output an opcode with an expected reference to the constant pool.  */
891static inline void tcg_out_modrm_pool(TCGContext *s, int opc, int r)
892{
893    tcg_out_opc(s, opc, r, 0, 0);
894    /* Absolute for 32-bit, pc-relative for 64-bit.  */
895    tcg_out8(s, LOWREGMASK(r) << 3 | 5);
896    tcg_out32(s, 0);
897}
898
899/* Output an opcode with an expected reference to the constant pool.  */
900static inline void tcg_out_vex_modrm_pool(TCGContext *s, int opc, int r)
901{
902    tcg_out_vex_opc(s, opc, r, 0, 0, 0);
903    /* Absolute for 32-bit, pc-relative for 64-bit.  */
904    tcg_out8(s, LOWREGMASK(r) << 3 | 5);
905    tcg_out32(s, 0);
906}
907
908/* Generate dest op= src.  Uses the same ARITH_* codes as tgen_arithi.  */
909static inline void tgen_arithr(TCGContext *s, int subop, int dest, int src)
910{
911    /* Propagate an opcode prefix, such as P_REXW.  */
912    int ext = subop & ~0x7;
913    subop &= 0x7;
914
915    tcg_out_modrm(s, OPC_ARITH_GvEv + (subop << 3) + ext, dest, src);
916}
917
918static bool tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
919{
920    int rexw = 0;
921
922    if (arg == ret) {
923        return true;
924    }
925    switch (type) {
926    case TCG_TYPE_I64:
927        rexw = P_REXW;
928        /* fallthru */
929    case TCG_TYPE_I32:
930        if (ret < 16) {
931            if (arg < 16) {
932                tcg_out_modrm(s, OPC_MOVL_GvEv + rexw, ret, arg);
933            } else {
934                tcg_out_vex_modrm(s, OPC_MOVD_EyVy + rexw, arg, 0, ret);
935            }
936        } else {
937            if (arg < 16) {
938                tcg_out_vex_modrm(s, OPC_MOVD_VyEy + rexw, ret, 0, arg);
939            } else {
940                tcg_out_vex_modrm(s, OPC_MOVQ_VqWq, ret, 0, arg);
941            }
942        }
943        break;
944
945    case TCG_TYPE_V64:
946        tcg_debug_assert(ret >= 16 && arg >= 16);
947        tcg_out_vex_modrm(s, OPC_MOVQ_VqWq, ret, 0, arg);
948        break;
949    case TCG_TYPE_V128:
950        tcg_debug_assert(ret >= 16 && arg >= 16);
951        tcg_out_vex_modrm(s, OPC_MOVDQA_VxWx, ret, 0, arg);
952        break;
953    case TCG_TYPE_V256:
954        tcg_debug_assert(ret >= 16 && arg >= 16);
955        tcg_out_vex_modrm(s, OPC_MOVDQA_VxWx | P_VEXL, ret, 0, arg);
956        break;
957
958    default:
959        g_assert_not_reached();
960    }
961    return true;
962}
963
964static const int avx2_dup_insn[4] = {
965    OPC_VPBROADCASTB, OPC_VPBROADCASTW,
966    OPC_VPBROADCASTD, OPC_VPBROADCASTQ,
967};
968
969static bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece,
970                            TCGReg r, TCGReg a)
971{
972    if (have_avx2) {
973        tcg_out_vex_modrm_type(s, avx2_dup_insn[vece], r, 0, a, type);
974    } else {
975        switch (vece) {
976        case MO_8:
977            /* ??? With zero in a register, use PSHUFB.  */
978            tcg_out_vex_modrm(s, OPC_PUNPCKLBW, r, a, a);
979            a = r;
980            /* FALLTHRU */
981        case MO_16:
982            tcg_out_vex_modrm(s, OPC_PUNPCKLWD, r, a, a);
983            a = r;
984            /* FALLTHRU */
985        case MO_32:
986            tcg_out_vex_modrm(s, OPC_PSHUFD, r, 0, a);
987            /* imm8 operand: all output lanes selected from input lane 0.  */
988            tcg_out8(s, 0);
989            break;
990        case MO_64:
991            tcg_out_vex_modrm(s, OPC_PUNPCKLQDQ, r, a, a);
992            break;
993        default:
994            g_assert_not_reached();
995        }
996    }
997    return true;
998}
999
1000static bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece,
1001                             TCGReg r, TCGReg base, intptr_t offset)
1002{
1003    if (have_avx2) {
1004        int vex_l = (type == TCG_TYPE_V256 ? P_VEXL : 0);
1005        tcg_out_vex_modrm_offset(s, avx2_dup_insn[vece] + vex_l,
1006                                 r, 0, base, offset);
1007    } else {
1008        switch (vece) {
1009        case MO_64:
1010            tcg_out_vex_modrm_offset(s, OPC_MOVDDUP, r, 0, base, offset);
1011            break;
1012        case MO_32:
1013            tcg_out_vex_modrm_offset(s, OPC_VBROADCASTSS, r, 0, base, offset);
1014            break;
1015        case MO_16:
1016            tcg_out_vex_modrm_offset(s, OPC_VPINSRW, r, r, base, offset);
1017            tcg_out8(s, 0); /* imm8 */
1018            tcg_out_dup_vec(s, type, vece, r, r);
1019            break;
1020        case MO_8:
1021            tcg_out_vex_modrm_offset(s, OPC_VPINSRB, r, r, base, offset);
1022            tcg_out8(s, 0); /* imm8 */
1023            tcg_out_dup_vec(s, type, vece, r, r);
1024            break;
1025        default:
1026            g_assert_not_reached();
1027        }
1028    }
1029    return true;
1030}
1031
1032static void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece,
1033                             TCGReg ret, int64_t arg)
1034{
1035    int vex_l = (type == TCG_TYPE_V256 ? P_VEXL : 0);
1036
1037    if (arg == 0) {
1038        tcg_out_vex_modrm(s, OPC_PXOR, ret, ret, ret);
1039        return;
1040    }
1041    if (arg == -1) {
1042        tcg_out_vex_modrm(s, OPC_PCMPEQB + vex_l, ret, ret, ret);
1043        return;
1044    }
1045
1046    if (TCG_TARGET_REG_BITS == 32 && vece < MO_64) {
1047        if (have_avx2) {
1048            tcg_out_vex_modrm_pool(s, OPC_VPBROADCASTD + vex_l, ret);
1049        } else {
1050            tcg_out_vex_modrm_pool(s, OPC_VBROADCASTSS, ret);
1051        }
1052        new_pool_label(s, arg, R_386_32, s->code_ptr - 4, 0);
1053    } else {
1054        if (type == TCG_TYPE_V64) {
1055            tcg_out_vex_modrm_pool(s, OPC_MOVQ_VqWq, ret);
1056        } else if (have_avx2) {
1057            tcg_out_vex_modrm_pool(s, OPC_VPBROADCASTQ + vex_l, ret);
1058        } else {
1059            tcg_out_vex_modrm_pool(s, OPC_MOVDDUP, ret);
1060        }
1061        if (TCG_TARGET_REG_BITS == 64) {
1062            new_pool_label(s, arg, R_386_PC32, s->code_ptr - 4, -4);
1063        } else {
1064            new_pool_l2(s, R_386_32, s->code_ptr - 4, 0, arg, arg >> 32);
1065        }
1066    }
1067}
1068
1069static void tcg_out_movi_vec(TCGContext *s, TCGType type,
1070                             TCGReg ret, tcg_target_long arg)
1071{
1072    if (arg == 0) {
1073        tcg_out_vex_modrm(s, OPC_PXOR, ret, ret, ret);
1074        return;
1075    }
1076    if (arg == -1) {
1077        tcg_out_vex_modrm(s, OPC_PCMPEQB, ret, ret, ret);
1078        return;
1079    }
1080
1081    int rexw = (type == TCG_TYPE_I32 ? 0 : P_REXW);
1082    tcg_out_vex_modrm_pool(s, OPC_MOVD_VyEy + rexw, ret);
1083    if (TCG_TARGET_REG_BITS == 64) {
1084        new_pool_label(s, arg, R_386_PC32, s->code_ptr - 4, -4);
1085    } else {
1086        new_pool_label(s, arg, R_386_32, s->code_ptr - 4, 0);
1087    }
1088}
1089
1090static void tcg_out_movi_int(TCGContext *s, TCGType type,
1091                             TCGReg ret, tcg_target_long arg)
1092{
1093    tcg_target_long diff;
1094
1095    if (arg == 0) {
1096        tgen_arithr(s, ARITH_XOR, ret, ret);
1097        return;
1098    }
1099    if (arg == (uint32_t)arg || type == TCG_TYPE_I32) {
1100        tcg_out_opc(s, OPC_MOVL_Iv + LOWREGMASK(ret), 0, ret, 0);
1101        tcg_out32(s, arg);
1102        return;
1103    }
1104    if (arg == (int32_t)arg) {
1105        tcg_out_modrm(s, OPC_MOVL_EvIz + P_REXW, 0, ret);
1106        tcg_out32(s, arg);
1107        return;
1108    }
1109
1110    /* Try a 7 byte pc-relative lea before the 10 byte movq.  */
1111    diff = tcg_pcrel_diff(s, (const void *)arg) - 7;
1112    if (diff == (int32_t)diff) {
1113        tcg_out_opc(s, OPC_LEA | P_REXW, ret, 0, 0);
1114        tcg_out8(s, (LOWREGMASK(ret) << 3) | 5);
1115        tcg_out32(s, diff);
1116        return;
1117    }
1118
1119    tcg_out_opc(s, OPC_MOVL_Iv + P_REXW + LOWREGMASK(ret), 0, ret, 0);
1120    tcg_out64(s, arg);
1121}
1122
1123static void tcg_out_movi(TCGContext *s, TCGType type,
1124                         TCGReg ret, tcg_target_long arg)
1125{
1126    switch (type) {
1127    case TCG_TYPE_I32:
1128#if TCG_TARGET_REG_BITS == 64
1129    case TCG_TYPE_I64:
1130#endif
1131        if (ret < 16) {
1132            tcg_out_movi_int(s, type, ret, arg);
1133        } else {
1134            tcg_out_movi_vec(s, type, ret, arg);
1135        }
1136        break;
1137    default:
1138        g_assert_not_reached();
1139    }
1140}
1141
1142static bool tcg_out_xchg(TCGContext *s, TCGType type, TCGReg r1, TCGReg r2)
1143{
1144    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1145    tcg_out_modrm(s, OPC_XCHG_EvGv + rexw, r1, r2);
1146    return true;
1147}
1148
1149static void tcg_out_addi_ptr(TCGContext *s, TCGReg rd, TCGReg rs,
1150                             tcg_target_long imm)
1151{
1152    /* This function is only used for passing structs by reference. */
1153    tcg_debug_assert(imm == (int32_t)imm);
1154    tcg_out_modrm_offset(s, OPC_LEA | P_REXW, rd, rs, imm);
1155}
1156
1157static inline void tcg_out_pushi(TCGContext *s, tcg_target_long val)
1158{
1159    if (val == (int8_t)val) {
1160        tcg_out_opc(s, OPC_PUSH_Ib, 0, 0, 0);
1161        tcg_out8(s, val);
1162    } else if (val == (int32_t)val) {
1163        tcg_out_opc(s, OPC_PUSH_Iv, 0, 0, 0);
1164        tcg_out32(s, val);
1165    } else {
1166        g_assert_not_reached();
1167    }
1168}
1169
1170static inline void tcg_out_mb(TCGContext *s, TCGArg a0)
1171{
1172    /* Given the strength of x86 memory ordering, we only need care for
1173       store-load ordering.  Experimentally, "lock orl $0,0(%esp)" is
1174       faster than "mfence", so don't bother with the sse insn.  */
1175    if (a0 & TCG_MO_ST_LD) {
1176        tcg_out8(s, 0xf0);
1177        tcg_out_modrm_offset(s, OPC_ARITH_EvIb, ARITH_OR, TCG_REG_ESP, 0);
1178        tcg_out8(s, 0);
1179    }
1180}
1181
1182static inline void tcg_out_push(TCGContext *s, int reg)
1183{
1184    tcg_out_opc(s, OPC_PUSH_r32 + LOWREGMASK(reg), 0, reg, 0);
1185}
1186
1187static inline void tcg_out_pop(TCGContext *s, int reg)
1188{
1189    tcg_out_opc(s, OPC_POP_r32 + LOWREGMASK(reg), 0, reg, 0);
1190}
1191
1192static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg ret,
1193                       TCGReg arg1, intptr_t arg2)
1194{
1195    switch (type) {
1196    case TCG_TYPE_I32:
1197        if (ret < 16) {
1198            tcg_out_modrm_offset(s, OPC_MOVL_GvEv, ret, arg1, arg2);
1199        } else {
1200            tcg_out_vex_modrm_offset(s, OPC_MOVD_VyEy, ret, 0, arg1, arg2);
1201        }
1202        break;
1203    case TCG_TYPE_I64:
1204        if (ret < 16) {
1205            tcg_out_modrm_offset(s, OPC_MOVL_GvEv | P_REXW, ret, arg1, arg2);
1206            break;
1207        }
1208        /* FALLTHRU */
1209    case TCG_TYPE_V64:
1210        /* There is no instruction that can validate 8-byte alignment.  */
1211        tcg_debug_assert(ret >= 16);
1212        tcg_out_vex_modrm_offset(s, OPC_MOVQ_VqWq, ret, 0, arg1, arg2);
1213        break;
1214    case TCG_TYPE_V128:
1215        /*
1216         * The gvec infrastructure is asserts that v128 vector loads
1217         * and stores use a 16-byte aligned offset.  Validate that the
1218         * final pointer is aligned by using an insn that will SIGSEGV.
1219         */
1220        tcg_debug_assert(ret >= 16);
1221        tcg_out_vex_modrm_offset(s, OPC_MOVDQA_VxWx, ret, 0, arg1, arg2);
1222        break;
1223    case TCG_TYPE_V256:
1224        /*
1225         * The gvec infrastructure only requires 16-byte alignment,
1226         * so here we must use an unaligned load.
1227         */
1228        tcg_debug_assert(ret >= 16);
1229        tcg_out_vex_modrm_offset(s, OPC_MOVDQU_VxWx | P_VEXL,
1230                                 ret, 0, arg1, arg2);
1231        break;
1232    default:
1233        g_assert_not_reached();
1234    }
1235}
1236
1237static void tcg_out_st(TCGContext *s, TCGType type, TCGReg arg,
1238                       TCGReg arg1, intptr_t arg2)
1239{
1240    switch (type) {
1241    case TCG_TYPE_I32:
1242        if (arg < 16) {
1243            tcg_out_modrm_offset(s, OPC_MOVL_EvGv, arg, arg1, arg2);
1244        } else {
1245            tcg_out_vex_modrm_offset(s, OPC_MOVD_EyVy, arg, 0, arg1, arg2);
1246        }
1247        break;
1248    case TCG_TYPE_I64:
1249        if (arg < 16) {
1250            tcg_out_modrm_offset(s, OPC_MOVL_EvGv | P_REXW, arg, arg1, arg2);
1251            break;
1252        }
1253        /* FALLTHRU */
1254    case TCG_TYPE_V64:
1255        /* There is no instruction that can validate 8-byte alignment.  */
1256        tcg_debug_assert(arg >= 16);
1257        tcg_out_vex_modrm_offset(s, OPC_MOVQ_WqVq, arg, 0, arg1, arg2);
1258        break;
1259    case TCG_TYPE_V128:
1260        /*
1261         * The gvec infrastructure is asserts that v128 vector loads
1262         * and stores use a 16-byte aligned offset.  Validate that the
1263         * final pointer is aligned by using an insn that will SIGSEGV.
1264         *
1265         * This specific instance is also used by TCG_CALL_RET_BY_VEC,
1266         * for _WIN64, which must have SSE2 but may not have AVX.
1267         */
1268        tcg_debug_assert(arg >= 16);
1269        if (have_avx1) {
1270            tcg_out_vex_modrm_offset(s, OPC_MOVDQA_WxVx, arg, 0, arg1, arg2);
1271        } else {
1272            tcg_out_modrm_offset(s, OPC_MOVDQA_WxVx, arg, arg1, arg2);
1273        }
1274        break;
1275    case TCG_TYPE_V256:
1276        /*
1277         * The gvec infrastructure only requires 16-byte alignment,
1278         * so here we must use an unaligned store.
1279         */
1280        tcg_debug_assert(arg >= 16);
1281        tcg_out_vex_modrm_offset(s, OPC_MOVDQU_WxVx | P_VEXL,
1282                                 arg, 0, arg1, arg2);
1283        break;
1284    default:
1285        g_assert_not_reached();
1286    }
1287}
1288
1289static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val,
1290                        TCGReg base, intptr_t ofs)
1291{
1292    int rexw = 0;
1293    if (TCG_TARGET_REG_BITS == 64 && type == TCG_TYPE_I64) {
1294        if (val != (int32_t)val) {
1295            return false;
1296        }
1297        rexw = P_REXW;
1298    } else if (type != TCG_TYPE_I32) {
1299        return false;
1300    }
1301    tcg_out_modrm_offset(s, OPC_MOVL_EvIz | rexw, 0, base, ofs);
1302    tcg_out32(s, val);
1303    return true;
1304}
1305
1306static void tcg_out_shifti(TCGContext *s, int subopc, int reg, int count)
1307{
1308    /* Propagate an opcode prefix, such as P_DATA16.  */
1309    int ext = subopc & ~0x7;
1310    subopc &= 0x7;
1311
1312    if (count == 1) {
1313        tcg_out_modrm(s, OPC_SHIFT_1 + ext, subopc, reg);
1314    } else {
1315        tcg_out_modrm(s, OPC_SHIFT_Ib + ext, subopc, reg);
1316        tcg_out8(s, count);
1317    }
1318}
1319
1320static inline void tcg_out_bswap32(TCGContext *s, int reg)
1321{
1322    tcg_out_opc(s, OPC_BSWAP + LOWREGMASK(reg), 0, reg, 0);
1323}
1324
1325static inline void tcg_out_rolw_8(TCGContext *s, int reg)
1326{
1327    tcg_out_shifti(s, SHIFT_ROL + P_DATA16, reg, 8);
1328}
1329
1330static void tcg_out_ext8u(TCGContext *s, TCGReg dest, TCGReg src)
1331{
1332    if (TCG_TARGET_REG_BITS == 32 && src >= 4) {
1333        tcg_out_mov(s, TCG_TYPE_I32, dest, src);
1334        if (dest >= 4) {
1335            tcg_out_modrm(s, OPC_ARITH_EvIz, ARITH_AND, dest);
1336            tcg_out32(s, 0xff);
1337            return;
1338        }
1339        src = dest;
1340    }
1341    tcg_out_modrm(s, OPC_MOVZBL + P_REXB_RM, dest, src);
1342}
1343
1344static void tcg_out_ext8s(TCGContext *s, TCGType type, TCGReg dest, TCGReg src)
1345{
1346    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1347
1348    if (TCG_TARGET_REG_BITS == 32 && src >= 4) {
1349        tcg_out_mov(s, TCG_TYPE_I32, dest, src);
1350        if (dest >= 4) {
1351            tcg_out_shifti(s, SHIFT_SHL, dest, 24);
1352            tcg_out_shifti(s, SHIFT_SAR, dest, 24);
1353            return;
1354        }
1355        src = dest;
1356    }
1357    tcg_out_modrm(s, OPC_MOVSBL + P_REXB_RM + rexw, dest, src);
1358}
1359
1360static void tcg_out_ext16u(TCGContext *s, TCGReg dest, TCGReg src)
1361{
1362    /* movzwl */
1363    tcg_out_modrm(s, OPC_MOVZWL, dest, src);
1364}
1365
1366static void tcg_out_ext16s(TCGContext *s, TCGType type, TCGReg dest, TCGReg src)
1367{
1368    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1369    /* movsw[lq] */
1370    tcg_out_modrm(s, OPC_MOVSWL + rexw, dest, src);
1371}
1372
1373static void tcg_out_ext32u(TCGContext *s, TCGReg dest, TCGReg src)
1374{
1375    /* 32-bit mov zero extends.  */
1376    tcg_out_modrm(s, OPC_MOVL_GvEv, dest, src);
1377}
1378
1379static void tcg_out_ext32s(TCGContext *s, TCGReg dest, TCGReg src)
1380{
1381    tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
1382    tcg_out_modrm(s, OPC_MOVSLQ, dest, src);
1383}
1384
1385static void tcg_out_exts_i32_i64(TCGContext *s, TCGReg dest, TCGReg src)
1386{
1387    tcg_out_ext32s(s, dest, src);
1388}
1389
1390static void tcg_out_extu_i32_i64(TCGContext *s, TCGReg dest, TCGReg src)
1391{
1392    if (dest != src) {
1393        tcg_out_ext32u(s, dest, src);
1394    }
1395}
1396
1397static void tcg_out_extrl_i64_i32(TCGContext *s, TCGReg dest, TCGReg src)
1398{
1399    tcg_out_ext32u(s, dest, src);
1400}
1401
1402static inline void tcg_out_bswap64(TCGContext *s, int reg)
1403{
1404    tcg_out_opc(s, OPC_BSWAP + P_REXW + LOWREGMASK(reg), 0, reg, 0);
1405}
1406
1407static void tgen_arithi(TCGContext *s, int c, int r0,
1408                        tcg_target_long val, int cf)
1409{
1410    int rexw = 0;
1411
1412    if (TCG_TARGET_REG_BITS == 64) {
1413        rexw = c & -8;
1414        c &= 7;
1415    }
1416
1417    switch (c) {
1418    case ARITH_ADD:
1419    case ARITH_SUB:
1420        if (!cf) {
1421            /*
1422             * ??? While INC is 2 bytes shorter than ADDL $1, they also induce
1423             * partial flags update stalls on Pentium4 and are not recommended
1424             * by current Intel optimization manuals.
1425             */
1426            if (val == 1 || val == -1) {
1427                int is_inc = (c == ARITH_ADD) ^ (val < 0);
1428                if (TCG_TARGET_REG_BITS == 64) {
1429                    /*
1430                     * The single-byte increment encodings are re-tasked
1431                     * as the REX prefixes.  Use the MODRM encoding.
1432                     */
1433                    tcg_out_modrm(s, OPC_GRP5 + rexw,
1434                                  (is_inc ? EXT5_INC_Ev : EXT5_DEC_Ev), r0);
1435                } else {
1436                    tcg_out8(s, (is_inc ? OPC_INC_r32 : OPC_DEC_r32) + r0);
1437                }
1438                return;
1439            }
1440            if (val == 128) {
1441                /*
1442                 * Facilitate using an 8-bit immediate.  Carry is inverted
1443                 * by this transformation, so do it only if cf == 0.
1444                 */
1445                c ^= ARITH_ADD ^ ARITH_SUB;
1446                val = -128;
1447            }
1448        }
1449        break;
1450
1451    case ARITH_AND:
1452        if (TCG_TARGET_REG_BITS == 64) {
1453            if (val == 0xffffffffu) {
1454                tcg_out_ext32u(s, r0, r0);
1455                return;
1456            }
1457            if (val == (uint32_t)val) {
1458                /* AND with no high bits set can use a 32-bit operation.  */
1459                rexw = 0;
1460            }
1461        }
1462        if (val == 0xffu && (r0 < 4 || TCG_TARGET_REG_BITS == 64)) {
1463            tcg_out_ext8u(s, r0, r0);
1464            return;
1465        }
1466        if (val == 0xffffu) {
1467            tcg_out_ext16u(s, r0, r0);
1468            return;
1469        }
1470        break;
1471
1472    case ARITH_OR:
1473    case ARITH_XOR:
1474        if (val >= 0x80 && val <= 0xff
1475            && (r0 < 4 || TCG_TARGET_REG_BITS == 64)) {
1476            tcg_out_modrm(s, OPC_ARITH_EbIb + P_REXB_RM, c, r0);
1477            tcg_out8(s, val);
1478            return;
1479        }
1480        break;
1481    }
1482
1483    if (val == (int8_t)val) {
1484        tcg_out_modrm(s, OPC_ARITH_EvIb + rexw, c, r0);
1485        tcg_out8(s, val);
1486        return;
1487    }
1488    if (rexw == 0 || val == (int32_t)val) {
1489        tcg_out_modrm(s, OPC_ARITH_EvIz + rexw, c, r0);
1490        tcg_out32(s, val);
1491        return;
1492    }
1493
1494    g_assert_not_reached();
1495}
1496
1497static void tcg_out_addi(TCGContext *s, int reg, tcg_target_long val)
1498{
1499    if (val != 0) {
1500        tgen_arithi(s, ARITH_ADD + P_REXW, reg, val, 0);
1501    }
1502}
1503
1504/* Set SMALL to force a short forward branch.  */
1505static void tcg_out_jxx(TCGContext *s, int opc, TCGLabel *l, bool small)
1506{
1507    int32_t val, val1;
1508
1509    if (l->has_value) {
1510        val = tcg_pcrel_diff(s, l->u.value_ptr);
1511        val1 = val - 2;
1512        if ((int8_t)val1 == val1) {
1513            if (opc == -1) {
1514                tcg_out8(s, OPC_JMP_short);
1515            } else {
1516                tcg_out8(s, OPC_JCC_short + opc);
1517            }
1518            tcg_out8(s, val1);
1519        } else {
1520            tcg_debug_assert(!small);
1521            if (opc == -1) {
1522                tcg_out8(s, OPC_JMP_long);
1523                tcg_out32(s, val - 5);
1524            } else {
1525                tcg_out_opc(s, OPC_JCC_long + opc, 0, 0, 0);
1526                tcg_out32(s, val - 6);
1527            }
1528        }
1529    } else if (small) {
1530        if (opc == -1) {
1531            tcg_out8(s, OPC_JMP_short);
1532        } else {
1533            tcg_out8(s, OPC_JCC_short + opc);
1534        }
1535        tcg_out_reloc(s, s->code_ptr, R_386_PC8, l, -1);
1536        s->code_ptr += 1;
1537    } else {
1538        if (opc == -1) {
1539            tcg_out8(s, OPC_JMP_long);
1540        } else {
1541            tcg_out_opc(s, OPC_JCC_long + opc, 0, 0, 0);
1542        }
1543        tcg_out_reloc(s, s->code_ptr, R_386_PC32, l, -4);
1544        s->code_ptr += 4;
1545    }
1546}
1547
1548static int tcg_out_cmp(TCGContext *s, TCGCond cond, TCGArg arg1,
1549                       TCGArg arg2, int const_arg2, int rexw)
1550{
1551    int jz, js;
1552
1553    if (!is_tst_cond(cond)) {
1554        if (!const_arg2) {
1555            tgen_arithr(s, ARITH_CMP + rexw, arg1, arg2);
1556        } else if (arg2 == 0) {
1557            tcg_out_modrm(s, OPC_TESTL + rexw, arg1, arg1);
1558        } else {
1559            tcg_debug_assert(!rexw || arg2 == (int32_t)arg2);
1560            tgen_arithi(s, ARITH_CMP + rexw, arg1, arg2, 0);
1561        }
1562        return tcg_cond_to_jcc[cond];
1563    }
1564
1565    jz = tcg_cond_to_jcc[cond];
1566    js = (cond == TCG_COND_TSTNE ? JCC_JS : JCC_JNS);
1567
1568    if (!const_arg2) {
1569        tcg_out_modrm(s, OPC_TESTL + rexw, arg1, arg2);
1570        return jz;
1571    }
1572
1573    if (arg2 <= 0xff && (TCG_TARGET_REG_BITS == 64 || arg1 < 4)) {
1574        if (arg2 == 0x80) {
1575            tcg_out_modrm(s, OPC_TESTB | P_REXB_R, arg1, arg1);
1576            return js;
1577        }
1578        if (arg2 == 0xff) {
1579            tcg_out_modrm(s, OPC_TESTB | P_REXB_R, arg1, arg1);
1580            return jz;
1581        }
1582        tcg_out_modrm(s, OPC_GRP3_Eb | P_REXB_RM, EXT3_TESTi, arg1);
1583        tcg_out8(s, arg2);
1584        return jz;
1585    }
1586
1587    if ((arg2 & ~0xff00) == 0 && arg1 < 4) {
1588        if (arg2 == 0x8000) {
1589            tcg_out_modrm(s, OPC_TESTB, arg1 + 4, arg1 + 4);
1590            return js;
1591        }
1592        if (arg2 == 0xff00) {
1593            tcg_out_modrm(s, OPC_TESTB, arg1 + 4, arg1 + 4);
1594            return jz;
1595        }
1596        tcg_out_modrm(s, OPC_GRP3_Eb, EXT3_TESTi, arg1 + 4);
1597        tcg_out8(s, arg2 >> 8);
1598        return jz;
1599    }
1600
1601    if (arg2 == 0xffff) {
1602        tcg_out_modrm(s, OPC_TESTL | P_DATA16, arg1, arg1);
1603        return jz;
1604    }
1605    if (arg2 == 0xffffffffu) {
1606        tcg_out_modrm(s, OPC_TESTL, arg1, arg1);
1607        return jz;
1608    }
1609
1610    if (is_power_of_2(rexw ? arg2 : (uint32_t)arg2)) {
1611        int jc = (cond == TCG_COND_TSTNE ? JCC_JB : JCC_JAE);
1612        int sh = ctz64(arg2);
1613
1614        rexw = (sh & 32 ? P_REXW : 0);
1615        if ((sh & 31) == 31) {
1616            tcg_out_modrm(s, OPC_TESTL | rexw, arg1, arg1);
1617            return js;
1618        } else {
1619            tcg_out_modrm(s, OPC_GRPBT | rexw, OPC_GRPBT_BT, arg1);
1620            tcg_out8(s, sh);
1621            return jc;
1622        }
1623    }
1624
1625    if (rexw) {
1626        if (arg2 == (uint32_t)arg2) {
1627            rexw = 0;
1628        } else {
1629            tcg_debug_assert(arg2 == (int32_t)arg2);
1630        }
1631    }
1632    tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_TESTi, arg1);
1633    tcg_out32(s, arg2);
1634    return jz;
1635}
1636
1637static void tcg_out_brcond(TCGContext *s, int rexw, TCGCond cond,
1638                           TCGArg arg1, TCGArg arg2, int const_arg2,
1639                           TCGLabel *label, bool small)
1640{
1641    int jcc = tcg_out_cmp(s, cond, arg1, arg2, const_arg2, rexw);
1642    tcg_out_jxx(s, jcc, label, small);
1643}
1644
1645#if TCG_TARGET_REG_BITS == 32
1646static void tcg_out_brcond2(TCGContext *s, const TCGArg *args,
1647                            const int *const_args, bool small)
1648{
1649    TCGLabel *label_next = gen_new_label();
1650    TCGLabel *label_this = arg_label(args[5]);
1651    TCGCond cond = args[4];
1652
1653    switch (cond) {
1654    case TCG_COND_EQ:
1655    case TCG_COND_TSTEQ:
1656        tcg_out_brcond(s, 0, tcg_invert_cond(cond),
1657                       args[0], args[2], const_args[2], label_next, 1);
1658        tcg_out_brcond(s, 0, cond, args[1], args[3], const_args[3],
1659                       label_this, small);
1660        break;
1661
1662    case TCG_COND_NE:
1663    case TCG_COND_TSTNE:
1664        tcg_out_brcond(s, 0, cond, args[0], args[2], const_args[2],
1665                       label_this, small);
1666        tcg_out_brcond(s, 0, cond, args[1], args[3], const_args[3],
1667                       label_this, small);
1668        break;
1669
1670    default:
1671        tcg_out_brcond(s, 0, tcg_high_cond(cond), args[1],
1672                       args[3], const_args[3], label_this, small);
1673        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1674        tcg_out_brcond(s, 0, tcg_unsigned_cond(cond), args[0],
1675                       args[2], const_args[2], label_this, small);
1676        break;
1677    }
1678    tcg_out_label(s, label_next);
1679}
1680#endif
1681
1682static void tcg_out_setcond(TCGContext *s, int rexw, TCGCond cond,
1683                            TCGArg dest, TCGArg arg1, TCGArg arg2,
1684                            int const_arg2, bool neg)
1685{
1686    int cmp_rexw = rexw;
1687    bool inv = false;
1688    bool cleared;
1689    int jcc;
1690
1691    switch (cond) {
1692    case TCG_COND_NE:
1693        inv = true;
1694        /* fall through */
1695    case TCG_COND_EQ:
1696        /* If arg2 is 0, convert to LTU/GEU vs 1. */
1697        if (const_arg2 && arg2 == 0) {
1698            arg2 = 1;
1699            goto do_ltu;
1700        }
1701        break;
1702
1703    case TCG_COND_TSTNE:
1704        inv = true;
1705        /* fall through */
1706    case TCG_COND_TSTEQ:
1707        /* If arg2 is -1, convert to LTU/GEU vs 1. */
1708        if (const_arg2 && arg2 == 0xffffffffu) {
1709            arg2 = 1;
1710            cmp_rexw = 0;
1711            goto do_ltu;
1712        }
1713        break;
1714
1715    case TCG_COND_LEU:
1716        inv = true;
1717        /* fall through */
1718    case TCG_COND_GTU:
1719        /* If arg2 is a register, swap for LTU/GEU. */
1720        if (!const_arg2) {
1721            TCGReg t = arg1;
1722            arg1 = arg2;
1723            arg2 = t;
1724            goto do_ltu;
1725        }
1726        break;
1727
1728    case TCG_COND_GEU:
1729        inv = true;
1730        /* fall through */
1731    case TCG_COND_LTU:
1732    do_ltu:
1733        /*
1734         * Relying on the carry bit, use SBB to produce -1 if LTU, 0 if GEU.
1735         * We can then use NEG or INC to produce the desired result.
1736         * This is always smaller than the SETCC expansion.
1737         */
1738        tcg_out_cmp(s, TCG_COND_LTU, arg1, arg2, const_arg2, cmp_rexw);
1739
1740        /* X - X - C = -C = (C ? -1 : 0) */
1741        tgen_arithr(s, ARITH_SBB + (neg ? rexw : 0), dest, dest);
1742        if (inv && neg) {
1743            /* ~(C ? -1 : 0) = (C ? 0 : -1) */
1744            tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NOT, dest);
1745        } else if (inv) {
1746            /* (C ? -1 : 0) + 1 = (C ? 0 : 1) */
1747            tgen_arithi(s, ARITH_ADD, dest, 1, 0);
1748        } else if (!neg) {
1749            /* -(C ? -1 : 0) = (C ? 1 : 0) */
1750            tcg_out_modrm(s, OPC_GRP3_Ev, EXT3_NEG, dest);
1751        }
1752        return;
1753
1754    case TCG_COND_GE:
1755        inv = true;
1756        /* fall through */
1757    case TCG_COND_LT:
1758        /* If arg2 is 0, extract the sign bit. */
1759        if (const_arg2 && arg2 == 0) {
1760            tcg_out_mov(s, rexw ? TCG_TYPE_I64 : TCG_TYPE_I32, dest, arg1);
1761            if (inv) {
1762                tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NOT, dest);
1763            }
1764            tcg_out_shifti(s, (neg ? SHIFT_SAR : SHIFT_SHR) + rexw,
1765                           dest, rexw ? 63 : 31);
1766            return;
1767        }
1768        break;
1769
1770    default:
1771        break;
1772    }
1773
1774    /*
1775     * If dest does not overlap the inputs, clearing it first is preferred.
1776     * The XOR breaks any false dependency for the low-byte write to dest,
1777     * and is also one byte smaller than MOVZBL.
1778     */
1779    cleared = false;
1780    if (dest != arg1 && (const_arg2 || dest != arg2)) {
1781        tgen_arithr(s, ARITH_XOR, dest, dest);
1782        cleared = true;
1783    }
1784
1785    jcc = tcg_out_cmp(s, cond, arg1, arg2, const_arg2, cmp_rexw);
1786    tcg_out_modrm(s, OPC_SETCC | jcc, 0, dest);
1787
1788    if (!cleared) {
1789        tcg_out_ext8u(s, dest, dest);
1790    }
1791    if (neg) {
1792        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NEG, dest);
1793    }
1794}
1795
1796#if TCG_TARGET_REG_BITS == 32
1797static void tcg_out_setcond2(TCGContext *s, const TCGArg *args,
1798                             const int *const_args)
1799{
1800    TCGArg new_args[6];
1801    TCGLabel *label_true, *label_over;
1802
1803    memcpy(new_args, args+1, 5*sizeof(TCGArg));
1804
1805    if (args[0] == args[1] || args[0] == args[2]
1806        || (!const_args[3] && args[0] == args[3])
1807        || (!const_args[4] && args[0] == args[4])) {
1808        /* When the destination overlaps with one of the argument
1809           registers, don't do anything tricky.  */
1810        label_true = gen_new_label();
1811        label_over = gen_new_label();
1812
1813        new_args[5] = label_arg(label_true);
1814        tcg_out_brcond2(s, new_args, const_args+1, 1);
1815
1816        tcg_out_movi(s, TCG_TYPE_I32, args[0], 0);
1817        tcg_out_jxx(s, JCC_JMP, label_over, 1);
1818        tcg_out_label(s, label_true);
1819
1820        tcg_out_movi(s, TCG_TYPE_I32, args[0], 1);
1821        tcg_out_label(s, label_over);
1822    } else {
1823        /* When the destination does not overlap one of the arguments,
1824           clear the destination first, jump if cond false, and emit an
1825           increment in the true case.  This results in smaller code.  */
1826
1827        tcg_out_movi(s, TCG_TYPE_I32, args[0], 0);
1828
1829        label_over = gen_new_label();
1830        new_args[4] = tcg_invert_cond(new_args[4]);
1831        new_args[5] = label_arg(label_over);
1832        tcg_out_brcond2(s, new_args, const_args+1, 1);
1833
1834        tgen_arithi(s, ARITH_ADD, args[0], 1, 0);
1835        tcg_out_label(s, label_over);
1836    }
1837}
1838#endif
1839
1840static void tcg_out_cmov(TCGContext *s, int jcc, int rexw,
1841                         TCGReg dest, TCGReg v1)
1842{
1843    tcg_out_modrm(s, OPC_CMOVCC | jcc | rexw, dest, v1);
1844}
1845
1846static void tcg_out_movcond(TCGContext *s, int rexw, TCGCond cond,
1847                            TCGReg dest, TCGReg c1, TCGArg c2, int const_c2,
1848                            TCGReg v1)
1849{
1850    int jcc = tcg_out_cmp(s, cond, c1, c2, const_c2, rexw);
1851    tcg_out_cmov(s, jcc, rexw, dest, v1);
1852}
1853
1854static void tcg_out_ctz(TCGContext *s, int rexw, TCGReg dest, TCGReg arg1,
1855                        TCGArg arg2, bool const_a2)
1856{
1857    if (have_bmi1) {
1858        tcg_out_modrm(s, OPC_TZCNT + rexw, dest, arg1);
1859        if (const_a2) {
1860            tcg_debug_assert(arg2 == (rexw ? 64 : 32));
1861        } else {
1862            tcg_debug_assert(dest != arg2);
1863            tcg_out_cmov(s, JCC_JB, rexw, dest, arg2);
1864        }
1865    } else {
1866        tcg_debug_assert(dest != arg2);
1867        tcg_out_modrm(s, OPC_BSF + rexw, dest, arg1);
1868        tcg_out_cmov(s, JCC_JE, rexw, dest, arg2);
1869    }
1870}
1871
1872static void tcg_out_clz(TCGContext *s, int rexw, TCGReg dest, TCGReg arg1,
1873                        TCGArg arg2, bool const_a2)
1874{
1875    if (have_lzcnt) {
1876        tcg_out_modrm(s, OPC_LZCNT + rexw, dest, arg1);
1877        if (const_a2) {
1878            tcg_debug_assert(arg2 == (rexw ? 64 : 32));
1879        } else {
1880            tcg_debug_assert(dest != arg2);
1881            tcg_out_cmov(s, JCC_JB, rexw, dest, arg2);
1882        }
1883    } else {
1884        tcg_debug_assert(!const_a2);
1885        tcg_debug_assert(dest != arg1);
1886        tcg_debug_assert(dest != arg2);
1887
1888        /* Recall that the output of BSR is the index not the count.  */
1889        tcg_out_modrm(s, OPC_BSR + rexw, dest, arg1);
1890        tgen_arithi(s, ARITH_XOR + rexw, dest, rexw ? 63 : 31, 0);
1891
1892        /* Since we have destroyed the flags from BSR, we have to re-test.  */
1893        int jcc = tcg_out_cmp(s, TCG_COND_EQ, arg1, 0, 1, rexw);
1894        tcg_out_cmov(s, jcc, rexw, dest, arg2);
1895    }
1896}
1897
1898static void tcg_out_branch(TCGContext *s, int call, const tcg_insn_unit *dest)
1899{
1900    intptr_t disp = tcg_pcrel_diff(s, dest) - 5;
1901
1902    if (disp == (int32_t)disp) {
1903        tcg_out_opc(s, call ? OPC_CALL_Jz : OPC_JMP_long, 0, 0, 0);
1904        tcg_out32(s, disp);
1905    } else {
1906        /* rip-relative addressing into the constant pool.
1907           This is 6 + 8 = 14 bytes, as compared to using an
1908           immediate load 10 + 6 = 16 bytes, plus we may
1909           be able to re-use the pool constant for more calls.  */
1910        tcg_out_opc(s, OPC_GRP5, 0, 0, 0);
1911        tcg_out8(s, (call ? EXT5_CALLN_Ev : EXT5_JMPN_Ev) << 3 | 5);
1912        new_pool_label(s, (uintptr_t)dest, R_386_PC32, s->code_ptr, -4);
1913        tcg_out32(s, 0);
1914    }
1915}
1916
1917static void tcg_out_call(TCGContext *s, const tcg_insn_unit *dest,
1918                         const TCGHelperInfo *info)
1919{
1920    tcg_out_branch(s, 1, dest);
1921
1922#ifndef _WIN32
1923    if (TCG_TARGET_REG_BITS == 32 && info->out_kind == TCG_CALL_RET_BY_REF) {
1924        /*
1925         * The sysv i386 abi for struct return places a reference as the
1926         * first argument of the stack, and pops that argument with the
1927         * return statement.  Since we want to retain the aligned stack
1928         * pointer for the callee, we do not want to actually push that
1929         * argument before the call but rely on the normal store to the
1930         * stack slot.  But we do need to compensate for the pop in order
1931         * to reset our correct stack pointer value.
1932         * Pushing a garbage value back onto the stack is quickest.
1933         */
1934        tcg_out_push(s, TCG_REG_EAX);
1935    }
1936#endif
1937}
1938
1939static void tcg_out_jmp(TCGContext *s, const tcg_insn_unit *dest)
1940{
1941    tcg_out_branch(s, 0, dest);
1942}
1943
1944static void tcg_out_nopn(TCGContext *s, int n)
1945{
1946    int i;
1947    /* Emit 1 or 2 operand size prefixes for the standard one byte nop,
1948     * "xchg %eax,%eax", forming "xchg %ax,%ax". All cores accept the
1949     * duplicate prefix, and all of the interesting recent cores can
1950     * decode and discard the duplicates in a single cycle.
1951     */
1952    tcg_debug_assert(n >= 1);
1953    for (i = 1; i < n; ++i) {
1954        tcg_out8(s, 0x66);
1955    }
1956    tcg_out8(s, 0x90);
1957}
1958
1959typedef struct {
1960    TCGReg base;
1961    int index;
1962    int ofs;
1963    int seg;
1964    TCGAtomAlign aa;
1965} HostAddress;
1966
1967bool tcg_target_has_memory_bswap(MemOp memop)
1968{
1969    TCGAtomAlign aa;
1970
1971    if (!have_movbe) {
1972        return false;
1973    }
1974    if ((memop & MO_SIZE) < MO_128) {
1975        return true;
1976    }
1977
1978    /*
1979     * Reject 16-byte memop with 16-byte atomicity, i.e. VMOVDQA,
1980     * but do allow a pair of 64-bit operations, i.e. MOVBEQ.
1981     */
1982    aa = atom_and_align_for_opc(tcg_ctx, memop, MO_ATOM_IFALIGN, true);
1983    return aa.atom < MO_128;
1984}
1985
1986/*
1987 * Because i686 has no register parameters and because x86_64 has xchg
1988 * to handle addr/data register overlap, we have placed all input arguments
1989 * before we need might need a scratch reg.
1990 *
1991 * Even then, a scratch is only needed for l->raddr.  Rather than expose
1992 * a general-purpose scratch when we don't actually know it's available,
1993 * use the ra_gen hook to load into RAX if needed.
1994 */
1995#if TCG_TARGET_REG_BITS == 64
1996static TCGReg ldst_ra_gen(TCGContext *s, const TCGLabelQemuLdst *l, int arg)
1997{
1998    if (arg < 0) {
1999        arg = TCG_REG_RAX;
2000    }
2001    tcg_out_movi(s, TCG_TYPE_PTR, arg, (uintptr_t)l->raddr);
2002    return arg;
2003}
2004static const TCGLdstHelperParam ldst_helper_param = {
2005    .ra_gen = ldst_ra_gen
2006};
2007#else
2008static const TCGLdstHelperParam ldst_helper_param = { };
2009#endif
2010
2011static void tcg_out_vec_to_pair(TCGContext *s, TCGType type,
2012                                TCGReg l, TCGReg h, TCGReg v)
2013{
2014    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2015
2016    /* vpmov{d,q} %v, %l */
2017    tcg_out_vex_modrm(s, OPC_MOVD_EyVy + rexw, v, 0, l);
2018    /* vpextr{d,q} $1, %v, %h */
2019    tcg_out_vex_modrm(s, OPC_PEXTRD + rexw, v, 0, h);
2020    tcg_out8(s, 1);
2021}
2022
2023static void tcg_out_pair_to_vec(TCGContext *s, TCGType type,
2024                                TCGReg v, TCGReg l, TCGReg h)
2025{
2026    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2027
2028    /* vmov{d,q} %l, %v */
2029    tcg_out_vex_modrm(s, OPC_MOVD_VyEy + rexw, v, 0, l);
2030    /* vpinsr{d,q} $1, %h, %v, %v */
2031    tcg_out_vex_modrm(s, OPC_PINSRD + rexw, v, v, h);
2032    tcg_out8(s, 1);
2033}
2034
2035/*
2036 * Generate code for the slow path for a load at the end of block
2037 */
2038static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
2039{
2040    MemOp opc = get_memop(l->oi);
2041    tcg_insn_unit **label_ptr = &l->label_ptr[0];
2042
2043    /* resolve label address */
2044    tcg_patch32(label_ptr[0], s->code_ptr - label_ptr[0] - 4);
2045    if (label_ptr[1]) {
2046        tcg_patch32(label_ptr[1], s->code_ptr - label_ptr[1] - 4);
2047    }
2048
2049    tcg_out_ld_helper_args(s, l, &ldst_helper_param);
2050    tcg_out_branch(s, 1, qemu_ld_helpers[opc & MO_SIZE]);
2051    tcg_out_ld_helper_ret(s, l, false, &ldst_helper_param);
2052
2053    tcg_out_jmp(s, l->raddr);
2054    return true;
2055}
2056
2057/*
2058 * Generate code for the slow path for a store at the end of block
2059 */
2060static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
2061{
2062    MemOp opc = get_memop(l->oi);
2063    tcg_insn_unit **label_ptr = &l->label_ptr[0];
2064
2065    /* resolve label address */
2066    tcg_patch32(label_ptr[0], s->code_ptr - label_ptr[0] - 4);
2067    if (label_ptr[1]) {
2068        tcg_patch32(label_ptr[1], s->code_ptr - label_ptr[1] - 4);
2069    }
2070
2071    tcg_out_st_helper_args(s, l, &ldst_helper_param);
2072    tcg_out_branch(s, 1, qemu_st_helpers[opc & MO_SIZE]);
2073
2074    tcg_out_jmp(s, l->raddr);
2075    return true;
2076}
2077
2078#ifdef CONFIG_USER_ONLY
2079static HostAddress x86_guest_base = {
2080    .index = -1
2081};
2082
2083#if defined(__x86_64__) && defined(__linux__)
2084# include <asm/prctl.h>
2085# include <sys/prctl.h>
2086int arch_prctl(int code, unsigned long addr);
2087static inline int setup_guest_base_seg(void)
2088{
2089    if (arch_prctl(ARCH_SET_GS, guest_base) == 0) {
2090        return P_GS;
2091    }
2092    return 0;
2093}
2094#define setup_guest_base_seg  setup_guest_base_seg
2095#elif defined(__x86_64__) && \
2096      (defined (__FreeBSD__) || defined (__FreeBSD_kernel__))
2097# include <machine/sysarch.h>
2098static inline int setup_guest_base_seg(void)
2099{
2100    if (sysarch(AMD64_SET_GSBASE, &guest_base) == 0) {
2101        return P_GS;
2102    }
2103    return 0;
2104}
2105#define setup_guest_base_seg  setup_guest_base_seg
2106#endif
2107#else
2108# define x86_guest_base (*(HostAddress *)({ qemu_build_not_reached(); NULL; }))
2109#endif /* CONFIG_USER_ONLY */
2110#ifndef setup_guest_base_seg
2111# define setup_guest_base_seg()  0
2112#endif
2113
2114#define MIN_TLB_MASK_TABLE_OFS  INT_MIN
2115
2116/*
2117 * For softmmu, perform the TLB load and compare.
2118 * For useronly, perform any required alignment tests.
2119 * In both cases, return a TCGLabelQemuLdst structure if the slow path
2120 * is required and fill in @h with the host address for the fast path.
2121 */
2122static TCGLabelQemuLdst *prepare_host_addr(TCGContext *s, HostAddress *h,
2123                                           TCGReg addr, MemOpIdx oi, bool is_ld)
2124{
2125    TCGLabelQemuLdst *ldst = NULL;
2126    MemOp opc = get_memop(oi);
2127    MemOp s_bits = opc & MO_SIZE;
2128    unsigned a_mask;
2129
2130    if (tcg_use_softmmu) {
2131        h->index = TCG_REG_L0;
2132        h->ofs = 0;
2133        h->seg = 0;
2134    } else {
2135        *h = x86_guest_base;
2136    }
2137    h->base = addr;
2138    h->aa = atom_and_align_for_opc(s, opc, MO_ATOM_IFALIGN, s_bits == MO_128);
2139    a_mask = (1 << h->aa.align) - 1;
2140
2141    if (tcg_use_softmmu) {
2142        int cmp_ofs = is_ld ? offsetof(CPUTLBEntry, addr_read)
2143                            : offsetof(CPUTLBEntry, addr_write);
2144        TCGType ttype = TCG_TYPE_I32;
2145        TCGType tlbtype = TCG_TYPE_I32;
2146        int trexw = 0, hrexw = 0, tlbrexw = 0;
2147        unsigned mem_index = get_mmuidx(oi);
2148        unsigned s_mask = (1 << s_bits) - 1;
2149        int fast_ofs = tlb_mask_table_ofs(s, mem_index);
2150        int tlb_mask;
2151
2152        ldst = new_ldst_label(s);
2153        ldst->is_ld = is_ld;
2154        ldst->oi = oi;
2155        ldst->addr_reg = addr;
2156
2157        if (TCG_TARGET_REG_BITS == 64) {
2158            ttype = s->addr_type;
2159            trexw = (ttype == TCG_TYPE_I32 ? 0 : P_REXW);
2160            if (TCG_TYPE_PTR == TCG_TYPE_I64) {
2161                hrexw = P_REXW;
2162                if (s->page_bits + s->tlb_dyn_max_bits > 32) {
2163                    tlbtype = TCG_TYPE_I64;
2164                    tlbrexw = P_REXW;
2165                }
2166            }
2167        }
2168
2169        tcg_out_mov(s, tlbtype, TCG_REG_L0, addr);
2170        tcg_out_shifti(s, SHIFT_SHR + tlbrexw, TCG_REG_L0,
2171                       s->page_bits - CPU_TLB_ENTRY_BITS);
2172
2173        tcg_out_modrm_offset(s, OPC_AND_GvEv + trexw, TCG_REG_L0, TCG_AREG0,
2174                             fast_ofs + offsetof(CPUTLBDescFast, mask));
2175
2176        tcg_out_modrm_offset(s, OPC_ADD_GvEv + hrexw, TCG_REG_L0, TCG_AREG0,
2177                             fast_ofs + offsetof(CPUTLBDescFast, table));
2178
2179        /*
2180         * If the required alignment is at least as large as the access,
2181         * simply copy the address and mask.  For lesser alignments,
2182         * check that we don't cross pages for the complete access.
2183         */
2184        if (a_mask >= s_mask) {
2185            tcg_out_mov(s, ttype, TCG_REG_L1, addr);
2186        } else {
2187            tcg_out_modrm_offset(s, OPC_LEA + trexw, TCG_REG_L1,
2188                                 addr, s_mask - a_mask);
2189        }
2190        tlb_mask = s->page_mask | a_mask;
2191        tgen_arithi(s, ARITH_AND + trexw, TCG_REG_L1, tlb_mask, 0);
2192
2193        /* cmp 0(TCG_REG_L0), TCG_REG_L1 */
2194        tcg_out_modrm_offset(s, OPC_CMP_GvEv + trexw,
2195                             TCG_REG_L1, TCG_REG_L0, cmp_ofs);
2196
2197        /* jne slow_path */
2198        tcg_out_opc(s, OPC_JCC_long + JCC_JNE, 0, 0, 0);
2199        ldst->label_ptr[0] = s->code_ptr;
2200        s->code_ptr += 4;
2201
2202        /* TLB Hit.  */
2203        tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_L0, TCG_REG_L0,
2204                   offsetof(CPUTLBEntry, addend));
2205    } else if (a_mask) {
2206        int jcc;
2207
2208        ldst = new_ldst_label(s);
2209        ldst->is_ld = is_ld;
2210        ldst->oi = oi;
2211        ldst->addr_reg = addr;
2212
2213        /* jne slow_path */
2214        jcc = tcg_out_cmp(s, TCG_COND_TSTNE, addr, a_mask, true, false);
2215        tcg_out_opc(s, OPC_JCC_long + jcc, 0, 0, 0);
2216        ldst->label_ptr[0] = s->code_ptr;
2217        s->code_ptr += 4;
2218    }
2219
2220    return ldst;
2221}
2222
2223static void tcg_out_qemu_ld_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
2224                                   HostAddress h, TCGType type, MemOp memop)
2225{
2226    bool use_movbe = false;
2227    int rexw = (type == TCG_TYPE_I32 ? 0 : P_REXW);
2228    int movop = OPC_MOVL_GvEv;
2229
2230    /* Do big-endian loads with movbe.  */
2231    if (memop & MO_BSWAP) {
2232        tcg_debug_assert(have_movbe);
2233        use_movbe = true;
2234        movop = OPC_MOVBE_GyMy;
2235    }
2236
2237    switch (memop & MO_SSIZE) {
2238    case MO_UB:
2239        tcg_out_modrm_sib_offset(s, OPC_MOVZBL + h.seg, datalo,
2240                                 h.base, h.index, 0, h.ofs);
2241        break;
2242    case MO_SB:
2243        tcg_out_modrm_sib_offset(s, OPC_MOVSBL + rexw + h.seg, datalo,
2244                                 h.base, h.index, 0, h.ofs);
2245        break;
2246    case MO_UW:
2247        if (use_movbe) {
2248            /* There is no extending movbe; only low 16-bits are modified.  */
2249            if (datalo != h.base && datalo != h.index) {
2250                /* XOR breaks dependency chains.  */
2251                tgen_arithr(s, ARITH_XOR, datalo, datalo);
2252                tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + P_DATA16 + h.seg,
2253                                         datalo, h.base, h.index, 0, h.ofs);
2254            } else {
2255                tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + P_DATA16 + h.seg,
2256                                         datalo, h.base, h.index, 0, h.ofs);
2257                tcg_out_ext16u(s, datalo, datalo);
2258            }
2259        } else {
2260            tcg_out_modrm_sib_offset(s, OPC_MOVZWL + h.seg, datalo,
2261                                     h.base, h.index, 0, h.ofs);
2262        }
2263        break;
2264    case MO_SW:
2265        if (use_movbe) {
2266            tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + P_DATA16 + h.seg,
2267                                     datalo, h.base, h.index, 0, h.ofs);
2268            tcg_out_ext16s(s, type, datalo, datalo);
2269        } else {
2270            tcg_out_modrm_sib_offset(s, OPC_MOVSWL + rexw + h.seg,
2271                                     datalo, h.base, h.index, 0, h.ofs);
2272        }
2273        break;
2274    case MO_UL:
2275        tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2276                                 h.base, h.index, 0, h.ofs);
2277        break;
2278#if TCG_TARGET_REG_BITS == 64
2279    case MO_SL:
2280        if (use_movbe) {
2281            tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + h.seg, datalo,
2282                                     h.base, h.index, 0, h.ofs);
2283            tcg_out_ext32s(s, datalo, datalo);
2284        } else {
2285            tcg_out_modrm_sib_offset(s, OPC_MOVSLQ + h.seg, datalo,
2286                                     h.base, h.index, 0, h.ofs);
2287        }
2288        break;
2289#endif
2290    case MO_UQ:
2291        if (TCG_TARGET_REG_BITS == 64) {
2292            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2293                                     h.base, h.index, 0, h.ofs);
2294            break;
2295        }
2296        if (use_movbe) {
2297            TCGReg t = datalo;
2298            datalo = datahi;
2299            datahi = t;
2300        }
2301        if (h.base == datalo || h.index == datalo) {
2302            tcg_out_modrm_sib_offset(s, OPC_LEA, datahi,
2303                                     h.base, h.index, 0, h.ofs);
2304            tcg_out_modrm_offset(s, movop + h.seg, datalo, datahi, 0);
2305            tcg_out_modrm_offset(s, movop + h.seg, datahi, datahi, 4);
2306        } else {
2307            tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2308                                     h.base, h.index, 0, h.ofs);
2309            tcg_out_modrm_sib_offset(s, movop + h.seg, datahi,
2310                                     h.base, h.index, 0, h.ofs + 4);
2311        }
2312        break;
2313
2314    case MO_128:
2315        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
2316
2317        /*
2318         * Without 16-byte atomicity, use integer regs.
2319         * That is where we want the data, and it allows bswaps.
2320         */
2321        if (h.aa.atom < MO_128) {
2322            if (use_movbe) {
2323                TCGReg t = datalo;
2324                datalo = datahi;
2325                datahi = t;
2326            }
2327            if (h.base == datalo || h.index == datalo) {
2328                tcg_out_modrm_sib_offset(s, OPC_LEA + P_REXW, datahi,
2329                                         h.base, h.index, 0, h.ofs);
2330                tcg_out_modrm_offset(s, movop + P_REXW + h.seg,
2331                                     datalo, datahi, 0);
2332                tcg_out_modrm_offset(s, movop + P_REXW + h.seg,
2333                                     datahi, datahi, 8);
2334            } else {
2335                tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2336                                         h.base, h.index, 0, h.ofs);
2337                tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datahi,
2338                                         h.base, h.index, 0, h.ofs + 8);
2339            }
2340            break;
2341        }
2342
2343        /*
2344         * With 16-byte atomicity, a vector load is required.
2345         * If we already have 16-byte alignment, then VMOVDQA always works.
2346         * Else if VMOVDQU has atomicity with dynamic alignment, use that.
2347         * Else use we require a runtime test for alignment for VMOVDQA;
2348         * use VMOVDQU on the unaligned nonatomic path for simplicity.
2349         */
2350        if (h.aa.align >= MO_128) {
2351            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_VxWx + h.seg,
2352                                         TCG_TMP_VEC, 0,
2353                                         h.base, h.index, 0, h.ofs);
2354        } else if (cpuinfo & CPUINFO_ATOMIC_VMOVDQU) {
2355            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_VxWx + h.seg,
2356                                         TCG_TMP_VEC, 0,
2357                                         h.base, h.index, 0, h.ofs);
2358        } else {
2359            TCGLabel *l1 = gen_new_label();
2360            TCGLabel *l2 = gen_new_label();
2361            int jcc;
2362
2363            jcc = tcg_out_cmp(s, TCG_COND_TSTNE, h.base, 15, true, false);
2364            tcg_out_jxx(s, jcc, l1, true);
2365
2366            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_VxWx + h.seg,
2367                                         TCG_TMP_VEC, 0,
2368                                         h.base, h.index, 0, h.ofs);
2369            tcg_out_jxx(s, JCC_JMP, l2, true);
2370
2371            tcg_out_label(s, l1);
2372            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_VxWx + h.seg,
2373                                         TCG_TMP_VEC, 0,
2374                                         h.base, h.index, 0, h.ofs);
2375            tcg_out_label(s, l2);
2376        }
2377        tcg_out_vec_to_pair(s, TCG_TYPE_I64, datalo, datahi, TCG_TMP_VEC);
2378        break;
2379
2380    default:
2381        g_assert_not_reached();
2382    }
2383}
2384
2385static void tcg_out_qemu_ld(TCGContext *s, TCGReg datalo, TCGReg datahi,
2386                            TCGReg addr, MemOpIdx oi, TCGType data_type)
2387{
2388    TCGLabelQemuLdst *ldst;
2389    HostAddress h;
2390
2391    ldst = prepare_host_addr(s, &h, addr, oi, true);
2392    tcg_out_qemu_ld_direct(s, datalo, datahi, h, data_type, get_memop(oi));
2393
2394    if (ldst) {
2395        ldst->type = data_type;
2396        ldst->datalo_reg = datalo;
2397        ldst->datahi_reg = datahi;
2398        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
2399    }
2400}
2401
2402static void tcg_out_qemu_st_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
2403                                   HostAddress h, MemOp memop)
2404{
2405    bool use_movbe = false;
2406    int movop = OPC_MOVL_EvGv;
2407
2408    /*
2409     * Do big-endian stores with movbe or system-mode.
2410     * User-only without movbe will have its swapping done generically.
2411     */
2412    if (memop & MO_BSWAP) {
2413        tcg_debug_assert(have_movbe);
2414        use_movbe = true;
2415        movop = OPC_MOVBE_MyGy;
2416    }
2417
2418    switch (memop & MO_SIZE) {
2419    case MO_8:
2420        /* This is handled with constraints on INDEX_op_qemu_st8_i32. */
2421        tcg_debug_assert(TCG_TARGET_REG_BITS == 64 || datalo < 4);
2422        tcg_out_modrm_sib_offset(s, OPC_MOVB_EvGv + P_REXB_R + h.seg,
2423                                 datalo, h.base, h.index, 0, h.ofs);
2424        break;
2425    case MO_16:
2426        tcg_out_modrm_sib_offset(s, movop + P_DATA16 + h.seg, datalo,
2427                                 h.base, h.index, 0, h.ofs);
2428        break;
2429    case MO_32:
2430        tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2431                                 h.base, h.index, 0, h.ofs);
2432        break;
2433    case MO_64:
2434        if (TCG_TARGET_REG_BITS == 64) {
2435            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2436                                     h.base, h.index, 0, h.ofs);
2437        } else {
2438            if (use_movbe) {
2439                TCGReg t = datalo;
2440                datalo = datahi;
2441                datahi = t;
2442            }
2443            tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2444                                     h.base, h.index, 0, h.ofs);
2445            tcg_out_modrm_sib_offset(s, movop + h.seg, datahi,
2446                                     h.base, h.index, 0, h.ofs + 4);
2447        }
2448        break;
2449
2450    case MO_128:
2451        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
2452
2453        /*
2454         * Without 16-byte atomicity, use integer regs.
2455         * That is where we have the data, and it allows bswaps.
2456         */
2457        if (h.aa.atom < MO_128) {
2458            if (use_movbe) {
2459                TCGReg t = datalo;
2460                datalo = datahi;
2461                datahi = t;
2462            }
2463            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2464                                     h.base, h.index, 0, h.ofs);
2465            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datahi,
2466                                     h.base, h.index, 0, h.ofs + 8);
2467            break;
2468        }
2469
2470        /*
2471         * With 16-byte atomicity, a vector store is required.
2472         * If we already have 16-byte alignment, then VMOVDQA always works.
2473         * Else if VMOVDQU has atomicity with dynamic alignment, use that.
2474         * Else use we require a runtime test for alignment for VMOVDQA;
2475         * use VMOVDQU on the unaligned nonatomic path for simplicity.
2476         */
2477        tcg_out_pair_to_vec(s, TCG_TYPE_I64, TCG_TMP_VEC, datalo, datahi);
2478        if (h.aa.align >= MO_128) {
2479            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_WxVx + h.seg,
2480                                         TCG_TMP_VEC, 0,
2481                                         h.base, h.index, 0, h.ofs);
2482        } else if (cpuinfo & CPUINFO_ATOMIC_VMOVDQU) {
2483            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_WxVx + h.seg,
2484                                         TCG_TMP_VEC, 0,
2485                                         h.base, h.index, 0, h.ofs);
2486        } else {
2487            TCGLabel *l1 = gen_new_label();
2488            TCGLabel *l2 = gen_new_label();
2489            int jcc;
2490
2491            jcc = tcg_out_cmp(s, TCG_COND_TSTNE, h.base, 15, true, false);
2492            tcg_out_jxx(s, jcc, l1, true);
2493
2494            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_WxVx + h.seg,
2495                                         TCG_TMP_VEC, 0,
2496                                         h.base, h.index, 0, h.ofs);
2497            tcg_out_jxx(s, JCC_JMP, l2, true);
2498
2499            tcg_out_label(s, l1);
2500            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_WxVx + h.seg,
2501                                         TCG_TMP_VEC, 0,
2502                                         h.base, h.index, 0, h.ofs);
2503            tcg_out_label(s, l2);
2504        }
2505        break;
2506
2507    default:
2508        g_assert_not_reached();
2509    }
2510}
2511
2512static void tcg_out_qemu_st(TCGContext *s, TCGReg datalo, TCGReg datahi,
2513                            TCGReg addr, MemOpIdx oi, TCGType data_type)
2514{
2515    TCGLabelQemuLdst *ldst;
2516    HostAddress h;
2517
2518    ldst = prepare_host_addr(s, &h, addr, oi, false);
2519    tcg_out_qemu_st_direct(s, datalo, datahi, h, get_memop(oi));
2520
2521    if (ldst) {
2522        ldst->type = data_type;
2523        ldst->datalo_reg = datalo;
2524        ldst->datahi_reg = datahi;
2525        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
2526    }
2527}
2528
2529static void tcg_out_exit_tb(TCGContext *s, uintptr_t a0)
2530{
2531    /* Reuse the zeroing that exists for goto_ptr.  */
2532    if (a0 == 0) {
2533        tcg_out_jmp(s, tcg_code_gen_epilogue);
2534    } else {
2535        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_EAX, a0);
2536        tcg_out_jmp(s, tb_ret_addr);
2537    }
2538}
2539
2540static void tcg_out_goto_tb(TCGContext *s, int which)
2541{
2542    /*
2543     * Jump displacement must be aligned for atomic patching;
2544     * see if we need to add extra nops before jump
2545     */
2546    int gap = QEMU_ALIGN_PTR_UP(s->code_ptr + 1, 4) - s->code_ptr;
2547    if (gap != 1) {
2548        tcg_out_nopn(s, gap - 1);
2549    }
2550    tcg_out8(s, OPC_JMP_long); /* jmp im */
2551    set_jmp_insn_offset(s, which);
2552    tcg_out32(s, 0);
2553    set_jmp_reset_offset(s, which);
2554}
2555
2556void tb_target_set_jmp_target(const TranslationBlock *tb, int n,
2557                              uintptr_t jmp_rx, uintptr_t jmp_rw)
2558{
2559    /* patch the branch destination */
2560    uintptr_t addr = tb->jmp_target_addr[n];
2561    qatomic_set((int32_t *)jmp_rw, addr - (jmp_rx + 4));
2562    /* no need to flush icache explicitly */
2563}
2564
2565
2566static void tgen_add(TCGContext *s, TCGType type,
2567                     TCGReg a0, TCGReg a1, TCGReg a2)
2568{
2569    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2570
2571    if (a0 == a1) {
2572        tgen_arithr(s, ARITH_ADD + rexw, a0, a2);
2573    } else if (a0 == a2) {
2574        tgen_arithr(s, ARITH_ADD + rexw, a0, a1);
2575    } else {
2576        tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, a1, a2, 0, 0);
2577    }
2578}
2579
2580static void tgen_addi(TCGContext *s, TCGType type,
2581                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2582{
2583    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2584
2585    if (a0 == a1) {
2586        tgen_arithi(s, ARITH_ADD + rexw, a0, a2, false);
2587    } else {
2588        tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, a1, -1, 0, a2);
2589    }
2590}
2591
2592static const TCGOutOpBinary outop_add = {
2593    .base.static_constraint = C_O1_I2(r, r, re),
2594    .out_rrr = tgen_add,
2595    .out_rri = tgen_addi,
2596};
2597
2598static void tgen_and(TCGContext *s, TCGType type,
2599                     TCGReg a0, TCGReg a1, TCGReg a2)
2600{
2601    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2602    tgen_arithr(s, ARITH_AND + rexw, a0, a2);
2603}
2604
2605static void tgen_andi(TCGContext *s, TCGType type,
2606                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2607{
2608    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2609    tgen_arithi(s, ARITH_AND + rexw, a0, a2, false);
2610}
2611
2612static const TCGOutOpBinary outop_and = {
2613    .base.static_constraint = C_O1_I2(r, 0, reZ),
2614    .out_rrr = tgen_and,
2615    .out_rri = tgen_andi,
2616};
2617
2618static void tgen_andc(TCGContext *s, TCGType type,
2619                      TCGReg a0, TCGReg a1, TCGReg a2)
2620{
2621    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2622    tcg_out_vex_modrm(s, OPC_ANDN + rexw, a0, a2, a1);
2623}
2624
2625static TCGConstraintSetIndex cset_andc(TCGType type, unsigned flags)
2626{
2627    return have_bmi1 ? C_O1_I2(r, r, r) : C_NotImplemented;
2628}
2629
2630static const TCGOutOpBinary outop_andc = {
2631    .base.static_constraint = C_Dynamic,
2632    .base.dynamic_constraint = cset_andc,
2633    .out_rrr = tgen_andc,
2634};
2635
2636static const TCGOutOpBinary outop_eqv = {
2637    .base.static_constraint = C_NotImplemented,
2638};
2639
2640static const TCGOutOpBinary outop_nand = {
2641    .base.static_constraint = C_NotImplemented,
2642};
2643
2644static const TCGOutOpBinary outop_nor = {
2645    .base.static_constraint = C_NotImplemented,
2646};
2647
2648static void tgen_or(TCGContext *s, TCGType type,
2649                    TCGReg a0, TCGReg a1, TCGReg a2)
2650{
2651    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2652    tgen_arithr(s, ARITH_OR + rexw, a0, a2);
2653}
2654
2655static void tgen_ori(TCGContext *s, TCGType type,
2656                     TCGReg a0, TCGReg a1, tcg_target_long a2)
2657{
2658    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2659    tgen_arithi(s, ARITH_OR + rexw, a0, a2, false);
2660}
2661
2662static const TCGOutOpBinary outop_or = {
2663    .base.static_constraint = C_O1_I2(r, 0, re),
2664    .out_rrr = tgen_or,
2665    .out_rri = tgen_ori,
2666};
2667
2668static const TCGOutOpBinary outop_orc = {
2669    .base.static_constraint = C_NotImplemented,
2670};
2671
2672static void tgen_xor(TCGContext *s, TCGType type,
2673                     TCGReg a0, TCGReg a1, TCGReg a2)
2674{
2675    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2676    tgen_arithr(s, ARITH_XOR + rexw, a0, a2);
2677}
2678
2679static void tgen_xori(TCGContext *s, TCGType type,
2680                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2681{
2682    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2683    tgen_arithi(s, ARITH_XOR + rexw, a0, a2, false);
2684}
2685
2686static const TCGOutOpBinary outop_xor = {
2687    .base.static_constraint = C_O1_I2(r, 0, re),
2688    .out_rrr = tgen_xor,
2689    .out_rri = tgen_xori,
2690};
2691
2692
2693static void tcg_out_op(TCGContext *s, TCGOpcode opc, TCGType type,
2694                       const TCGArg args[TCG_MAX_OP_ARGS],
2695                       const int const_args[TCG_MAX_OP_ARGS])
2696{
2697    TCGArg a0, a1, a2;
2698    int c, const_a2, vexop, rexw;
2699
2700#if TCG_TARGET_REG_BITS == 64
2701# define OP_32_64(x) \
2702        case glue(glue(INDEX_op_, x), _i64): \
2703        case glue(glue(INDEX_op_, x), _i32)
2704#else
2705# define OP_32_64(x) \
2706        case glue(glue(INDEX_op_, x), _i32)
2707#endif
2708
2709    /* Hoist the loads of the most common arguments.  */
2710    a0 = args[0];
2711    a1 = args[1];
2712    a2 = args[2];
2713    const_a2 = const_args[2];
2714    rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2715
2716    switch (opc) {
2717    case INDEX_op_goto_ptr:
2718        /* jmp to the given host address (could be epilogue) */
2719        tcg_out_modrm(s, OPC_GRP5, EXT5_JMPN_Ev, a0);
2720        break;
2721    case INDEX_op_br:
2722        tcg_out_jxx(s, JCC_JMP, arg_label(a0), 0);
2723        break;
2724    OP_32_64(ld8u):
2725        /* Note that we can ignore REXW for the zero-extend to 64-bit.  */
2726        tcg_out_modrm_offset(s, OPC_MOVZBL, a0, a1, a2);
2727        break;
2728    OP_32_64(ld8s):
2729        tcg_out_modrm_offset(s, OPC_MOVSBL + rexw, a0, a1, a2);
2730        break;
2731    OP_32_64(ld16u):
2732        /* Note that we can ignore REXW for the zero-extend to 64-bit.  */
2733        tcg_out_modrm_offset(s, OPC_MOVZWL, a0, a1, a2);
2734        break;
2735    OP_32_64(ld16s):
2736        tcg_out_modrm_offset(s, OPC_MOVSWL + rexw, a0, a1, a2);
2737        break;
2738#if TCG_TARGET_REG_BITS == 64
2739    case INDEX_op_ld32u_i64:
2740#endif
2741    case INDEX_op_ld_i32:
2742        tcg_out_ld(s, TCG_TYPE_I32, a0, a1, a2);
2743        break;
2744
2745    OP_32_64(st8):
2746        if (const_args[0]) {
2747            tcg_out_modrm_offset(s, OPC_MOVB_EvIz, 0, a1, a2);
2748            tcg_out8(s, a0);
2749        } else {
2750            tcg_out_modrm_offset(s, OPC_MOVB_EvGv | P_REXB_R, a0, a1, a2);
2751        }
2752        break;
2753    OP_32_64(st16):
2754        if (const_args[0]) {
2755            tcg_out_modrm_offset(s, OPC_MOVL_EvIz | P_DATA16, 0, a1, a2);
2756            tcg_out16(s, a0);
2757        } else {
2758            tcg_out_modrm_offset(s, OPC_MOVL_EvGv | P_DATA16, a0, a1, a2);
2759        }
2760        break;
2761#if TCG_TARGET_REG_BITS == 64
2762    case INDEX_op_st32_i64:
2763#endif
2764    case INDEX_op_st_i32:
2765        if (const_args[0]) {
2766            tcg_out_modrm_offset(s, OPC_MOVL_EvIz, 0, a1, a2);
2767            tcg_out32(s, a0);
2768        } else {
2769            tcg_out_st(s, TCG_TYPE_I32, a0, a1, a2);
2770        }
2771        break;
2772
2773    OP_32_64(sub):
2774        c = ARITH_SUB;
2775        if (const_a2) {
2776            tgen_arithi(s, c + rexw, a0, a2, 0);
2777        } else {
2778            tgen_arithr(s, c + rexw, a0, a2);
2779        }
2780        break;
2781
2782    OP_32_64(mul):
2783        if (const_a2) {
2784            int32_t val;
2785            val = a2;
2786            if (val == (int8_t)val) {
2787                tcg_out_modrm(s, OPC_IMUL_GvEvIb + rexw, a0, a0);
2788                tcg_out8(s, val);
2789            } else {
2790                tcg_out_modrm(s, OPC_IMUL_GvEvIz + rexw, a0, a0);
2791                tcg_out32(s, val);
2792            }
2793        } else {
2794            tcg_out_modrm(s, OPC_IMUL_GvEv + rexw, a0, a2);
2795        }
2796        break;
2797
2798    OP_32_64(div2):
2799        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_IDIV, args[4]);
2800        break;
2801    OP_32_64(divu2):
2802        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_DIV, args[4]);
2803        break;
2804
2805    OP_32_64(shl):
2806        /* For small constant 3-operand shift, use LEA.  */
2807        if (const_a2 && a0 != a1 && (a2 - 1) < 3) {
2808            if (a2 - 1 == 0) {
2809                /* shl $1,a1,a0 -> lea (a1,a1),a0 */
2810                tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, a1, a1, 0, 0);
2811            } else {
2812                /* shl $n,a1,a0 -> lea 0(,a1,n),a0 */
2813                tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, -1, a1, a2, 0);
2814            }
2815            break;
2816        }
2817        c = SHIFT_SHL;
2818        vexop = OPC_SHLX;
2819        goto gen_shift_maybe_vex;
2820    OP_32_64(shr):
2821        c = SHIFT_SHR;
2822        vexop = OPC_SHRX;
2823        goto gen_shift_maybe_vex;
2824    OP_32_64(sar):
2825        c = SHIFT_SAR;
2826        vexop = OPC_SARX;
2827        goto gen_shift_maybe_vex;
2828    OP_32_64(rotl):
2829        c = SHIFT_ROL;
2830        goto gen_shift;
2831    OP_32_64(rotr):
2832        c = SHIFT_ROR;
2833        goto gen_shift;
2834    gen_shift_maybe_vex:
2835        if (have_bmi2) {
2836            if (!const_a2) {
2837                tcg_out_vex_modrm(s, vexop + rexw, a0, a2, a1);
2838                break;
2839            }
2840            tcg_out_mov(s, rexw ? TCG_TYPE_I64 : TCG_TYPE_I32, a0, a1);
2841        }
2842        /* FALLTHRU */
2843    gen_shift:
2844        if (const_a2) {
2845            tcg_out_shifti(s, c + rexw, a0, a2);
2846        } else {
2847            tcg_out_modrm(s, OPC_SHIFT_cl + rexw, c, a0);
2848        }
2849        break;
2850
2851    OP_32_64(ctz):
2852        tcg_out_ctz(s, rexw, args[0], args[1], args[2], const_args[2]);
2853        break;
2854    OP_32_64(clz):
2855        tcg_out_clz(s, rexw, args[0], args[1], args[2], const_args[2]);
2856        break;
2857    OP_32_64(ctpop):
2858        tcg_out_modrm(s, OPC_POPCNT + rexw, a0, a1);
2859        break;
2860
2861    OP_32_64(brcond):
2862        tcg_out_brcond(s, rexw, a2, a0, a1, const_args[1],
2863                       arg_label(args[3]), 0);
2864        break;
2865    OP_32_64(setcond):
2866        tcg_out_setcond(s, rexw, args[3], a0, a1, a2, const_a2, false);
2867        break;
2868    OP_32_64(negsetcond):
2869        tcg_out_setcond(s, rexw, args[3], a0, a1, a2, const_a2, true);
2870        break;
2871    OP_32_64(movcond):
2872        tcg_out_movcond(s, rexw, args[5], a0, a1, a2, const_a2, args[3]);
2873        break;
2874
2875    OP_32_64(bswap16):
2876        if (a2 & TCG_BSWAP_OS) {
2877            /* Output must be sign-extended. */
2878            if (rexw) {
2879                tcg_out_bswap64(s, a0);
2880                tcg_out_shifti(s, SHIFT_SAR + rexw, a0, 48);
2881            } else {
2882                tcg_out_bswap32(s, a0);
2883                tcg_out_shifti(s, SHIFT_SAR, a0, 16);
2884            }
2885        } else if ((a2 & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
2886            /* Output must be zero-extended, but input isn't. */
2887            tcg_out_bswap32(s, a0);
2888            tcg_out_shifti(s, SHIFT_SHR, a0, 16);
2889        } else {
2890            tcg_out_rolw_8(s, a0);
2891        }
2892        break;
2893    OP_32_64(bswap32):
2894        tcg_out_bswap32(s, a0);
2895        if (rexw && (a2 & TCG_BSWAP_OS)) {
2896            tcg_out_ext32s(s, a0, a0);
2897        }
2898        break;
2899
2900    OP_32_64(neg):
2901        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NEG, a0);
2902        break;
2903    OP_32_64(not):
2904        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NOT, a0);
2905        break;
2906
2907    case INDEX_op_qemu_ld_i32:
2908        tcg_out_qemu_ld(s, a0, -1, a1, a2, TCG_TYPE_I32);
2909        break;
2910    case INDEX_op_qemu_ld_i64:
2911        if (TCG_TARGET_REG_BITS == 64) {
2912            tcg_out_qemu_ld(s, a0, -1, a1, a2, TCG_TYPE_I64);
2913        } else {
2914            tcg_out_qemu_ld(s, a0, a1, a2, args[3], TCG_TYPE_I64);
2915        }
2916        break;
2917    case INDEX_op_qemu_ld_i128:
2918        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
2919        tcg_out_qemu_ld(s, a0, a1, a2, args[3], TCG_TYPE_I128);
2920        break;
2921
2922    case INDEX_op_qemu_st_i32:
2923    case INDEX_op_qemu_st8_i32:
2924        tcg_out_qemu_st(s, a0, -1, a1, a2, TCG_TYPE_I32);
2925        break;
2926    case INDEX_op_qemu_st_i64:
2927        if (TCG_TARGET_REG_BITS == 64) {
2928            tcg_out_qemu_st(s, a0, -1, a1, a2, TCG_TYPE_I64);
2929        } else {
2930            tcg_out_qemu_st(s, a0, a1, a2, args[3], TCG_TYPE_I64);
2931        }
2932        break;
2933    case INDEX_op_qemu_st_i128:
2934        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
2935        tcg_out_qemu_st(s, a0, a1, a2, args[3], TCG_TYPE_I128);
2936        break;
2937
2938    OP_32_64(mulu2):
2939        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_MUL, args[3]);
2940        break;
2941    OP_32_64(muls2):
2942        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_IMUL, args[3]);
2943        break;
2944    OP_32_64(add2):
2945        if (const_args[4]) {
2946            tgen_arithi(s, ARITH_ADD + rexw, a0, args[4], 1);
2947        } else {
2948            tgen_arithr(s, ARITH_ADD + rexw, a0, args[4]);
2949        }
2950        if (const_args[5]) {
2951            tgen_arithi(s, ARITH_ADC + rexw, a1, args[5], 1);
2952        } else {
2953            tgen_arithr(s, ARITH_ADC + rexw, a1, args[5]);
2954        }
2955        break;
2956    OP_32_64(sub2):
2957        if (const_args[4]) {
2958            tgen_arithi(s, ARITH_SUB + rexw, a0, args[4], 1);
2959        } else {
2960            tgen_arithr(s, ARITH_SUB + rexw, a0, args[4]);
2961        }
2962        if (const_args[5]) {
2963            tgen_arithi(s, ARITH_SBB + rexw, a1, args[5], 1);
2964        } else {
2965            tgen_arithr(s, ARITH_SBB + rexw, a1, args[5]);
2966        }
2967        break;
2968
2969#if TCG_TARGET_REG_BITS == 32
2970    case INDEX_op_brcond2_i32:
2971        tcg_out_brcond2(s, args, const_args, 0);
2972        break;
2973    case INDEX_op_setcond2_i32:
2974        tcg_out_setcond2(s, args, const_args);
2975        break;
2976#else /* TCG_TARGET_REG_BITS == 64 */
2977    case INDEX_op_ld32s_i64:
2978        tcg_out_modrm_offset(s, OPC_MOVSLQ, a0, a1, a2);
2979        break;
2980    case INDEX_op_ld_i64:
2981        tcg_out_ld(s, TCG_TYPE_I64, a0, a1, a2);
2982        break;
2983    case INDEX_op_st_i64:
2984        if (const_args[0]) {
2985            tcg_out_modrm_offset(s, OPC_MOVL_EvIz | P_REXW, 0, a1, a2);
2986            tcg_out32(s, a0);
2987        } else {
2988            tcg_out_st(s, TCG_TYPE_I64, a0, a1, a2);
2989        }
2990        break;
2991
2992    case INDEX_op_bswap64_i64:
2993        tcg_out_bswap64(s, a0);
2994        break;
2995    case INDEX_op_extrh_i64_i32:
2996        tcg_out_shifti(s, SHIFT_SHR + P_REXW, a0, 32);
2997        break;
2998#endif
2999
3000    OP_32_64(deposit):
3001        if (args[3] == 0 && args[4] == 8) {
3002            /* load bits 0..7 */
3003            if (const_a2) {
3004                tcg_out_opc(s, OPC_MOVB_Ib | P_REXB_RM | LOWREGMASK(a0),
3005                            0, a0, 0);
3006                tcg_out8(s, a2);
3007            } else {
3008                tcg_out_modrm(s, OPC_MOVB_EvGv | P_REXB_R | P_REXB_RM, a2, a0);
3009            }
3010        } else if (TCG_TARGET_REG_BITS == 32 && args[3] == 8 && args[4] == 8) {
3011            /* load bits 8..15 */
3012            if (const_a2) {
3013                tcg_out8(s, OPC_MOVB_Ib + a0 + 4);
3014                tcg_out8(s, a2);
3015            } else {
3016                tcg_out_modrm(s, OPC_MOVB_EvGv, a2, a0 + 4);
3017            }
3018        } else if (args[3] == 0 && args[4] == 16) {
3019            /* load bits 0..15 */
3020            if (const_a2) {
3021                tcg_out_opc(s, OPC_MOVL_Iv | P_DATA16 | LOWREGMASK(a0),
3022                            0, a0, 0);
3023                tcg_out16(s, a2);
3024            } else {
3025                tcg_out_modrm(s, OPC_MOVL_EvGv | P_DATA16, a2, a0);
3026            }
3027        } else {
3028            g_assert_not_reached();
3029        }
3030        break;
3031
3032    case INDEX_op_extract_i64:
3033        if (a2 + args[3] == 32) {
3034            if (a2 == 0) {
3035                tcg_out_ext32u(s, a0, a1);
3036                break;
3037            }
3038            /* This is a 32-bit zero-extending right shift.  */
3039            tcg_out_mov(s, TCG_TYPE_I32, a0, a1);
3040            tcg_out_shifti(s, SHIFT_SHR, a0, a2);
3041            break;
3042        }
3043        /* FALLTHRU */
3044    case INDEX_op_extract_i32:
3045        if (a2 == 0 && args[3] == 8) {
3046            tcg_out_ext8u(s, a0, a1);
3047        } else if (a2 == 0 && args[3] == 16) {
3048            tcg_out_ext16u(s, a0, a1);
3049        } else if (a2 == 8 && args[3] == 8) {
3050            /*
3051             * On the off-chance that we can use the high-byte registers.
3052             * Otherwise we emit the same ext16 + shift pattern that we
3053             * would have gotten from the normal tcg-op.c expansion.
3054             */
3055            if (a1 < 4 && a0 < 8) {
3056                tcg_out_modrm(s, OPC_MOVZBL, a0, a1 + 4);
3057            } else {
3058                tcg_out_ext16u(s, a0, a1);
3059                tcg_out_shifti(s, SHIFT_SHR, a0, 8);
3060            }
3061        } else {
3062            g_assert_not_reached();
3063        }
3064        break;
3065
3066    case INDEX_op_sextract_i64:
3067        if (a2 == 0 && args[3] == 8) {
3068            tcg_out_ext8s(s, TCG_TYPE_I64, a0, a1);
3069        } else if (a2 == 0 && args[3] == 16) {
3070            tcg_out_ext16s(s, TCG_TYPE_I64, a0, a1);
3071        } else if (a2 == 0 && args[3] == 32) {
3072            tcg_out_ext32s(s, a0, a1);
3073        } else {
3074            g_assert_not_reached();
3075        }
3076        break;
3077
3078    case INDEX_op_sextract_i32:
3079        if (a2 == 0 && args[3] == 8) {
3080            tcg_out_ext8s(s, TCG_TYPE_I32, a0, a1);
3081        } else if (a2 == 0 && args[3] == 16) {
3082            tcg_out_ext16s(s, TCG_TYPE_I32, a0, a1);
3083        } else if (a2 == 8 && args[3] == 8) {
3084            if (a1 < 4 && a0 < 8) {
3085                tcg_out_modrm(s, OPC_MOVSBL, a0, a1 + 4);
3086            } else {
3087                tcg_out_ext16s(s, TCG_TYPE_I32, a0, a1);
3088                tcg_out_shifti(s, SHIFT_SAR, a0, 8);
3089            }
3090        } else {
3091            g_assert_not_reached();
3092        }
3093        break;
3094
3095    OP_32_64(extract2):
3096        /* Note that SHRD outputs to the r/m operand.  */
3097        tcg_out_modrm(s, OPC_SHRD_Ib + rexw, a2, a0);
3098        tcg_out8(s, args[3]);
3099        break;
3100
3101    case INDEX_op_mb:
3102        tcg_out_mb(s, a0);
3103        break;
3104    case INDEX_op_call:     /* Always emitted via tcg_out_call.  */
3105    case INDEX_op_exit_tb:  /* Always emitted via tcg_out_exit_tb.  */
3106    case INDEX_op_goto_tb:  /* Always emitted via tcg_out_goto_tb.  */
3107    case INDEX_op_ext_i32_i64:  /* Always emitted via tcg_reg_alloc_op.  */
3108    case INDEX_op_extu_i32_i64:
3109    case INDEX_op_extrl_i64_i32:
3110    default:
3111        g_assert_not_reached();
3112    }
3113
3114#undef OP_32_64
3115}
3116
3117static int const umin_insn[4] = {
3118    OPC_PMINUB, OPC_PMINUW, OPC_PMINUD, OPC_VPMINUQ
3119};
3120
3121static int const umax_insn[4] = {
3122    OPC_PMAXUB, OPC_PMAXUW, OPC_PMAXUD, OPC_VPMAXUQ
3123};
3124
3125static bool tcg_out_cmp_vec_noinv(TCGContext *s, TCGType type, unsigned vece,
3126                                  TCGReg v0, TCGReg v1, TCGReg v2, TCGCond cond)
3127{
3128    static int const cmpeq_insn[4] = {
3129        OPC_PCMPEQB, OPC_PCMPEQW, OPC_PCMPEQD, OPC_PCMPEQQ
3130    };
3131    static int const cmpgt_insn[4] = {
3132        OPC_PCMPGTB, OPC_PCMPGTW, OPC_PCMPGTD, OPC_PCMPGTQ
3133    };
3134
3135    enum {
3136        NEED_INV  = 1,
3137        NEED_SWAP = 2,
3138        NEED_UMIN = 4,
3139        NEED_UMAX = 8,
3140        INVALID   = 16,
3141    };
3142    static const uint8_t cond_fixup[16] = {
3143        [0 ... 15] = INVALID,
3144        [TCG_COND_EQ] = 0,
3145        [TCG_COND_GT] = 0,
3146        [TCG_COND_NE] = NEED_INV,
3147        [TCG_COND_LE] = NEED_INV,
3148        [TCG_COND_LT] = NEED_SWAP,
3149        [TCG_COND_GE] = NEED_SWAP | NEED_INV,
3150        [TCG_COND_LEU] = NEED_UMIN,
3151        [TCG_COND_GTU] = NEED_UMIN | NEED_INV,
3152        [TCG_COND_GEU] = NEED_UMAX,
3153        [TCG_COND_LTU] = NEED_UMAX | NEED_INV,
3154    };
3155    int fixup = cond_fixup[cond];
3156
3157    assert(!(fixup & INVALID));
3158
3159    if (fixup & NEED_INV) {
3160        cond = tcg_invert_cond(cond);
3161    }
3162
3163    if (fixup & NEED_SWAP) {
3164        TCGReg swap = v1;
3165        v1 = v2;
3166        v2 = swap;
3167        cond = tcg_swap_cond(cond);
3168    }
3169
3170    if (fixup & (NEED_UMIN | NEED_UMAX)) {
3171        int op = (fixup & NEED_UMIN ? umin_insn[vece] : umax_insn[vece]);
3172
3173        /* avx2 does not have 64-bit min/max; adjusted during expand. */
3174        assert(vece <= MO_32);
3175
3176        tcg_out_vex_modrm_type(s, op, TCG_TMP_VEC, v1, v2, type);
3177        v2 = TCG_TMP_VEC;
3178        cond = TCG_COND_EQ;
3179    }
3180
3181    switch (cond) {
3182    case TCG_COND_EQ:
3183        tcg_out_vex_modrm_type(s, cmpeq_insn[vece], v0, v1, v2, type);
3184        break;
3185    case TCG_COND_GT:
3186        tcg_out_vex_modrm_type(s, cmpgt_insn[vece], v0, v1, v2, type);
3187        break;
3188    default:
3189        g_assert_not_reached();
3190    }
3191    return fixup & NEED_INV;
3192}
3193
3194static void tcg_out_cmp_vec_k1(TCGContext *s, TCGType type, unsigned vece,
3195                               TCGReg v1, TCGReg v2, TCGCond cond)
3196{
3197    static const int cmpm_insn[2][4] = {
3198        { OPC_VPCMPB, OPC_VPCMPW, OPC_VPCMPD, OPC_VPCMPQ },
3199        { OPC_VPCMPUB, OPC_VPCMPUW, OPC_VPCMPUD, OPC_VPCMPUQ }
3200    };
3201    static const int testm_insn[4] = {
3202        OPC_VPTESTMB, OPC_VPTESTMW, OPC_VPTESTMD, OPC_VPTESTMQ
3203    };
3204    static const int testnm_insn[4] = {
3205        OPC_VPTESTNMB, OPC_VPTESTNMW, OPC_VPTESTNMD, OPC_VPTESTNMQ
3206    };
3207
3208    static const int cond_ext[16] = {
3209        [TCG_COND_EQ] = 0,
3210        [TCG_COND_NE] = 4,
3211        [TCG_COND_LT] = 1,
3212        [TCG_COND_LTU] = 1,
3213        [TCG_COND_LE] = 2,
3214        [TCG_COND_LEU] = 2,
3215        [TCG_COND_NEVER] = 3,
3216        [TCG_COND_GE] = 5,
3217        [TCG_COND_GEU] = 5,
3218        [TCG_COND_GT] = 6,
3219        [TCG_COND_GTU] = 6,
3220        [TCG_COND_ALWAYS] = 7,
3221    };
3222
3223    switch (cond) {
3224    case TCG_COND_TSTNE:
3225        tcg_out_vex_modrm_type(s, testm_insn[vece], /* k1 */ 1, v1, v2, type);
3226        break;
3227    case TCG_COND_TSTEQ:
3228        tcg_out_vex_modrm_type(s, testnm_insn[vece], /* k1 */ 1, v1, v2, type);
3229        break;
3230    default:
3231        tcg_out_vex_modrm_type(s, cmpm_insn[is_unsigned_cond(cond)][vece],
3232                               /* k1 */ 1, v1, v2, type);
3233        tcg_out8(s, cond_ext[cond]);
3234        break;
3235    }
3236}
3237
3238static void tcg_out_k1_to_vec(TCGContext *s, TCGType type,
3239                              unsigned vece, TCGReg dest)
3240{
3241    static const int movm_insn[] = {
3242        OPC_VPMOVM2B, OPC_VPMOVM2W, OPC_VPMOVM2D, OPC_VPMOVM2Q
3243    };
3244    tcg_out_vex_modrm_type(s, movm_insn[vece], dest, 0, /* k1 */ 1, type);
3245}
3246
3247static void tcg_out_cmp_vec(TCGContext *s, TCGType type, unsigned vece,
3248                            TCGReg v0, TCGReg v1, TCGReg v2, TCGCond cond)
3249{
3250    /*
3251     * With avx512, we have a complete set of comparisons into mask.
3252     * Unless there's a single insn expansion for the comparision,
3253     * expand via a mask in k1.
3254     */
3255    if ((vece <= MO_16 ? have_avx512bw : have_avx512dq)
3256        && cond != TCG_COND_EQ
3257        && cond != TCG_COND_LT
3258        && cond != TCG_COND_GT) {
3259        tcg_out_cmp_vec_k1(s, type, vece, v1, v2, cond);
3260        tcg_out_k1_to_vec(s, type, vece, v0);
3261        return;
3262    }
3263
3264    if (tcg_out_cmp_vec_noinv(s, type, vece, v0, v1, v2, cond)) {
3265        tcg_out_dupi_vec(s, type, vece, TCG_TMP_VEC, -1);
3266        tcg_out_vex_modrm_type(s, OPC_PXOR, v0, v0, TCG_TMP_VEC, type);
3267    }
3268}
3269
3270static void tcg_out_cmpsel_vec_k1(TCGContext *s, TCGType type, unsigned vece,
3271                                  TCGReg v0, TCGReg c1, TCGReg c2,
3272                                  TCGReg v3, TCGReg v4, TCGCond cond)
3273{
3274    static const int vpblendm_insn[] = {
3275        OPC_VPBLENDMB, OPC_VPBLENDMW, OPC_VPBLENDMD, OPC_VPBLENDMQ
3276    };
3277    bool z = false;
3278
3279    /* Swap to place constant in V4 to take advantage of zero-masking. */
3280    if (!v3) {
3281        z = true;
3282        v3 = v4;
3283        cond = tcg_invert_cond(cond);
3284    }
3285
3286    tcg_out_cmp_vec_k1(s, type, vece, c1, c2, cond);
3287    tcg_out_evex_modrm_type(s, vpblendm_insn[vece], v0, v4, v3,
3288                            /* k1 */1, z, type);
3289}
3290
3291static void tcg_out_cmpsel_vec(TCGContext *s, TCGType type, unsigned vece,
3292                               TCGReg v0, TCGReg c1, TCGReg c2,
3293                               TCGReg v3, TCGReg v4, TCGCond cond)
3294{
3295    bool inv;
3296
3297    if (vece <= MO_16 ? have_avx512bw : have_avx512vl) {
3298        tcg_out_cmpsel_vec_k1(s, type, vece, v0, c1, c2, v3, v4, cond);
3299        return;
3300    }
3301
3302    inv = tcg_out_cmp_vec_noinv(s, type, vece, TCG_TMP_VEC, c1, c2, cond);
3303
3304    /*
3305     * Since XMM0 is 16, the only way we get 0 into V3
3306     * is via the constant zero constraint.
3307     */
3308    if (!v3) {
3309        if (inv) {
3310            tcg_out_vex_modrm_type(s, OPC_PAND, v0, TCG_TMP_VEC, v4, type);
3311        } else {
3312            tcg_out_vex_modrm_type(s, OPC_PANDN, v0, TCG_TMP_VEC, v4, type);
3313        }
3314    } else {
3315        if (inv) {
3316            TCGReg swap = v3;
3317            v3 = v4;
3318            v4 = swap;
3319        }
3320        tcg_out_vex_modrm_type(s, OPC_VPBLENDVB, v0, v4, v3, type);
3321        tcg_out8(s, (TCG_TMP_VEC - TCG_REG_XMM0) << 4);
3322    }
3323}
3324
3325static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
3326                           unsigned vecl, unsigned vece,
3327                           const TCGArg args[TCG_MAX_OP_ARGS],
3328                           const int const_args[TCG_MAX_OP_ARGS])
3329{
3330    static int const add_insn[4] = {
3331        OPC_PADDB, OPC_PADDW, OPC_PADDD, OPC_PADDQ
3332    };
3333    static int const ssadd_insn[4] = {
3334        OPC_PADDSB, OPC_PADDSW, OPC_UD2, OPC_UD2
3335    };
3336    static int const usadd_insn[4] = {
3337        OPC_PADDUB, OPC_PADDUW, OPC_UD2, OPC_UD2
3338    };
3339    static int const sub_insn[4] = {
3340        OPC_PSUBB, OPC_PSUBW, OPC_PSUBD, OPC_PSUBQ
3341    };
3342    static int const sssub_insn[4] = {
3343        OPC_PSUBSB, OPC_PSUBSW, OPC_UD2, OPC_UD2
3344    };
3345    static int const ussub_insn[4] = {
3346        OPC_PSUBUB, OPC_PSUBUW, OPC_UD2, OPC_UD2
3347    };
3348    static int const mul_insn[4] = {
3349        OPC_UD2, OPC_PMULLW, OPC_PMULLD, OPC_VPMULLQ
3350    };
3351    static int const shift_imm_insn[4] = {
3352        OPC_UD2, OPC_PSHIFTW_Ib, OPC_PSHIFTD_Ib, OPC_PSHIFTQ_Ib
3353    };
3354    static int const punpckl_insn[4] = {
3355        OPC_PUNPCKLBW, OPC_PUNPCKLWD, OPC_PUNPCKLDQ, OPC_PUNPCKLQDQ
3356    };
3357    static int const punpckh_insn[4] = {
3358        OPC_PUNPCKHBW, OPC_PUNPCKHWD, OPC_PUNPCKHDQ, OPC_PUNPCKHQDQ
3359    };
3360    static int const packss_insn[4] = {
3361        OPC_PACKSSWB, OPC_PACKSSDW, OPC_UD2, OPC_UD2
3362    };
3363    static int const packus_insn[4] = {
3364        OPC_PACKUSWB, OPC_PACKUSDW, OPC_UD2, OPC_UD2
3365    };
3366    static int const smin_insn[4] = {
3367        OPC_PMINSB, OPC_PMINSW, OPC_PMINSD, OPC_VPMINSQ
3368    };
3369    static int const smax_insn[4] = {
3370        OPC_PMAXSB, OPC_PMAXSW, OPC_PMAXSD, OPC_VPMAXSQ
3371    };
3372    static int const rotlv_insn[4] = {
3373        OPC_UD2, OPC_UD2, OPC_VPROLVD, OPC_VPROLVQ
3374    };
3375    static int const rotrv_insn[4] = {
3376        OPC_UD2, OPC_UD2, OPC_VPRORVD, OPC_VPRORVQ
3377    };
3378    static int const shlv_insn[4] = {
3379        OPC_UD2, OPC_VPSLLVW, OPC_VPSLLVD, OPC_VPSLLVQ
3380    };
3381    static int const shrv_insn[4] = {
3382        OPC_UD2, OPC_VPSRLVW, OPC_VPSRLVD, OPC_VPSRLVQ
3383    };
3384    static int const sarv_insn[4] = {
3385        OPC_UD2, OPC_VPSRAVW, OPC_VPSRAVD, OPC_VPSRAVQ
3386    };
3387    static int const shls_insn[4] = {
3388        OPC_UD2, OPC_PSLLW, OPC_PSLLD, OPC_PSLLQ
3389    };
3390    static int const shrs_insn[4] = {
3391        OPC_UD2, OPC_PSRLW, OPC_PSRLD, OPC_PSRLQ
3392    };
3393    static int const sars_insn[4] = {
3394        OPC_UD2, OPC_PSRAW, OPC_PSRAD, OPC_VPSRAQ
3395    };
3396    static int const vpshldi_insn[4] = {
3397        OPC_UD2, OPC_VPSHLDW, OPC_VPSHLDD, OPC_VPSHLDQ
3398    };
3399    static int const vpshldv_insn[4] = {
3400        OPC_UD2, OPC_VPSHLDVW, OPC_VPSHLDVD, OPC_VPSHLDVQ
3401    };
3402    static int const vpshrdv_insn[4] = {
3403        OPC_UD2, OPC_VPSHRDVW, OPC_VPSHRDVD, OPC_VPSHRDVQ
3404    };
3405    static int const abs_insn[4] = {
3406        OPC_PABSB, OPC_PABSW, OPC_PABSD, OPC_VPABSQ
3407    };
3408
3409    TCGType type = vecl + TCG_TYPE_V64;
3410    int insn, sub;
3411    TCGArg a0, a1, a2, a3;
3412
3413    a0 = args[0];
3414    a1 = args[1];
3415    a2 = args[2];
3416
3417    switch (opc) {
3418    case INDEX_op_add_vec:
3419        insn = add_insn[vece];
3420        goto gen_simd;
3421    case INDEX_op_ssadd_vec:
3422        insn = ssadd_insn[vece];
3423        goto gen_simd;
3424    case INDEX_op_usadd_vec:
3425        insn = usadd_insn[vece];
3426        goto gen_simd;
3427    case INDEX_op_sub_vec:
3428        insn = sub_insn[vece];
3429        goto gen_simd;
3430    case INDEX_op_sssub_vec:
3431        insn = sssub_insn[vece];
3432        goto gen_simd;
3433    case INDEX_op_ussub_vec:
3434        insn = ussub_insn[vece];
3435        goto gen_simd;
3436    case INDEX_op_mul_vec:
3437        insn = mul_insn[vece];
3438        goto gen_simd;
3439    case INDEX_op_and_vec:
3440        insn = OPC_PAND;
3441        goto gen_simd;
3442    case INDEX_op_or_vec:
3443        insn = OPC_POR;
3444        goto gen_simd;
3445    case INDEX_op_xor_vec:
3446        insn = OPC_PXOR;
3447        goto gen_simd;
3448    case INDEX_op_smin_vec:
3449        insn = smin_insn[vece];
3450        goto gen_simd;
3451    case INDEX_op_umin_vec:
3452        insn = umin_insn[vece];
3453        goto gen_simd;
3454    case INDEX_op_smax_vec:
3455        insn = smax_insn[vece];
3456        goto gen_simd;
3457    case INDEX_op_umax_vec:
3458        insn = umax_insn[vece];
3459        goto gen_simd;
3460    case INDEX_op_shlv_vec:
3461        insn = shlv_insn[vece];
3462        goto gen_simd;
3463    case INDEX_op_shrv_vec:
3464        insn = shrv_insn[vece];
3465        goto gen_simd;
3466    case INDEX_op_sarv_vec:
3467        insn = sarv_insn[vece];
3468        goto gen_simd;
3469    case INDEX_op_rotlv_vec:
3470        insn = rotlv_insn[vece];
3471        goto gen_simd;
3472    case INDEX_op_rotrv_vec:
3473        insn = rotrv_insn[vece];
3474        goto gen_simd;
3475    case INDEX_op_shls_vec:
3476        insn = shls_insn[vece];
3477        goto gen_simd;
3478    case INDEX_op_shrs_vec:
3479        insn = shrs_insn[vece];
3480        goto gen_simd;
3481    case INDEX_op_sars_vec:
3482        insn = sars_insn[vece];
3483        goto gen_simd;
3484    case INDEX_op_x86_punpckl_vec:
3485        insn = punpckl_insn[vece];
3486        goto gen_simd;
3487    case INDEX_op_x86_punpckh_vec:
3488        insn = punpckh_insn[vece];
3489        goto gen_simd;
3490    case INDEX_op_x86_packss_vec:
3491        insn = packss_insn[vece];
3492        goto gen_simd;
3493    case INDEX_op_x86_packus_vec:
3494        insn = packus_insn[vece];
3495        goto gen_simd;
3496    case INDEX_op_x86_vpshldv_vec:
3497        insn = vpshldv_insn[vece];
3498        a1 = a2;
3499        a2 = args[3];
3500        goto gen_simd;
3501    case INDEX_op_x86_vpshrdv_vec:
3502        insn = vpshrdv_insn[vece];
3503        a1 = a2;
3504        a2 = args[3];
3505        goto gen_simd;
3506#if TCG_TARGET_REG_BITS == 32
3507    case INDEX_op_dup2_vec:
3508        /* First merge the two 32-bit inputs to a single 64-bit element. */
3509        tcg_out_vex_modrm(s, OPC_PUNPCKLDQ, a0, a1, a2);
3510        /* Then replicate the 64-bit elements across the rest of the vector. */
3511        if (type != TCG_TYPE_V64) {
3512            tcg_out_dup_vec(s, type, MO_64, a0, a0);
3513        }
3514        break;
3515#endif
3516    case INDEX_op_abs_vec:
3517        insn = abs_insn[vece];
3518        a2 = a1;
3519        a1 = 0;
3520        goto gen_simd;
3521    gen_simd:
3522        tcg_debug_assert(insn != OPC_UD2);
3523        tcg_out_vex_modrm_type(s, insn, a0, a1, a2, type);
3524        break;
3525
3526    case INDEX_op_cmp_vec:
3527        tcg_out_cmp_vec(s, type, vece, a0, a1, a2, args[3]);
3528        break;
3529
3530    case INDEX_op_cmpsel_vec:
3531        tcg_out_cmpsel_vec(s, type, vece, a0, a1, a2,
3532                           args[3], args[4], args[5]);
3533        break;
3534
3535    case INDEX_op_andc_vec:
3536        insn = OPC_PANDN;
3537        tcg_out_vex_modrm_type(s, insn, a0, a2, a1, type);
3538        break;
3539
3540    case INDEX_op_shli_vec:
3541        insn = shift_imm_insn[vece];
3542        sub = 6;
3543        goto gen_shift;
3544    case INDEX_op_shri_vec:
3545        insn = shift_imm_insn[vece];
3546        sub = 2;
3547        goto gen_shift;
3548    case INDEX_op_sari_vec:
3549        if (vece == MO_64) {
3550            insn = OPC_PSHIFTD_Ib | P_VEXW | P_EVEX;
3551        } else {
3552            insn = shift_imm_insn[vece];
3553        }
3554        sub = 4;
3555        goto gen_shift;
3556    case INDEX_op_rotli_vec:
3557        insn = OPC_PSHIFTD_Ib | P_EVEX;  /* VPROL[DQ] */
3558        if (vece == MO_64) {
3559            insn |= P_VEXW;
3560        }
3561        sub = 1;
3562        goto gen_shift;
3563    gen_shift:
3564        tcg_debug_assert(vece != MO_8);
3565        tcg_out_vex_modrm_type(s, insn, sub, a0, a1, type);
3566        tcg_out8(s, a2);
3567        break;
3568
3569    case INDEX_op_ld_vec:
3570        tcg_out_ld(s, type, a0, a1, a2);
3571        break;
3572    case INDEX_op_st_vec:
3573        tcg_out_st(s, type, a0, a1, a2);
3574        break;
3575    case INDEX_op_dupm_vec:
3576        tcg_out_dupm_vec(s, type, vece, a0, a1, a2);
3577        break;
3578
3579    case INDEX_op_x86_shufps_vec:
3580        insn = OPC_SHUFPS;
3581        sub = args[3];
3582        goto gen_simd_imm8;
3583    case INDEX_op_x86_blend_vec:
3584        if (vece == MO_16) {
3585            insn = OPC_PBLENDW;
3586        } else if (vece == MO_32) {
3587            insn = (have_avx2 ? OPC_VPBLENDD : OPC_BLENDPS);
3588        } else {
3589            g_assert_not_reached();
3590        }
3591        sub = args[3];
3592        goto gen_simd_imm8;
3593    case INDEX_op_x86_vperm2i128_vec:
3594        insn = OPC_VPERM2I128;
3595        sub = args[3];
3596        goto gen_simd_imm8;
3597    case INDEX_op_x86_vpshldi_vec:
3598        insn = vpshldi_insn[vece];
3599        sub = args[3];
3600        goto gen_simd_imm8;
3601
3602    case INDEX_op_not_vec:
3603        insn = OPC_VPTERNLOGQ;
3604        a2 = a1;
3605        sub = 0x33; /* !B */
3606        goto gen_simd_imm8;
3607    case INDEX_op_nor_vec:
3608        insn = OPC_VPTERNLOGQ;
3609        sub = 0x11; /* norCB */
3610        goto gen_simd_imm8;
3611    case INDEX_op_nand_vec:
3612        insn = OPC_VPTERNLOGQ;
3613        sub = 0x77; /* nandCB */
3614        goto gen_simd_imm8;
3615    case INDEX_op_eqv_vec:
3616        insn = OPC_VPTERNLOGQ;
3617        sub = 0x99; /* xnorCB */
3618        goto gen_simd_imm8;
3619    case INDEX_op_orc_vec:
3620        insn = OPC_VPTERNLOGQ;
3621        sub = 0xdd; /* orB!C */
3622        goto gen_simd_imm8;
3623
3624    case INDEX_op_bitsel_vec:
3625        insn = OPC_VPTERNLOGQ;
3626        a3 = args[3];
3627        if (a0 == a1) {
3628            a1 = a2;
3629            a2 = a3;
3630            sub = 0xca; /* A?B:C */
3631        } else if (a0 == a2) {
3632            a2 = a3;
3633            sub = 0xe2; /* B?A:C */
3634        } else {
3635            tcg_out_mov(s, type, a0, a3);
3636            sub = 0xb8; /* B?C:A */
3637        }
3638        goto gen_simd_imm8;
3639
3640    gen_simd_imm8:
3641        tcg_debug_assert(insn != OPC_UD2);
3642        tcg_out_vex_modrm_type(s, insn, a0, a1, a2, type);
3643        tcg_out8(s, sub);
3644        break;
3645
3646    case INDEX_op_x86_psrldq_vec:
3647        tcg_out_vex_modrm(s, OPC_GRP14, 3, a0, a1);
3648        tcg_out8(s, a2);
3649        break;
3650
3651    case INDEX_op_mov_vec:  /* Always emitted via tcg_out_mov.  */
3652    case INDEX_op_dup_vec:  /* Always emitted via tcg_out_dup_vec.  */
3653    default:
3654        g_assert_not_reached();
3655    }
3656}
3657
3658static TCGConstraintSetIndex
3659tcg_target_op_def(TCGOpcode op, TCGType type, unsigned flags)
3660{
3661    switch (op) {
3662    case INDEX_op_goto_ptr:
3663        return C_O0_I1(r);
3664
3665    case INDEX_op_ld8u_i32:
3666    case INDEX_op_ld8u_i64:
3667    case INDEX_op_ld8s_i32:
3668    case INDEX_op_ld8s_i64:
3669    case INDEX_op_ld16u_i32:
3670    case INDEX_op_ld16u_i64:
3671    case INDEX_op_ld16s_i32:
3672    case INDEX_op_ld16s_i64:
3673    case INDEX_op_ld_i32:
3674    case INDEX_op_ld32u_i64:
3675    case INDEX_op_ld32s_i64:
3676    case INDEX_op_ld_i64:
3677        return C_O1_I1(r, r);
3678
3679    case INDEX_op_st8_i32:
3680    case INDEX_op_st8_i64:
3681        return C_O0_I2(qi, r);
3682
3683    case INDEX_op_st16_i32:
3684    case INDEX_op_st16_i64:
3685    case INDEX_op_st_i32:
3686    case INDEX_op_st32_i64:
3687        return C_O0_I2(ri, r);
3688
3689    case INDEX_op_st_i64:
3690        return C_O0_I2(re, r);
3691
3692    case INDEX_op_sub_i32:
3693    case INDEX_op_sub_i64:
3694    case INDEX_op_mul_i32:
3695    case INDEX_op_mul_i64:
3696        return C_O1_I2(r, 0, re);
3697
3698    case INDEX_op_shl_i32:
3699    case INDEX_op_shl_i64:
3700    case INDEX_op_shr_i32:
3701    case INDEX_op_shr_i64:
3702    case INDEX_op_sar_i32:
3703    case INDEX_op_sar_i64:
3704        return have_bmi2 ? C_O1_I2(r, r, ri) : C_O1_I2(r, 0, ci);
3705
3706    case INDEX_op_rotl_i32:
3707    case INDEX_op_rotl_i64:
3708    case INDEX_op_rotr_i32:
3709    case INDEX_op_rotr_i64:
3710        return C_O1_I2(r, 0, ci);
3711
3712    case INDEX_op_brcond_i32:
3713    case INDEX_op_brcond_i64:
3714        return C_O0_I2(r, reT);
3715
3716    case INDEX_op_bswap16_i32:
3717    case INDEX_op_bswap16_i64:
3718    case INDEX_op_bswap32_i32:
3719    case INDEX_op_bswap32_i64:
3720    case INDEX_op_bswap64_i64:
3721    case INDEX_op_neg_i32:
3722    case INDEX_op_neg_i64:
3723    case INDEX_op_not_i32:
3724    case INDEX_op_not_i64:
3725    case INDEX_op_extrh_i64_i32:
3726        return C_O1_I1(r, 0);
3727
3728    case INDEX_op_ext_i32_i64:
3729    case INDEX_op_extu_i32_i64:
3730    case INDEX_op_extrl_i64_i32:
3731    case INDEX_op_extract_i32:
3732    case INDEX_op_extract_i64:
3733    case INDEX_op_sextract_i32:
3734    case INDEX_op_sextract_i64:
3735    case INDEX_op_ctpop_i32:
3736    case INDEX_op_ctpop_i64:
3737        return C_O1_I1(r, r);
3738
3739    case INDEX_op_extract2_i32:
3740    case INDEX_op_extract2_i64:
3741        return C_O1_I2(r, 0, r);
3742
3743    case INDEX_op_deposit_i32:
3744    case INDEX_op_deposit_i64:
3745        return C_O1_I2(q, 0, qi);
3746
3747    case INDEX_op_setcond_i32:
3748    case INDEX_op_setcond_i64:
3749    case INDEX_op_negsetcond_i32:
3750    case INDEX_op_negsetcond_i64:
3751        return C_O1_I2(q, r, reT);
3752
3753    case INDEX_op_movcond_i32:
3754    case INDEX_op_movcond_i64:
3755        return C_O1_I4(r, r, reT, r, 0);
3756
3757    case INDEX_op_div2_i32:
3758    case INDEX_op_div2_i64:
3759    case INDEX_op_divu2_i32:
3760    case INDEX_op_divu2_i64:
3761        return C_O2_I3(a, d, 0, 1, r);
3762
3763    case INDEX_op_mulu2_i32:
3764    case INDEX_op_mulu2_i64:
3765    case INDEX_op_muls2_i32:
3766    case INDEX_op_muls2_i64:
3767        return C_O2_I2(a, d, a, r);
3768
3769    case INDEX_op_add2_i32:
3770    case INDEX_op_add2_i64:
3771    case INDEX_op_sub2_i32:
3772    case INDEX_op_sub2_i64:
3773        return C_N1_O1_I4(r, r, 0, 1, re, re);
3774
3775    case INDEX_op_ctz_i32:
3776    case INDEX_op_ctz_i64:
3777        return have_bmi1 ? C_N1_I2(r, r, rW) : C_N1_I2(r, r, r);
3778
3779    case INDEX_op_clz_i32:
3780    case INDEX_op_clz_i64:
3781        return have_lzcnt ? C_N1_I2(r, r, rW) : C_N1_I2(r, r, r);
3782
3783    case INDEX_op_qemu_ld_i32:
3784        return C_O1_I1(r, L);
3785
3786    case INDEX_op_qemu_st_i32:
3787        return C_O0_I2(L, L);
3788    case INDEX_op_qemu_st8_i32:
3789        return C_O0_I2(s, L);
3790
3791    case INDEX_op_qemu_ld_i64:
3792        return TCG_TARGET_REG_BITS == 64 ? C_O1_I1(r, L) : C_O2_I1(r, r, L);
3793
3794    case INDEX_op_qemu_st_i64:
3795        return TCG_TARGET_REG_BITS == 64 ? C_O0_I2(L, L) : C_O0_I3(L, L, L);
3796
3797    case INDEX_op_qemu_ld_i128:
3798        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
3799        return C_O2_I1(r, r, L);
3800    case INDEX_op_qemu_st_i128:
3801        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
3802        return C_O0_I3(L, L, L);
3803
3804    case INDEX_op_brcond2_i32:
3805        return C_O0_I4(r, r, ri, ri);
3806
3807    case INDEX_op_setcond2_i32:
3808        return C_O1_I4(r, r, r, ri, ri);
3809
3810    case INDEX_op_ld_vec:
3811    case INDEX_op_dupm_vec:
3812        return C_O1_I1(x, r);
3813
3814    case INDEX_op_st_vec:
3815        return C_O0_I2(x, r);
3816
3817    case INDEX_op_add_vec:
3818    case INDEX_op_sub_vec:
3819    case INDEX_op_mul_vec:
3820    case INDEX_op_and_vec:
3821    case INDEX_op_or_vec:
3822    case INDEX_op_xor_vec:
3823    case INDEX_op_andc_vec:
3824    case INDEX_op_orc_vec:
3825    case INDEX_op_nand_vec:
3826    case INDEX_op_nor_vec:
3827    case INDEX_op_eqv_vec:
3828    case INDEX_op_ssadd_vec:
3829    case INDEX_op_usadd_vec:
3830    case INDEX_op_sssub_vec:
3831    case INDEX_op_ussub_vec:
3832    case INDEX_op_smin_vec:
3833    case INDEX_op_umin_vec:
3834    case INDEX_op_smax_vec:
3835    case INDEX_op_umax_vec:
3836    case INDEX_op_shlv_vec:
3837    case INDEX_op_shrv_vec:
3838    case INDEX_op_sarv_vec:
3839    case INDEX_op_rotlv_vec:
3840    case INDEX_op_rotrv_vec:
3841    case INDEX_op_shls_vec:
3842    case INDEX_op_shrs_vec:
3843    case INDEX_op_sars_vec:
3844    case INDEX_op_cmp_vec:
3845    case INDEX_op_x86_shufps_vec:
3846    case INDEX_op_x86_blend_vec:
3847    case INDEX_op_x86_packss_vec:
3848    case INDEX_op_x86_packus_vec:
3849    case INDEX_op_x86_vperm2i128_vec:
3850    case INDEX_op_x86_punpckl_vec:
3851    case INDEX_op_x86_punpckh_vec:
3852    case INDEX_op_x86_vpshldi_vec:
3853#if TCG_TARGET_REG_BITS == 32
3854    case INDEX_op_dup2_vec:
3855#endif
3856        return C_O1_I2(x, x, x);
3857
3858    case INDEX_op_abs_vec:
3859    case INDEX_op_dup_vec:
3860    case INDEX_op_not_vec:
3861    case INDEX_op_shli_vec:
3862    case INDEX_op_shri_vec:
3863    case INDEX_op_sari_vec:
3864    case INDEX_op_rotli_vec:
3865    case INDEX_op_x86_psrldq_vec:
3866        return C_O1_I1(x, x);
3867
3868    case INDEX_op_x86_vpshldv_vec:
3869    case INDEX_op_x86_vpshrdv_vec:
3870        return C_O1_I3(x, 0, x, x);
3871
3872    case INDEX_op_bitsel_vec:
3873        return C_O1_I3(x, x, x, x);
3874    case INDEX_op_cmpsel_vec:
3875        return C_O1_I4(x, x, x, xO, x);
3876
3877    default:
3878        return C_NotImplemented;
3879    }
3880}
3881
3882int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
3883{
3884    switch (opc) {
3885    case INDEX_op_add_vec:
3886    case INDEX_op_sub_vec:
3887    case INDEX_op_and_vec:
3888    case INDEX_op_or_vec:
3889    case INDEX_op_xor_vec:
3890    case INDEX_op_andc_vec:
3891    case INDEX_op_orc_vec:
3892    case INDEX_op_nand_vec:
3893    case INDEX_op_nor_vec:
3894    case INDEX_op_eqv_vec:
3895    case INDEX_op_not_vec:
3896    case INDEX_op_bitsel_vec:
3897        return 1;
3898    case INDEX_op_cmp_vec:
3899    case INDEX_op_cmpsel_vec:
3900        return -1;
3901
3902    case INDEX_op_rotli_vec:
3903        return have_avx512vl && vece >= MO_32 ? 1 : -1;
3904
3905    case INDEX_op_shli_vec:
3906    case INDEX_op_shri_vec:
3907        /* We must expand the operation for MO_8.  */
3908        return vece == MO_8 ? -1 : 1;
3909
3910    case INDEX_op_sari_vec:
3911        switch (vece) {
3912        case MO_8:
3913            return -1;
3914        case MO_16:
3915        case MO_32:
3916            return 1;
3917        case MO_64:
3918            if (have_avx512vl) {
3919                return 1;
3920            }
3921            /*
3922             * We can emulate this for MO_64, but it does not pay off
3923             * unless we're producing at least 4 values.
3924             */
3925            return type >= TCG_TYPE_V256 ? -1 : 0;
3926        }
3927        return 0;
3928
3929    case INDEX_op_shls_vec:
3930    case INDEX_op_shrs_vec:
3931        return vece >= MO_16;
3932    case INDEX_op_sars_vec:
3933        switch (vece) {
3934        case MO_16:
3935        case MO_32:
3936            return 1;
3937        case MO_64:
3938            return have_avx512vl;
3939        }
3940        return 0;
3941    case INDEX_op_rotls_vec:
3942        return vece >= MO_16 ? -1 : 0;
3943
3944    case INDEX_op_shlv_vec:
3945    case INDEX_op_shrv_vec:
3946        switch (vece) {
3947        case MO_16:
3948            return have_avx512bw;
3949        case MO_32:
3950        case MO_64:
3951            return have_avx2;
3952        }
3953        return 0;
3954    case INDEX_op_sarv_vec:
3955        switch (vece) {
3956        case MO_16:
3957            return have_avx512bw;
3958        case MO_32:
3959            return have_avx2;
3960        case MO_64:
3961            return have_avx512vl;
3962        }
3963        return 0;
3964    case INDEX_op_rotlv_vec:
3965    case INDEX_op_rotrv_vec:
3966        switch (vece) {
3967        case MO_16:
3968            return have_avx512vbmi2 ? -1 : 0;
3969        case MO_32:
3970        case MO_64:
3971            return have_avx512vl ? 1 : have_avx2 ? -1 : 0;
3972        }
3973        return 0;
3974
3975    case INDEX_op_mul_vec:
3976        switch (vece) {
3977        case MO_8:
3978            return -1;
3979        case MO_64:
3980            return have_avx512dq;
3981        }
3982        return 1;
3983
3984    case INDEX_op_ssadd_vec:
3985    case INDEX_op_usadd_vec:
3986    case INDEX_op_sssub_vec:
3987    case INDEX_op_ussub_vec:
3988        return vece <= MO_16;
3989    case INDEX_op_smin_vec:
3990    case INDEX_op_smax_vec:
3991    case INDEX_op_umin_vec:
3992    case INDEX_op_umax_vec:
3993    case INDEX_op_abs_vec:
3994        return vece <= MO_32 || have_avx512vl;
3995
3996    default:
3997        return 0;
3998    }
3999}
4000
4001static void expand_vec_shi(TCGType type, unsigned vece, bool right,
4002                           TCGv_vec v0, TCGv_vec v1, TCGArg imm)
4003{
4004    uint8_t mask;
4005
4006    tcg_debug_assert(vece == MO_8);
4007    if (right) {
4008        mask = 0xff >> imm;
4009        tcg_gen_shri_vec(MO_16, v0, v1, imm);
4010    } else {
4011        mask = 0xff << imm;
4012        tcg_gen_shli_vec(MO_16, v0, v1, imm);
4013    }
4014    tcg_gen_and_vec(MO_8, v0, v0, tcg_constant_vec(type, MO_8, mask));
4015}
4016
4017static void expand_vec_sari(TCGType type, unsigned vece,
4018                            TCGv_vec v0, TCGv_vec v1, TCGArg imm)
4019{
4020    TCGv_vec t1, t2;
4021
4022    switch (vece) {
4023    case MO_8:
4024        /* Unpack to 16-bit, shift, and repack.  */
4025        t1 = tcg_temp_new_vec(type);
4026        t2 = tcg_temp_new_vec(type);
4027        vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
4028                  tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
4029        vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
4030                  tcgv_vec_arg(t2), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
4031        tcg_gen_sari_vec(MO_16, t1, t1, imm + 8);
4032        tcg_gen_sari_vec(MO_16, t2, t2, imm + 8);
4033        vec_gen_3(INDEX_op_x86_packss_vec, type, MO_8,
4034                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t2));
4035        tcg_temp_free_vec(t1);
4036        tcg_temp_free_vec(t2);
4037        break;
4038
4039    case MO_64:
4040        t1 = tcg_temp_new_vec(type);
4041        if (imm <= 32) {
4042            /*
4043             * We can emulate a small sign extend by performing an arithmetic
4044             * 32-bit shift and overwriting the high half of a 64-bit logical
4045             * shift.  Note that the ISA says shift of 32 is valid, but TCG
4046             * does not, so we have to bound the smaller shift -- we get the
4047             * same result in the high half either way.
4048             */
4049            tcg_gen_sari_vec(MO_32, t1, v1, MIN(imm, 31));
4050            tcg_gen_shri_vec(MO_64, v0, v1, imm);
4051            vec_gen_4(INDEX_op_x86_blend_vec, type, MO_32,
4052                      tcgv_vec_arg(v0), tcgv_vec_arg(v0),
4053                      tcgv_vec_arg(t1), 0xaa);
4054        } else {
4055            /* Otherwise we will need to use a compare vs 0 to produce
4056             * the sign-extend, shift and merge.
4057             */
4058            tcg_gen_cmp_vec(TCG_COND_GT, MO_64, t1,
4059                            tcg_constant_vec(type, MO_64, 0), v1);
4060            tcg_gen_shri_vec(MO_64, v0, v1, imm);
4061            tcg_gen_shli_vec(MO_64, t1, t1, 64 - imm);
4062            tcg_gen_or_vec(MO_64, v0, v0, t1);
4063        }
4064        tcg_temp_free_vec(t1);
4065        break;
4066
4067    default:
4068        g_assert_not_reached();
4069    }
4070}
4071
4072static void expand_vec_rotli(TCGType type, unsigned vece,
4073                             TCGv_vec v0, TCGv_vec v1, TCGArg imm)
4074{
4075    TCGv_vec t;
4076
4077    if (vece != MO_8 && have_avx512vbmi2) {
4078        vec_gen_4(INDEX_op_x86_vpshldi_vec, type, vece,
4079                  tcgv_vec_arg(v0), tcgv_vec_arg(v1), tcgv_vec_arg(v1), imm);
4080        return;
4081    }
4082
4083    t = tcg_temp_new_vec(type);
4084    tcg_gen_shli_vec(vece, t, v1, imm);
4085    tcg_gen_shri_vec(vece, v0, v1, (8 << vece) - imm);
4086    tcg_gen_or_vec(vece, v0, v0, t);
4087    tcg_temp_free_vec(t);
4088}
4089
4090static void expand_vec_rotv(TCGType type, unsigned vece, TCGv_vec v0,
4091                            TCGv_vec v1, TCGv_vec sh, bool right)
4092{
4093    TCGv_vec t;
4094
4095    if (have_avx512vbmi2) {
4096        vec_gen_4(right ? INDEX_op_x86_vpshrdv_vec : INDEX_op_x86_vpshldv_vec,
4097                  type, vece, tcgv_vec_arg(v0), tcgv_vec_arg(v1),
4098                  tcgv_vec_arg(v1), tcgv_vec_arg(sh));
4099        return;
4100    }
4101
4102    t = tcg_temp_new_vec(type);
4103    tcg_gen_dupi_vec(vece, t, 8 << vece);
4104    tcg_gen_sub_vec(vece, t, t, sh);
4105    if (right) {
4106        tcg_gen_shlv_vec(vece, t, v1, t);
4107        tcg_gen_shrv_vec(vece, v0, v1, sh);
4108    } else {
4109        tcg_gen_shrv_vec(vece, t, v1, t);
4110        tcg_gen_shlv_vec(vece, v0, v1, sh);
4111    }
4112    tcg_gen_or_vec(vece, v0, v0, t);
4113    tcg_temp_free_vec(t);
4114}
4115
4116static void expand_vec_rotls(TCGType type, unsigned vece,
4117                             TCGv_vec v0, TCGv_vec v1, TCGv_i32 lsh)
4118{
4119    TCGv_vec t = tcg_temp_new_vec(type);
4120
4121    tcg_debug_assert(vece != MO_8);
4122
4123    if (vece >= MO_32 ? have_avx512vl : have_avx512vbmi2) {
4124        tcg_gen_dup_i32_vec(vece, t, lsh);
4125        if (vece >= MO_32) {
4126            tcg_gen_rotlv_vec(vece, v0, v1, t);
4127        } else {
4128            expand_vec_rotv(type, vece, v0, v1, t, false);
4129        }
4130    } else {
4131        TCGv_i32 rsh = tcg_temp_new_i32();
4132
4133        tcg_gen_neg_i32(rsh, lsh);
4134        tcg_gen_andi_i32(rsh, rsh, (8 << vece) - 1);
4135        tcg_gen_shls_vec(vece, t, v1, lsh);
4136        tcg_gen_shrs_vec(vece, v0, v1, rsh);
4137        tcg_gen_or_vec(vece, v0, v0, t);
4138
4139        tcg_temp_free_i32(rsh);
4140    }
4141
4142    tcg_temp_free_vec(t);
4143}
4144
4145static void expand_vec_mul(TCGType type, unsigned vece,
4146                           TCGv_vec v0, TCGv_vec v1, TCGv_vec v2)
4147{
4148    TCGv_vec t1, t2, t3, t4, zero;
4149
4150    tcg_debug_assert(vece == MO_8);
4151
4152    /*
4153     * Unpack v1 bytes to words, 0 | x.
4154     * Unpack v2 bytes to words, y | 0.
4155     * This leaves the 8-bit result, x * y, with 8 bits of right padding.
4156     * Shift logical right by 8 bits to clear the high 8 bytes before
4157     * using an unsigned saturated pack.
4158     *
4159     * The difference between the V64, V128 and V256 cases is merely how
4160     * we distribute the expansion between temporaries.
4161     */
4162    switch (type) {
4163    case TCG_TYPE_V64:
4164        t1 = tcg_temp_new_vec(TCG_TYPE_V128);
4165        t2 = tcg_temp_new_vec(TCG_TYPE_V128);
4166        zero = tcg_constant_vec(TCG_TYPE_V128, MO_8, 0);
4167        vec_gen_3(INDEX_op_x86_punpckl_vec, TCG_TYPE_V128, MO_8,
4168                  tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
4169        vec_gen_3(INDEX_op_x86_punpckl_vec, TCG_TYPE_V128, MO_8,
4170                  tcgv_vec_arg(t2), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
4171        tcg_gen_mul_vec(MO_16, t1, t1, t2);
4172        tcg_gen_shri_vec(MO_16, t1, t1, 8);
4173        vec_gen_3(INDEX_op_x86_packus_vec, TCG_TYPE_V128, MO_8,
4174                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t1));
4175        tcg_temp_free_vec(t1);
4176        tcg_temp_free_vec(t2);
4177        break;
4178
4179    case TCG_TYPE_V128:
4180    case TCG_TYPE_V256:
4181        t1 = tcg_temp_new_vec(type);
4182        t2 = tcg_temp_new_vec(type);
4183        t3 = tcg_temp_new_vec(type);
4184        t4 = tcg_temp_new_vec(type);
4185        zero = tcg_constant_vec(TCG_TYPE_V128, MO_8, 0);
4186        vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
4187                  tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
4188        vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
4189                  tcgv_vec_arg(t2), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
4190        vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
4191                  tcgv_vec_arg(t3), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
4192        vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
4193                  tcgv_vec_arg(t4), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
4194        tcg_gen_mul_vec(MO_16, t1, t1, t2);
4195        tcg_gen_mul_vec(MO_16, t3, t3, t4);
4196        tcg_gen_shri_vec(MO_16, t1, t1, 8);
4197        tcg_gen_shri_vec(MO_16, t3, t3, 8);
4198        vec_gen_3(INDEX_op_x86_packus_vec, type, MO_8,
4199                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t3));
4200        tcg_temp_free_vec(t1);
4201        tcg_temp_free_vec(t2);
4202        tcg_temp_free_vec(t3);
4203        tcg_temp_free_vec(t4);
4204        break;
4205
4206    default:
4207        g_assert_not_reached();
4208    }
4209}
4210
4211static TCGCond expand_vec_cond(TCGType type, unsigned vece,
4212                               TCGArg *a1, TCGArg *a2, TCGCond cond)
4213{
4214    /*
4215     * Without AVX512, there are no 64-bit unsigned comparisons.
4216     * We must bias the inputs so that they become signed.
4217     * All other swapping and inversion are handled during code generation.
4218     */
4219    if (vece == MO_64 && !have_avx512dq && is_unsigned_cond(cond)) {
4220        TCGv_vec v1 = temp_tcgv_vec(arg_temp(*a1));
4221        TCGv_vec v2 = temp_tcgv_vec(arg_temp(*a2));
4222        TCGv_vec t1 = tcg_temp_new_vec(type);
4223        TCGv_vec t2 = tcg_temp_new_vec(type);
4224        TCGv_vec t3 = tcg_constant_vec(type, vece, 1ull << ((8 << vece) - 1));
4225
4226        tcg_gen_sub_vec(vece, t1, v1, t3);
4227        tcg_gen_sub_vec(vece, t2, v2, t3);
4228        *a1 = tcgv_vec_arg(t1);
4229        *a2 = tcgv_vec_arg(t2);
4230        cond = tcg_signed_cond(cond);
4231    }
4232    return cond;
4233}
4234
4235static void expand_vec_cmp(TCGType type, unsigned vece, TCGArg a0,
4236                           TCGArg a1, TCGArg a2, TCGCond cond)
4237{
4238    cond = expand_vec_cond(type, vece, &a1, &a2, cond);
4239    /* Expand directly; do not recurse.  */
4240    vec_gen_4(INDEX_op_cmp_vec, type, vece, a0, a1, a2, cond);
4241}
4242
4243static void expand_vec_cmpsel(TCGType type, unsigned vece, TCGArg a0,
4244                              TCGArg a1, TCGArg a2,
4245                              TCGArg a3, TCGArg a4, TCGCond cond)
4246{
4247    cond = expand_vec_cond(type, vece, &a1, &a2, cond);
4248    /* Expand directly; do not recurse.  */
4249    vec_gen_6(INDEX_op_cmpsel_vec, type, vece, a0, a1, a2, a3, a4, cond);
4250}
4251
4252void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
4253                       TCGArg a0, ...)
4254{
4255    va_list va;
4256    TCGArg a1, a2, a3, a4, a5;
4257    TCGv_vec v0, v1, v2;
4258
4259    va_start(va, a0);
4260    a1 = va_arg(va, TCGArg);
4261    a2 = va_arg(va, TCGArg);
4262    v0 = temp_tcgv_vec(arg_temp(a0));
4263    v1 = temp_tcgv_vec(arg_temp(a1));
4264
4265    switch (opc) {
4266    case INDEX_op_shli_vec:
4267        expand_vec_shi(type, vece, false, v0, v1, a2);
4268        break;
4269    case INDEX_op_shri_vec:
4270        expand_vec_shi(type, vece, true, v0, v1, a2);
4271        break;
4272    case INDEX_op_sari_vec:
4273        expand_vec_sari(type, vece, v0, v1, a2);
4274        break;
4275
4276    case INDEX_op_rotli_vec:
4277        expand_vec_rotli(type, vece, v0, v1, a2);
4278        break;
4279
4280    case INDEX_op_rotls_vec:
4281        expand_vec_rotls(type, vece, v0, v1, temp_tcgv_i32(arg_temp(a2)));
4282        break;
4283
4284    case INDEX_op_rotlv_vec:
4285        v2 = temp_tcgv_vec(arg_temp(a2));
4286        expand_vec_rotv(type, vece, v0, v1, v2, false);
4287        break;
4288    case INDEX_op_rotrv_vec:
4289        v2 = temp_tcgv_vec(arg_temp(a2));
4290        expand_vec_rotv(type, vece, v0, v1, v2, true);
4291        break;
4292
4293    case INDEX_op_mul_vec:
4294        v2 = temp_tcgv_vec(arg_temp(a2));
4295        expand_vec_mul(type, vece, v0, v1, v2);
4296        break;
4297
4298    case INDEX_op_cmp_vec:
4299        a3 = va_arg(va, TCGArg);
4300        expand_vec_cmp(type, vece, a0, a1, a2, a3);
4301        break;
4302
4303    case INDEX_op_cmpsel_vec:
4304        a3 = va_arg(va, TCGArg);
4305        a4 = va_arg(va, TCGArg);
4306        a5 = va_arg(va, TCGArg);
4307        expand_vec_cmpsel(type, vece, a0, a1, a2, a3, a4, a5);
4308        break;
4309
4310    default:
4311        break;
4312    }
4313
4314    va_end(va);
4315}
4316
4317static const int tcg_target_callee_save_regs[] = {
4318#if TCG_TARGET_REG_BITS == 64
4319    TCG_REG_RBP,
4320    TCG_REG_RBX,
4321#if defined(_WIN64)
4322    TCG_REG_RDI,
4323    TCG_REG_RSI,
4324#endif
4325    TCG_REG_R12,
4326    TCG_REG_R13,
4327    TCG_REG_R14, /* Currently used for the global env. */
4328    TCG_REG_R15,
4329#else
4330    TCG_REG_EBP, /* Currently used for the global env. */
4331    TCG_REG_EBX,
4332    TCG_REG_ESI,
4333    TCG_REG_EDI,
4334#endif
4335};
4336
4337/* Compute frame size via macros, to share between tcg_target_qemu_prologue
4338   and tcg_register_jit.  */
4339
4340#define PUSH_SIZE \
4341    ((1 + ARRAY_SIZE(tcg_target_callee_save_regs)) \
4342     * (TCG_TARGET_REG_BITS / 8))
4343
4344#define FRAME_SIZE \
4345    ((PUSH_SIZE \
4346      + TCG_STATIC_CALL_ARGS_SIZE \
4347      + CPU_TEMP_BUF_NLONGS * sizeof(long) \
4348      + TCG_TARGET_STACK_ALIGN - 1) \
4349     & ~(TCG_TARGET_STACK_ALIGN - 1))
4350
4351/* Generate global QEMU prologue and epilogue code */
4352static void tcg_target_qemu_prologue(TCGContext *s)
4353{
4354    int i, stack_addend;
4355
4356    /* TB prologue */
4357
4358    /* Reserve some stack space, also for TCG temps.  */
4359    stack_addend = FRAME_SIZE - PUSH_SIZE;
4360    tcg_set_frame(s, TCG_REG_CALL_STACK, TCG_STATIC_CALL_ARGS_SIZE,
4361                  CPU_TEMP_BUF_NLONGS * sizeof(long));
4362
4363    /* Save all callee saved registers.  */
4364    for (i = 0; i < ARRAY_SIZE(tcg_target_callee_save_regs); i++) {
4365        tcg_out_push(s, tcg_target_callee_save_regs[i]);
4366    }
4367
4368    if (!tcg_use_softmmu && guest_base) {
4369        int seg = setup_guest_base_seg();
4370        if (seg != 0) {
4371            x86_guest_base.seg = seg;
4372        } else if (guest_base == (int32_t)guest_base) {
4373            x86_guest_base.ofs = guest_base;
4374        } else {
4375            assert(TCG_TARGET_REG_BITS == 64);
4376            /* Choose R12 because, as a base, it requires a SIB byte. */
4377            x86_guest_base.index = TCG_REG_R12;
4378            tcg_out_movi(s, TCG_TYPE_PTR, x86_guest_base.index, guest_base);
4379            tcg_regset_set_reg(s->reserved_regs, x86_guest_base.index);
4380        }
4381    }
4382
4383    if (TCG_TARGET_REG_BITS == 32) {
4384        tcg_out_ld(s, TCG_TYPE_PTR, TCG_AREG0, TCG_REG_ESP,
4385                   (ARRAY_SIZE(tcg_target_callee_save_regs) + 1) * 4);
4386        tcg_out_addi(s, TCG_REG_ESP, -stack_addend);
4387        /* jmp *tb.  */
4388        tcg_out_modrm_offset(s, OPC_GRP5, EXT5_JMPN_Ev, TCG_REG_ESP,
4389                             (ARRAY_SIZE(tcg_target_callee_save_regs) + 2) * 4
4390                             + stack_addend);
4391    } else {
4392        tcg_out_mov(s, TCG_TYPE_PTR, TCG_AREG0, tcg_target_call_iarg_regs[0]);
4393        tcg_out_addi(s, TCG_REG_ESP, -stack_addend);
4394        /* jmp *tb.  */
4395        tcg_out_modrm(s, OPC_GRP5, EXT5_JMPN_Ev, tcg_target_call_iarg_regs[1]);
4396    }
4397
4398    /*
4399     * Return path for goto_ptr. Set return value to 0, a-la exit_tb,
4400     * and fall through to the rest of the epilogue.
4401     */
4402    tcg_code_gen_epilogue = tcg_splitwx_to_rx(s->code_ptr);
4403    tcg_out_movi(s, TCG_TYPE_REG, TCG_REG_EAX, 0);
4404
4405    /* TB epilogue */
4406    tb_ret_addr = tcg_splitwx_to_rx(s->code_ptr);
4407
4408    tcg_out_addi(s, TCG_REG_CALL_STACK, stack_addend);
4409
4410    if (have_avx2) {
4411        tcg_out_vex_opc(s, OPC_VZEROUPPER, 0, 0, 0, 0);
4412    }
4413    for (i = ARRAY_SIZE(tcg_target_callee_save_regs) - 1; i >= 0; i--) {
4414        tcg_out_pop(s, tcg_target_callee_save_regs[i]);
4415    }
4416    tcg_out_opc(s, OPC_RET, 0, 0, 0);
4417}
4418
4419static void tcg_out_tb_start(TCGContext *s)
4420{
4421    /* nothing to do */
4422}
4423
4424static void tcg_out_nop_fill(tcg_insn_unit *p, int count)
4425{
4426    memset(p, 0x90, count);
4427}
4428
4429static void tcg_target_init(TCGContext *s)
4430{
4431    tcg_target_available_regs[TCG_TYPE_I32] = ALL_GENERAL_REGS;
4432    if (TCG_TARGET_REG_BITS == 64) {
4433        tcg_target_available_regs[TCG_TYPE_I64] = ALL_GENERAL_REGS;
4434    }
4435    if (have_avx1) {
4436        tcg_target_available_regs[TCG_TYPE_V64] = ALL_VECTOR_REGS;
4437        tcg_target_available_regs[TCG_TYPE_V128] = ALL_VECTOR_REGS;
4438    }
4439    if (have_avx2) {
4440        tcg_target_available_regs[TCG_TYPE_V256] = ALL_VECTOR_REGS;
4441    }
4442
4443    tcg_target_call_clobber_regs = ALL_VECTOR_REGS;
4444    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_EAX);
4445    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_EDX);
4446    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_ECX);
4447    if (TCG_TARGET_REG_BITS == 64) {
4448#if !defined(_WIN64)
4449        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_RDI);
4450        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_RSI);
4451#endif
4452        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R8);
4453        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R9);
4454        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R10);
4455        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R11);
4456    }
4457
4458    s->reserved_regs = 0;
4459    tcg_regset_set_reg(s->reserved_regs, TCG_REG_CALL_STACK);
4460    tcg_regset_set_reg(s->reserved_regs, TCG_TMP_VEC);
4461#ifdef _WIN64
4462    /* These are call saved, and we don't save them, so don't use them. */
4463    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM6);
4464    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM7);
4465    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM8);
4466    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM9);
4467    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM10);
4468    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM11);
4469    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM12);
4470    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM13);
4471    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM14);
4472    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM15);
4473#endif
4474}
4475
4476typedef struct {
4477    DebugFrameHeader h;
4478    uint8_t fde_def_cfa[4];
4479    uint8_t fde_reg_ofs[14];
4480} DebugFrame;
4481
4482/* We're expecting a 2 byte uleb128 encoded value.  */
4483QEMU_BUILD_BUG_ON(FRAME_SIZE >= (1 << 14));
4484
4485#if !defined(__ELF__)
4486    /* Host machine without ELF. */
4487#elif TCG_TARGET_REG_BITS == 64
4488#define ELF_HOST_MACHINE EM_X86_64
4489static const DebugFrame debug_frame = {
4490    .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
4491    .h.cie.id = -1,
4492    .h.cie.version = 1,
4493    .h.cie.code_align = 1,
4494    .h.cie.data_align = 0x78,             /* sleb128 -8 */
4495    .h.cie.return_column = 16,
4496
4497    /* Total FDE size does not include the "len" member.  */
4498    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
4499
4500    .fde_def_cfa = {
4501        12, 7,                          /* DW_CFA_def_cfa %rsp, ... */
4502        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
4503        (FRAME_SIZE >> 7)
4504    },
4505    .fde_reg_ofs = {
4506        0x90, 1,                        /* DW_CFA_offset, %rip, -8 */
4507        /* The following ordering must match tcg_target_callee_save_regs.  */
4508        0x86, 2,                        /* DW_CFA_offset, %rbp, -16 */
4509        0x83, 3,                        /* DW_CFA_offset, %rbx, -24 */
4510        0x8c, 4,                        /* DW_CFA_offset, %r12, -32 */
4511        0x8d, 5,                        /* DW_CFA_offset, %r13, -40 */
4512        0x8e, 6,                        /* DW_CFA_offset, %r14, -48 */
4513        0x8f, 7,                        /* DW_CFA_offset, %r15, -56 */
4514    }
4515};
4516#else
4517#define ELF_HOST_MACHINE EM_386
4518static const DebugFrame debug_frame = {
4519    .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
4520    .h.cie.id = -1,
4521    .h.cie.version = 1,
4522    .h.cie.code_align = 1,
4523    .h.cie.data_align = 0x7c,             /* sleb128 -4 */
4524    .h.cie.return_column = 8,
4525
4526    /* Total FDE size does not include the "len" member.  */
4527    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
4528
4529    .fde_def_cfa = {
4530        12, 4,                          /* DW_CFA_def_cfa %esp, ... */
4531        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
4532        (FRAME_SIZE >> 7)
4533    },
4534    .fde_reg_ofs = {
4535        0x88, 1,                        /* DW_CFA_offset, %eip, -4 */
4536        /* The following ordering must match tcg_target_callee_save_regs.  */
4537        0x85, 2,                        /* DW_CFA_offset, %ebp, -8 */
4538        0x83, 3,                        /* DW_CFA_offset, %ebx, -12 */
4539        0x86, 4,                        /* DW_CFA_offset, %esi, -16 */
4540        0x87, 5,                        /* DW_CFA_offset, %edi, -20 */
4541    }
4542};
4543#endif
4544
4545#if defined(ELF_HOST_MACHINE)
4546void tcg_register_jit(const void *buf, size_t buf_size)
4547{
4548    tcg_register_jit_int(buf, buf_size, &debug_frame, sizeof(debug_frame));
4549}
4550#endif
4551