xref: /openbmc/qemu/tcg/i386/tcg-target.c.inc (revision 6b8abd244b9355bc840bc14182aae9043f86f2f6)
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    case TCG_COND_NE:
1662    case TCG_COND_TSTNE:
1663        tcg_out_brcond(s, 0, cond, args[0], args[2], const_args[2],
1664                       label_this, small);
1665        tcg_out_brcond(s, 0, cond, args[1], args[3], const_args[3],
1666                       label_this, small);
1667        break;
1668    case TCG_COND_LT:
1669        tcg_out_brcond(s, 0, TCG_COND_LT, args[1], args[3], const_args[3],
1670                       label_this, small);
1671        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1672        tcg_out_brcond(s, 0, TCG_COND_LTU, args[0], args[2], const_args[2],
1673                       label_this, small);
1674        break;
1675    case TCG_COND_LE:
1676        tcg_out_brcond(s, 0, TCG_COND_LT, args[1], args[3], const_args[3],
1677                       label_this, small);
1678        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1679        tcg_out_brcond(s, 0, TCG_COND_LEU, args[0], args[2], const_args[2],
1680                       label_this, small);
1681        break;
1682    case TCG_COND_GT:
1683        tcg_out_brcond(s, 0, TCG_COND_GT, args[1], args[3], const_args[3],
1684                       label_this, small);
1685        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1686        tcg_out_brcond(s, 0, TCG_COND_GTU, args[0], args[2], const_args[2],
1687                       label_this, small);
1688        break;
1689    case TCG_COND_GE:
1690        tcg_out_brcond(s, 0, TCG_COND_GT, args[1], args[3], const_args[3],
1691                       label_this, small);
1692        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1693        tcg_out_brcond(s, 0, TCG_COND_GEU, args[0], args[2], const_args[2],
1694                       label_this, small);
1695        break;
1696    case TCG_COND_LTU:
1697        tcg_out_brcond(s, 0, TCG_COND_LTU, args[1], args[3], const_args[3],
1698                       label_this, small);
1699        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1700        tcg_out_brcond(s, 0, TCG_COND_LTU, args[0], args[2], const_args[2],
1701                       label_this, small);
1702        break;
1703    case TCG_COND_LEU:
1704        tcg_out_brcond(s, 0, TCG_COND_LTU, args[1], args[3], const_args[3],
1705                       label_this, small);
1706        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1707        tcg_out_brcond(s, 0, TCG_COND_LEU, args[0], args[2], const_args[2],
1708                       label_this, small);
1709        break;
1710    case TCG_COND_GTU:
1711        tcg_out_brcond(s, 0, TCG_COND_GTU, args[1], args[3], const_args[3],
1712                       label_this, small);
1713        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1714        tcg_out_brcond(s, 0, TCG_COND_GTU, args[0], args[2], const_args[2],
1715                       label_this, small);
1716        break;
1717    case TCG_COND_GEU:
1718        tcg_out_brcond(s, 0, TCG_COND_GTU, args[1], args[3], const_args[3],
1719                       label_this, small);
1720        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1721        tcg_out_brcond(s, 0, TCG_COND_GEU, args[0], args[2], const_args[2],
1722                       label_this, small);
1723        break;
1724    default:
1725        g_assert_not_reached();
1726    }
1727    tcg_out_label(s, label_next);
1728}
1729#endif
1730
1731static void tcg_out_setcond(TCGContext *s, int rexw, TCGCond cond,
1732                            TCGArg dest, TCGArg arg1, TCGArg arg2,
1733                            int const_arg2, bool neg)
1734{
1735    int cmp_rexw = rexw;
1736    bool inv = false;
1737    bool cleared;
1738    int jcc;
1739
1740    switch (cond) {
1741    case TCG_COND_NE:
1742        inv = true;
1743        /* fall through */
1744    case TCG_COND_EQ:
1745        /* If arg2 is 0, convert to LTU/GEU vs 1. */
1746        if (const_arg2 && arg2 == 0) {
1747            arg2 = 1;
1748            goto do_ltu;
1749        }
1750        break;
1751
1752    case TCG_COND_TSTNE:
1753        inv = true;
1754        /* fall through */
1755    case TCG_COND_TSTEQ:
1756        /* If arg2 is -1, convert to LTU/GEU vs 1. */
1757        if (const_arg2 && arg2 == 0xffffffffu) {
1758            arg2 = 1;
1759            cmp_rexw = 0;
1760            goto do_ltu;
1761        }
1762        break;
1763
1764    case TCG_COND_LEU:
1765        inv = true;
1766        /* fall through */
1767    case TCG_COND_GTU:
1768        /* If arg2 is a register, swap for LTU/GEU. */
1769        if (!const_arg2) {
1770            TCGReg t = arg1;
1771            arg1 = arg2;
1772            arg2 = t;
1773            goto do_ltu;
1774        }
1775        break;
1776
1777    case TCG_COND_GEU:
1778        inv = true;
1779        /* fall through */
1780    case TCG_COND_LTU:
1781    do_ltu:
1782        /*
1783         * Relying on the carry bit, use SBB to produce -1 if LTU, 0 if GEU.
1784         * We can then use NEG or INC to produce the desired result.
1785         * This is always smaller than the SETCC expansion.
1786         */
1787        tcg_out_cmp(s, TCG_COND_LTU, arg1, arg2, const_arg2, cmp_rexw);
1788
1789        /* X - X - C = -C = (C ? -1 : 0) */
1790        tgen_arithr(s, ARITH_SBB + (neg ? rexw : 0), dest, dest);
1791        if (inv && neg) {
1792            /* ~(C ? -1 : 0) = (C ? 0 : -1) */
1793            tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NOT, dest);
1794        } else if (inv) {
1795            /* (C ? -1 : 0) + 1 = (C ? 0 : 1) */
1796            tgen_arithi(s, ARITH_ADD, dest, 1, 0);
1797        } else if (!neg) {
1798            /* -(C ? -1 : 0) = (C ? 1 : 0) */
1799            tcg_out_modrm(s, OPC_GRP3_Ev, EXT3_NEG, dest);
1800        }
1801        return;
1802
1803    case TCG_COND_GE:
1804        inv = true;
1805        /* fall through */
1806    case TCG_COND_LT:
1807        /* If arg2 is 0, extract the sign bit. */
1808        if (const_arg2 && arg2 == 0) {
1809            tcg_out_mov(s, rexw ? TCG_TYPE_I64 : TCG_TYPE_I32, dest, arg1);
1810            if (inv) {
1811                tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NOT, dest);
1812            }
1813            tcg_out_shifti(s, (neg ? SHIFT_SAR : SHIFT_SHR) + rexw,
1814                           dest, rexw ? 63 : 31);
1815            return;
1816        }
1817        break;
1818
1819    default:
1820        break;
1821    }
1822
1823    /*
1824     * If dest does not overlap the inputs, clearing it first is preferred.
1825     * The XOR breaks any false dependency for the low-byte write to dest,
1826     * and is also one byte smaller than MOVZBL.
1827     */
1828    cleared = false;
1829    if (dest != arg1 && (const_arg2 || dest != arg2)) {
1830        tgen_arithr(s, ARITH_XOR, dest, dest);
1831        cleared = true;
1832    }
1833
1834    jcc = tcg_out_cmp(s, cond, arg1, arg2, const_arg2, cmp_rexw);
1835    tcg_out_modrm(s, OPC_SETCC | jcc, 0, dest);
1836
1837    if (!cleared) {
1838        tcg_out_ext8u(s, dest, dest);
1839    }
1840    if (neg) {
1841        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NEG, dest);
1842    }
1843}
1844
1845#if TCG_TARGET_REG_BITS == 32
1846static void tcg_out_setcond2(TCGContext *s, const TCGArg *args,
1847                             const int *const_args)
1848{
1849    TCGArg new_args[6];
1850    TCGLabel *label_true, *label_over;
1851
1852    memcpy(new_args, args+1, 5*sizeof(TCGArg));
1853
1854    if (args[0] == args[1] || args[0] == args[2]
1855        || (!const_args[3] && args[0] == args[3])
1856        || (!const_args[4] && args[0] == args[4])) {
1857        /* When the destination overlaps with one of the argument
1858           registers, don't do anything tricky.  */
1859        label_true = gen_new_label();
1860        label_over = gen_new_label();
1861
1862        new_args[5] = label_arg(label_true);
1863        tcg_out_brcond2(s, new_args, const_args+1, 1);
1864
1865        tcg_out_movi(s, TCG_TYPE_I32, args[0], 0);
1866        tcg_out_jxx(s, JCC_JMP, label_over, 1);
1867        tcg_out_label(s, label_true);
1868
1869        tcg_out_movi(s, TCG_TYPE_I32, args[0], 1);
1870        tcg_out_label(s, label_over);
1871    } else {
1872        /* When the destination does not overlap one of the arguments,
1873           clear the destination first, jump if cond false, and emit an
1874           increment in the true case.  This results in smaller code.  */
1875
1876        tcg_out_movi(s, TCG_TYPE_I32, args[0], 0);
1877
1878        label_over = gen_new_label();
1879        new_args[4] = tcg_invert_cond(new_args[4]);
1880        new_args[5] = label_arg(label_over);
1881        tcg_out_brcond2(s, new_args, const_args+1, 1);
1882
1883        tgen_arithi(s, ARITH_ADD, args[0], 1, 0);
1884        tcg_out_label(s, label_over);
1885    }
1886}
1887#endif
1888
1889static void tcg_out_cmov(TCGContext *s, int jcc, int rexw,
1890                         TCGReg dest, TCGReg v1)
1891{
1892    tcg_out_modrm(s, OPC_CMOVCC | jcc | rexw, dest, v1);
1893}
1894
1895static void tcg_out_movcond(TCGContext *s, int rexw, TCGCond cond,
1896                            TCGReg dest, TCGReg c1, TCGArg c2, int const_c2,
1897                            TCGReg v1)
1898{
1899    int jcc = tcg_out_cmp(s, cond, c1, c2, const_c2, rexw);
1900    tcg_out_cmov(s, jcc, rexw, dest, v1);
1901}
1902
1903static void tcg_out_ctz(TCGContext *s, int rexw, TCGReg dest, TCGReg arg1,
1904                        TCGArg arg2, bool const_a2)
1905{
1906    if (have_bmi1) {
1907        tcg_out_modrm(s, OPC_TZCNT + rexw, dest, arg1);
1908        if (const_a2) {
1909            tcg_debug_assert(arg2 == (rexw ? 64 : 32));
1910        } else {
1911            tcg_debug_assert(dest != arg2);
1912            tcg_out_cmov(s, JCC_JB, rexw, dest, arg2);
1913        }
1914    } else {
1915        tcg_debug_assert(dest != arg2);
1916        tcg_out_modrm(s, OPC_BSF + rexw, dest, arg1);
1917        tcg_out_cmov(s, JCC_JE, rexw, dest, arg2);
1918    }
1919}
1920
1921static void tcg_out_clz(TCGContext *s, int rexw, TCGReg dest, TCGReg arg1,
1922                        TCGArg arg2, bool const_a2)
1923{
1924    if (have_lzcnt) {
1925        tcg_out_modrm(s, OPC_LZCNT + rexw, dest, arg1);
1926        if (const_a2) {
1927            tcg_debug_assert(arg2 == (rexw ? 64 : 32));
1928        } else {
1929            tcg_debug_assert(dest != arg2);
1930            tcg_out_cmov(s, JCC_JB, rexw, dest, arg2);
1931        }
1932    } else {
1933        tcg_debug_assert(!const_a2);
1934        tcg_debug_assert(dest != arg1);
1935        tcg_debug_assert(dest != arg2);
1936
1937        /* Recall that the output of BSR is the index not the count.  */
1938        tcg_out_modrm(s, OPC_BSR + rexw, dest, arg1);
1939        tgen_arithi(s, ARITH_XOR + rexw, dest, rexw ? 63 : 31, 0);
1940
1941        /* Since we have destroyed the flags from BSR, we have to re-test.  */
1942        int jcc = tcg_out_cmp(s, TCG_COND_EQ, arg1, 0, 1, rexw);
1943        tcg_out_cmov(s, jcc, rexw, dest, arg2);
1944    }
1945}
1946
1947static void tcg_out_branch(TCGContext *s, int call, const tcg_insn_unit *dest)
1948{
1949    intptr_t disp = tcg_pcrel_diff(s, dest) - 5;
1950
1951    if (disp == (int32_t)disp) {
1952        tcg_out_opc(s, call ? OPC_CALL_Jz : OPC_JMP_long, 0, 0, 0);
1953        tcg_out32(s, disp);
1954    } else {
1955        /* rip-relative addressing into the constant pool.
1956           This is 6 + 8 = 14 bytes, as compared to using an
1957           immediate load 10 + 6 = 16 bytes, plus we may
1958           be able to re-use the pool constant for more calls.  */
1959        tcg_out_opc(s, OPC_GRP5, 0, 0, 0);
1960        tcg_out8(s, (call ? EXT5_CALLN_Ev : EXT5_JMPN_Ev) << 3 | 5);
1961        new_pool_label(s, (uintptr_t)dest, R_386_PC32, s->code_ptr, -4);
1962        tcg_out32(s, 0);
1963    }
1964}
1965
1966static void tcg_out_call(TCGContext *s, const tcg_insn_unit *dest,
1967                         const TCGHelperInfo *info)
1968{
1969    tcg_out_branch(s, 1, dest);
1970
1971#ifndef _WIN32
1972    if (TCG_TARGET_REG_BITS == 32 && info->out_kind == TCG_CALL_RET_BY_REF) {
1973        /*
1974         * The sysv i386 abi for struct return places a reference as the
1975         * first argument of the stack, and pops that argument with the
1976         * return statement.  Since we want to retain the aligned stack
1977         * pointer for the callee, we do not want to actually push that
1978         * argument before the call but rely on the normal store to the
1979         * stack slot.  But we do need to compensate for the pop in order
1980         * to reset our correct stack pointer value.
1981         * Pushing a garbage value back onto the stack is quickest.
1982         */
1983        tcg_out_push(s, TCG_REG_EAX);
1984    }
1985#endif
1986}
1987
1988static void tcg_out_jmp(TCGContext *s, const tcg_insn_unit *dest)
1989{
1990    tcg_out_branch(s, 0, dest);
1991}
1992
1993static void tcg_out_nopn(TCGContext *s, int n)
1994{
1995    int i;
1996    /* Emit 1 or 2 operand size prefixes for the standard one byte nop,
1997     * "xchg %eax,%eax", forming "xchg %ax,%ax". All cores accept the
1998     * duplicate prefix, and all of the interesting recent cores can
1999     * decode and discard the duplicates in a single cycle.
2000     */
2001    tcg_debug_assert(n >= 1);
2002    for (i = 1; i < n; ++i) {
2003        tcg_out8(s, 0x66);
2004    }
2005    tcg_out8(s, 0x90);
2006}
2007
2008typedef struct {
2009    TCGReg base;
2010    int index;
2011    int ofs;
2012    int seg;
2013    TCGAtomAlign aa;
2014} HostAddress;
2015
2016bool tcg_target_has_memory_bswap(MemOp memop)
2017{
2018    TCGAtomAlign aa;
2019
2020    if (!have_movbe) {
2021        return false;
2022    }
2023    if ((memop & MO_SIZE) < MO_128) {
2024        return true;
2025    }
2026
2027    /*
2028     * Reject 16-byte memop with 16-byte atomicity, i.e. VMOVDQA,
2029     * but do allow a pair of 64-bit operations, i.e. MOVBEQ.
2030     */
2031    aa = atom_and_align_for_opc(tcg_ctx, memop, MO_ATOM_IFALIGN, true);
2032    return aa.atom < MO_128;
2033}
2034
2035/*
2036 * Because i686 has no register parameters and because x86_64 has xchg
2037 * to handle addr/data register overlap, we have placed all input arguments
2038 * before we need might need a scratch reg.
2039 *
2040 * Even then, a scratch is only needed for l->raddr.  Rather than expose
2041 * a general-purpose scratch when we don't actually know it's available,
2042 * use the ra_gen hook to load into RAX if needed.
2043 */
2044#if TCG_TARGET_REG_BITS == 64
2045static TCGReg ldst_ra_gen(TCGContext *s, const TCGLabelQemuLdst *l, int arg)
2046{
2047    if (arg < 0) {
2048        arg = TCG_REG_RAX;
2049    }
2050    tcg_out_movi(s, TCG_TYPE_PTR, arg, (uintptr_t)l->raddr);
2051    return arg;
2052}
2053static const TCGLdstHelperParam ldst_helper_param = {
2054    .ra_gen = ldst_ra_gen
2055};
2056#else
2057static const TCGLdstHelperParam ldst_helper_param = { };
2058#endif
2059
2060static void tcg_out_vec_to_pair(TCGContext *s, TCGType type,
2061                                TCGReg l, TCGReg h, TCGReg v)
2062{
2063    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2064
2065    /* vpmov{d,q} %v, %l */
2066    tcg_out_vex_modrm(s, OPC_MOVD_EyVy + rexw, v, 0, l);
2067    /* vpextr{d,q} $1, %v, %h */
2068    tcg_out_vex_modrm(s, OPC_PEXTRD + rexw, v, 0, h);
2069    tcg_out8(s, 1);
2070}
2071
2072static void tcg_out_pair_to_vec(TCGContext *s, TCGType type,
2073                                TCGReg v, TCGReg l, TCGReg h)
2074{
2075    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2076
2077    /* vmov{d,q} %l, %v */
2078    tcg_out_vex_modrm(s, OPC_MOVD_VyEy + rexw, v, 0, l);
2079    /* vpinsr{d,q} $1, %h, %v, %v */
2080    tcg_out_vex_modrm(s, OPC_PINSRD + rexw, v, v, h);
2081    tcg_out8(s, 1);
2082}
2083
2084/*
2085 * Generate code for the slow path for a load at the end of block
2086 */
2087static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
2088{
2089    MemOp opc = get_memop(l->oi);
2090    tcg_insn_unit **label_ptr = &l->label_ptr[0];
2091
2092    /* resolve label address */
2093    tcg_patch32(label_ptr[0], s->code_ptr - label_ptr[0] - 4);
2094    if (label_ptr[1]) {
2095        tcg_patch32(label_ptr[1], s->code_ptr - label_ptr[1] - 4);
2096    }
2097
2098    tcg_out_ld_helper_args(s, l, &ldst_helper_param);
2099    tcg_out_branch(s, 1, qemu_ld_helpers[opc & MO_SIZE]);
2100    tcg_out_ld_helper_ret(s, l, false, &ldst_helper_param);
2101
2102    tcg_out_jmp(s, l->raddr);
2103    return true;
2104}
2105
2106/*
2107 * Generate code for the slow path for a store at the end of block
2108 */
2109static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
2110{
2111    MemOp opc = get_memop(l->oi);
2112    tcg_insn_unit **label_ptr = &l->label_ptr[0];
2113
2114    /* resolve label address */
2115    tcg_patch32(label_ptr[0], s->code_ptr - label_ptr[0] - 4);
2116    if (label_ptr[1]) {
2117        tcg_patch32(label_ptr[1], s->code_ptr - label_ptr[1] - 4);
2118    }
2119
2120    tcg_out_st_helper_args(s, l, &ldst_helper_param);
2121    tcg_out_branch(s, 1, qemu_st_helpers[opc & MO_SIZE]);
2122
2123    tcg_out_jmp(s, l->raddr);
2124    return true;
2125}
2126
2127#ifdef CONFIG_USER_ONLY
2128static HostAddress x86_guest_base = {
2129    .index = -1
2130};
2131
2132#if defined(__x86_64__) && defined(__linux__)
2133# include <asm/prctl.h>
2134# include <sys/prctl.h>
2135int arch_prctl(int code, unsigned long addr);
2136static inline int setup_guest_base_seg(void)
2137{
2138    if (arch_prctl(ARCH_SET_GS, guest_base) == 0) {
2139        return P_GS;
2140    }
2141    return 0;
2142}
2143#define setup_guest_base_seg  setup_guest_base_seg
2144#elif defined(__x86_64__) && \
2145      (defined (__FreeBSD__) || defined (__FreeBSD_kernel__))
2146# include <machine/sysarch.h>
2147static inline int setup_guest_base_seg(void)
2148{
2149    if (sysarch(AMD64_SET_GSBASE, &guest_base) == 0) {
2150        return P_GS;
2151    }
2152    return 0;
2153}
2154#define setup_guest_base_seg  setup_guest_base_seg
2155#endif
2156#else
2157# define x86_guest_base (*(HostAddress *)({ qemu_build_not_reached(); NULL; }))
2158#endif /* CONFIG_USER_ONLY */
2159#ifndef setup_guest_base_seg
2160# define setup_guest_base_seg()  0
2161#endif
2162
2163#define MIN_TLB_MASK_TABLE_OFS  INT_MIN
2164
2165/*
2166 * For softmmu, perform the TLB load and compare.
2167 * For useronly, perform any required alignment tests.
2168 * In both cases, return a TCGLabelQemuLdst structure if the slow path
2169 * is required and fill in @h with the host address for the fast path.
2170 */
2171static TCGLabelQemuLdst *prepare_host_addr(TCGContext *s, HostAddress *h,
2172                                           TCGReg addr, MemOpIdx oi, bool is_ld)
2173{
2174    TCGLabelQemuLdst *ldst = NULL;
2175    MemOp opc = get_memop(oi);
2176    MemOp s_bits = opc & MO_SIZE;
2177    unsigned a_mask;
2178
2179    if (tcg_use_softmmu) {
2180        h->index = TCG_REG_L0;
2181        h->ofs = 0;
2182        h->seg = 0;
2183    } else {
2184        *h = x86_guest_base;
2185    }
2186    h->base = addr;
2187    h->aa = atom_and_align_for_opc(s, opc, MO_ATOM_IFALIGN, s_bits == MO_128);
2188    a_mask = (1 << h->aa.align) - 1;
2189
2190    if (tcg_use_softmmu) {
2191        int cmp_ofs = is_ld ? offsetof(CPUTLBEntry, addr_read)
2192                            : offsetof(CPUTLBEntry, addr_write);
2193        TCGType ttype = TCG_TYPE_I32;
2194        TCGType tlbtype = TCG_TYPE_I32;
2195        int trexw = 0, hrexw = 0, tlbrexw = 0;
2196        unsigned mem_index = get_mmuidx(oi);
2197        unsigned s_mask = (1 << s_bits) - 1;
2198        int fast_ofs = tlb_mask_table_ofs(s, mem_index);
2199        int tlb_mask;
2200
2201        ldst = new_ldst_label(s);
2202        ldst->is_ld = is_ld;
2203        ldst->oi = oi;
2204        ldst->addr_reg = addr;
2205
2206        if (TCG_TARGET_REG_BITS == 64) {
2207            ttype = s->addr_type;
2208            trexw = (ttype == TCG_TYPE_I32 ? 0 : P_REXW);
2209            if (TCG_TYPE_PTR == TCG_TYPE_I64) {
2210                hrexw = P_REXW;
2211                if (s->page_bits + s->tlb_dyn_max_bits > 32) {
2212                    tlbtype = TCG_TYPE_I64;
2213                    tlbrexw = P_REXW;
2214                }
2215            }
2216        }
2217
2218        tcg_out_mov(s, tlbtype, TCG_REG_L0, addr);
2219        tcg_out_shifti(s, SHIFT_SHR + tlbrexw, TCG_REG_L0,
2220                       s->page_bits - CPU_TLB_ENTRY_BITS);
2221
2222        tcg_out_modrm_offset(s, OPC_AND_GvEv + trexw, TCG_REG_L0, TCG_AREG0,
2223                             fast_ofs + offsetof(CPUTLBDescFast, mask));
2224
2225        tcg_out_modrm_offset(s, OPC_ADD_GvEv + hrexw, TCG_REG_L0, TCG_AREG0,
2226                             fast_ofs + offsetof(CPUTLBDescFast, table));
2227
2228        /*
2229         * If the required alignment is at least as large as the access,
2230         * simply copy the address and mask.  For lesser alignments,
2231         * check that we don't cross pages for the complete access.
2232         */
2233        if (a_mask >= s_mask) {
2234            tcg_out_mov(s, ttype, TCG_REG_L1, addr);
2235        } else {
2236            tcg_out_modrm_offset(s, OPC_LEA + trexw, TCG_REG_L1,
2237                                 addr, s_mask - a_mask);
2238        }
2239        tlb_mask = s->page_mask | a_mask;
2240        tgen_arithi(s, ARITH_AND + trexw, TCG_REG_L1, tlb_mask, 0);
2241
2242        /* cmp 0(TCG_REG_L0), TCG_REG_L1 */
2243        tcg_out_modrm_offset(s, OPC_CMP_GvEv + trexw,
2244                             TCG_REG_L1, TCG_REG_L0, cmp_ofs);
2245
2246        /* jne slow_path */
2247        tcg_out_opc(s, OPC_JCC_long + JCC_JNE, 0, 0, 0);
2248        ldst->label_ptr[0] = s->code_ptr;
2249        s->code_ptr += 4;
2250
2251        /* TLB Hit.  */
2252        tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_L0, TCG_REG_L0,
2253                   offsetof(CPUTLBEntry, addend));
2254    } else if (a_mask) {
2255        int jcc;
2256
2257        ldst = new_ldst_label(s);
2258        ldst->is_ld = is_ld;
2259        ldst->oi = oi;
2260        ldst->addr_reg = addr;
2261
2262        /* jne slow_path */
2263        jcc = tcg_out_cmp(s, TCG_COND_TSTNE, addr, a_mask, true, false);
2264        tcg_out_opc(s, OPC_JCC_long + jcc, 0, 0, 0);
2265        ldst->label_ptr[0] = s->code_ptr;
2266        s->code_ptr += 4;
2267    }
2268
2269    return ldst;
2270}
2271
2272static void tcg_out_qemu_ld_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
2273                                   HostAddress h, TCGType type, MemOp memop)
2274{
2275    bool use_movbe = false;
2276    int rexw = (type == TCG_TYPE_I32 ? 0 : P_REXW);
2277    int movop = OPC_MOVL_GvEv;
2278
2279    /* Do big-endian loads with movbe.  */
2280    if (memop & MO_BSWAP) {
2281        tcg_debug_assert(have_movbe);
2282        use_movbe = true;
2283        movop = OPC_MOVBE_GyMy;
2284    }
2285
2286    switch (memop & MO_SSIZE) {
2287    case MO_UB:
2288        tcg_out_modrm_sib_offset(s, OPC_MOVZBL + h.seg, datalo,
2289                                 h.base, h.index, 0, h.ofs);
2290        break;
2291    case MO_SB:
2292        tcg_out_modrm_sib_offset(s, OPC_MOVSBL + rexw + h.seg, datalo,
2293                                 h.base, h.index, 0, h.ofs);
2294        break;
2295    case MO_UW:
2296        if (use_movbe) {
2297            /* There is no extending movbe; only low 16-bits are modified.  */
2298            if (datalo != h.base && datalo != h.index) {
2299                /* XOR breaks dependency chains.  */
2300                tgen_arithr(s, ARITH_XOR, datalo, datalo);
2301                tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + P_DATA16 + h.seg,
2302                                         datalo, h.base, h.index, 0, h.ofs);
2303            } else {
2304                tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + P_DATA16 + h.seg,
2305                                         datalo, h.base, h.index, 0, h.ofs);
2306                tcg_out_ext16u(s, datalo, datalo);
2307            }
2308        } else {
2309            tcg_out_modrm_sib_offset(s, OPC_MOVZWL + h.seg, datalo,
2310                                     h.base, h.index, 0, h.ofs);
2311        }
2312        break;
2313    case MO_SW:
2314        if (use_movbe) {
2315            tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + P_DATA16 + h.seg,
2316                                     datalo, h.base, h.index, 0, h.ofs);
2317            tcg_out_ext16s(s, type, datalo, datalo);
2318        } else {
2319            tcg_out_modrm_sib_offset(s, OPC_MOVSWL + rexw + h.seg,
2320                                     datalo, h.base, h.index, 0, h.ofs);
2321        }
2322        break;
2323    case MO_UL:
2324        tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2325                                 h.base, h.index, 0, h.ofs);
2326        break;
2327#if TCG_TARGET_REG_BITS == 64
2328    case MO_SL:
2329        if (use_movbe) {
2330            tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + h.seg, datalo,
2331                                     h.base, h.index, 0, h.ofs);
2332            tcg_out_ext32s(s, datalo, datalo);
2333        } else {
2334            tcg_out_modrm_sib_offset(s, OPC_MOVSLQ + h.seg, datalo,
2335                                     h.base, h.index, 0, h.ofs);
2336        }
2337        break;
2338#endif
2339    case MO_UQ:
2340        if (TCG_TARGET_REG_BITS == 64) {
2341            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2342                                     h.base, h.index, 0, h.ofs);
2343            break;
2344        }
2345        if (use_movbe) {
2346            TCGReg t = datalo;
2347            datalo = datahi;
2348            datahi = t;
2349        }
2350        if (h.base == datalo || h.index == datalo) {
2351            tcg_out_modrm_sib_offset(s, OPC_LEA, datahi,
2352                                     h.base, h.index, 0, h.ofs);
2353            tcg_out_modrm_offset(s, movop + h.seg, datalo, datahi, 0);
2354            tcg_out_modrm_offset(s, movop + h.seg, datahi, datahi, 4);
2355        } else {
2356            tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2357                                     h.base, h.index, 0, h.ofs);
2358            tcg_out_modrm_sib_offset(s, movop + h.seg, datahi,
2359                                     h.base, h.index, 0, h.ofs + 4);
2360        }
2361        break;
2362
2363    case MO_128:
2364        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
2365
2366        /*
2367         * Without 16-byte atomicity, use integer regs.
2368         * That is where we want the data, and it allows bswaps.
2369         */
2370        if (h.aa.atom < MO_128) {
2371            if (use_movbe) {
2372                TCGReg t = datalo;
2373                datalo = datahi;
2374                datahi = t;
2375            }
2376            if (h.base == datalo || h.index == datalo) {
2377                tcg_out_modrm_sib_offset(s, OPC_LEA + P_REXW, datahi,
2378                                         h.base, h.index, 0, h.ofs);
2379                tcg_out_modrm_offset(s, movop + P_REXW + h.seg,
2380                                     datalo, datahi, 0);
2381                tcg_out_modrm_offset(s, movop + P_REXW + h.seg,
2382                                     datahi, datahi, 8);
2383            } else {
2384                tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2385                                         h.base, h.index, 0, h.ofs);
2386                tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datahi,
2387                                         h.base, h.index, 0, h.ofs + 8);
2388            }
2389            break;
2390        }
2391
2392        /*
2393         * With 16-byte atomicity, a vector load is required.
2394         * If we already have 16-byte alignment, then VMOVDQA always works.
2395         * Else if VMOVDQU has atomicity with dynamic alignment, use that.
2396         * Else use we require a runtime test for alignment for VMOVDQA;
2397         * use VMOVDQU on the unaligned nonatomic path for simplicity.
2398         */
2399        if (h.aa.align >= MO_128) {
2400            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_VxWx + h.seg,
2401                                         TCG_TMP_VEC, 0,
2402                                         h.base, h.index, 0, h.ofs);
2403        } else if (cpuinfo & CPUINFO_ATOMIC_VMOVDQU) {
2404            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_VxWx + h.seg,
2405                                         TCG_TMP_VEC, 0,
2406                                         h.base, h.index, 0, h.ofs);
2407        } else {
2408            TCGLabel *l1 = gen_new_label();
2409            TCGLabel *l2 = gen_new_label();
2410            int jcc;
2411
2412            jcc = tcg_out_cmp(s, TCG_COND_TSTNE, h.base, 15, true, false);
2413            tcg_out_jxx(s, jcc, l1, true);
2414
2415            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_VxWx + h.seg,
2416                                         TCG_TMP_VEC, 0,
2417                                         h.base, h.index, 0, h.ofs);
2418            tcg_out_jxx(s, JCC_JMP, l2, true);
2419
2420            tcg_out_label(s, l1);
2421            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_VxWx + h.seg,
2422                                         TCG_TMP_VEC, 0,
2423                                         h.base, h.index, 0, h.ofs);
2424            tcg_out_label(s, l2);
2425        }
2426        tcg_out_vec_to_pair(s, TCG_TYPE_I64, datalo, datahi, TCG_TMP_VEC);
2427        break;
2428
2429    default:
2430        g_assert_not_reached();
2431    }
2432}
2433
2434static void tcg_out_qemu_ld(TCGContext *s, TCGReg datalo, TCGReg datahi,
2435                            TCGReg addr, MemOpIdx oi, TCGType data_type)
2436{
2437    TCGLabelQemuLdst *ldst;
2438    HostAddress h;
2439
2440    ldst = prepare_host_addr(s, &h, addr, oi, true);
2441    tcg_out_qemu_ld_direct(s, datalo, datahi, h, data_type, get_memop(oi));
2442
2443    if (ldst) {
2444        ldst->type = data_type;
2445        ldst->datalo_reg = datalo;
2446        ldst->datahi_reg = datahi;
2447        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
2448    }
2449}
2450
2451static void tcg_out_qemu_st_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
2452                                   HostAddress h, MemOp memop)
2453{
2454    bool use_movbe = false;
2455    int movop = OPC_MOVL_EvGv;
2456
2457    /*
2458     * Do big-endian stores with movbe or system-mode.
2459     * User-only without movbe will have its swapping done generically.
2460     */
2461    if (memop & MO_BSWAP) {
2462        tcg_debug_assert(have_movbe);
2463        use_movbe = true;
2464        movop = OPC_MOVBE_MyGy;
2465    }
2466
2467    switch (memop & MO_SIZE) {
2468    case MO_8:
2469        /* This is handled with constraints on INDEX_op_qemu_st8_i32. */
2470        tcg_debug_assert(TCG_TARGET_REG_BITS == 64 || datalo < 4);
2471        tcg_out_modrm_sib_offset(s, OPC_MOVB_EvGv + P_REXB_R + h.seg,
2472                                 datalo, h.base, h.index, 0, h.ofs);
2473        break;
2474    case MO_16:
2475        tcg_out_modrm_sib_offset(s, movop + P_DATA16 + h.seg, datalo,
2476                                 h.base, h.index, 0, h.ofs);
2477        break;
2478    case MO_32:
2479        tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2480                                 h.base, h.index, 0, h.ofs);
2481        break;
2482    case MO_64:
2483        if (TCG_TARGET_REG_BITS == 64) {
2484            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2485                                     h.base, h.index, 0, h.ofs);
2486        } else {
2487            if (use_movbe) {
2488                TCGReg t = datalo;
2489                datalo = datahi;
2490                datahi = t;
2491            }
2492            tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2493                                     h.base, h.index, 0, h.ofs);
2494            tcg_out_modrm_sib_offset(s, movop + h.seg, datahi,
2495                                     h.base, h.index, 0, h.ofs + 4);
2496        }
2497        break;
2498
2499    case MO_128:
2500        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
2501
2502        /*
2503         * Without 16-byte atomicity, use integer regs.
2504         * That is where we have the data, and it allows bswaps.
2505         */
2506        if (h.aa.atom < MO_128) {
2507            if (use_movbe) {
2508                TCGReg t = datalo;
2509                datalo = datahi;
2510                datahi = t;
2511            }
2512            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2513                                     h.base, h.index, 0, h.ofs);
2514            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datahi,
2515                                     h.base, h.index, 0, h.ofs + 8);
2516            break;
2517        }
2518
2519        /*
2520         * With 16-byte atomicity, a vector store is required.
2521         * If we already have 16-byte alignment, then VMOVDQA always works.
2522         * Else if VMOVDQU has atomicity with dynamic alignment, use that.
2523         * Else use we require a runtime test for alignment for VMOVDQA;
2524         * use VMOVDQU on the unaligned nonatomic path for simplicity.
2525         */
2526        tcg_out_pair_to_vec(s, TCG_TYPE_I64, TCG_TMP_VEC, datalo, datahi);
2527        if (h.aa.align >= MO_128) {
2528            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_WxVx + h.seg,
2529                                         TCG_TMP_VEC, 0,
2530                                         h.base, h.index, 0, h.ofs);
2531        } else if (cpuinfo & CPUINFO_ATOMIC_VMOVDQU) {
2532            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_WxVx + h.seg,
2533                                         TCG_TMP_VEC, 0,
2534                                         h.base, h.index, 0, h.ofs);
2535        } else {
2536            TCGLabel *l1 = gen_new_label();
2537            TCGLabel *l2 = gen_new_label();
2538            int jcc;
2539
2540            jcc = tcg_out_cmp(s, TCG_COND_TSTNE, h.base, 15, true, false);
2541            tcg_out_jxx(s, jcc, l1, true);
2542
2543            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_WxVx + h.seg,
2544                                         TCG_TMP_VEC, 0,
2545                                         h.base, h.index, 0, h.ofs);
2546            tcg_out_jxx(s, JCC_JMP, l2, true);
2547
2548            tcg_out_label(s, l1);
2549            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_WxVx + h.seg,
2550                                         TCG_TMP_VEC, 0,
2551                                         h.base, h.index, 0, h.ofs);
2552            tcg_out_label(s, l2);
2553        }
2554        break;
2555
2556    default:
2557        g_assert_not_reached();
2558    }
2559}
2560
2561static void tcg_out_qemu_st(TCGContext *s, TCGReg datalo, TCGReg datahi,
2562                            TCGReg addr, MemOpIdx oi, TCGType data_type)
2563{
2564    TCGLabelQemuLdst *ldst;
2565    HostAddress h;
2566
2567    ldst = prepare_host_addr(s, &h, addr, oi, false);
2568    tcg_out_qemu_st_direct(s, datalo, datahi, h, get_memop(oi));
2569
2570    if (ldst) {
2571        ldst->type = data_type;
2572        ldst->datalo_reg = datalo;
2573        ldst->datahi_reg = datahi;
2574        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
2575    }
2576}
2577
2578static void tcg_out_exit_tb(TCGContext *s, uintptr_t a0)
2579{
2580    /* Reuse the zeroing that exists for goto_ptr.  */
2581    if (a0 == 0) {
2582        tcg_out_jmp(s, tcg_code_gen_epilogue);
2583    } else {
2584        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_EAX, a0);
2585        tcg_out_jmp(s, tb_ret_addr);
2586    }
2587}
2588
2589static void tcg_out_goto_tb(TCGContext *s, int which)
2590{
2591    /*
2592     * Jump displacement must be aligned for atomic patching;
2593     * see if we need to add extra nops before jump
2594     */
2595    int gap = QEMU_ALIGN_PTR_UP(s->code_ptr + 1, 4) - s->code_ptr;
2596    if (gap != 1) {
2597        tcg_out_nopn(s, gap - 1);
2598    }
2599    tcg_out8(s, OPC_JMP_long); /* jmp im */
2600    set_jmp_insn_offset(s, which);
2601    tcg_out32(s, 0);
2602    set_jmp_reset_offset(s, which);
2603}
2604
2605void tb_target_set_jmp_target(const TranslationBlock *tb, int n,
2606                              uintptr_t jmp_rx, uintptr_t jmp_rw)
2607{
2608    /* patch the branch destination */
2609    uintptr_t addr = tb->jmp_target_addr[n];
2610    qatomic_set((int32_t *)jmp_rw, addr - (jmp_rx + 4));
2611    /* no need to flush icache explicitly */
2612}
2613
2614static void tcg_out_op(TCGContext *s, TCGOpcode opc, TCGType type,
2615                       const TCGArg args[TCG_MAX_OP_ARGS],
2616                       const int const_args[TCG_MAX_OP_ARGS])
2617{
2618    TCGArg a0, a1, a2;
2619    int c, const_a2, vexop, rexw;
2620
2621#if TCG_TARGET_REG_BITS == 64
2622# define OP_32_64(x) \
2623        case glue(glue(INDEX_op_, x), _i64): \
2624        case glue(glue(INDEX_op_, x), _i32)
2625#else
2626# define OP_32_64(x) \
2627        case glue(glue(INDEX_op_, x), _i32)
2628#endif
2629
2630    /* Hoist the loads of the most common arguments.  */
2631    a0 = args[0];
2632    a1 = args[1];
2633    a2 = args[2];
2634    const_a2 = const_args[2];
2635    rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2636
2637    switch (opc) {
2638    case INDEX_op_goto_ptr:
2639        /* jmp to the given host address (could be epilogue) */
2640        tcg_out_modrm(s, OPC_GRP5, EXT5_JMPN_Ev, a0);
2641        break;
2642    case INDEX_op_br:
2643        tcg_out_jxx(s, JCC_JMP, arg_label(a0), 0);
2644        break;
2645    OP_32_64(ld8u):
2646        /* Note that we can ignore REXW for the zero-extend to 64-bit.  */
2647        tcg_out_modrm_offset(s, OPC_MOVZBL, a0, a1, a2);
2648        break;
2649    OP_32_64(ld8s):
2650        tcg_out_modrm_offset(s, OPC_MOVSBL + rexw, a0, a1, a2);
2651        break;
2652    OP_32_64(ld16u):
2653        /* Note that we can ignore REXW for the zero-extend to 64-bit.  */
2654        tcg_out_modrm_offset(s, OPC_MOVZWL, a0, a1, a2);
2655        break;
2656    OP_32_64(ld16s):
2657        tcg_out_modrm_offset(s, OPC_MOVSWL + rexw, a0, a1, a2);
2658        break;
2659#if TCG_TARGET_REG_BITS == 64
2660    case INDEX_op_ld32u_i64:
2661#endif
2662    case INDEX_op_ld_i32:
2663        tcg_out_ld(s, TCG_TYPE_I32, a0, a1, a2);
2664        break;
2665
2666    OP_32_64(st8):
2667        if (const_args[0]) {
2668            tcg_out_modrm_offset(s, OPC_MOVB_EvIz, 0, a1, a2);
2669            tcg_out8(s, a0);
2670        } else {
2671            tcg_out_modrm_offset(s, OPC_MOVB_EvGv | P_REXB_R, a0, a1, a2);
2672        }
2673        break;
2674    OP_32_64(st16):
2675        if (const_args[0]) {
2676            tcg_out_modrm_offset(s, OPC_MOVL_EvIz | P_DATA16, 0, a1, a2);
2677            tcg_out16(s, a0);
2678        } else {
2679            tcg_out_modrm_offset(s, OPC_MOVL_EvGv | P_DATA16, a0, a1, a2);
2680        }
2681        break;
2682#if TCG_TARGET_REG_BITS == 64
2683    case INDEX_op_st32_i64:
2684#endif
2685    case INDEX_op_st_i32:
2686        if (const_args[0]) {
2687            tcg_out_modrm_offset(s, OPC_MOVL_EvIz, 0, a1, a2);
2688            tcg_out32(s, a0);
2689        } else {
2690            tcg_out_st(s, TCG_TYPE_I32, a0, a1, a2);
2691        }
2692        break;
2693
2694    OP_32_64(add):
2695        /* For 3-operand addition, use LEA.  */
2696        if (a0 != a1) {
2697            TCGArg c3 = 0;
2698            if (const_a2) {
2699                c3 = a2, a2 = -1;
2700            } else if (a0 == a2) {
2701                /* Watch out for dest = src + dest, since we've removed
2702                   the matching constraint on the add.  */
2703                tgen_arithr(s, ARITH_ADD + rexw, a0, a1);
2704                break;
2705            }
2706
2707            tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, a1, a2, 0, c3);
2708            break;
2709        }
2710        c = ARITH_ADD;
2711        goto gen_arith;
2712    OP_32_64(sub):
2713        c = ARITH_SUB;
2714        goto gen_arith;
2715    OP_32_64(and):
2716        c = ARITH_AND;
2717        goto gen_arith;
2718    OP_32_64(or):
2719        c = ARITH_OR;
2720        goto gen_arith;
2721    OP_32_64(xor):
2722        c = ARITH_XOR;
2723        goto gen_arith;
2724    gen_arith:
2725        if (const_a2) {
2726            tgen_arithi(s, c + rexw, a0, a2, 0);
2727        } else {
2728            tgen_arithr(s, c + rexw, a0, a2);
2729        }
2730        break;
2731
2732    OP_32_64(andc):
2733        if (const_a2) {
2734            tcg_out_mov(s, rexw ? TCG_TYPE_I64 : TCG_TYPE_I32, a0, a1);
2735            tgen_arithi(s, ARITH_AND + rexw, a0, ~a2, 0);
2736        } else {
2737            tcg_out_vex_modrm(s, OPC_ANDN + rexw, a0, a2, a1);
2738        }
2739        break;
2740
2741    OP_32_64(mul):
2742        if (const_a2) {
2743            int32_t val;
2744            val = a2;
2745            if (val == (int8_t)val) {
2746                tcg_out_modrm(s, OPC_IMUL_GvEvIb + rexw, a0, a0);
2747                tcg_out8(s, val);
2748            } else {
2749                tcg_out_modrm(s, OPC_IMUL_GvEvIz + rexw, a0, a0);
2750                tcg_out32(s, val);
2751            }
2752        } else {
2753            tcg_out_modrm(s, OPC_IMUL_GvEv + rexw, a0, a2);
2754        }
2755        break;
2756
2757    OP_32_64(div2):
2758        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_IDIV, args[4]);
2759        break;
2760    OP_32_64(divu2):
2761        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_DIV, args[4]);
2762        break;
2763
2764    OP_32_64(shl):
2765        /* For small constant 3-operand shift, use LEA.  */
2766        if (const_a2 && a0 != a1 && (a2 - 1) < 3) {
2767            if (a2 - 1 == 0) {
2768                /* shl $1,a1,a0 -> lea (a1,a1),a0 */
2769                tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, a1, a1, 0, 0);
2770            } else {
2771                /* shl $n,a1,a0 -> lea 0(,a1,n),a0 */
2772                tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, -1, a1, a2, 0);
2773            }
2774            break;
2775        }
2776        c = SHIFT_SHL;
2777        vexop = OPC_SHLX;
2778        goto gen_shift_maybe_vex;
2779    OP_32_64(shr):
2780        c = SHIFT_SHR;
2781        vexop = OPC_SHRX;
2782        goto gen_shift_maybe_vex;
2783    OP_32_64(sar):
2784        c = SHIFT_SAR;
2785        vexop = OPC_SARX;
2786        goto gen_shift_maybe_vex;
2787    OP_32_64(rotl):
2788        c = SHIFT_ROL;
2789        goto gen_shift;
2790    OP_32_64(rotr):
2791        c = SHIFT_ROR;
2792        goto gen_shift;
2793    gen_shift_maybe_vex:
2794        if (have_bmi2) {
2795            if (!const_a2) {
2796                tcg_out_vex_modrm(s, vexop + rexw, a0, a2, a1);
2797                break;
2798            }
2799            tcg_out_mov(s, rexw ? TCG_TYPE_I64 : TCG_TYPE_I32, a0, a1);
2800        }
2801        /* FALLTHRU */
2802    gen_shift:
2803        if (const_a2) {
2804            tcg_out_shifti(s, c + rexw, a0, a2);
2805        } else {
2806            tcg_out_modrm(s, OPC_SHIFT_cl + rexw, c, a0);
2807        }
2808        break;
2809
2810    OP_32_64(ctz):
2811        tcg_out_ctz(s, rexw, args[0], args[1], args[2], const_args[2]);
2812        break;
2813    OP_32_64(clz):
2814        tcg_out_clz(s, rexw, args[0], args[1], args[2], const_args[2]);
2815        break;
2816    OP_32_64(ctpop):
2817        tcg_out_modrm(s, OPC_POPCNT + rexw, a0, a1);
2818        break;
2819
2820    OP_32_64(brcond):
2821        tcg_out_brcond(s, rexw, a2, a0, a1, const_args[1],
2822                       arg_label(args[3]), 0);
2823        break;
2824    OP_32_64(setcond):
2825        tcg_out_setcond(s, rexw, args[3], a0, a1, a2, const_a2, false);
2826        break;
2827    OP_32_64(negsetcond):
2828        tcg_out_setcond(s, rexw, args[3], a0, a1, a2, const_a2, true);
2829        break;
2830    OP_32_64(movcond):
2831        tcg_out_movcond(s, rexw, args[5], a0, a1, a2, const_a2, args[3]);
2832        break;
2833
2834    OP_32_64(bswap16):
2835        if (a2 & TCG_BSWAP_OS) {
2836            /* Output must be sign-extended. */
2837            if (rexw) {
2838                tcg_out_bswap64(s, a0);
2839                tcg_out_shifti(s, SHIFT_SAR + rexw, a0, 48);
2840            } else {
2841                tcg_out_bswap32(s, a0);
2842                tcg_out_shifti(s, SHIFT_SAR, a0, 16);
2843            }
2844        } else if ((a2 & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
2845            /* Output must be zero-extended, but input isn't. */
2846            tcg_out_bswap32(s, a0);
2847            tcg_out_shifti(s, SHIFT_SHR, a0, 16);
2848        } else {
2849            tcg_out_rolw_8(s, a0);
2850        }
2851        break;
2852    OP_32_64(bswap32):
2853        tcg_out_bswap32(s, a0);
2854        if (rexw && (a2 & TCG_BSWAP_OS)) {
2855            tcg_out_ext32s(s, a0, a0);
2856        }
2857        break;
2858
2859    OP_32_64(neg):
2860        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NEG, a0);
2861        break;
2862    OP_32_64(not):
2863        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NOT, a0);
2864        break;
2865
2866    case INDEX_op_qemu_ld_i32:
2867        tcg_out_qemu_ld(s, a0, -1, a1, a2, TCG_TYPE_I32);
2868        break;
2869    case INDEX_op_qemu_ld_i64:
2870        if (TCG_TARGET_REG_BITS == 64) {
2871            tcg_out_qemu_ld(s, a0, -1, a1, a2, TCG_TYPE_I64);
2872        } else {
2873            tcg_out_qemu_ld(s, a0, a1, a2, args[3], TCG_TYPE_I64);
2874        }
2875        break;
2876    case INDEX_op_qemu_ld_i128:
2877        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
2878        tcg_out_qemu_ld(s, a0, a1, a2, args[3], TCG_TYPE_I128);
2879        break;
2880
2881    case INDEX_op_qemu_st_i32:
2882    case INDEX_op_qemu_st8_i32:
2883        tcg_out_qemu_st(s, a0, -1, a1, a2, TCG_TYPE_I32);
2884        break;
2885    case INDEX_op_qemu_st_i64:
2886        if (TCG_TARGET_REG_BITS == 64) {
2887            tcg_out_qemu_st(s, a0, -1, a1, a2, TCG_TYPE_I64);
2888        } else {
2889            tcg_out_qemu_st(s, a0, a1, a2, args[3], TCG_TYPE_I64);
2890        }
2891        break;
2892    case INDEX_op_qemu_st_i128:
2893        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
2894        tcg_out_qemu_st(s, a0, a1, a2, args[3], TCG_TYPE_I128);
2895        break;
2896
2897    OP_32_64(mulu2):
2898        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_MUL, args[3]);
2899        break;
2900    OP_32_64(muls2):
2901        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_IMUL, args[3]);
2902        break;
2903    OP_32_64(add2):
2904        if (const_args[4]) {
2905            tgen_arithi(s, ARITH_ADD + rexw, a0, args[4], 1);
2906        } else {
2907            tgen_arithr(s, ARITH_ADD + rexw, a0, args[4]);
2908        }
2909        if (const_args[5]) {
2910            tgen_arithi(s, ARITH_ADC + rexw, a1, args[5], 1);
2911        } else {
2912            tgen_arithr(s, ARITH_ADC + rexw, a1, args[5]);
2913        }
2914        break;
2915    OP_32_64(sub2):
2916        if (const_args[4]) {
2917            tgen_arithi(s, ARITH_SUB + rexw, a0, args[4], 1);
2918        } else {
2919            tgen_arithr(s, ARITH_SUB + rexw, a0, args[4]);
2920        }
2921        if (const_args[5]) {
2922            tgen_arithi(s, ARITH_SBB + rexw, a1, args[5], 1);
2923        } else {
2924            tgen_arithr(s, ARITH_SBB + rexw, a1, args[5]);
2925        }
2926        break;
2927
2928#if TCG_TARGET_REG_BITS == 32
2929    case INDEX_op_brcond2_i32:
2930        tcg_out_brcond2(s, args, const_args, 0);
2931        break;
2932    case INDEX_op_setcond2_i32:
2933        tcg_out_setcond2(s, args, const_args);
2934        break;
2935#else /* TCG_TARGET_REG_BITS == 64 */
2936    case INDEX_op_ld32s_i64:
2937        tcg_out_modrm_offset(s, OPC_MOVSLQ, a0, a1, a2);
2938        break;
2939    case INDEX_op_ld_i64:
2940        tcg_out_ld(s, TCG_TYPE_I64, a0, a1, a2);
2941        break;
2942    case INDEX_op_st_i64:
2943        if (const_args[0]) {
2944            tcg_out_modrm_offset(s, OPC_MOVL_EvIz | P_REXW, 0, a1, a2);
2945            tcg_out32(s, a0);
2946        } else {
2947            tcg_out_st(s, TCG_TYPE_I64, a0, a1, a2);
2948        }
2949        break;
2950
2951    case INDEX_op_bswap64_i64:
2952        tcg_out_bswap64(s, a0);
2953        break;
2954    case INDEX_op_extrh_i64_i32:
2955        tcg_out_shifti(s, SHIFT_SHR + P_REXW, a0, 32);
2956        break;
2957#endif
2958
2959    OP_32_64(deposit):
2960        if (args[3] == 0 && args[4] == 8) {
2961            /* load bits 0..7 */
2962            if (const_a2) {
2963                tcg_out_opc(s, OPC_MOVB_Ib | P_REXB_RM | LOWREGMASK(a0),
2964                            0, a0, 0);
2965                tcg_out8(s, a2);
2966            } else {
2967                tcg_out_modrm(s, OPC_MOVB_EvGv | P_REXB_R | P_REXB_RM, a2, a0);
2968            }
2969        } else if (TCG_TARGET_REG_BITS == 32 && args[3] == 8 && args[4] == 8) {
2970            /* load bits 8..15 */
2971            if (const_a2) {
2972                tcg_out8(s, OPC_MOVB_Ib + a0 + 4);
2973                tcg_out8(s, a2);
2974            } else {
2975                tcg_out_modrm(s, OPC_MOVB_EvGv, a2, a0 + 4);
2976            }
2977        } else if (args[3] == 0 && args[4] == 16) {
2978            /* load bits 0..15 */
2979            if (const_a2) {
2980                tcg_out_opc(s, OPC_MOVL_Iv | P_DATA16 | LOWREGMASK(a0),
2981                            0, a0, 0);
2982                tcg_out16(s, a2);
2983            } else {
2984                tcg_out_modrm(s, OPC_MOVL_EvGv | P_DATA16, a2, a0);
2985            }
2986        } else {
2987            g_assert_not_reached();
2988        }
2989        break;
2990
2991    case INDEX_op_extract_i64:
2992        if (a2 + args[3] == 32) {
2993            if (a2 == 0) {
2994                tcg_out_ext32u(s, a0, a1);
2995                break;
2996            }
2997            /* This is a 32-bit zero-extending right shift.  */
2998            tcg_out_mov(s, TCG_TYPE_I32, a0, a1);
2999            tcg_out_shifti(s, SHIFT_SHR, a0, a2);
3000            break;
3001        }
3002        /* FALLTHRU */
3003    case INDEX_op_extract_i32:
3004        if (a2 == 0 && args[3] == 8) {
3005            tcg_out_ext8u(s, a0, a1);
3006        } else if (a2 == 0 && args[3] == 16) {
3007            tcg_out_ext16u(s, a0, a1);
3008        } else if (a2 == 8 && args[3] == 8) {
3009            /*
3010             * On the off-chance that we can use the high-byte registers.
3011             * Otherwise we emit the same ext16 + shift pattern that we
3012             * would have gotten from the normal tcg-op.c expansion.
3013             */
3014            if (a1 < 4 && a0 < 8) {
3015                tcg_out_modrm(s, OPC_MOVZBL, a0, a1 + 4);
3016            } else {
3017                tcg_out_ext16u(s, a0, a1);
3018                tcg_out_shifti(s, SHIFT_SHR, a0, 8);
3019            }
3020        } else {
3021            g_assert_not_reached();
3022        }
3023        break;
3024
3025    case INDEX_op_sextract_i64:
3026        if (a2 == 0 && args[3] == 8) {
3027            tcg_out_ext8s(s, TCG_TYPE_I64, a0, a1);
3028        } else if (a2 == 0 && args[3] == 16) {
3029            tcg_out_ext16s(s, TCG_TYPE_I64, a0, a1);
3030        } else if (a2 == 0 && args[3] == 32) {
3031            tcg_out_ext32s(s, a0, a1);
3032        } else {
3033            g_assert_not_reached();
3034        }
3035        break;
3036
3037    case INDEX_op_sextract_i32:
3038        if (a2 == 0 && args[3] == 8) {
3039            tcg_out_ext8s(s, TCG_TYPE_I32, a0, a1);
3040        } else if (a2 == 0 && args[3] == 16) {
3041            tcg_out_ext16s(s, TCG_TYPE_I32, a0, a1);
3042        } else if (a2 == 8 && args[3] == 8) {
3043            if (a1 < 4 && a0 < 8) {
3044                tcg_out_modrm(s, OPC_MOVSBL, a0, a1 + 4);
3045            } else {
3046                tcg_out_ext16s(s, TCG_TYPE_I32, a0, a1);
3047                tcg_out_shifti(s, SHIFT_SAR, a0, 8);
3048            }
3049        } else {
3050            g_assert_not_reached();
3051        }
3052        break;
3053
3054    OP_32_64(extract2):
3055        /* Note that SHRD outputs to the r/m operand.  */
3056        tcg_out_modrm(s, OPC_SHRD_Ib + rexw, a2, a0);
3057        tcg_out8(s, args[3]);
3058        break;
3059
3060    case INDEX_op_mb:
3061        tcg_out_mb(s, a0);
3062        break;
3063    case INDEX_op_mov_i32:  /* Always emitted via tcg_out_mov.  */
3064    case INDEX_op_mov_i64:
3065    case INDEX_op_call:     /* Always emitted via tcg_out_call.  */
3066    case INDEX_op_exit_tb:  /* Always emitted via tcg_out_exit_tb.  */
3067    case INDEX_op_goto_tb:  /* Always emitted via tcg_out_goto_tb.  */
3068    case INDEX_op_ext8s_i32:  /* Always emitted via tcg_reg_alloc_op.  */
3069    case INDEX_op_ext8s_i64:
3070    case INDEX_op_ext8u_i32:
3071    case INDEX_op_ext8u_i64:
3072    case INDEX_op_ext16s_i32:
3073    case INDEX_op_ext16s_i64:
3074    case INDEX_op_ext16u_i32:
3075    case INDEX_op_ext16u_i64:
3076    case INDEX_op_ext32s_i64:
3077    case INDEX_op_ext32u_i64:
3078    case INDEX_op_ext_i32_i64:
3079    case INDEX_op_extu_i32_i64:
3080    case INDEX_op_extrl_i64_i32:
3081    default:
3082        g_assert_not_reached();
3083    }
3084
3085#undef OP_32_64
3086}
3087
3088static int const umin_insn[4] = {
3089    OPC_PMINUB, OPC_PMINUW, OPC_PMINUD, OPC_VPMINUQ
3090};
3091
3092static int const umax_insn[4] = {
3093    OPC_PMAXUB, OPC_PMAXUW, OPC_PMAXUD, OPC_VPMAXUQ
3094};
3095
3096static bool tcg_out_cmp_vec_noinv(TCGContext *s, TCGType type, unsigned vece,
3097                                  TCGReg v0, TCGReg v1, TCGReg v2, TCGCond cond)
3098{
3099    static int const cmpeq_insn[4] = {
3100        OPC_PCMPEQB, OPC_PCMPEQW, OPC_PCMPEQD, OPC_PCMPEQQ
3101    };
3102    static int const cmpgt_insn[4] = {
3103        OPC_PCMPGTB, OPC_PCMPGTW, OPC_PCMPGTD, OPC_PCMPGTQ
3104    };
3105
3106    enum {
3107        NEED_INV  = 1,
3108        NEED_SWAP = 2,
3109        NEED_UMIN = 4,
3110        NEED_UMAX = 8,
3111        INVALID   = 16,
3112    };
3113    static const uint8_t cond_fixup[16] = {
3114        [0 ... 15] = INVALID,
3115        [TCG_COND_EQ] = 0,
3116        [TCG_COND_GT] = 0,
3117        [TCG_COND_NE] = NEED_INV,
3118        [TCG_COND_LE] = NEED_INV,
3119        [TCG_COND_LT] = NEED_SWAP,
3120        [TCG_COND_GE] = NEED_SWAP | NEED_INV,
3121        [TCG_COND_LEU] = NEED_UMIN,
3122        [TCG_COND_GTU] = NEED_UMIN | NEED_INV,
3123        [TCG_COND_GEU] = NEED_UMAX,
3124        [TCG_COND_LTU] = NEED_UMAX | NEED_INV,
3125    };
3126    int fixup = cond_fixup[cond];
3127
3128    assert(!(fixup & INVALID));
3129
3130    if (fixup & NEED_INV) {
3131        cond = tcg_invert_cond(cond);
3132    }
3133
3134    if (fixup & NEED_SWAP) {
3135        TCGReg swap = v1;
3136        v1 = v2;
3137        v2 = swap;
3138        cond = tcg_swap_cond(cond);
3139    }
3140
3141    if (fixup & (NEED_UMIN | NEED_UMAX)) {
3142        int op = (fixup & NEED_UMIN ? umin_insn[vece] : umax_insn[vece]);
3143
3144        /* avx2 does not have 64-bit min/max; adjusted during expand. */
3145        assert(vece <= MO_32);
3146
3147        tcg_out_vex_modrm_type(s, op, TCG_TMP_VEC, v1, v2, type);
3148        v2 = TCG_TMP_VEC;
3149        cond = TCG_COND_EQ;
3150    }
3151
3152    switch (cond) {
3153    case TCG_COND_EQ:
3154        tcg_out_vex_modrm_type(s, cmpeq_insn[vece], v0, v1, v2, type);
3155        break;
3156    case TCG_COND_GT:
3157        tcg_out_vex_modrm_type(s, cmpgt_insn[vece], v0, v1, v2, type);
3158        break;
3159    default:
3160        g_assert_not_reached();
3161    }
3162    return fixup & NEED_INV;
3163}
3164
3165static void tcg_out_cmp_vec_k1(TCGContext *s, TCGType type, unsigned vece,
3166                               TCGReg v1, TCGReg v2, TCGCond cond)
3167{
3168    static const int cmpm_insn[2][4] = {
3169        { OPC_VPCMPB, OPC_VPCMPW, OPC_VPCMPD, OPC_VPCMPQ },
3170        { OPC_VPCMPUB, OPC_VPCMPUW, OPC_VPCMPUD, OPC_VPCMPUQ }
3171    };
3172    static const int testm_insn[4] = {
3173        OPC_VPTESTMB, OPC_VPTESTMW, OPC_VPTESTMD, OPC_VPTESTMQ
3174    };
3175    static const int testnm_insn[4] = {
3176        OPC_VPTESTNMB, OPC_VPTESTNMW, OPC_VPTESTNMD, OPC_VPTESTNMQ
3177    };
3178
3179    static const int cond_ext[16] = {
3180        [TCG_COND_EQ] = 0,
3181        [TCG_COND_NE] = 4,
3182        [TCG_COND_LT] = 1,
3183        [TCG_COND_LTU] = 1,
3184        [TCG_COND_LE] = 2,
3185        [TCG_COND_LEU] = 2,
3186        [TCG_COND_NEVER] = 3,
3187        [TCG_COND_GE] = 5,
3188        [TCG_COND_GEU] = 5,
3189        [TCG_COND_GT] = 6,
3190        [TCG_COND_GTU] = 6,
3191        [TCG_COND_ALWAYS] = 7,
3192    };
3193
3194    switch (cond) {
3195    case TCG_COND_TSTNE:
3196        tcg_out_vex_modrm_type(s, testm_insn[vece], /* k1 */ 1, v1, v2, type);
3197        break;
3198    case TCG_COND_TSTEQ:
3199        tcg_out_vex_modrm_type(s, testnm_insn[vece], /* k1 */ 1, v1, v2, type);
3200        break;
3201    default:
3202        tcg_out_vex_modrm_type(s, cmpm_insn[is_unsigned_cond(cond)][vece],
3203                               /* k1 */ 1, v1, v2, type);
3204        tcg_out8(s, cond_ext[cond]);
3205        break;
3206    }
3207}
3208
3209static void tcg_out_k1_to_vec(TCGContext *s, TCGType type,
3210                              unsigned vece, TCGReg dest)
3211{
3212    static const int movm_insn[] = {
3213        OPC_VPMOVM2B, OPC_VPMOVM2W, OPC_VPMOVM2D, OPC_VPMOVM2Q
3214    };
3215    tcg_out_vex_modrm_type(s, movm_insn[vece], dest, 0, /* k1 */ 1, type);
3216}
3217
3218static void tcg_out_cmp_vec(TCGContext *s, TCGType type, unsigned vece,
3219                            TCGReg v0, TCGReg v1, TCGReg v2, TCGCond cond)
3220{
3221    /*
3222     * With avx512, we have a complete set of comparisons into mask.
3223     * Unless there's a single insn expansion for the comparision,
3224     * expand via a mask in k1.
3225     */
3226    if ((vece <= MO_16 ? have_avx512bw : have_avx512dq)
3227        && cond != TCG_COND_EQ
3228        && cond != TCG_COND_LT
3229        && cond != TCG_COND_GT) {
3230        tcg_out_cmp_vec_k1(s, type, vece, v1, v2, cond);
3231        tcg_out_k1_to_vec(s, type, vece, v0);
3232        return;
3233    }
3234
3235    if (tcg_out_cmp_vec_noinv(s, type, vece, v0, v1, v2, cond)) {
3236        tcg_out_dupi_vec(s, type, vece, TCG_TMP_VEC, -1);
3237        tcg_out_vex_modrm_type(s, OPC_PXOR, v0, v0, TCG_TMP_VEC, type);
3238    }
3239}
3240
3241static void tcg_out_cmpsel_vec_k1(TCGContext *s, TCGType type, unsigned vece,
3242                                  TCGReg v0, TCGReg c1, TCGReg c2,
3243                                  TCGReg v3, TCGReg v4, TCGCond cond)
3244{
3245    static const int vpblendm_insn[] = {
3246        OPC_VPBLENDMB, OPC_VPBLENDMW, OPC_VPBLENDMD, OPC_VPBLENDMQ
3247    };
3248    bool z = false;
3249
3250    /* Swap to place constant in V4 to take advantage of zero-masking. */
3251    if (!v3) {
3252        z = true;
3253        v3 = v4;
3254        cond = tcg_invert_cond(cond);
3255    }
3256
3257    tcg_out_cmp_vec_k1(s, type, vece, c1, c2, cond);
3258    tcg_out_evex_modrm_type(s, vpblendm_insn[vece], v0, v4, v3,
3259                            /* k1 */1, z, type);
3260}
3261
3262static void tcg_out_cmpsel_vec(TCGContext *s, TCGType type, unsigned vece,
3263                               TCGReg v0, TCGReg c1, TCGReg c2,
3264                               TCGReg v3, TCGReg v4, TCGCond cond)
3265{
3266    bool inv;
3267
3268    if (vece <= MO_16 ? have_avx512bw : have_avx512vl) {
3269        tcg_out_cmpsel_vec_k1(s, type, vece, v0, c1, c2, v3, v4, cond);
3270        return;
3271    }
3272
3273    inv = tcg_out_cmp_vec_noinv(s, type, vece, TCG_TMP_VEC, c1, c2, cond);
3274
3275    /*
3276     * Since XMM0 is 16, the only way we get 0 into V3
3277     * is via the constant zero constraint.
3278     */
3279    if (!v3) {
3280        if (inv) {
3281            tcg_out_vex_modrm_type(s, OPC_PAND, v0, TCG_TMP_VEC, v4, type);
3282        } else {
3283            tcg_out_vex_modrm_type(s, OPC_PANDN, v0, TCG_TMP_VEC, v4, type);
3284        }
3285    } else {
3286        if (inv) {
3287            TCGReg swap = v3;
3288            v3 = v4;
3289            v4 = swap;
3290        }
3291        tcg_out_vex_modrm_type(s, OPC_VPBLENDVB, v0, v4, v3, type);
3292        tcg_out8(s, (TCG_TMP_VEC - TCG_REG_XMM0) << 4);
3293    }
3294}
3295
3296static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
3297                           unsigned vecl, unsigned vece,
3298                           const TCGArg args[TCG_MAX_OP_ARGS],
3299                           const int const_args[TCG_MAX_OP_ARGS])
3300{
3301    static int const add_insn[4] = {
3302        OPC_PADDB, OPC_PADDW, OPC_PADDD, OPC_PADDQ
3303    };
3304    static int const ssadd_insn[4] = {
3305        OPC_PADDSB, OPC_PADDSW, OPC_UD2, OPC_UD2
3306    };
3307    static int const usadd_insn[4] = {
3308        OPC_PADDUB, OPC_PADDUW, OPC_UD2, OPC_UD2
3309    };
3310    static int const sub_insn[4] = {
3311        OPC_PSUBB, OPC_PSUBW, OPC_PSUBD, OPC_PSUBQ
3312    };
3313    static int const sssub_insn[4] = {
3314        OPC_PSUBSB, OPC_PSUBSW, OPC_UD2, OPC_UD2
3315    };
3316    static int const ussub_insn[4] = {
3317        OPC_PSUBUB, OPC_PSUBUW, OPC_UD2, OPC_UD2
3318    };
3319    static int const mul_insn[4] = {
3320        OPC_UD2, OPC_PMULLW, OPC_PMULLD, OPC_VPMULLQ
3321    };
3322    static int const shift_imm_insn[4] = {
3323        OPC_UD2, OPC_PSHIFTW_Ib, OPC_PSHIFTD_Ib, OPC_PSHIFTQ_Ib
3324    };
3325    static int const punpckl_insn[4] = {
3326        OPC_PUNPCKLBW, OPC_PUNPCKLWD, OPC_PUNPCKLDQ, OPC_PUNPCKLQDQ
3327    };
3328    static int const punpckh_insn[4] = {
3329        OPC_PUNPCKHBW, OPC_PUNPCKHWD, OPC_PUNPCKHDQ, OPC_PUNPCKHQDQ
3330    };
3331    static int const packss_insn[4] = {
3332        OPC_PACKSSWB, OPC_PACKSSDW, OPC_UD2, OPC_UD2
3333    };
3334    static int const packus_insn[4] = {
3335        OPC_PACKUSWB, OPC_PACKUSDW, OPC_UD2, OPC_UD2
3336    };
3337    static int const smin_insn[4] = {
3338        OPC_PMINSB, OPC_PMINSW, OPC_PMINSD, OPC_VPMINSQ
3339    };
3340    static int const smax_insn[4] = {
3341        OPC_PMAXSB, OPC_PMAXSW, OPC_PMAXSD, OPC_VPMAXSQ
3342    };
3343    static int const rotlv_insn[4] = {
3344        OPC_UD2, OPC_UD2, OPC_VPROLVD, OPC_VPROLVQ
3345    };
3346    static int const rotrv_insn[4] = {
3347        OPC_UD2, OPC_UD2, OPC_VPRORVD, OPC_VPRORVQ
3348    };
3349    static int const shlv_insn[4] = {
3350        OPC_UD2, OPC_VPSLLVW, OPC_VPSLLVD, OPC_VPSLLVQ
3351    };
3352    static int const shrv_insn[4] = {
3353        OPC_UD2, OPC_VPSRLVW, OPC_VPSRLVD, OPC_VPSRLVQ
3354    };
3355    static int const sarv_insn[4] = {
3356        OPC_UD2, OPC_VPSRAVW, OPC_VPSRAVD, OPC_VPSRAVQ
3357    };
3358    static int const shls_insn[4] = {
3359        OPC_UD2, OPC_PSLLW, OPC_PSLLD, OPC_PSLLQ
3360    };
3361    static int const shrs_insn[4] = {
3362        OPC_UD2, OPC_PSRLW, OPC_PSRLD, OPC_PSRLQ
3363    };
3364    static int const sars_insn[4] = {
3365        OPC_UD2, OPC_PSRAW, OPC_PSRAD, OPC_VPSRAQ
3366    };
3367    static int const vpshldi_insn[4] = {
3368        OPC_UD2, OPC_VPSHLDW, OPC_VPSHLDD, OPC_VPSHLDQ
3369    };
3370    static int const vpshldv_insn[4] = {
3371        OPC_UD2, OPC_VPSHLDVW, OPC_VPSHLDVD, OPC_VPSHLDVQ
3372    };
3373    static int const vpshrdv_insn[4] = {
3374        OPC_UD2, OPC_VPSHRDVW, OPC_VPSHRDVD, OPC_VPSHRDVQ
3375    };
3376    static int const abs_insn[4] = {
3377        OPC_PABSB, OPC_PABSW, OPC_PABSD, OPC_VPABSQ
3378    };
3379
3380    TCGType type = vecl + TCG_TYPE_V64;
3381    int insn, sub;
3382    TCGArg a0, a1, a2, a3;
3383
3384    a0 = args[0];
3385    a1 = args[1];
3386    a2 = args[2];
3387
3388    switch (opc) {
3389    case INDEX_op_add_vec:
3390        insn = add_insn[vece];
3391        goto gen_simd;
3392    case INDEX_op_ssadd_vec:
3393        insn = ssadd_insn[vece];
3394        goto gen_simd;
3395    case INDEX_op_usadd_vec:
3396        insn = usadd_insn[vece];
3397        goto gen_simd;
3398    case INDEX_op_sub_vec:
3399        insn = sub_insn[vece];
3400        goto gen_simd;
3401    case INDEX_op_sssub_vec:
3402        insn = sssub_insn[vece];
3403        goto gen_simd;
3404    case INDEX_op_ussub_vec:
3405        insn = ussub_insn[vece];
3406        goto gen_simd;
3407    case INDEX_op_mul_vec:
3408        insn = mul_insn[vece];
3409        goto gen_simd;
3410    case INDEX_op_and_vec:
3411        insn = OPC_PAND;
3412        goto gen_simd;
3413    case INDEX_op_or_vec:
3414        insn = OPC_POR;
3415        goto gen_simd;
3416    case INDEX_op_xor_vec:
3417        insn = OPC_PXOR;
3418        goto gen_simd;
3419    case INDEX_op_smin_vec:
3420        insn = smin_insn[vece];
3421        goto gen_simd;
3422    case INDEX_op_umin_vec:
3423        insn = umin_insn[vece];
3424        goto gen_simd;
3425    case INDEX_op_smax_vec:
3426        insn = smax_insn[vece];
3427        goto gen_simd;
3428    case INDEX_op_umax_vec:
3429        insn = umax_insn[vece];
3430        goto gen_simd;
3431    case INDEX_op_shlv_vec:
3432        insn = shlv_insn[vece];
3433        goto gen_simd;
3434    case INDEX_op_shrv_vec:
3435        insn = shrv_insn[vece];
3436        goto gen_simd;
3437    case INDEX_op_sarv_vec:
3438        insn = sarv_insn[vece];
3439        goto gen_simd;
3440    case INDEX_op_rotlv_vec:
3441        insn = rotlv_insn[vece];
3442        goto gen_simd;
3443    case INDEX_op_rotrv_vec:
3444        insn = rotrv_insn[vece];
3445        goto gen_simd;
3446    case INDEX_op_shls_vec:
3447        insn = shls_insn[vece];
3448        goto gen_simd;
3449    case INDEX_op_shrs_vec:
3450        insn = shrs_insn[vece];
3451        goto gen_simd;
3452    case INDEX_op_sars_vec:
3453        insn = sars_insn[vece];
3454        goto gen_simd;
3455    case INDEX_op_x86_punpckl_vec:
3456        insn = punpckl_insn[vece];
3457        goto gen_simd;
3458    case INDEX_op_x86_punpckh_vec:
3459        insn = punpckh_insn[vece];
3460        goto gen_simd;
3461    case INDEX_op_x86_packss_vec:
3462        insn = packss_insn[vece];
3463        goto gen_simd;
3464    case INDEX_op_x86_packus_vec:
3465        insn = packus_insn[vece];
3466        goto gen_simd;
3467    case INDEX_op_x86_vpshldv_vec:
3468        insn = vpshldv_insn[vece];
3469        a1 = a2;
3470        a2 = args[3];
3471        goto gen_simd;
3472    case INDEX_op_x86_vpshrdv_vec:
3473        insn = vpshrdv_insn[vece];
3474        a1 = a2;
3475        a2 = args[3];
3476        goto gen_simd;
3477#if TCG_TARGET_REG_BITS == 32
3478    case INDEX_op_dup2_vec:
3479        /* First merge the two 32-bit inputs to a single 64-bit element. */
3480        tcg_out_vex_modrm(s, OPC_PUNPCKLDQ, a0, a1, a2);
3481        /* Then replicate the 64-bit elements across the rest of the vector. */
3482        if (type != TCG_TYPE_V64) {
3483            tcg_out_dup_vec(s, type, MO_64, a0, a0);
3484        }
3485        break;
3486#endif
3487    case INDEX_op_abs_vec:
3488        insn = abs_insn[vece];
3489        a2 = a1;
3490        a1 = 0;
3491        goto gen_simd;
3492    gen_simd:
3493        tcg_debug_assert(insn != OPC_UD2);
3494        tcg_out_vex_modrm_type(s, insn, a0, a1, a2, type);
3495        break;
3496
3497    case INDEX_op_cmp_vec:
3498        tcg_out_cmp_vec(s, type, vece, a0, a1, a2, args[3]);
3499        break;
3500
3501    case INDEX_op_cmpsel_vec:
3502        tcg_out_cmpsel_vec(s, type, vece, a0, a1, a2,
3503                           args[3], args[4], args[5]);
3504        break;
3505
3506    case INDEX_op_andc_vec:
3507        insn = OPC_PANDN;
3508        tcg_out_vex_modrm_type(s, insn, a0, a2, a1, type);
3509        break;
3510
3511    case INDEX_op_shli_vec:
3512        insn = shift_imm_insn[vece];
3513        sub = 6;
3514        goto gen_shift;
3515    case INDEX_op_shri_vec:
3516        insn = shift_imm_insn[vece];
3517        sub = 2;
3518        goto gen_shift;
3519    case INDEX_op_sari_vec:
3520        if (vece == MO_64) {
3521            insn = OPC_PSHIFTD_Ib | P_VEXW | P_EVEX;
3522        } else {
3523            insn = shift_imm_insn[vece];
3524        }
3525        sub = 4;
3526        goto gen_shift;
3527    case INDEX_op_rotli_vec:
3528        insn = OPC_PSHIFTD_Ib | P_EVEX;  /* VPROL[DQ] */
3529        if (vece == MO_64) {
3530            insn |= P_VEXW;
3531        }
3532        sub = 1;
3533        goto gen_shift;
3534    gen_shift:
3535        tcg_debug_assert(vece != MO_8);
3536        tcg_out_vex_modrm_type(s, insn, sub, a0, a1, type);
3537        tcg_out8(s, a2);
3538        break;
3539
3540    case INDEX_op_ld_vec:
3541        tcg_out_ld(s, type, a0, a1, a2);
3542        break;
3543    case INDEX_op_st_vec:
3544        tcg_out_st(s, type, a0, a1, a2);
3545        break;
3546    case INDEX_op_dupm_vec:
3547        tcg_out_dupm_vec(s, type, vece, a0, a1, a2);
3548        break;
3549
3550    case INDEX_op_x86_shufps_vec:
3551        insn = OPC_SHUFPS;
3552        sub = args[3];
3553        goto gen_simd_imm8;
3554    case INDEX_op_x86_blend_vec:
3555        if (vece == MO_16) {
3556            insn = OPC_PBLENDW;
3557        } else if (vece == MO_32) {
3558            insn = (have_avx2 ? OPC_VPBLENDD : OPC_BLENDPS);
3559        } else {
3560            g_assert_not_reached();
3561        }
3562        sub = args[3];
3563        goto gen_simd_imm8;
3564    case INDEX_op_x86_vperm2i128_vec:
3565        insn = OPC_VPERM2I128;
3566        sub = args[3];
3567        goto gen_simd_imm8;
3568    case INDEX_op_x86_vpshldi_vec:
3569        insn = vpshldi_insn[vece];
3570        sub = args[3];
3571        goto gen_simd_imm8;
3572
3573    case INDEX_op_not_vec:
3574        insn = OPC_VPTERNLOGQ;
3575        a2 = a1;
3576        sub = 0x33; /* !B */
3577        goto gen_simd_imm8;
3578    case INDEX_op_nor_vec:
3579        insn = OPC_VPTERNLOGQ;
3580        sub = 0x11; /* norCB */
3581        goto gen_simd_imm8;
3582    case INDEX_op_nand_vec:
3583        insn = OPC_VPTERNLOGQ;
3584        sub = 0x77; /* nandCB */
3585        goto gen_simd_imm8;
3586    case INDEX_op_eqv_vec:
3587        insn = OPC_VPTERNLOGQ;
3588        sub = 0x99; /* xnorCB */
3589        goto gen_simd_imm8;
3590    case INDEX_op_orc_vec:
3591        insn = OPC_VPTERNLOGQ;
3592        sub = 0xdd; /* orB!C */
3593        goto gen_simd_imm8;
3594
3595    case INDEX_op_bitsel_vec:
3596        insn = OPC_VPTERNLOGQ;
3597        a3 = args[3];
3598        if (a0 == a1) {
3599            a1 = a2;
3600            a2 = a3;
3601            sub = 0xca; /* A?B:C */
3602        } else if (a0 == a2) {
3603            a2 = a3;
3604            sub = 0xe2; /* B?A:C */
3605        } else {
3606            tcg_out_mov(s, type, a0, a3);
3607            sub = 0xb8; /* B?C:A */
3608        }
3609        goto gen_simd_imm8;
3610
3611    gen_simd_imm8:
3612        tcg_debug_assert(insn != OPC_UD2);
3613        tcg_out_vex_modrm_type(s, insn, a0, a1, a2, type);
3614        tcg_out8(s, sub);
3615        break;
3616
3617    case INDEX_op_x86_psrldq_vec:
3618        tcg_out_vex_modrm(s, OPC_GRP14, 3, a0, a1);
3619        tcg_out8(s, a2);
3620        break;
3621
3622    case INDEX_op_mov_vec:  /* Always emitted via tcg_out_mov.  */
3623    case INDEX_op_dup_vec:  /* Always emitted via tcg_out_dup_vec.  */
3624    default:
3625        g_assert_not_reached();
3626    }
3627}
3628
3629static TCGConstraintSetIndex
3630tcg_target_op_def(TCGOpcode op, TCGType type, unsigned flags)
3631{
3632    switch (op) {
3633    case INDEX_op_goto_ptr:
3634        return C_O0_I1(r);
3635
3636    case INDEX_op_ld8u_i32:
3637    case INDEX_op_ld8u_i64:
3638    case INDEX_op_ld8s_i32:
3639    case INDEX_op_ld8s_i64:
3640    case INDEX_op_ld16u_i32:
3641    case INDEX_op_ld16u_i64:
3642    case INDEX_op_ld16s_i32:
3643    case INDEX_op_ld16s_i64:
3644    case INDEX_op_ld_i32:
3645    case INDEX_op_ld32u_i64:
3646    case INDEX_op_ld32s_i64:
3647    case INDEX_op_ld_i64:
3648        return C_O1_I1(r, r);
3649
3650    case INDEX_op_st8_i32:
3651    case INDEX_op_st8_i64:
3652        return C_O0_I2(qi, r);
3653
3654    case INDEX_op_st16_i32:
3655    case INDEX_op_st16_i64:
3656    case INDEX_op_st_i32:
3657    case INDEX_op_st32_i64:
3658        return C_O0_I2(ri, r);
3659
3660    case INDEX_op_st_i64:
3661        return C_O0_I2(re, r);
3662
3663    case INDEX_op_add_i32:
3664    case INDEX_op_add_i64:
3665        return C_O1_I2(r, r, re);
3666
3667    case INDEX_op_sub_i32:
3668    case INDEX_op_sub_i64:
3669    case INDEX_op_mul_i32:
3670    case INDEX_op_mul_i64:
3671    case INDEX_op_or_i32:
3672    case INDEX_op_or_i64:
3673    case INDEX_op_xor_i32:
3674    case INDEX_op_xor_i64:
3675        return C_O1_I2(r, 0, re);
3676
3677    case INDEX_op_and_i32:
3678    case INDEX_op_and_i64:
3679        return C_O1_I2(r, 0, reZ);
3680
3681    case INDEX_op_andc_i32:
3682    case INDEX_op_andc_i64:
3683        return C_O1_I2(r, r, rI);
3684
3685    case INDEX_op_shl_i32:
3686    case INDEX_op_shl_i64:
3687    case INDEX_op_shr_i32:
3688    case INDEX_op_shr_i64:
3689    case INDEX_op_sar_i32:
3690    case INDEX_op_sar_i64:
3691        return have_bmi2 ? C_O1_I2(r, r, ri) : C_O1_I2(r, 0, ci);
3692
3693    case INDEX_op_rotl_i32:
3694    case INDEX_op_rotl_i64:
3695    case INDEX_op_rotr_i32:
3696    case INDEX_op_rotr_i64:
3697        return C_O1_I2(r, 0, ci);
3698
3699    case INDEX_op_brcond_i32:
3700    case INDEX_op_brcond_i64:
3701        return C_O0_I2(r, reT);
3702
3703    case INDEX_op_bswap16_i32:
3704    case INDEX_op_bswap16_i64:
3705    case INDEX_op_bswap32_i32:
3706    case INDEX_op_bswap32_i64:
3707    case INDEX_op_bswap64_i64:
3708    case INDEX_op_neg_i32:
3709    case INDEX_op_neg_i64:
3710    case INDEX_op_not_i32:
3711    case INDEX_op_not_i64:
3712    case INDEX_op_extrh_i64_i32:
3713        return C_O1_I1(r, 0);
3714
3715    case INDEX_op_ext8s_i32:
3716    case INDEX_op_ext8s_i64:
3717    case INDEX_op_ext8u_i32:
3718    case INDEX_op_ext8u_i64:
3719        return C_O1_I1(r, q);
3720
3721    case INDEX_op_ext16s_i32:
3722    case INDEX_op_ext16s_i64:
3723    case INDEX_op_ext16u_i32:
3724    case INDEX_op_ext16u_i64:
3725    case INDEX_op_ext32s_i64:
3726    case INDEX_op_ext32u_i64:
3727    case INDEX_op_ext_i32_i64:
3728    case INDEX_op_extu_i32_i64:
3729    case INDEX_op_extrl_i64_i32:
3730    case INDEX_op_extract_i32:
3731    case INDEX_op_extract_i64:
3732    case INDEX_op_sextract_i32:
3733    case INDEX_op_sextract_i64:
3734    case INDEX_op_ctpop_i32:
3735    case INDEX_op_ctpop_i64:
3736        return C_O1_I1(r, r);
3737
3738    case INDEX_op_extract2_i32:
3739    case INDEX_op_extract2_i64:
3740        return C_O1_I2(r, 0, r);
3741
3742    case INDEX_op_deposit_i32:
3743    case INDEX_op_deposit_i64:
3744        return C_O1_I2(q, 0, qi);
3745
3746    case INDEX_op_setcond_i32:
3747    case INDEX_op_setcond_i64:
3748    case INDEX_op_negsetcond_i32:
3749    case INDEX_op_negsetcond_i64:
3750        return C_O1_I2(q, r, reT);
3751
3752    case INDEX_op_movcond_i32:
3753    case INDEX_op_movcond_i64:
3754        return C_O1_I4(r, r, reT, r, 0);
3755
3756    case INDEX_op_div2_i32:
3757    case INDEX_op_div2_i64:
3758    case INDEX_op_divu2_i32:
3759    case INDEX_op_divu2_i64:
3760        return C_O2_I3(a, d, 0, 1, r);
3761
3762    case INDEX_op_mulu2_i32:
3763    case INDEX_op_mulu2_i64:
3764    case INDEX_op_muls2_i32:
3765    case INDEX_op_muls2_i64:
3766        return C_O2_I2(a, d, a, r);
3767
3768    case INDEX_op_add2_i32:
3769    case INDEX_op_add2_i64:
3770    case INDEX_op_sub2_i32:
3771    case INDEX_op_sub2_i64:
3772        return C_N1_O1_I4(r, r, 0, 1, re, re);
3773
3774    case INDEX_op_ctz_i32:
3775    case INDEX_op_ctz_i64:
3776        return have_bmi1 ? C_N1_I2(r, r, rW) : C_N1_I2(r, r, r);
3777
3778    case INDEX_op_clz_i32:
3779    case INDEX_op_clz_i64:
3780        return have_lzcnt ? C_N1_I2(r, r, rW) : C_N1_I2(r, r, r);
3781
3782    case INDEX_op_qemu_ld_i32:
3783        return C_O1_I1(r, L);
3784
3785    case INDEX_op_qemu_st_i32:
3786        return C_O0_I2(L, L);
3787    case INDEX_op_qemu_st8_i32:
3788        return C_O0_I2(s, L);
3789
3790    case INDEX_op_qemu_ld_i64:
3791        return TCG_TARGET_REG_BITS == 64 ? C_O1_I1(r, L) : C_O2_I1(r, r, L);
3792
3793    case INDEX_op_qemu_st_i64:
3794        return TCG_TARGET_REG_BITS == 64 ? C_O0_I2(L, L) : C_O0_I3(L, L, L);
3795
3796    case INDEX_op_qemu_ld_i128:
3797        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
3798        return C_O2_I1(r, r, L);
3799    case INDEX_op_qemu_st_i128:
3800        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
3801        return C_O0_I3(L, L, L);
3802
3803    case INDEX_op_brcond2_i32:
3804        return C_O0_I4(r, r, ri, ri);
3805
3806    case INDEX_op_setcond2_i32:
3807        return C_O1_I4(r, r, r, ri, ri);
3808
3809    case INDEX_op_ld_vec:
3810    case INDEX_op_dupm_vec:
3811        return C_O1_I1(x, r);
3812
3813    case INDEX_op_st_vec:
3814        return C_O0_I2(x, r);
3815
3816    case INDEX_op_add_vec:
3817    case INDEX_op_sub_vec:
3818    case INDEX_op_mul_vec:
3819    case INDEX_op_and_vec:
3820    case INDEX_op_or_vec:
3821    case INDEX_op_xor_vec:
3822    case INDEX_op_andc_vec:
3823    case INDEX_op_orc_vec:
3824    case INDEX_op_nand_vec:
3825    case INDEX_op_nor_vec:
3826    case INDEX_op_eqv_vec:
3827    case INDEX_op_ssadd_vec:
3828    case INDEX_op_usadd_vec:
3829    case INDEX_op_sssub_vec:
3830    case INDEX_op_ussub_vec:
3831    case INDEX_op_smin_vec:
3832    case INDEX_op_umin_vec:
3833    case INDEX_op_smax_vec:
3834    case INDEX_op_umax_vec:
3835    case INDEX_op_shlv_vec:
3836    case INDEX_op_shrv_vec:
3837    case INDEX_op_sarv_vec:
3838    case INDEX_op_rotlv_vec:
3839    case INDEX_op_rotrv_vec:
3840    case INDEX_op_shls_vec:
3841    case INDEX_op_shrs_vec:
3842    case INDEX_op_sars_vec:
3843    case INDEX_op_cmp_vec:
3844    case INDEX_op_x86_shufps_vec:
3845    case INDEX_op_x86_blend_vec:
3846    case INDEX_op_x86_packss_vec:
3847    case INDEX_op_x86_packus_vec:
3848    case INDEX_op_x86_vperm2i128_vec:
3849    case INDEX_op_x86_punpckl_vec:
3850    case INDEX_op_x86_punpckh_vec:
3851    case INDEX_op_x86_vpshldi_vec:
3852#if TCG_TARGET_REG_BITS == 32
3853    case INDEX_op_dup2_vec:
3854#endif
3855        return C_O1_I2(x, x, x);
3856
3857    case INDEX_op_abs_vec:
3858    case INDEX_op_dup_vec:
3859    case INDEX_op_not_vec:
3860    case INDEX_op_shli_vec:
3861    case INDEX_op_shri_vec:
3862    case INDEX_op_sari_vec:
3863    case INDEX_op_rotli_vec:
3864    case INDEX_op_x86_psrldq_vec:
3865        return C_O1_I1(x, x);
3866
3867    case INDEX_op_x86_vpshldv_vec:
3868    case INDEX_op_x86_vpshrdv_vec:
3869        return C_O1_I3(x, 0, x, x);
3870
3871    case INDEX_op_bitsel_vec:
3872        return C_O1_I3(x, x, x, x);
3873    case INDEX_op_cmpsel_vec:
3874        return C_O1_I4(x, x, x, xO, x);
3875
3876    default:
3877        return C_NotImplemented;
3878    }
3879}
3880
3881int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
3882{
3883    switch (opc) {
3884    case INDEX_op_add_vec:
3885    case INDEX_op_sub_vec:
3886    case INDEX_op_and_vec:
3887    case INDEX_op_or_vec:
3888    case INDEX_op_xor_vec:
3889    case INDEX_op_andc_vec:
3890    case INDEX_op_orc_vec:
3891    case INDEX_op_nand_vec:
3892    case INDEX_op_nor_vec:
3893    case INDEX_op_eqv_vec:
3894    case INDEX_op_not_vec:
3895    case INDEX_op_bitsel_vec:
3896        return 1;
3897    case INDEX_op_cmp_vec:
3898    case INDEX_op_cmpsel_vec:
3899        return -1;
3900
3901    case INDEX_op_rotli_vec:
3902        return have_avx512vl && vece >= MO_32 ? 1 : -1;
3903
3904    case INDEX_op_shli_vec:
3905    case INDEX_op_shri_vec:
3906        /* We must expand the operation for MO_8.  */
3907        return vece == MO_8 ? -1 : 1;
3908
3909    case INDEX_op_sari_vec:
3910        switch (vece) {
3911        case MO_8:
3912            return -1;
3913        case MO_16:
3914        case MO_32:
3915            return 1;
3916        case MO_64:
3917            if (have_avx512vl) {
3918                return 1;
3919            }
3920            /*
3921             * We can emulate this for MO_64, but it does not pay off
3922             * unless we're producing at least 4 values.
3923             */
3924            return type >= TCG_TYPE_V256 ? -1 : 0;
3925        }
3926        return 0;
3927
3928    case INDEX_op_shls_vec:
3929    case INDEX_op_shrs_vec:
3930        return vece >= MO_16;
3931    case INDEX_op_sars_vec:
3932        switch (vece) {
3933        case MO_16:
3934        case MO_32:
3935            return 1;
3936        case MO_64:
3937            return have_avx512vl;
3938        }
3939        return 0;
3940    case INDEX_op_rotls_vec:
3941        return vece >= MO_16 ? -1 : 0;
3942
3943    case INDEX_op_shlv_vec:
3944    case INDEX_op_shrv_vec:
3945        switch (vece) {
3946        case MO_16:
3947            return have_avx512bw;
3948        case MO_32:
3949        case MO_64:
3950            return have_avx2;
3951        }
3952        return 0;
3953    case INDEX_op_sarv_vec:
3954        switch (vece) {
3955        case MO_16:
3956            return have_avx512bw;
3957        case MO_32:
3958            return have_avx2;
3959        case MO_64:
3960            return have_avx512vl;
3961        }
3962        return 0;
3963    case INDEX_op_rotlv_vec:
3964    case INDEX_op_rotrv_vec:
3965        switch (vece) {
3966        case MO_16:
3967            return have_avx512vbmi2 ? -1 : 0;
3968        case MO_32:
3969        case MO_64:
3970            return have_avx512vl ? 1 : have_avx2 ? -1 : 0;
3971        }
3972        return 0;
3973
3974    case INDEX_op_mul_vec:
3975        switch (vece) {
3976        case MO_8:
3977            return -1;
3978        case MO_64:
3979            return have_avx512dq;
3980        }
3981        return 1;
3982
3983    case INDEX_op_ssadd_vec:
3984    case INDEX_op_usadd_vec:
3985    case INDEX_op_sssub_vec:
3986    case INDEX_op_ussub_vec:
3987        return vece <= MO_16;
3988    case INDEX_op_smin_vec:
3989    case INDEX_op_smax_vec:
3990    case INDEX_op_umin_vec:
3991    case INDEX_op_umax_vec:
3992    case INDEX_op_abs_vec:
3993        return vece <= MO_32 || have_avx512vl;
3994
3995    default:
3996        return 0;
3997    }
3998}
3999
4000static void expand_vec_shi(TCGType type, unsigned vece, bool right,
4001                           TCGv_vec v0, TCGv_vec v1, TCGArg imm)
4002{
4003    uint8_t mask;
4004
4005    tcg_debug_assert(vece == MO_8);
4006    if (right) {
4007        mask = 0xff >> imm;
4008        tcg_gen_shri_vec(MO_16, v0, v1, imm);
4009    } else {
4010        mask = 0xff << imm;
4011        tcg_gen_shli_vec(MO_16, v0, v1, imm);
4012    }
4013    tcg_gen_and_vec(MO_8, v0, v0, tcg_constant_vec(type, MO_8, mask));
4014}
4015
4016static void expand_vec_sari(TCGType type, unsigned vece,
4017                            TCGv_vec v0, TCGv_vec v1, TCGArg imm)
4018{
4019    TCGv_vec t1, t2;
4020
4021    switch (vece) {
4022    case MO_8:
4023        /* Unpack to 16-bit, shift, and repack.  */
4024        t1 = tcg_temp_new_vec(type);
4025        t2 = tcg_temp_new_vec(type);
4026        vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
4027                  tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
4028        vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
4029                  tcgv_vec_arg(t2), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
4030        tcg_gen_sari_vec(MO_16, t1, t1, imm + 8);
4031        tcg_gen_sari_vec(MO_16, t2, t2, imm + 8);
4032        vec_gen_3(INDEX_op_x86_packss_vec, type, MO_8,
4033                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t2));
4034        tcg_temp_free_vec(t1);
4035        tcg_temp_free_vec(t2);
4036        break;
4037
4038    case MO_64:
4039        t1 = tcg_temp_new_vec(type);
4040        if (imm <= 32) {
4041            /*
4042             * We can emulate a small sign extend by performing an arithmetic
4043             * 32-bit shift and overwriting the high half of a 64-bit logical
4044             * shift.  Note that the ISA says shift of 32 is valid, but TCG
4045             * does not, so we have to bound the smaller shift -- we get the
4046             * same result in the high half either way.
4047             */
4048            tcg_gen_sari_vec(MO_32, t1, v1, MIN(imm, 31));
4049            tcg_gen_shri_vec(MO_64, v0, v1, imm);
4050            vec_gen_4(INDEX_op_x86_blend_vec, type, MO_32,
4051                      tcgv_vec_arg(v0), tcgv_vec_arg(v0),
4052                      tcgv_vec_arg(t1), 0xaa);
4053        } else {
4054            /* Otherwise we will need to use a compare vs 0 to produce
4055             * the sign-extend, shift and merge.
4056             */
4057            tcg_gen_cmp_vec(TCG_COND_GT, MO_64, t1,
4058                            tcg_constant_vec(type, MO_64, 0), v1);
4059            tcg_gen_shri_vec(MO_64, v0, v1, imm);
4060            tcg_gen_shli_vec(MO_64, t1, t1, 64 - imm);
4061            tcg_gen_or_vec(MO_64, v0, v0, t1);
4062        }
4063        tcg_temp_free_vec(t1);
4064        break;
4065
4066    default:
4067        g_assert_not_reached();
4068    }
4069}
4070
4071static void expand_vec_rotli(TCGType type, unsigned vece,
4072                             TCGv_vec v0, TCGv_vec v1, TCGArg imm)
4073{
4074    TCGv_vec t;
4075
4076    if (vece != MO_8 && have_avx512vbmi2) {
4077        vec_gen_4(INDEX_op_x86_vpshldi_vec, type, vece,
4078                  tcgv_vec_arg(v0), tcgv_vec_arg(v1), tcgv_vec_arg(v1), imm);
4079        return;
4080    }
4081
4082    t = tcg_temp_new_vec(type);
4083    tcg_gen_shli_vec(vece, t, v1, imm);
4084    tcg_gen_shri_vec(vece, v0, v1, (8 << vece) - imm);
4085    tcg_gen_or_vec(vece, v0, v0, t);
4086    tcg_temp_free_vec(t);
4087}
4088
4089static void expand_vec_rotv(TCGType type, unsigned vece, TCGv_vec v0,
4090                            TCGv_vec v1, TCGv_vec sh, bool right)
4091{
4092    TCGv_vec t;
4093
4094    if (have_avx512vbmi2) {
4095        vec_gen_4(right ? INDEX_op_x86_vpshrdv_vec : INDEX_op_x86_vpshldv_vec,
4096                  type, vece, tcgv_vec_arg(v0), tcgv_vec_arg(v1),
4097                  tcgv_vec_arg(v1), tcgv_vec_arg(sh));
4098        return;
4099    }
4100
4101    t = tcg_temp_new_vec(type);
4102    tcg_gen_dupi_vec(vece, t, 8 << vece);
4103    tcg_gen_sub_vec(vece, t, t, sh);
4104    if (right) {
4105        tcg_gen_shlv_vec(vece, t, v1, t);
4106        tcg_gen_shrv_vec(vece, v0, v1, sh);
4107    } else {
4108        tcg_gen_shrv_vec(vece, t, v1, t);
4109        tcg_gen_shlv_vec(vece, v0, v1, sh);
4110    }
4111    tcg_gen_or_vec(vece, v0, v0, t);
4112    tcg_temp_free_vec(t);
4113}
4114
4115static void expand_vec_rotls(TCGType type, unsigned vece,
4116                             TCGv_vec v0, TCGv_vec v1, TCGv_i32 lsh)
4117{
4118    TCGv_vec t = tcg_temp_new_vec(type);
4119
4120    tcg_debug_assert(vece != MO_8);
4121
4122    if (vece >= MO_32 ? have_avx512vl : have_avx512vbmi2) {
4123        tcg_gen_dup_i32_vec(vece, t, lsh);
4124        if (vece >= MO_32) {
4125            tcg_gen_rotlv_vec(vece, v0, v1, t);
4126        } else {
4127            expand_vec_rotv(type, vece, v0, v1, t, false);
4128        }
4129    } else {
4130        TCGv_i32 rsh = tcg_temp_new_i32();
4131
4132        tcg_gen_neg_i32(rsh, lsh);
4133        tcg_gen_andi_i32(rsh, rsh, (8 << vece) - 1);
4134        tcg_gen_shls_vec(vece, t, v1, lsh);
4135        tcg_gen_shrs_vec(vece, v0, v1, rsh);
4136        tcg_gen_or_vec(vece, v0, v0, t);
4137
4138        tcg_temp_free_i32(rsh);
4139    }
4140
4141    tcg_temp_free_vec(t);
4142}
4143
4144static void expand_vec_mul(TCGType type, unsigned vece,
4145                           TCGv_vec v0, TCGv_vec v1, TCGv_vec v2)
4146{
4147    TCGv_vec t1, t2, t3, t4, zero;
4148
4149    tcg_debug_assert(vece == MO_8);
4150
4151    /*
4152     * Unpack v1 bytes to words, 0 | x.
4153     * Unpack v2 bytes to words, y | 0.
4154     * This leaves the 8-bit result, x * y, with 8 bits of right padding.
4155     * Shift logical right by 8 bits to clear the high 8 bytes before
4156     * using an unsigned saturated pack.
4157     *
4158     * The difference between the V64, V128 and V256 cases is merely how
4159     * we distribute the expansion between temporaries.
4160     */
4161    switch (type) {
4162    case TCG_TYPE_V64:
4163        t1 = tcg_temp_new_vec(TCG_TYPE_V128);
4164        t2 = tcg_temp_new_vec(TCG_TYPE_V128);
4165        zero = tcg_constant_vec(TCG_TYPE_V128, MO_8, 0);
4166        vec_gen_3(INDEX_op_x86_punpckl_vec, TCG_TYPE_V128, MO_8,
4167                  tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
4168        vec_gen_3(INDEX_op_x86_punpckl_vec, TCG_TYPE_V128, MO_8,
4169                  tcgv_vec_arg(t2), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
4170        tcg_gen_mul_vec(MO_16, t1, t1, t2);
4171        tcg_gen_shri_vec(MO_16, t1, t1, 8);
4172        vec_gen_3(INDEX_op_x86_packus_vec, TCG_TYPE_V128, MO_8,
4173                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t1));
4174        tcg_temp_free_vec(t1);
4175        tcg_temp_free_vec(t2);
4176        break;
4177
4178    case TCG_TYPE_V128:
4179    case TCG_TYPE_V256:
4180        t1 = tcg_temp_new_vec(type);
4181        t2 = tcg_temp_new_vec(type);
4182        t3 = tcg_temp_new_vec(type);
4183        t4 = tcg_temp_new_vec(type);
4184        zero = tcg_constant_vec(TCG_TYPE_V128, MO_8, 0);
4185        vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
4186                  tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
4187        vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
4188                  tcgv_vec_arg(t2), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
4189        vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
4190                  tcgv_vec_arg(t3), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
4191        vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
4192                  tcgv_vec_arg(t4), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
4193        tcg_gen_mul_vec(MO_16, t1, t1, t2);
4194        tcg_gen_mul_vec(MO_16, t3, t3, t4);
4195        tcg_gen_shri_vec(MO_16, t1, t1, 8);
4196        tcg_gen_shri_vec(MO_16, t3, t3, 8);
4197        vec_gen_3(INDEX_op_x86_packus_vec, type, MO_8,
4198                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t3));
4199        tcg_temp_free_vec(t1);
4200        tcg_temp_free_vec(t2);
4201        tcg_temp_free_vec(t3);
4202        tcg_temp_free_vec(t4);
4203        break;
4204
4205    default:
4206        g_assert_not_reached();
4207    }
4208}
4209
4210static TCGCond expand_vec_cond(TCGType type, unsigned vece,
4211                               TCGArg *a1, TCGArg *a2, TCGCond cond)
4212{
4213    /*
4214     * Without AVX512, there are no 64-bit unsigned comparisons.
4215     * We must bias the inputs so that they become signed.
4216     * All other swapping and inversion are handled during code generation.
4217     */
4218    if (vece == MO_64 && !have_avx512dq && is_unsigned_cond(cond)) {
4219        TCGv_vec v1 = temp_tcgv_vec(arg_temp(*a1));
4220        TCGv_vec v2 = temp_tcgv_vec(arg_temp(*a2));
4221        TCGv_vec t1 = tcg_temp_new_vec(type);
4222        TCGv_vec t2 = tcg_temp_new_vec(type);
4223        TCGv_vec t3 = tcg_constant_vec(type, vece, 1ull << ((8 << vece) - 1));
4224
4225        tcg_gen_sub_vec(vece, t1, v1, t3);
4226        tcg_gen_sub_vec(vece, t2, v2, t3);
4227        *a1 = tcgv_vec_arg(t1);
4228        *a2 = tcgv_vec_arg(t2);
4229        cond = tcg_signed_cond(cond);
4230    }
4231    return cond;
4232}
4233
4234static void expand_vec_cmp(TCGType type, unsigned vece, TCGArg a0,
4235                           TCGArg a1, TCGArg a2, TCGCond cond)
4236{
4237    cond = expand_vec_cond(type, vece, &a1, &a2, cond);
4238    /* Expand directly; do not recurse.  */
4239    vec_gen_4(INDEX_op_cmp_vec, type, vece, a0, a1, a2, cond);
4240}
4241
4242static void expand_vec_cmpsel(TCGType type, unsigned vece, TCGArg a0,
4243                              TCGArg a1, TCGArg a2,
4244                              TCGArg a3, TCGArg a4, TCGCond cond)
4245{
4246    cond = expand_vec_cond(type, vece, &a1, &a2, cond);
4247    /* Expand directly; do not recurse.  */
4248    vec_gen_6(INDEX_op_cmpsel_vec, type, vece, a0, a1, a2, a3, a4, cond);
4249}
4250
4251void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
4252                       TCGArg a0, ...)
4253{
4254    va_list va;
4255    TCGArg a1, a2, a3, a4, a5;
4256    TCGv_vec v0, v1, v2;
4257
4258    va_start(va, a0);
4259    a1 = va_arg(va, TCGArg);
4260    a2 = va_arg(va, TCGArg);
4261    v0 = temp_tcgv_vec(arg_temp(a0));
4262    v1 = temp_tcgv_vec(arg_temp(a1));
4263
4264    switch (opc) {
4265    case INDEX_op_shli_vec:
4266        expand_vec_shi(type, vece, false, v0, v1, a2);
4267        break;
4268    case INDEX_op_shri_vec:
4269        expand_vec_shi(type, vece, true, v0, v1, a2);
4270        break;
4271    case INDEX_op_sari_vec:
4272        expand_vec_sari(type, vece, v0, v1, a2);
4273        break;
4274
4275    case INDEX_op_rotli_vec:
4276        expand_vec_rotli(type, vece, v0, v1, a2);
4277        break;
4278
4279    case INDEX_op_rotls_vec:
4280        expand_vec_rotls(type, vece, v0, v1, temp_tcgv_i32(arg_temp(a2)));
4281        break;
4282
4283    case INDEX_op_rotlv_vec:
4284        v2 = temp_tcgv_vec(arg_temp(a2));
4285        expand_vec_rotv(type, vece, v0, v1, v2, false);
4286        break;
4287    case INDEX_op_rotrv_vec:
4288        v2 = temp_tcgv_vec(arg_temp(a2));
4289        expand_vec_rotv(type, vece, v0, v1, v2, true);
4290        break;
4291
4292    case INDEX_op_mul_vec:
4293        v2 = temp_tcgv_vec(arg_temp(a2));
4294        expand_vec_mul(type, vece, v0, v1, v2);
4295        break;
4296
4297    case INDEX_op_cmp_vec:
4298        a3 = va_arg(va, TCGArg);
4299        expand_vec_cmp(type, vece, a0, a1, a2, a3);
4300        break;
4301
4302    case INDEX_op_cmpsel_vec:
4303        a3 = va_arg(va, TCGArg);
4304        a4 = va_arg(va, TCGArg);
4305        a5 = va_arg(va, TCGArg);
4306        expand_vec_cmpsel(type, vece, a0, a1, a2, a3, a4, a5);
4307        break;
4308
4309    default:
4310        break;
4311    }
4312
4313    va_end(va);
4314}
4315
4316static const int tcg_target_callee_save_regs[] = {
4317#if TCG_TARGET_REG_BITS == 64
4318    TCG_REG_RBP,
4319    TCG_REG_RBX,
4320#if defined(_WIN64)
4321    TCG_REG_RDI,
4322    TCG_REG_RSI,
4323#endif
4324    TCG_REG_R12,
4325    TCG_REG_R13,
4326    TCG_REG_R14, /* Currently used for the global env. */
4327    TCG_REG_R15,
4328#else
4329    TCG_REG_EBP, /* Currently used for the global env. */
4330    TCG_REG_EBX,
4331    TCG_REG_ESI,
4332    TCG_REG_EDI,
4333#endif
4334};
4335
4336/* Compute frame size via macros, to share between tcg_target_qemu_prologue
4337   and tcg_register_jit.  */
4338
4339#define PUSH_SIZE \
4340    ((1 + ARRAY_SIZE(tcg_target_callee_save_regs)) \
4341     * (TCG_TARGET_REG_BITS / 8))
4342
4343#define FRAME_SIZE \
4344    ((PUSH_SIZE \
4345      + TCG_STATIC_CALL_ARGS_SIZE \
4346      + CPU_TEMP_BUF_NLONGS * sizeof(long) \
4347      + TCG_TARGET_STACK_ALIGN - 1) \
4348     & ~(TCG_TARGET_STACK_ALIGN - 1))
4349
4350/* Generate global QEMU prologue and epilogue code */
4351static void tcg_target_qemu_prologue(TCGContext *s)
4352{
4353    int i, stack_addend;
4354
4355    /* TB prologue */
4356
4357    /* Reserve some stack space, also for TCG temps.  */
4358    stack_addend = FRAME_SIZE - PUSH_SIZE;
4359    tcg_set_frame(s, TCG_REG_CALL_STACK, TCG_STATIC_CALL_ARGS_SIZE,
4360                  CPU_TEMP_BUF_NLONGS * sizeof(long));
4361
4362    /* Save all callee saved registers.  */
4363    for (i = 0; i < ARRAY_SIZE(tcg_target_callee_save_regs); i++) {
4364        tcg_out_push(s, tcg_target_callee_save_regs[i]);
4365    }
4366
4367    if (!tcg_use_softmmu && guest_base) {
4368        int seg = setup_guest_base_seg();
4369        if (seg != 0) {
4370            x86_guest_base.seg = seg;
4371        } else if (guest_base == (int32_t)guest_base) {
4372            x86_guest_base.ofs = guest_base;
4373        } else {
4374            assert(TCG_TARGET_REG_BITS == 64);
4375            /* Choose R12 because, as a base, it requires a SIB byte. */
4376            x86_guest_base.index = TCG_REG_R12;
4377            tcg_out_movi(s, TCG_TYPE_PTR, x86_guest_base.index, guest_base);
4378            tcg_regset_set_reg(s->reserved_regs, x86_guest_base.index);
4379        }
4380    }
4381
4382    if (TCG_TARGET_REG_BITS == 32) {
4383        tcg_out_ld(s, TCG_TYPE_PTR, TCG_AREG0, TCG_REG_ESP,
4384                   (ARRAY_SIZE(tcg_target_callee_save_regs) + 1) * 4);
4385        tcg_out_addi(s, TCG_REG_ESP, -stack_addend);
4386        /* jmp *tb.  */
4387        tcg_out_modrm_offset(s, OPC_GRP5, EXT5_JMPN_Ev, TCG_REG_ESP,
4388                             (ARRAY_SIZE(tcg_target_callee_save_regs) + 2) * 4
4389                             + stack_addend);
4390    } else {
4391        tcg_out_mov(s, TCG_TYPE_PTR, TCG_AREG0, tcg_target_call_iarg_regs[0]);
4392        tcg_out_addi(s, TCG_REG_ESP, -stack_addend);
4393        /* jmp *tb.  */
4394        tcg_out_modrm(s, OPC_GRP5, EXT5_JMPN_Ev, tcg_target_call_iarg_regs[1]);
4395    }
4396
4397    /*
4398     * Return path for goto_ptr. Set return value to 0, a-la exit_tb,
4399     * and fall through to the rest of the epilogue.
4400     */
4401    tcg_code_gen_epilogue = tcg_splitwx_to_rx(s->code_ptr);
4402    tcg_out_movi(s, TCG_TYPE_REG, TCG_REG_EAX, 0);
4403
4404    /* TB epilogue */
4405    tb_ret_addr = tcg_splitwx_to_rx(s->code_ptr);
4406
4407    tcg_out_addi(s, TCG_REG_CALL_STACK, stack_addend);
4408
4409    if (have_avx2) {
4410        tcg_out_vex_opc(s, OPC_VZEROUPPER, 0, 0, 0, 0);
4411    }
4412    for (i = ARRAY_SIZE(tcg_target_callee_save_regs) - 1; i >= 0; i--) {
4413        tcg_out_pop(s, tcg_target_callee_save_regs[i]);
4414    }
4415    tcg_out_opc(s, OPC_RET, 0, 0, 0);
4416}
4417
4418static void tcg_out_tb_start(TCGContext *s)
4419{
4420    /* nothing to do */
4421}
4422
4423static void tcg_out_nop_fill(tcg_insn_unit *p, int count)
4424{
4425    memset(p, 0x90, count);
4426}
4427
4428static void tcg_target_init(TCGContext *s)
4429{
4430    tcg_target_available_regs[TCG_TYPE_I32] = ALL_GENERAL_REGS;
4431    if (TCG_TARGET_REG_BITS == 64) {
4432        tcg_target_available_regs[TCG_TYPE_I64] = ALL_GENERAL_REGS;
4433    }
4434    if (have_avx1) {
4435        tcg_target_available_regs[TCG_TYPE_V64] = ALL_VECTOR_REGS;
4436        tcg_target_available_regs[TCG_TYPE_V128] = ALL_VECTOR_REGS;
4437    }
4438    if (have_avx2) {
4439        tcg_target_available_regs[TCG_TYPE_V256] = ALL_VECTOR_REGS;
4440    }
4441
4442    tcg_target_call_clobber_regs = ALL_VECTOR_REGS;
4443    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_EAX);
4444    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_EDX);
4445    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_ECX);
4446    if (TCG_TARGET_REG_BITS == 64) {
4447#if !defined(_WIN64)
4448        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_RDI);
4449        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_RSI);
4450#endif
4451        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R8);
4452        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R9);
4453        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R10);
4454        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R11);
4455    }
4456
4457    s->reserved_regs = 0;
4458    tcg_regset_set_reg(s->reserved_regs, TCG_REG_CALL_STACK);
4459    tcg_regset_set_reg(s->reserved_regs, TCG_TMP_VEC);
4460#ifdef _WIN64
4461    /* These are call saved, and we don't save them, so don't use them. */
4462    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM6);
4463    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM7);
4464    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM8);
4465    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM9);
4466    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM10);
4467    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM11);
4468    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM12);
4469    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM13);
4470    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM14);
4471    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM15);
4472#endif
4473}
4474
4475typedef struct {
4476    DebugFrameHeader h;
4477    uint8_t fde_def_cfa[4];
4478    uint8_t fde_reg_ofs[14];
4479} DebugFrame;
4480
4481/* We're expecting a 2 byte uleb128 encoded value.  */
4482QEMU_BUILD_BUG_ON(FRAME_SIZE >= (1 << 14));
4483
4484#if !defined(__ELF__)
4485    /* Host machine without ELF. */
4486#elif TCG_TARGET_REG_BITS == 64
4487#define ELF_HOST_MACHINE EM_X86_64
4488static const DebugFrame debug_frame = {
4489    .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
4490    .h.cie.id = -1,
4491    .h.cie.version = 1,
4492    .h.cie.code_align = 1,
4493    .h.cie.data_align = 0x78,             /* sleb128 -8 */
4494    .h.cie.return_column = 16,
4495
4496    /* Total FDE size does not include the "len" member.  */
4497    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
4498
4499    .fde_def_cfa = {
4500        12, 7,                          /* DW_CFA_def_cfa %rsp, ... */
4501        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
4502        (FRAME_SIZE >> 7)
4503    },
4504    .fde_reg_ofs = {
4505        0x90, 1,                        /* DW_CFA_offset, %rip, -8 */
4506        /* The following ordering must match tcg_target_callee_save_regs.  */
4507        0x86, 2,                        /* DW_CFA_offset, %rbp, -16 */
4508        0x83, 3,                        /* DW_CFA_offset, %rbx, -24 */
4509        0x8c, 4,                        /* DW_CFA_offset, %r12, -32 */
4510        0x8d, 5,                        /* DW_CFA_offset, %r13, -40 */
4511        0x8e, 6,                        /* DW_CFA_offset, %r14, -48 */
4512        0x8f, 7,                        /* DW_CFA_offset, %r15, -56 */
4513    }
4514};
4515#else
4516#define ELF_HOST_MACHINE EM_386
4517static const DebugFrame debug_frame = {
4518    .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
4519    .h.cie.id = -1,
4520    .h.cie.version = 1,
4521    .h.cie.code_align = 1,
4522    .h.cie.data_align = 0x7c,             /* sleb128 -4 */
4523    .h.cie.return_column = 8,
4524
4525    /* Total FDE size does not include the "len" member.  */
4526    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
4527
4528    .fde_def_cfa = {
4529        12, 4,                          /* DW_CFA_def_cfa %esp, ... */
4530        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
4531        (FRAME_SIZE >> 7)
4532    },
4533    .fde_reg_ofs = {
4534        0x88, 1,                        /* DW_CFA_offset, %eip, -4 */
4535        /* The following ordering must match tcg_target_callee_save_regs.  */
4536        0x85, 2,                        /* DW_CFA_offset, %ebp, -8 */
4537        0x83, 3,                        /* DW_CFA_offset, %ebx, -12 */
4538        0x86, 4,                        /* DW_CFA_offset, %esi, -16 */
4539        0x87, 5,                        /* DW_CFA_offset, %edi, -20 */
4540    }
4541};
4542#endif
4543
4544#if defined(ELF_HOST_MACHINE)
4545void tcg_register_jit(const void *buf, size_t buf_size)
4546{
4547    tcg_register_jit_int(buf, buf_size, &debug_frame, sizeof(debug_frame));
4548}
4549#endif
4550