xref: /openbmc/qemu/tcg/i386/tcg-target.c.inc (revision 89aafcf2)
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#include "../tcg-ldst.c.inc"
26#include "../tcg-pool.c.inc"
27
28#ifdef CONFIG_DEBUG_TCG
29static const char * const tcg_target_reg_names[TCG_TARGET_NB_REGS] = {
30#if TCG_TARGET_REG_BITS == 64
31    "%rax", "%rcx", "%rdx", "%rbx", "%rsp", "%rbp", "%rsi", "%rdi",
32#else
33    "%eax", "%ecx", "%edx", "%ebx", "%esp", "%ebp", "%esi", "%edi",
34#endif
35    "%r8",  "%r9",  "%r10", "%r11", "%r12", "%r13", "%r14", "%r15",
36    "%xmm0", "%xmm1", "%xmm2", "%xmm3", "%xmm4", "%xmm5", "%xmm6", "%xmm7",
37#if TCG_TARGET_REG_BITS == 64
38    "%xmm8", "%xmm9", "%xmm10", "%xmm11",
39    "%xmm12", "%xmm13", "%xmm14", "%xmm15",
40#endif
41};
42#endif
43
44static const int tcg_target_reg_alloc_order[] = {
45#if TCG_TARGET_REG_BITS == 64
46    TCG_REG_RBP,
47    TCG_REG_RBX,
48    TCG_REG_R12,
49    TCG_REG_R13,
50    TCG_REG_R14,
51    TCG_REG_R15,
52    TCG_REG_R10,
53    TCG_REG_R11,
54    TCG_REG_R9,
55    TCG_REG_R8,
56    TCG_REG_RCX,
57    TCG_REG_RDX,
58    TCG_REG_RSI,
59    TCG_REG_RDI,
60    TCG_REG_RAX,
61#else
62    TCG_REG_EBX,
63    TCG_REG_ESI,
64    TCG_REG_EDI,
65    TCG_REG_EBP,
66    TCG_REG_ECX,
67    TCG_REG_EDX,
68    TCG_REG_EAX,
69#endif
70    TCG_REG_XMM0,
71    TCG_REG_XMM1,
72    TCG_REG_XMM2,
73    TCG_REG_XMM3,
74    TCG_REG_XMM4,
75    TCG_REG_XMM5,
76#ifndef _WIN64
77    /* The Win64 ABI has xmm6-xmm15 as caller-saves, and we do not save
78       any of them.  Therefore only allow xmm0-xmm5 to be allocated.  */
79    TCG_REG_XMM6,
80    TCG_REG_XMM7,
81#if TCG_TARGET_REG_BITS == 64
82    TCG_REG_XMM8,
83    TCG_REG_XMM9,
84    TCG_REG_XMM10,
85    TCG_REG_XMM11,
86    TCG_REG_XMM12,
87    TCG_REG_XMM13,
88    TCG_REG_XMM14,
89    TCG_REG_XMM15,
90#endif
91#endif
92};
93
94#define TCG_TMP_VEC  TCG_REG_XMM5
95
96static const int tcg_target_call_iarg_regs[] = {
97#if TCG_TARGET_REG_BITS == 64
98#if defined(_WIN64)
99    TCG_REG_RCX,
100    TCG_REG_RDX,
101#else
102    TCG_REG_RDI,
103    TCG_REG_RSI,
104    TCG_REG_RDX,
105    TCG_REG_RCX,
106#endif
107    TCG_REG_R8,
108    TCG_REG_R9,
109#else
110    /* 32 bit mode uses stack based calling convention (GCC default). */
111#endif
112};
113
114static TCGReg tcg_target_call_oarg_reg(TCGCallReturnKind kind, int slot)
115{
116    switch (kind) {
117    case TCG_CALL_RET_NORMAL:
118        tcg_debug_assert(slot >= 0 && slot <= 1);
119        return slot ? TCG_REG_EDX : TCG_REG_EAX;
120#ifdef _WIN64
121    case TCG_CALL_RET_BY_VEC:
122        tcg_debug_assert(slot == 0);
123        return TCG_REG_XMM0;
124#endif
125    default:
126        g_assert_not_reached();
127    }
128}
129
130/* Constants we accept.  */
131#define TCG_CT_CONST_S32 0x100
132#define TCG_CT_CONST_U32 0x200
133#define TCG_CT_CONST_I32 0x400
134#define TCG_CT_CONST_WSZ 0x800
135
136/* Registers used with L constraint, which are the first argument
137   registers on x86_64, and two random call clobbered registers on
138   i386. */
139#if TCG_TARGET_REG_BITS == 64
140# define TCG_REG_L0 tcg_target_call_iarg_regs[0]
141# define TCG_REG_L1 tcg_target_call_iarg_regs[1]
142#else
143# define TCG_REG_L0 TCG_REG_EAX
144# define TCG_REG_L1 TCG_REG_EDX
145#endif
146
147#define ALL_BYTEH_REGS         0x0000000fu
148#if TCG_TARGET_REG_BITS == 64
149# define ALL_GENERAL_REGS      0x0000ffffu
150# define ALL_VECTOR_REGS       0xffff0000u
151# define ALL_BYTEL_REGS        ALL_GENERAL_REGS
152#else
153# define ALL_GENERAL_REGS      0x000000ffu
154# define ALL_VECTOR_REGS       0x00ff0000u
155# define ALL_BYTEL_REGS        ALL_BYTEH_REGS
156#endif
157#ifdef CONFIG_SOFTMMU
158# define SOFTMMU_RESERVE_REGS  ((1 << TCG_REG_L0) | (1 << TCG_REG_L1))
159#else
160# define SOFTMMU_RESERVE_REGS  0
161#endif
162
163/* For 64-bit, we always know that CMOV is available.  */
164#if TCG_TARGET_REG_BITS == 64
165# define have_cmov      true
166#else
167# define have_cmov      (cpuinfo & CPUINFO_CMOV)
168#endif
169#define have_bmi2       (cpuinfo & CPUINFO_BMI2)
170#define have_lzcnt      (cpuinfo & CPUINFO_LZCNT)
171
172static const tcg_insn_unit *tb_ret_addr;
173
174static bool patch_reloc(tcg_insn_unit *code_ptr, int type,
175                        intptr_t value, intptr_t addend)
176{
177    value += addend;
178    switch(type) {
179    case R_386_PC32:
180        value -= (uintptr_t)tcg_splitwx_to_rx(code_ptr);
181        if (value != (int32_t)value) {
182            return false;
183        }
184        /* FALLTHRU */
185    case R_386_32:
186        tcg_patch32(code_ptr, value);
187        break;
188    case R_386_PC8:
189        value -= (uintptr_t)tcg_splitwx_to_rx(code_ptr);
190        if (value != (int8_t)value) {
191            return false;
192        }
193        tcg_patch8(code_ptr, value);
194        break;
195    default:
196        g_assert_not_reached();
197    }
198    return true;
199}
200
201/* test if a constant matches the constraint */
202static bool tcg_target_const_match(int64_t val, TCGType type, int ct)
203{
204    if (ct & TCG_CT_CONST) {
205        return 1;
206    }
207    if (type == TCG_TYPE_I32) {
208        if (ct & (TCG_CT_CONST_S32 | TCG_CT_CONST_U32 | TCG_CT_CONST_I32)) {
209            return 1;
210        }
211    } else {
212        if ((ct & TCG_CT_CONST_S32) && val == (int32_t)val) {
213            return 1;
214        }
215        if ((ct & TCG_CT_CONST_U32) && val == (uint32_t)val) {
216            return 1;
217        }
218        if ((ct & TCG_CT_CONST_I32) && ~val == (int32_t)~val) {
219            return 1;
220        }
221    }
222    if ((ct & TCG_CT_CONST_WSZ) && val == (type == TCG_TYPE_I32 ? 32 : 64)) {
223        return 1;
224    }
225    return 0;
226}
227
228# define LOWREGMASK(x)	((x) & 7)
229
230#define P_EXT		0x100		/* 0x0f opcode prefix */
231#define P_EXT38         0x200           /* 0x0f 0x38 opcode prefix */
232#define P_DATA16        0x400           /* 0x66 opcode prefix */
233#define P_VEXW          0x1000          /* Set VEX.W = 1 */
234#if TCG_TARGET_REG_BITS == 64
235# define P_REXW         P_VEXW          /* Set REX.W = 1; match VEXW */
236# define P_REXB_R       0x2000          /* REG field as byte register */
237# define P_REXB_RM      0x4000          /* R/M field as byte register */
238# define P_GS           0x8000          /* gs segment override */
239#else
240# define P_REXW		0
241# define P_REXB_R	0
242# define P_REXB_RM	0
243# define P_GS           0
244#endif
245#define P_EXT3A         0x10000         /* 0x0f 0x3a opcode prefix */
246#define P_SIMDF3        0x20000         /* 0xf3 opcode prefix */
247#define P_SIMDF2        0x40000         /* 0xf2 opcode prefix */
248#define P_VEXL          0x80000         /* Set VEX.L = 1 */
249#define P_EVEX          0x100000        /* Requires EVEX encoding */
250
251#define OPC_ARITH_EvIz	(0x81)
252#define OPC_ARITH_EvIb	(0x83)
253#define OPC_ARITH_GvEv	(0x03)		/* ... plus (ARITH_FOO << 3) */
254#define OPC_ANDN        (0xf2 | P_EXT38)
255#define OPC_ADD_GvEv	(OPC_ARITH_GvEv | (ARITH_ADD << 3))
256#define OPC_AND_GvEv    (OPC_ARITH_GvEv | (ARITH_AND << 3))
257#define OPC_BLENDPS     (0x0c | P_EXT3A | P_DATA16)
258#define OPC_BSF         (0xbc | P_EXT)
259#define OPC_BSR         (0xbd | P_EXT)
260#define OPC_BSWAP	(0xc8 | P_EXT)
261#define OPC_CALL_Jz	(0xe8)
262#define OPC_CMOVCC      (0x40 | P_EXT)  /* ... plus condition code */
263#define OPC_CMP_GvEv	(OPC_ARITH_GvEv | (ARITH_CMP << 3))
264#define OPC_DEC_r32	(0x48)
265#define OPC_IMUL_GvEv	(0xaf | P_EXT)
266#define OPC_IMUL_GvEvIb	(0x6b)
267#define OPC_IMUL_GvEvIz	(0x69)
268#define OPC_INC_r32	(0x40)
269#define OPC_JCC_long	(0x80 | P_EXT)	/* ... plus condition code */
270#define OPC_JCC_short	(0x70)		/* ... plus condition code */
271#define OPC_JMP_long	(0xe9)
272#define OPC_JMP_short	(0xeb)
273#define OPC_LEA         (0x8d)
274#define OPC_LZCNT       (0xbd | P_EXT | P_SIMDF3)
275#define OPC_MOVB_EvGv	(0x88)		/* stores, more or less */
276#define OPC_MOVL_EvGv	(0x89)		/* stores, more or less */
277#define OPC_MOVL_GvEv	(0x8b)		/* loads, more or less */
278#define OPC_MOVB_EvIz   (0xc6)
279#define OPC_MOVL_EvIz	(0xc7)
280#define OPC_MOVL_Iv     (0xb8)
281#define OPC_MOVBE_GyMy  (0xf0 | P_EXT38)
282#define OPC_MOVBE_MyGy  (0xf1 | P_EXT38)
283#define OPC_MOVD_VyEy   (0x6e | P_EXT | P_DATA16)
284#define OPC_MOVD_EyVy   (0x7e | P_EXT | P_DATA16)
285#define OPC_MOVDDUP     (0x12 | P_EXT | P_SIMDF2)
286#define OPC_MOVDQA_VxWx (0x6f | P_EXT | P_DATA16)
287#define OPC_MOVDQA_WxVx (0x7f | P_EXT | P_DATA16)
288#define OPC_MOVDQU_VxWx (0x6f | P_EXT | P_SIMDF3)
289#define OPC_MOVDQU_WxVx (0x7f | P_EXT | P_SIMDF3)
290#define OPC_MOVQ_VqWq   (0x7e | P_EXT | P_SIMDF3)
291#define OPC_MOVQ_WqVq   (0xd6 | P_EXT | P_DATA16)
292#define OPC_MOVSBL	(0xbe | P_EXT)
293#define OPC_MOVSWL	(0xbf | P_EXT)
294#define OPC_MOVSLQ	(0x63 | P_REXW)
295#define OPC_MOVZBL	(0xb6 | P_EXT)
296#define OPC_MOVZWL	(0xb7 | P_EXT)
297#define OPC_PABSB       (0x1c | P_EXT38 | P_DATA16)
298#define OPC_PABSW       (0x1d | P_EXT38 | P_DATA16)
299#define OPC_PABSD       (0x1e | P_EXT38 | P_DATA16)
300#define OPC_VPABSQ      (0x1f | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
301#define OPC_PACKSSDW    (0x6b | P_EXT | P_DATA16)
302#define OPC_PACKSSWB    (0x63 | P_EXT | P_DATA16)
303#define OPC_PACKUSDW    (0x2b | P_EXT38 | P_DATA16)
304#define OPC_PACKUSWB    (0x67 | P_EXT | P_DATA16)
305#define OPC_PADDB       (0xfc | P_EXT | P_DATA16)
306#define OPC_PADDW       (0xfd | P_EXT | P_DATA16)
307#define OPC_PADDD       (0xfe | P_EXT | P_DATA16)
308#define OPC_PADDQ       (0xd4 | P_EXT | P_DATA16)
309#define OPC_PADDSB      (0xec | P_EXT | P_DATA16)
310#define OPC_PADDSW      (0xed | P_EXT | P_DATA16)
311#define OPC_PADDUB      (0xdc | P_EXT | P_DATA16)
312#define OPC_PADDUW      (0xdd | P_EXT | P_DATA16)
313#define OPC_PAND        (0xdb | P_EXT | P_DATA16)
314#define OPC_PANDN       (0xdf | P_EXT | P_DATA16)
315#define OPC_PBLENDW     (0x0e | P_EXT3A | P_DATA16)
316#define OPC_PCMPEQB     (0x74 | P_EXT | P_DATA16)
317#define OPC_PCMPEQW     (0x75 | P_EXT | P_DATA16)
318#define OPC_PCMPEQD     (0x76 | P_EXT | P_DATA16)
319#define OPC_PCMPEQQ     (0x29 | P_EXT38 | P_DATA16)
320#define OPC_PCMPGTB     (0x64 | P_EXT | P_DATA16)
321#define OPC_PCMPGTW     (0x65 | P_EXT | P_DATA16)
322#define OPC_PCMPGTD     (0x66 | P_EXT | P_DATA16)
323#define OPC_PCMPGTQ     (0x37 | P_EXT38 | P_DATA16)
324#define OPC_PEXTRD      (0x16 | P_EXT3A | P_DATA16)
325#define OPC_PINSRD      (0x22 | P_EXT3A | P_DATA16)
326#define OPC_PMAXSB      (0x3c | P_EXT38 | P_DATA16)
327#define OPC_PMAXSW      (0xee | P_EXT | P_DATA16)
328#define OPC_PMAXSD      (0x3d | P_EXT38 | P_DATA16)
329#define OPC_VPMAXSQ     (0x3d | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
330#define OPC_PMAXUB      (0xde | P_EXT | P_DATA16)
331#define OPC_PMAXUW      (0x3e | P_EXT38 | P_DATA16)
332#define OPC_PMAXUD      (0x3f | P_EXT38 | P_DATA16)
333#define OPC_VPMAXUQ     (0x3f | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
334#define OPC_PMINSB      (0x38 | P_EXT38 | P_DATA16)
335#define OPC_PMINSW      (0xea | P_EXT | P_DATA16)
336#define OPC_PMINSD      (0x39 | P_EXT38 | P_DATA16)
337#define OPC_VPMINSQ     (0x39 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
338#define OPC_PMINUB      (0xda | P_EXT | P_DATA16)
339#define OPC_PMINUW      (0x3a | P_EXT38 | P_DATA16)
340#define OPC_PMINUD      (0x3b | P_EXT38 | P_DATA16)
341#define OPC_VPMINUQ     (0x3b | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
342#define OPC_PMOVSXBW    (0x20 | P_EXT38 | P_DATA16)
343#define OPC_PMOVSXWD    (0x23 | P_EXT38 | P_DATA16)
344#define OPC_PMOVSXDQ    (0x25 | P_EXT38 | P_DATA16)
345#define OPC_PMOVZXBW    (0x30 | P_EXT38 | P_DATA16)
346#define OPC_PMOVZXWD    (0x33 | P_EXT38 | P_DATA16)
347#define OPC_PMOVZXDQ    (0x35 | P_EXT38 | P_DATA16)
348#define OPC_PMULLW      (0xd5 | P_EXT | P_DATA16)
349#define OPC_PMULLD      (0x40 | P_EXT38 | P_DATA16)
350#define OPC_VPMULLQ     (0x40 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
351#define OPC_POR         (0xeb | P_EXT | P_DATA16)
352#define OPC_PSHUFB      (0x00 | P_EXT38 | P_DATA16)
353#define OPC_PSHUFD      (0x70 | P_EXT | P_DATA16)
354#define OPC_PSHUFLW     (0x70 | P_EXT | P_SIMDF2)
355#define OPC_PSHUFHW     (0x70 | P_EXT | P_SIMDF3)
356#define OPC_PSHIFTW_Ib  (0x71 | P_EXT | P_DATA16) /* /2 /6 /4 */
357#define OPC_PSHIFTD_Ib  (0x72 | P_EXT | P_DATA16) /* /1 /2 /6 /4 */
358#define OPC_PSHIFTQ_Ib  (0x73 | P_EXT | P_DATA16) /* /2 /6 /4 */
359#define OPC_PSLLW       (0xf1 | P_EXT | P_DATA16)
360#define OPC_PSLLD       (0xf2 | P_EXT | P_DATA16)
361#define OPC_PSLLQ       (0xf3 | P_EXT | P_DATA16)
362#define OPC_PSRAW       (0xe1 | P_EXT | P_DATA16)
363#define OPC_PSRAD       (0xe2 | P_EXT | P_DATA16)
364#define OPC_VPSRAQ      (0xe2 | P_EXT | P_DATA16 | P_VEXW | P_EVEX)
365#define OPC_PSRLW       (0xd1 | P_EXT | P_DATA16)
366#define OPC_PSRLD       (0xd2 | P_EXT | P_DATA16)
367#define OPC_PSRLQ       (0xd3 | P_EXT | P_DATA16)
368#define OPC_PSUBB       (0xf8 | P_EXT | P_DATA16)
369#define OPC_PSUBW       (0xf9 | P_EXT | P_DATA16)
370#define OPC_PSUBD       (0xfa | P_EXT | P_DATA16)
371#define OPC_PSUBQ       (0xfb | P_EXT | P_DATA16)
372#define OPC_PSUBSB      (0xe8 | P_EXT | P_DATA16)
373#define OPC_PSUBSW      (0xe9 | P_EXT | P_DATA16)
374#define OPC_PSUBUB      (0xd8 | P_EXT | P_DATA16)
375#define OPC_PSUBUW      (0xd9 | P_EXT | P_DATA16)
376#define OPC_PUNPCKLBW   (0x60 | P_EXT | P_DATA16)
377#define OPC_PUNPCKLWD   (0x61 | P_EXT | P_DATA16)
378#define OPC_PUNPCKLDQ   (0x62 | P_EXT | P_DATA16)
379#define OPC_PUNPCKLQDQ  (0x6c | P_EXT | P_DATA16)
380#define OPC_PUNPCKHBW   (0x68 | P_EXT | P_DATA16)
381#define OPC_PUNPCKHWD   (0x69 | P_EXT | P_DATA16)
382#define OPC_PUNPCKHDQ   (0x6a | P_EXT | P_DATA16)
383#define OPC_PUNPCKHQDQ  (0x6d | P_EXT | P_DATA16)
384#define OPC_PXOR        (0xef | P_EXT | P_DATA16)
385#define OPC_POP_r32	(0x58)
386#define OPC_POPCNT      (0xb8 | P_EXT | P_SIMDF3)
387#define OPC_PUSH_r32	(0x50)
388#define OPC_PUSH_Iv	(0x68)
389#define OPC_PUSH_Ib	(0x6a)
390#define OPC_RET		(0xc3)
391#define OPC_SETCC	(0x90 | P_EXT | P_REXB_RM) /* ... plus cc */
392#define OPC_SHIFT_1	(0xd1)
393#define OPC_SHIFT_Ib	(0xc1)
394#define OPC_SHIFT_cl	(0xd3)
395#define OPC_SARX        (0xf7 | P_EXT38 | P_SIMDF3)
396#define OPC_SHUFPS      (0xc6 | P_EXT)
397#define OPC_SHLX        (0xf7 | P_EXT38 | P_DATA16)
398#define OPC_SHRX        (0xf7 | P_EXT38 | P_SIMDF2)
399#define OPC_SHRD_Ib     (0xac | P_EXT)
400#define OPC_TESTL	(0x85)
401#define OPC_TZCNT       (0xbc | P_EXT | P_SIMDF3)
402#define OPC_UD2         (0x0b | P_EXT)
403#define OPC_VPBLENDD    (0x02 | P_EXT3A | P_DATA16)
404#define OPC_VPBLENDVB   (0x4c | P_EXT3A | P_DATA16)
405#define OPC_VPINSRB     (0x20 | P_EXT3A | P_DATA16)
406#define OPC_VPINSRW     (0xc4 | P_EXT | P_DATA16)
407#define OPC_VBROADCASTSS (0x18 | P_EXT38 | P_DATA16)
408#define OPC_VBROADCASTSD (0x19 | P_EXT38 | P_DATA16)
409#define OPC_VPBROADCASTB (0x78 | P_EXT38 | P_DATA16)
410#define OPC_VPBROADCASTW (0x79 | P_EXT38 | P_DATA16)
411#define OPC_VPBROADCASTD (0x58 | P_EXT38 | P_DATA16)
412#define OPC_VPBROADCASTQ (0x59 | P_EXT38 | P_DATA16)
413#define OPC_VPERMQ      (0x00 | P_EXT3A | P_DATA16 | P_VEXW)
414#define OPC_VPERM2I128  (0x46 | P_EXT3A | P_DATA16 | P_VEXL)
415#define OPC_VPROLVD     (0x15 | P_EXT38 | P_DATA16 | P_EVEX)
416#define OPC_VPROLVQ     (0x15 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
417#define OPC_VPRORVD     (0x14 | P_EXT38 | P_DATA16 | P_EVEX)
418#define OPC_VPRORVQ     (0x14 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
419#define OPC_VPSHLDW     (0x70 | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
420#define OPC_VPSHLDD     (0x71 | P_EXT3A | P_DATA16 | P_EVEX)
421#define OPC_VPSHLDQ     (0x71 | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
422#define OPC_VPSHLDVW    (0x70 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
423#define OPC_VPSHLDVD    (0x71 | P_EXT38 | P_DATA16 | P_EVEX)
424#define OPC_VPSHLDVQ    (0x71 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
425#define OPC_VPSHRDVW    (0x72 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
426#define OPC_VPSHRDVD    (0x73 | P_EXT38 | P_DATA16 | P_EVEX)
427#define OPC_VPSHRDVQ    (0x73 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
428#define OPC_VPSLLVW     (0x12 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
429#define OPC_VPSLLVD     (0x47 | P_EXT38 | P_DATA16)
430#define OPC_VPSLLVQ     (0x47 | P_EXT38 | P_DATA16 | P_VEXW)
431#define OPC_VPSRAVW     (0x11 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
432#define OPC_VPSRAVD     (0x46 | P_EXT38 | P_DATA16)
433#define OPC_VPSRAVQ     (0x46 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
434#define OPC_VPSRLVW     (0x10 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
435#define OPC_VPSRLVD     (0x45 | P_EXT38 | P_DATA16)
436#define OPC_VPSRLVQ     (0x45 | P_EXT38 | P_DATA16 | P_VEXW)
437#define OPC_VPTERNLOGQ  (0x25 | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
438#define OPC_VZEROUPPER  (0x77 | P_EXT)
439#define OPC_XCHG_ax_r32	(0x90)
440#define OPC_XCHG_EvGv   (0x87)
441
442#define OPC_GRP3_Eb     (0xf6)
443#define OPC_GRP3_Ev     (0xf7)
444#define OPC_GRP5        (0xff)
445#define OPC_GRP14       (0x73 | P_EXT | P_DATA16)
446
447/* Group 1 opcode extensions for 0x80-0x83.
448   These are also used as modifiers for OPC_ARITH.  */
449#define ARITH_ADD 0
450#define ARITH_OR  1
451#define ARITH_ADC 2
452#define ARITH_SBB 3
453#define ARITH_AND 4
454#define ARITH_SUB 5
455#define ARITH_XOR 6
456#define ARITH_CMP 7
457
458/* Group 2 opcode extensions for 0xc0, 0xc1, 0xd0-0xd3.  */
459#define SHIFT_ROL 0
460#define SHIFT_ROR 1
461#define SHIFT_SHL 4
462#define SHIFT_SHR 5
463#define SHIFT_SAR 7
464
465/* Group 3 opcode extensions for 0xf6, 0xf7.  To be used with OPC_GRP3.  */
466#define EXT3_TESTi 0
467#define EXT3_NOT   2
468#define EXT3_NEG   3
469#define EXT3_MUL   4
470#define EXT3_IMUL  5
471#define EXT3_DIV   6
472#define EXT3_IDIV  7
473
474/* Group 5 opcode extensions for 0xff.  To be used with OPC_GRP5.  */
475#define EXT5_INC_Ev	0
476#define EXT5_DEC_Ev	1
477#define EXT5_CALLN_Ev	2
478#define EXT5_JMPN_Ev	4
479
480/* Condition codes to be added to OPC_JCC_{long,short}.  */
481#define JCC_JMP (-1)
482#define JCC_JO  0x0
483#define JCC_JNO 0x1
484#define JCC_JB  0x2
485#define JCC_JAE 0x3
486#define JCC_JE  0x4
487#define JCC_JNE 0x5
488#define JCC_JBE 0x6
489#define JCC_JA  0x7
490#define JCC_JS  0x8
491#define JCC_JNS 0x9
492#define JCC_JP  0xa
493#define JCC_JNP 0xb
494#define JCC_JL  0xc
495#define JCC_JGE 0xd
496#define JCC_JLE 0xe
497#define JCC_JG  0xf
498
499static const uint8_t tcg_cond_to_jcc[] = {
500    [TCG_COND_EQ] = JCC_JE,
501    [TCG_COND_NE] = JCC_JNE,
502    [TCG_COND_LT] = JCC_JL,
503    [TCG_COND_GE] = JCC_JGE,
504    [TCG_COND_LE] = JCC_JLE,
505    [TCG_COND_GT] = JCC_JG,
506    [TCG_COND_LTU] = JCC_JB,
507    [TCG_COND_GEU] = JCC_JAE,
508    [TCG_COND_LEU] = JCC_JBE,
509    [TCG_COND_GTU] = JCC_JA,
510};
511
512#if TCG_TARGET_REG_BITS == 64
513static void tcg_out_opc(TCGContext *s, int opc, int r, int rm, int x)
514{
515    int rex;
516
517    if (opc & P_GS) {
518        tcg_out8(s, 0x65);
519    }
520    if (opc & P_DATA16) {
521        /* We should never be asking for both 16 and 64-bit operation.  */
522        tcg_debug_assert((opc & P_REXW) == 0);
523        tcg_out8(s, 0x66);
524    }
525    if (opc & P_SIMDF3) {
526        tcg_out8(s, 0xf3);
527    } else if (opc & P_SIMDF2) {
528        tcg_out8(s, 0xf2);
529    }
530
531    rex = 0;
532    rex |= (opc & P_REXW) ? 0x8 : 0x0;  /* REX.W */
533    rex |= (r & 8) >> 1;                /* REX.R */
534    rex |= (x & 8) >> 2;                /* REX.X */
535    rex |= (rm & 8) >> 3;               /* REX.B */
536
537    /* P_REXB_{R,RM} indicates that the given register is the low byte.
538       For %[abcd]l we need no REX prefix, but for %{si,di,bp,sp}l we do,
539       as otherwise the encoding indicates %[abcd]h.  Note that the values
540       that are ORed in merely indicate that the REX byte must be present;
541       those bits get discarded in output.  */
542    rex |= opc & (r >= 4 ? P_REXB_R : 0);
543    rex |= opc & (rm >= 4 ? P_REXB_RM : 0);
544
545    if (rex) {
546        tcg_out8(s, (uint8_t)(rex | 0x40));
547    }
548
549    if (opc & (P_EXT | P_EXT38 | P_EXT3A)) {
550        tcg_out8(s, 0x0f);
551        if (opc & P_EXT38) {
552            tcg_out8(s, 0x38);
553        } else if (opc & P_EXT3A) {
554            tcg_out8(s, 0x3a);
555        }
556    }
557
558    tcg_out8(s, opc);
559}
560#else
561static void tcg_out_opc(TCGContext *s, int opc)
562{
563    if (opc & P_DATA16) {
564        tcg_out8(s, 0x66);
565    }
566    if (opc & P_SIMDF3) {
567        tcg_out8(s, 0xf3);
568    } else if (opc & P_SIMDF2) {
569        tcg_out8(s, 0xf2);
570    }
571    if (opc & (P_EXT | P_EXT38 | P_EXT3A)) {
572        tcg_out8(s, 0x0f);
573        if (opc & P_EXT38) {
574            tcg_out8(s, 0x38);
575        } else if (opc & P_EXT3A) {
576            tcg_out8(s, 0x3a);
577        }
578    }
579    tcg_out8(s, opc);
580}
581/* Discard the register arguments to tcg_out_opc early, so as not to penalize
582   the 32-bit compilation paths.  This method works with all versions of gcc,
583   whereas relying on optimization may not be able to exclude them.  */
584#define tcg_out_opc(s, opc, r, rm, x)  (tcg_out_opc)(s, opc)
585#endif
586
587static void tcg_out_modrm(TCGContext *s, int opc, int r, int rm)
588{
589    tcg_out_opc(s, opc, r, rm, 0);
590    tcg_out8(s, 0xc0 | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
591}
592
593static void tcg_out_vex_opc(TCGContext *s, int opc, int r, int v,
594                            int rm, int index)
595{
596    int tmp;
597
598    /* Use the two byte form if possible, which cannot encode
599       VEX.W, VEX.B, VEX.X, or an m-mmmm field other than P_EXT.  */
600    if ((opc & (P_EXT | P_EXT38 | P_EXT3A | P_VEXW)) == P_EXT
601        && ((rm | index) & 8) == 0) {
602        /* Two byte VEX prefix.  */
603        tcg_out8(s, 0xc5);
604
605        tmp = (r & 8 ? 0 : 0x80);              /* VEX.R */
606    } else {
607        /* Three byte VEX prefix.  */
608        tcg_out8(s, 0xc4);
609
610        /* VEX.m-mmmm */
611        if (opc & P_EXT3A) {
612            tmp = 3;
613        } else if (opc & P_EXT38) {
614            tmp = 2;
615        } else if (opc & P_EXT) {
616            tmp = 1;
617        } else {
618            g_assert_not_reached();
619        }
620        tmp |= (r & 8 ? 0 : 0x80);             /* VEX.R */
621        tmp |= (index & 8 ? 0 : 0x40);         /* VEX.X */
622        tmp |= (rm & 8 ? 0 : 0x20);            /* VEX.B */
623        tcg_out8(s, tmp);
624
625        tmp = (opc & P_VEXW ? 0x80 : 0);       /* VEX.W */
626    }
627
628    tmp |= (opc & P_VEXL ? 0x04 : 0);      /* VEX.L */
629    /* VEX.pp */
630    if (opc & P_DATA16) {
631        tmp |= 1;                          /* 0x66 */
632    } else if (opc & P_SIMDF3) {
633        tmp |= 2;                          /* 0xf3 */
634    } else if (opc & P_SIMDF2) {
635        tmp |= 3;                          /* 0xf2 */
636    }
637    tmp |= (~v & 15) << 3;                 /* VEX.vvvv */
638    tcg_out8(s, tmp);
639    tcg_out8(s, opc);
640}
641
642static void tcg_out_evex_opc(TCGContext *s, int opc, int r, int v,
643                             int rm, int index)
644{
645    /* The entire 4-byte evex prefix; with R' and V' set. */
646    uint32_t p = 0x08041062;
647    int mm, pp;
648
649    tcg_debug_assert(have_avx512vl);
650
651    /* EVEX.mm */
652    if (opc & P_EXT3A) {
653        mm = 3;
654    } else if (opc & P_EXT38) {
655        mm = 2;
656    } else if (opc & P_EXT) {
657        mm = 1;
658    } else {
659        g_assert_not_reached();
660    }
661
662    /* EVEX.pp */
663    if (opc & P_DATA16) {
664        pp = 1;                          /* 0x66 */
665    } else if (opc & P_SIMDF3) {
666        pp = 2;                          /* 0xf3 */
667    } else if (opc & P_SIMDF2) {
668        pp = 3;                          /* 0xf2 */
669    } else {
670        pp = 0;
671    }
672
673    p = deposit32(p, 8, 2, mm);
674    p = deposit32(p, 13, 1, (rm & 8) == 0);             /* EVEX.RXB.B */
675    p = deposit32(p, 14, 1, (index & 8) == 0);          /* EVEX.RXB.X */
676    p = deposit32(p, 15, 1, (r & 8) == 0);              /* EVEX.RXB.R */
677    p = deposit32(p, 16, 2, pp);
678    p = deposit32(p, 19, 4, ~v);
679    p = deposit32(p, 23, 1, (opc & P_VEXW) != 0);
680    p = deposit32(p, 29, 2, (opc & P_VEXL) != 0);
681
682    tcg_out32(s, p);
683    tcg_out8(s, opc);
684}
685
686static void tcg_out_vex_modrm(TCGContext *s, int opc, int r, int v, int rm)
687{
688    if (opc & P_EVEX) {
689        tcg_out_evex_opc(s, opc, r, v, rm, 0);
690    } else {
691        tcg_out_vex_opc(s, opc, r, v, rm, 0);
692    }
693    tcg_out8(s, 0xc0 | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
694}
695
696/* Output an opcode with a full "rm + (index<<shift) + offset" address mode.
697   We handle either RM and INDEX missing with a negative value.  In 64-bit
698   mode for absolute addresses, ~RM is the size of the immediate operand
699   that will follow the instruction.  */
700
701static void tcg_out_sib_offset(TCGContext *s, int r, int rm, int index,
702                               int shift, intptr_t offset)
703{
704    int mod, len;
705
706    if (index < 0 && rm < 0) {
707        if (TCG_TARGET_REG_BITS == 64) {
708            /* Try for a rip-relative addressing mode.  This has replaced
709               the 32-bit-mode absolute addressing encoding.  */
710            intptr_t pc = (intptr_t)s->code_ptr + 5 + ~rm;
711            intptr_t disp = offset - pc;
712            if (disp == (int32_t)disp) {
713                tcg_out8(s, (LOWREGMASK(r) << 3) | 5);
714                tcg_out32(s, disp);
715                return;
716            }
717
718            /* Try for an absolute address encoding.  This requires the
719               use of the MODRM+SIB encoding and is therefore larger than
720               rip-relative addressing.  */
721            if (offset == (int32_t)offset) {
722                tcg_out8(s, (LOWREGMASK(r) << 3) | 4);
723                tcg_out8(s, (4 << 3) | 5);
724                tcg_out32(s, offset);
725                return;
726            }
727
728            /* ??? The memory isn't directly addressable.  */
729            g_assert_not_reached();
730        } else {
731            /* Absolute address.  */
732            tcg_out8(s, (r << 3) | 5);
733            tcg_out32(s, offset);
734            return;
735        }
736    }
737
738    /* Find the length of the immediate addend.  Note that the encoding
739       that would be used for (%ebp) indicates absolute addressing.  */
740    if (rm < 0) {
741        mod = 0, len = 4, rm = 5;
742    } else if (offset == 0 && LOWREGMASK(rm) != TCG_REG_EBP) {
743        mod = 0, len = 0;
744    } else if (offset == (int8_t)offset) {
745        mod = 0x40, len = 1;
746    } else {
747        mod = 0x80, len = 4;
748    }
749
750    /* Use a single byte MODRM format if possible.  Note that the encoding
751       that would be used for %esp is the escape to the two byte form.  */
752    if (index < 0 && LOWREGMASK(rm) != TCG_REG_ESP) {
753        /* Single byte MODRM format.  */
754        tcg_out8(s, mod | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
755    } else {
756        /* Two byte MODRM+SIB format.  */
757
758        /* Note that the encoding that would place %esp into the index
759           field indicates no index register.  In 64-bit mode, the REX.X
760           bit counts, so %r12 can be used as the index.  */
761        if (index < 0) {
762            index = 4;
763        } else {
764            tcg_debug_assert(index != TCG_REG_ESP);
765        }
766
767        tcg_out8(s, mod | (LOWREGMASK(r) << 3) | 4);
768        tcg_out8(s, (shift << 6) | (LOWREGMASK(index) << 3) | LOWREGMASK(rm));
769    }
770
771    if (len == 1) {
772        tcg_out8(s, offset);
773    } else if (len == 4) {
774        tcg_out32(s, offset);
775    }
776}
777
778static void tcg_out_modrm_sib_offset(TCGContext *s, int opc, int r, int rm,
779                                     int index, int shift, intptr_t offset)
780{
781    tcg_out_opc(s, opc, r, rm < 0 ? 0 : rm, index < 0 ? 0 : index);
782    tcg_out_sib_offset(s, r, rm, index, shift, offset);
783}
784
785static void tcg_out_vex_modrm_sib_offset(TCGContext *s, int opc, int r, int v,
786                                         int rm, int index, int shift,
787                                         intptr_t offset)
788{
789    tcg_out_vex_opc(s, opc, r, v, rm < 0 ? 0 : rm, index < 0 ? 0 : index);
790    tcg_out_sib_offset(s, r, rm, index, shift, offset);
791}
792
793/* A simplification of the above with no index or shift.  */
794static inline void tcg_out_modrm_offset(TCGContext *s, int opc, int r,
795                                        int rm, intptr_t offset)
796{
797    tcg_out_modrm_sib_offset(s, opc, r, rm, -1, 0, offset);
798}
799
800static inline void tcg_out_vex_modrm_offset(TCGContext *s, int opc, int r,
801                                            int v, int rm, intptr_t offset)
802{
803    tcg_out_vex_modrm_sib_offset(s, opc, r, v, rm, -1, 0, offset);
804}
805
806/* Output an opcode with an expected reference to the constant pool.  */
807static inline void tcg_out_modrm_pool(TCGContext *s, int opc, int r)
808{
809    tcg_out_opc(s, opc, r, 0, 0);
810    /* Absolute for 32-bit, pc-relative for 64-bit.  */
811    tcg_out8(s, LOWREGMASK(r) << 3 | 5);
812    tcg_out32(s, 0);
813}
814
815/* Output an opcode with an expected reference to the constant pool.  */
816static inline void tcg_out_vex_modrm_pool(TCGContext *s, int opc, int r)
817{
818    tcg_out_vex_opc(s, opc, r, 0, 0, 0);
819    /* Absolute for 32-bit, pc-relative for 64-bit.  */
820    tcg_out8(s, LOWREGMASK(r) << 3 | 5);
821    tcg_out32(s, 0);
822}
823
824/* Generate dest op= src.  Uses the same ARITH_* codes as tgen_arithi.  */
825static inline void tgen_arithr(TCGContext *s, int subop, int dest, int src)
826{
827    /* Propagate an opcode prefix, such as P_REXW.  */
828    int ext = subop & ~0x7;
829    subop &= 0x7;
830
831    tcg_out_modrm(s, OPC_ARITH_GvEv + (subop << 3) + ext, dest, src);
832}
833
834static bool tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
835{
836    int rexw = 0;
837
838    if (arg == ret) {
839        return true;
840    }
841    switch (type) {
842    case TCG_TYPE_I64:
843        rexw = P_REXW;
844        /* fallthru */
845    case TCG_TYPE_I32:
846        if (ret < 16) {
847            if (arg < 16) {
848                tcg_out_modrm(s, OPC_MOVL_GvEv + rexw, ret, arg);
849            } else {
850                tcg_out_vex_modrm(s, OPC_MOVD_EyVy + rexw, arg, 0, ret);
851            }
852        } else {
853            if (arg < 16) {
854                tcg_out_vex_modrm(s, OPC_MOVD_VyEy + rexw, ret, 0, arg);
855            } else {
856                tcg_out_vex_modrm(s, OPC_MOVQ_VqWq, ret, 0, arg);
857            }
858        }
859        break;
860
861    case TCG_TYPE_V64:
862        tcg_debug_assert(ret >= 16 && arg >= 16);
863        tcg_out_vex_modrm(s, OPC_MOVQ_VqWq, ret, 0, arg);
864        break;
865    case TCG_TYPE_V128:
866        tcg_debug_assert(ret >= 16 && arg >= 16);
867        tcg_out_vex_modrm(s, OPC_MOVDQA_VxWx, ret, 0, arg);
868        break;
869    case TCG_TYPE_V256:
870        tcg_debug_assert(ret >= 16 && arg >= 16);
871        tcg_out_vex_modrm(s, OPC_MOVDQA_VxWx | P_VEXL, ret, 0, arg);
872        break;
873
874    default:
875        g_assert_not_reached();
876    }
877    return true;
878}
879
880static const int avx2_dup_insn[4] = {
881    OPC_VPBROADCASTB, OPC_VPBROADCASTW,
882    OPC_VPBROADCASTD, OPC_VPBROADCASTQ,
883};
884
885static bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece,
886                            TCGReg r, TCGReg a)
887{
888    if (have_avx2) {
889        int vex_l = (type == TCG_TYPE_V256 ? P_VEXL : 0);
890        tcg_out_vex_modrm(s, avx2_dup_insn[vece] + vex_l, r, 0, a);
891    } else {
892        switch (vece) {
893        case MO_8:
894            /* ??? With zero in a register, use PSHUFB.  */
895            tcg_out_vex_modrm(s, OPC_PUNPCKLBW, r, a, a);
896            a = r;
897            /* FALLTHRU */
898        case MO_16:
899            tcg_out_vex_modrm(s, OPC_PUNPCKLWD, r, a, a);
900            a = r;
901            /* FALLTHRU */
902        case MO_32:
903            tcg_out_vex_modrm(s, OPC_PSHUFD, r, 0, a);
904            /* imm8 operand: all output lanes selected from input lane 0.  */
905            tcg_out8(s, 0);
906            break;
907        case MO_64:
908            tcg_out_vex_modrm(s, OPC_PUNPCKLQDQ, r, a, a);
909            break;
910        default:
911            g_assert_not_reached();
912        }
913    }
914    return true;
915}
916
917static bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece,
918                             TCGReg r, TCGReg base, intptr_t offset)
919{
920    if (have_avx2) {
921        int vex_l = (type == TCG_TYPE_V256 ? P_VEXL : 0);
922        tcg_out_vex_modrm_offset(s, avx2_dup_insn[vece] + vex_l,
923                                 r, 0, base, offset);
924    } else {
925        switch (vece) {
926        case MO_64:
927            tcg_out_vex_modrm_offset(s, OPC_MOVDDUP, r, 0, base, offset);
928            break;
929        case MO_32:
930            tcg_out_vex_modrm_offset(s, OPC_VBROADCASTSS, r, 0, base, offset);
931            break;
932        case MO_16:
933            tcg_out_vex_modrm_offset(s, OPC_VPINSRW, r, r, base, offset);
934            tcg_out8(s, 0); /* imm8 */
935            tcg_out_dup_vec(s, type, vece, r, r);
936            break;
937        case MO_8:
938            tcg_out_vex_modrm_offset(s, OPC_VPINSRB, r, r, base, offset);
939            tcg_out8(s, 0); /* imm8 */
940            tcg_out_dup_vec(s, type, vece, r, r);
941            break;
942        default:
943            g_assert_not_reached();
944        }
945    }
946    return true;
947}
948
949static void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece,
950                             TCGReg ret, int64_t arg)
951{
952    int vex_l = (type == TCG_TYPE_V256 ? P_VEXL : 0);
953
954    if (arg == 0) {
955        tcg_out_vex_modrm(s, OPC_PXOR, ret, ret, ret);
956        return;
957    }
958    if (arg == -1) {
959        tcg_out_vex_modrm(s, OPC_PCMPEQB + vex_l, ret, ret, ret);
960        return;
961    }
962
963    if (TCG_TARGET_REG_BITS == 32 && vece < MO_64) {
964        if (have_avx2) {
965            tcg_out_vex_modrm_pool(s, OPC_VPBROADCASTD + vex_l, ret);
966        } else {
967            tcg_out_vex_modrm_pool(s, OPC_VBROADCASTSS, ret);
968        }
969        new_pool_label(s, arg, R_386_32, s->code_ptr - 4, 0);
970    } else {
971        if (type == TCG_TYPE_V64) {
972            tcg_out_vex_modrm_pool(s, OPC_MOVQ_VqWq, ret);
973        } else if (have_avx2) {
974            tcg_out_vex_modrm_pool(s, OPC_VPBROADCASTQ + vex_l, ret);
975        } else {
976            tcg_out_vex_modrm_pool(s, OPC_MOVDDUP, ret);
977        }
978        if (TCG_TARGET_REG_BITS == 64) {
979            new_pool_label(s, arg, R_386_PC32, s->code_ptr - 4, -4);
980        } else {
981            new_pool_l2(s, R_386_32, s->code_ptr - 4, 0, arg, arg >> 32);
982        }
983    }
984}
985
986static void tcg_out_movi_vec(TCGContext *s, TCGType type,
987                             TCGReg ret, tcg_target_long arg)
988{
989    if (arg == 0) {
990        tcg_out_vex_modrm(s, OPC_PXOR, ret, ret, ret);
991        return;
992    }
993    if (arg == -1) {
994        tcg_out_vex_modrm(s, OPC_PCMPEQB, ret, ret, ret);
995        return;
996    }
997
998    int rexw = (type == TCG_TYPE_I32 ? 0 : P_REXW);
999    tcg_out_vex_modrm_pool(s, OPC_MOVD_VyEy + rexw, ret);
1000    if (TCG_TARGET_REG_BITS == 64) {
1001        new_pool_label(s, arg, R_386_PC32, s->code_ptr - 4, -4);
1002    } else {
1003        new_pool_label(s, arg, R_386_32, s->code_ptr - 4, 0);
1004    }
1005}
1006
1007static void tcg_out_movi_int(TCGContext *s, TCGType type,
1008                             TCGReg ret, tcg_target_long arg)
1009{
1010    tcg_target_long diff;
1011
1012    if (arg == 0) {
1013        tgen_arithr(s, ARITH_XOR, ret, ret);
1014        return;
1015    }
1016    if (arg == (uint32_t)arg || type == TCG_TYPE_I32) {
1017        tcg_out_opc(s, OPC_MOVL_Iv + LOWREGMASK(ret), 0, ret, 0);
1018        tcg_out32(s, arg);
1019        return;
1020    }
1021    if (arg == (int32_t)arg) {
1022        tcg_out_modrm(s, OPC_MOVL_EvIz + P_REXW, 0, ret);
1023        tcg_out32(s, arg);
1024        return;
1025    }
1026
1027    /* Try a 7 byte pc-relative lea before the 10 byte movq.  */
1028    diff = tcg_pcrel_diff(s, (const void *)arg) - 7;
1029    if (diff == (int32_t)diff) {
1030        tcg_out_opc(s, OPC_LEA | P_REXW, ret, 0, 0);
1031        tcg_out8(s, (LOWREGMASK(ret) << 3) | 5);
1032        tcg_out32(s, diff);
1033        return;
1034    }
1035
1036    tcg_out_opc(s, OPC_MOVL_Iv + P_REXW + LOWREGMASK(ret), 0, ret, 0);
1037    tcg_out64(s, arg);
1038}
1039
1040static void tcg_out_movi(TCGContext *s, TCGType type,
1041                         TCGReg ret, tcg_target_long arg)
1042{
1043    switch (type) {
1044    case TCG_TYPE_I32:
1045#if TCG_TARGET_REG_BITS == 64
1046    case TCG_TYPE_I64:
1047#endif
1048        if (ret < 16) {
1049            tcg_out_movi_int(s, type, ret, arg);
1050        } else {
1051            tcg_out_movi_vec(s, type, ret, arg);
1052        }
1053        break;
1054    default:
1055        g_assert_not_reached();
1056    }
1057}
1058
1059static bool tcg_out_xchg(TCGContext *s, TCGType type, TCGReg r1, TCGReg r2)
1060{
1061    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1062    tcg_out_modrm(s, OPC_XCHG_EvGv + rexw, r1, r2);
1063    return true;
1064}
1065
1066static void tcg_out_addi_ptr(TCGContext *s, TCGReg rd, TCGReg rs,
1067                             tcg_target_long imm)
1068{
1069    /* This function is only used for passing structs by reference. */
1070    tcg_debug_assert(imm == (int32_t)imm);
1071    tcg_out_modrm_offset(s, OPC_LEA | P_REXW, rd, rs, imm);
1072}
1073
1074static inline void tcg_out_pushi(TCGContext *s, tcg_target_long val)
1075{
1076    if (val == (int8_t)val) {
1077        tcg_out_opc(s, OPC_PUSH_Ib, 0, 0, 0);
1078        tcg_out8(s, val);
1079    } else if (val == (int32_t)val) {
1080        tcg_out_opc(s, OPC_PUSH_Iv, 0, 0, 0);
1081        tcg_out32(s, val);
1082    } else {
1083        g_assert_not_reached();
1084    }
1085}
1086
1087static inline void tcg_out_mb(TCGContext *s, TCGArg a0)
1088{
1089    /* Given the strength of x86 memory ordering, we only need care for
1090       store-load ordering.  Experimentally, "lock orl $0,0(%esp)" is
1091       faster than "mfence", so don't bother with the sse insn.  */
1092    if (a0 & TCG_MO_ST_LD) {
1093        tcg_out8(s, 0xf0);
1094        tcg_out_modrm_offset(s, OPC_ARITH_EvIb, ARITH_OR, TCG_REG_ESP, 0);
1095        tcg_out8(s, 0);
1096    }
1097}
1098
1099static inline void tcg_out_push(TCGContext *s, int reg)
1100{
1101    tcg_out_opc(s, OPC_PUSH_r32 + LOWREGMASK(reg), 0, reg, 0);
1102}
1103
1104static inline void tcg_out_pop(TCGContext *s, int reg)
1105{
1106    tcg_out_opc(s, OPC_POP_r32 + LOWREGMASK(reg), 0, reg, 0);
1107}
1108
1109static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg ret,
1110                       TCGReg arg1, intptr_t arg2)
1111{
1112    switch (type) {
1113    case TCG_TYPE_I32:
1114        if (ret < 16) {
1115            tcg_out_modrm_offset(s, OPC_MOVL_GvEv, ret, arg1, arg2);
1116        } else {
1117            tcg_out_vex_modrm_offset(s, OPC_MOVD_VyEy, ret, 0, arg1, arg2);
1118        }
1119        break;
1120    case TCG_TYPE_I64:
1121        if (ret < 16) {
1122            tcg_out_modrm_offset(s, OPC_MOVL_GvEv | P_REXW, ret, arg1, arg2);
1123            break;
1124        }
1125        /* FALLTHRU */
1126    case TCG_TYPE_V64:
1127        /* There is no instruction that can validate 8-byte alignment.  */
1128        tcg_debug_assert(ret >= 16);
1129        tcg_out_vex_modrm_offset(s, OPC_MOVQ_VqWq, ret, 0, arg1, arg2);
1130        break;
1131    case TCG_TYPE_V128:
1132        /*
1133         * The gvec infrastructure is asserts that v128 vector loads
1134         * and stores use a 16-byte aligned offset.  Validate that the
1135         * final pointer is aligned by using an insn that will SIGSEGV.
1136         */
1137        tcg_debug_assert(ret >= 16);
1138        tcg_out_vex_modrm_offset(s, OPC_MOVDQA_VxWx, ret, 0, arg1, arg2);
1139        break;
1140    case TCG_TYPE_V256:
1141        /*
1142         * The gvec infrastructure only requires 16-byte alignment,
1143         * so here we must use an unaligned load.
1144         */
1145        tcg_debug_assert(ret >= 16);
1146        tcg_out_vex_modrm_offset(s, OPC_MOVDQU_VxWx | P_VEXL,
1147                                 ret, 0, arg1, arg2);
1148        break;
1149    default:
1150        g_assert_not_reached();
1151    }
1152}
1153
1154static void tcg_out_st(TCGContext *s, TCGType type, TCGReg arg,
1155                       TCGReg arg1, intptr_t arg2)
1156{
1157    switch (type) {
1158    case TCG_TYPE_I32:
1159        if (arg < 16) {
1160            tcg_out_modrm_offset(s, OPC_MOVL_EvGv, arg, arg1, arg2);
1161        } else {
1162            tcg_out_vex_modrm_offset(s, OPC_MOVD_EyVy, arg, 0, arg1, arg2);
1163        }
1164        break;
1165    case TCG_TYPE_I64:
1166        if (arg < 16) {
1167            tcg_out_modrm_offset(s, OPC_MOVL_EvGv | P_REXW, arg, arg1, arg2);
1168            break;
1169        }
1170        /* FALLTHRU */
1171    case TCG_TYPE_V64:
1172        /* There is no instruction that can validate 8-byte alignment.  */
1173        tcg_debug_assert(arg >= 16);
1174        tcg_out_vex_modrm_offset(s, OPC_MOVQ_WqVq, arg, 0, arg1, arg2);
1175        break;
1176    case TCG_TYPE_V128:
1177        /*
1178         * The gvec infrastructure is asserts that v128 vector loads
1179         * and stores use a 16-byte aligned offset.  Validate that the
1180         * final pointer is aligned by using an insn that will SIGSEGV.
1181         *
1182         * This specific instance is also used by TCG_CALL_RET_BY_VEC,
1183         * for _WIN64, which must have SSE2 but may not have AVX.
1184         */
1185        tcg_debug_assert(arg >= 16);
1186        if (have_avx1) {
1187            tcg_out_vex_modrm_offset(s, OPC_MOVDQA_WxVx, arg, 0, arg1, arg2);
1188        } else {
1189            tcg_out_modrm_offset(s, OPC_MOVDQA_WxVx, arg, arg1, arg2);
1190        }
1191        break;
1192    case TCG_TYPE_V256:
1193        /*
1194         * The gvec infrastructure only requires 16-byte alignment,
1195         * so here we must use an unaligned store.
1196         */
1197        tcg_debug_assert(arg >= 16);
1198        tcg_out_vex_modrm_offset(s, OPC_MOVDQU_WxVx | P_VEXL,
1199                                 arg, 0, arg1, arg2);
1200        break;
1201    default:
1202        g_assert_not_reached();
1203    }
1204}
1205
1206static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val,
1207                        TCGReg base, intptr_t ofs)
1208{
1209    int rexw = 0;
1210    if (TCG_TARGET_REG_BITS == 64 && type == TCG_TYPE_I64) {
1211        if (val != (int32_t)val) {
1212            return false;
1213        }
1214        rexw = P_REXW;
1215    } else if (type != TCG_TYPE_I32) {
1216        return false;
1217    }
1218    tcg_out_modrm_offset(s, OPC_MOVL_EvIz | rexw, 0, base, ofs);
1219    tcg_out32(s, val);
1220    return true;
1221}
1222
1223static void tcg_out_shifti(TCGContext *s, int subopc, int reg, int count)
1224{
1225    /* Propagate an opcode prefix, such as P_DATA16.  */
1226    int ext = subopc & ~0x7;
1227    subopc &= 0x7;
1228
1229    if (count == 1) {
1230        tcg_out_modrm(s, OPC_SHIFT_1 + ext, subopc, reg);
1231    } else {
1232        tcg_out_modrm(s, OPC_SHIFT_Ib + ext, subopc, reg);
1233        tcg_out8(s, count);
1234    }
1235}
1236
1237static inline void tcg_out_bswap32(TCGContext *s, int reg)
1238{
1239    tcg_out_opc(s, OPC_BSWAP + LOWREGMASK(reg), 0, reg, 0);
1240}
1241
1242static inline void tcg_out_rolw_8(TCGContext *s, int reg)
1243{
1244    tcg_out_shifti(s, SHIFT_ROL + P_DATA16, reg, 8);
1245}
1246
1247static void tcg_out_ext8u(TCGContext *s, TCGReg dest, TCGReg src)
1248{
1249    /* movzbl */
1250    tcg_debug_assert(src < 4 || TCG_TARGET_REG_BITS == 64);
1251    tcg_out_modrm(s, OPC_MOVZBL + P_REXB_RM, dest, src);
1252}
1253
1254static void tcg_out_ext8s(TCGContext *s, TCGType type, TCGReg dest, TCGReg src)
1255{
1256    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1257    /* movsbl */
1258    tcg_debug_assert(src < 4 || TCG_TARGET_REG_BITS == 64);
1259    tcg_out_modrm(s, OPC_MOVSBL + P_REXB_RM + rexw, dest, src);
1260}
1261
1262static void tcg_out_ext16u(TCGContext *s, TCGReg dest, TCGReg src)
1263{
1264    /* movzwl */
1265    tcg_out_modrm(s, OPC_MOVZWL, dest, src);
1266}
1267
1268static void tcg_out_ext16s(TCGContext *s, TCGType type, TCGReg dest, TCGReg src)
1269{
1270    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1271    /* movsw[lq] */
1272    tcg_out_modrm(s, OPC_MOVSWL + rexw, dest, src);
1273}
1274
1275static void tcg_out_ext32u(TCGContext *s, TCGReg dest, TCGReg src)
1276{
1277    /* 32-bit mov zero extends.  */
1278    tcg_out_modrm(s, OPC_MOVL_GvEv, dest, src);
1279}
1280
1281static void tcg_out_ext32s(TCGContext *s, TCGReg dest, TCGReg src)
1282{
1283    tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
1284    tcg_out_modrm(s, OPC_MOVSLQ, dest, src);
1285}
1286
1287static void tcg_out_exts_i32_i64(TCGContext *s, TCGReg dest, TCGReg src)
1288{
1289    tcg_out_ext32s(s, dest, src);
1290}
1291
1292static void tcg_out_extu_i32_i64(TCGContext *s, TCGReg dest, TCGReg src)
1293{
1294    if (dest != src) {
1295        tcg_out_ext32u(s, dest, src);
1296    }
1297}
1298
1299static void tcg_out_extrl_i64_i32(TCGContext *s, TCGReg dest, TCGReg src)
1300{
1301    tcg_out_ext32u(s, dest, src);
1302}
1303
1304static inline void tcg_out_bswap64(TCGContext *s, int reg)
1305{
1306    tcg_out_opc(s, OPC_BSWAP + P_REXW + LOWREGMASK(reg), 0, reg, 0);
1307}
1308
1309static void tgen_arithi(TCGContext *s, int c, int r0,
1310                        tcg_target_long val, int cf)
1311{
1312    int rexw = 0;
1313
1314    if (TCG_TARGET_REG_BITS == 64) {
1315        rexw = c & -8;
1316        c &= 7;
1317    }
1318
1319    /* ??? While INC is 2 bytes shorter than ADDL $1, they also induce
1320       partial flags update stalls on Pentium4 and are not recommended
1321       by current Intel optimization manuals.  */
1322    if (!cf && (c == ARITH_ADD || c == ARITH_SUB) && (val == 1 || val == -1)) {
1323        int is_inc = (c == ARITH_ADD) ^ (val < 0);
1324        if (TCG_TARGET_REG_BITS == 64) {
1325            /* The single-byte increment encodings are re-tasked as the
1326               REX prefixes.  Use the MODRM encoding.  */
1327            tcg_out_modrm(s, OPC_GRP5 + rexw,
1328                          (is_inc ? EXT5_INC_Ev : EXT5_DEC_Ev), r0);
1329        } else {
1330            tcg_out8(s, (is_inc ? OPC_INC_r32 : OPC_DEC_r32) + r0);
1331        }
1332        return;
1333    }
1334
1335    if (c == ARITH_AND) {
1336        if (TCG_TARGET_REG_BITS == 64) {
1337            if (val == 0xffffffffu) {
1338                tcg_out_ext32u(s, r0, r0);
1339                return;
1340            }
1341            if (val == (uint32_t)val) {
1342                /* AND with no high bits set can use a 32-bit operation.  */
1343                rexw = 0;
1344            }
1345        }
1346        if (val == 0xffu && (r0 < 4 || TCG_TARGET_REG_BITS == 64)) {
1347            tcg_out_ext8u(s, r0, r0);
1348            return;
1349        }
1350        if (val == 0xffffu) {
1351            tcg_out_ext16u(s, r0, r0);
1352            return;
1353        }
1354    }
1355
1356    if (val == (int8_t)val) {
1357        tcg_out_modrm(s, OPC_ARITH_EvIb + rexw, c, r0);
1358        tcg_out8(s, val);
1359        return;
1360    }
1361    if (rexw == 0 || val == (int32_t)val) {
1362        tcg_out_modrm(s, OPC_ARITH_EvIz + rexw, c, r0);
1363        tcg_out32(s, val);
1364        return;
1365    }
1366
1367    g_assert_not_reached();
1368}
1369
1370static void tcg_out_addi(TCGContext *s, int reg, tcg_target_long val)
1371{
1372    if (val != 0) {
1373        tgen_arithi(s, ARITH_ADD + P_REXW, reg, val, 0);
1374    }
1375}
1376
1377/* Set SMALL to force a short forward branch.  */
1378static void tcg_out_jxx(TCGContext *s, int opc, TCGLabel *l, bool small)
1379{
1380    int32_t val, val1;
1381
1382    if (l->has_value) {
1383        val = tcg_pcrel_diff(s, l->u.value_ptr);
1384        val1 = val - 2;
1385        if ((int8_t)val1 == val1) {
1386            if (opc == -1) {
1387                tcg_out8(s, OPC_JMP_short);
1388            } else {
1389                tcg_out8(s, OPC_JCC_short + opc);
1390            }
1391            tcg_out8(s, val1);
1392        } else {
1393            tcg_debug_assert(!small);
1394            if (opc == -1) {
1395                tcg_out8(s, OPC_JMP_long);
1396                tcg_out32(s, val - 5);
1397            } else {
1398                tcg_out_opc(s, OPC_JCC_long + opc, 0, 0, 0);
1399                tcg_out32(s, val - 6);
1400            }
1401        }
1402    } else if (small) {
1403        if (opc == -1) {
1404            tcg_out8(s, OPC_JMP_short);
1405        } else {
1406            tcg_out8(s, OPC_JCC_short + opc);
1407        }
1408        tcg_out_reloc(s, s->code_ptr, R_386_PC8, l, -1);
1409        s->code_ptr += 1;
1410    } else {
1411        if (opc == -1) {
1412            tcg_out8(s, OPC_JMP_long);
1413        } else {
1414            tcg_out_opc(s, OPC_JCC_long + opc, 0, 0, 0);
1415        }
1416        tcg_out_reloc(s, s->code_ptr, R_386_PC32, l, -4);
1417        s->code_ptr += 4;
1418    }
1419}
1420
1421static void tcg_out_cmp(TCGContext *s, TCGArg arg1, TCGArg arg2,
1422                        int const_arg2, int rexw)
1423{
1424    if (const_arg2) {
1425        if (arg2 == 0) {
1426            /* test r, r */
1427            tcg_out_modrm(s, OPC_TESTL + rexw, arg1, arg1);
1428        } else {
1429            tgen_arithi(s, ARITH_CMP + rexw, arg1, arg2, 0);
1430        }
1431    } else {
1432        tgen_arithr(s, ARITH_CMP + rexw, arg1, arg2);
1433    }
1434}
1435
1436static void tcg_out_brcond32(TCGContext *s, TCGCond cond,
1437                             TCGArg arg1, TCGArg arg2, int const_arg2,
1438                             TCGLabel *label, int small)
1439{
1440    tcg_out_cmp(s, arg1, arg2, const_arg2, 0);
1441    tcg_out_jxx(s, tcg_cond_to_jcc[cond], label, small);
1442}
1443
1444#if TCG_TARGET_REG_BITS == 64
1445static void tcg_out_brcond64(TCGContext *s, TCGCond cond,
1446                             TCGArg arg1, TCGArg arg2, int const_arg2,
1447                             TCGLabel *label, int small)
1448{
1449    tcg_out_cmp(s, arg1, arg2, const_arg2, P_REXW);
1450    tcg_out_jxx(s, tcg_cond_to_jcc[cond], label, small);
1451}
1452#else
1453/* XXX: we implement it at the target level to avoid having to
1454   handle cross basic blocks temporaries */
1455static void tcg_out_brcond2(TCGContext *s, const TCGArg *args,
1456                            const int *const_args, int small)
1457{
1458    TCGLabel *label_next = gen_new_label();
1459    TCGLabel *label_this = arg_label(args[5]);
1460
1461    switch(args[4]) {
1462    case TCG_COND_EQ:
1463        tcg_out_brcond32(s, TCG_COND_NE, args[0], args[2], const_args[2],
1464                         label_next, 1);
1465        tcg_out_brcond32(s, TCG_COND_EQ, args[1], args[3], const_args[3],
1466                         label_this, small);
1467        break;
1468    case TCG_COND_NE:
1469        tcg_out_brcond32(s, TCG_COND_NE, args[0], args[2], const_args[2],
1470                         label_this, small);
1471        tcg_out_brcond32(s, TCG_COND_NE, args[1], args[3], const_args[3],
1472                         label_this, small);
1473        break;
1474    case TCG_COND_LT:
1475        tcg_out_brcond32(s, TCG_COND_LT, args[1], args[3], const_args[3],
1476                         label_this, small);
1477        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1478        tcg_out_brcond32(s, TCG_COND_LTU, args[0], args[2], const_args[2],
1479                         label_this, small);
1480        break;
1481    case TCG_COND_LE:
1482        tcg_out_brcond32(s, TCG_COND_LT, args[1], args[3], const_args[3],
1483                         label_this, small);
1484        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1485        tcg_out_brcond32(s, TCG_COND_LEU, args[0], args[2], const_args[2],
1486                         label_this, small);
1487        break;
1488    case TCG_COND_GT:
1489        tcg_out_brcond32(s, TCG_COND_GT, args[1], args[3], const_args[3],
1490                         label_this, small);
1491        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1492        tcg_out_brcond32(s, TCG_COND_GTU, args[0], args[2], const_args[2],
1493                         label_this, small);
1494        break;
1495    case TCG_COND_GE:
1496        tcg_out_brcond32(s, TCG_COND_GT, args[1], args[3], const_args[3],
1497                         label_this, small);
1498        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1499        tcg_out_brcond32(s, TCG_COND_GEU, args[0], args[2], const_args[2],
1500                         label_this, small);
1501        break;
1502    case TCG_COND_LTU:
1503        tcg_out_brcond32(s, TCG_COND_LTU, args[1], args[3], const_args[3],
1504                         label_this, small);
1505        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1506        tcg_out_brcond32(s, TCG_COND_LTU, args[0], args[2], const_args[2],
1507                         label_this, small);
1508        break;
1509    case TCG_COND_LEU:
1510        tcg_out_brcond32(s, TCG_COND_LTU, args[1], args[3], const_args[3],
1511                         label_this, small);
1512        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1513        tcg_out_brcond32(s, TCG_COND_LEU, args[0], args[2], const_args[2],
1514                         label_this, small);
1515        break;
1516    case TCG_COND_GTU:
1517        tcg_out_brcond32(s, TCG_COND_GTU, args[1], args[3], const_args[3],
1518                         label_this, small);
1519        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1520        tcg_out_brcond32(s, TCG_COND_GTU, args[0], args[2], const_args[2],
1521                         label_this, small);
1522        break;
1523    case TCG_COND_GEU:
1524        tcg_out_brcond32(s, TCG_COND_GTU, args[1], args[3], const_args[3],
1525                         label_this, small);
1526        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1527        tcg_out_brcond32(s, TCG_COND_GEU, args[0], args[2], const_args[2],
1528                         label_this, small);
1529        break;
1530    default:
1531        g_assert_not_reached();
1532    }
1533    tcg_out_label(s, label_next);
1534}
1535#endif
1536
1537static void tcg_out_setcond32(TCGContext *s, TCGCond cond, TCGArg dest,
1538                              TCGArg arg1, TCGArg arg2, int const_arg2)
1539{
1540    tcg_out_cmp(s, arg1, arg2, const_arg2, 0);
1541    tcg_out_modrm(s, OPC_SETCC | tcg_cond_to_jcc[cond], 0, dest);
1542    tcg_out_ext8u(s, dest, dest);
1543}
1544
1545#if TCG_TARGET_REG_BITS == 64
1546static void tcg_out_setcond64(TCGContext *s, TCGCond cond, TCGArg dest,
1547                              TCGArg arg1, TCGArg arg2, int const_arg2)
1548{
1549    tcg_out_cmp(s, arg1, arg2, const_arg2, P_REXW);
1550    tcg_out_modrm(s, OPC_SETCC | tcg_cond_to_jcc[cond], 0, dest);
1551    tcg_out_ext8u(s, dest, dest);
1552}
1553#else
1554static void tcg_out_setcond2(TCGContext *s, const TCGArg *args,
1555                             const int *const_args)
1556{
1557    TCGArg new_args[6];
1558    TCGLabel *label_true, *label_over;
1559
1560    memcpy(new_args, args+1, 5*sizeof(TCGArg));
1561
1562    if (args[0] == args[1] || args[0] == args[2]
1563        || (!const_args[3] && args[0] == args[3])
1564        || (!const_args[4] && args[0] == args[4])) {
1565        /* When the destination overlaps with one of the argument
1566           registers, don't do anything tricky.  */
1567        label_true = gen_new_label();
1568        label_over = gen_new_label();
1569
1570        new_args[5] = label_arg(label_true);
1571        tcg_out_brcond2(s, new_args, const_args+1, 1);
1572
1573        tcg_out_movi(s, TCG_TYPE_I32, args[0], 0);
1574        tcg_out_jxx(s, JCC_JMP, label_over, 1);
1575        tcg_out_label(s, label_true);
1576
1577        tcg_out_movi(s, TCG_TYPE_I32, args[0], 1);
1578        tcg_out_label(s, label_over);
1579    } else {
1580        /* When the destination does not overlap one of the arguments,
1581           clear the destination first, jump if cond false, and emit an
1582           increment in the true case.  This results in smaller code.  */
1583
1584        tcg_out_movi(s, TCG_TYPE_I32, args[0], 0);
1585
1586        label_over = gen_new_label();
1587        new_args[4] = tcg_invert_cond(new_args[4]);
1588        new_args[5] = label_arg(label_over);
1589        tcg_out_brcond2(s, new_args, const_args+1, 1);
1590
1591        tgen_arithi(s, ARITH_ADD, args[0], 1, 0);
1592        tcg_out_label(s, label_over);
1593    }
1594}
1595#endif
1596
1597static void tcg_out_cmov(TCGContext *s, TCGCond cond, int rexw,
1598                         TCGReg dest, TCGReg v1)
1599{
1600    if (have_cmov) {
1601        tcg_out_modrm(s, OPC_CMOVCC | tcg_cond_to_jcc[cond] | rexw, dest, v1);
1602    } else {
1603        TCGLabel *over = gen_new_label();
1604        tcg_out_jxx(s, tcg_cond_to_jcc[tcg_invert_cond(cond)], over, 1);
1605        tcg_out_mov(s, TCG_TYPE_I32, dest, v1);
1606        tcg_out_label(s, over);
1607    }
1608}
1609
1610static void tcg_out_movcond32(TCGContext *s, TCGCond cond, TCGReg dest,
1611                              TCGReg c1, TCGArg c2, int const_c2,
1612                              TCGReg v1)
1613{
1614    tcg_out_cmp(s, c1, c2, const_c2, 0);
1615    tcg_out_cmov(s, cond, 0, dest, v1);
1616}
1617
1618#if TCG_TARGET_REG_BITS == 64
1619static void tcg_out_movcond64(TCGContext *s, TCGCond cond, TCGReg dest,
1620                              TCGReg c1, TCGArg c2, int const_c2,
1621                              TCGReg v1)
1622{
1623    tcg_out_cmp(s, c1, c2, const_c2, P_REXW);
1624    tcg_out_cmov(s, cond, P_REXW, dest, v1);
1625}
1626#endif
1627
1628static void tcg_out_ctz(TCGContext *s, int rexw, TCGReg dest, TCGReg arg1,
1629                        TCGArg arg2, bool const_a2)
1630{
1631    if (have_bmi1) {
1632        tcg_out_modrm(s, OPC_TZCNT + rexw, dest, arg1);
1633        if (const_a2) {
1634            tcg_debug_assert(arg2 == (rexw ? 64 : 32));
1635        } else {
1636            tcg_debug_assert(dest != arg2);
1637            tcg_out_cmov(s, TCG_COND_LTU, rexw, dest, arg2);
1638        }
1639    } else {
1640        tcg_debug_assert(dest != arg2);
1641        tcg_out_modrm(s, OPC_BSF + rexw, dest, arg1);
1642        tcg_out_cmov(s, TCG_COND_EQ, rexw, dest, arg2);
1643    }
1644}
1645
1646static void tcg_out_clz(TCGContext *s, int rexw, TCGReg dest, TCGReg arg1,
1647                        TCGArg arg2, bool const_a2)
1648{
1649    if (have_lzcnt) {
1650        tcg_out_modrm(s, OPC_LZCNT + rexw, dest, arg1);
1651        if (const_a2) {
1652            tcg_debug_assert(arg2 == (rexw ? 64 : 32));
1653        } else {
1654            tcg_debug_assert(dest != arg2);
1655            tcg_out_cmov(s, TCG_COND_LTU, rexw, dest, arg2);
1656        }
1657    } else {
1658        tcg_debug_assert(!const_a2);
1659        tcg_debug_assert(dest != arg1);
1660        tcg_debug_assert(dest != arg2);
1661
1662        /* Recall that the output of BSR is the index not the count.  */
1663        tcg_out_modrm(s, OPC_BSR + rexw, dest, arg1);
1664        tgen_arithi(s, ARITH_XOR + rexw, dest, rexw ? 63 : 31, 0);
1665
1666        /* Since we have destroyed the flags from BSR, we have to re-test.  */
1667        tcg_out_cmp(s, arg1, 0, 1, rexw);
1668        tcg_out_cmov(s, TCG_COND_EQ, rexw, dest, arg2);
1669    }
1670}
1671
1672static void tcg_out_branch(TCGContext *s, int call, const tcg_insn_unit *dest)
1673{
1674    intptr_t disp = tcg_pcrel_diff(s, dest) - 5;
1675
1676    if (disp == (int32_t)disp) {
1677        tcg_out_opc(s, call ? OPC_CALL_Jz : OPC_JMP_long, 0, 0, 0);
1678        tcg_out32(s, disp);
1679    } else {
1680        /* rip-relative addressing into the constant pool.
1681           This is 6 + 8 = 14 bytes, as compared to using an
1682           immediate load 10 + 6 = 16 bytes, plus we may
1683           be able to re-use the pool constant for more calls.  */
1684        tcg_out_opc(s, OPC_GRP5, 0, 0, 0);
1685        tcg_out8(s, (call ? EXT5_CALLN_Ev : EXT5_JMPN_Ev) << 3 | 5);
1686        new_pool_label(s, (uintptr_t)dest, R_386_PC32, s->code_ptr, -4);
1687        tcg_out32(s, 0);
1688    }
1689}
1690
1691static void tcg_out_call(TCGContext *s, const tcg_insn_unit *dest,
1692                         const TCGHelperInfo *info)
1693{
1694    tcg_out_branch(s, 1, dest);
1695
1696#ifndef _WIN32
1697    if (TCG_TARGET_REG_BITS == 32 && info->out_kind == TCG_CALL_RET_BY_REF) {
1698        /*
1699         * The sysv i386 abi for struct return places a reference as the
1700         * first argument of the stack, and pops that argument with the
1701         * return statement.  Since we want to retain the aligned stack
1702         * pointer for the callee, we do not want to actually push that
1703         * argument before the call but rely on the normal store to the
1704         * stack slot.  But we do need to compensate for the pop in order
1705         * to reset our correct stack pointer value.
1706         * Pushing a garbage value back onto the stack is quickest.
1707         */
1708        tcg_out_push(s, TCG_REG_EAX);
1709    }
1710#endif
1711}
1712
1713static void tcg_out_jmp(TCGContext *s, const tcg_insn_unit *dest)
1714{
1715    tcg_out_branch(s, 0, dest);
1716}
1717
1718static void tcg_out_nopn(TCGContext *s, int n)
1719{
1720    int i;
1721    /* Emit 1 or 2 operand size prefixes for the standard one byte nop,
1722     * "xchg %eax,%eax", forming "xchg %ax,%ax". All cores accept the
1723     * duplicate prefix, and all of the interesting recent cores can
1724     * decode and discard the duplicates in a single cycle.
1725     */
1726    tcg_debug_assert(n >= 1);
1727    for (i = 1; i < n; ++i) {
1728        tcg_out8(s, 0x66);
1729    }
1730    tcg_out8(s, 0x90);
1731}
1732
1733/* Test register R vs immediate bits I, setting Z flag for EQ/NE. */
1734static void __attribute__((unused))
1735tcg_out_testi(TCGContext *s, TCGReg r, uint32_t i)
1736{
1737    /*
1738     * This is used for testing alignment, so we can usually use testb.
1739     * For i686, we have to use testl for %esi/%edi.
1740     */
1741    if (i <= 0xff && (TCG_TARGET_REG_BITS == 64 || r < 4)) {
1742        tcg_out_modrm(s, OPC_GRP3_Eb | P_REXB_RM, EXT3_TESTi, r);
1743        tcg_out8(s, i);
1744    } else {
1745        tcg_out_modrm(s, OPC_GRP3_Ev, EXT3_TESTi, r);
1746        tcg_out32(s, i);
1747    }
1748}
1749
1750typedef struct {
1751    TCGReg base;
1752    int index;
1753    int ofs;
1754    int seg;
1755    TCGAtomAlign aa;
1756} HostAddress;
1757
1758bool tcg_target_has_memory_bswap(MemOp memop)
1759{
1760    TCGAtomAlign aa;
1761
1762    if (!have_movbe) {
1763        return false;
1764    }
1765    if ((memop & MO_SIZE) < MO_128) {
1766        return true;
1767    }
1768
1769    /*
1770     * Reject 16-byte memop with 16-byte atomicity, i.e. VMOVDQA,
1771     * but do allow a pair of 64-bit operations, i.e. MOVBEQ.
1772     */
1773    aa = atom_and_align_for_opc(tcg_ctx, memop, MO_ATOM_IFALIGN, true);
1774    return aa.atom < MO_128;
1775}
1776
1777/*
1778 * Because i686 has no register parameters and because x86_64 has xchg
1779 * to handle addr/data register overlap, we have placed all input arguments
1780 * before we need might need a scratch reg.
1781 *
1782 * Even then, a scratch is only needed for l->raddr.  Rather than expose
1783 * a general-purpose scratch when we don't actually know it's available,
1784 * use the ra_gen hook to load into RAX if needed.
1785 */
1786#if TCG_TARGET_REG_BITS == 64
1787static TCGReg ldst_ra_gen(TCGContext *s, const TCGLabelQemuLdst *l, int arg)
1788{
1789    if (arg < 0) {
1790        arg = TCG_REG_RAX;
1791    }
1792    tcg_out_movi(s, TCG_TYPE_PTR, arg, (uintptr_t)l->raddr);
1793    return arg;
1794}
1795static const TCGLdstHelperParam ldst_helper_param = {
1796    .ra_gen = ldst_ra_gen
1797};
1798#else
1799static const TCGLdstHelperParam ldst_helper_param = { };
1800#endif
1801
1802static void tcg_out_vec_to_pair(TCGContext *s, TCGType type,
1803                                TCGReg l, TCGReg h, TCGReg v)
1804{
1805    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1806
1807    /* vpmov{d,q} %v, %l */
1808    tcg_out_vex_modrm(s, OPC_MOVD_EyVy + rexw, v, 0, l);
1809    /* vpextr{d,q} $1, %v, %h */
1810    tcg_out_vex_modrm(s, OPC_PEXTRD + rexw, v, 0, h);
1811    tcg_out8(s, 1);
1812}
1813
1814static void tcg_out_pair_to_vec(TCGContext *s, TCGType type,
1815                                TCGReg v, TCGReg l, TCGReg h)
1816{
1817    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1818
1819    /* vmov{d,q} %l, %v */
1820    tcg_out_vex_modrm(s, OPC_MOVD_VyEy + rexw, v, 0, l);
1821    /* vpinsr{d,q} $1, %h, %v, %v */
1822    tcg_out_vex_modrm(s, OPC_PINSRD + rexw, v, v, h);
1823    tcg_out8(s, 1);
1824}
1825
1826/*
1827 * Generate code for the slow path for a load at the end of block
1828 */
1829static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
1830{
1831    MemOp opc = get_memop(l->oi);
1832    tcg_insn_unit **label_ptr = &l->label_ptr[0];
1833
1834    /* resolve label address */
1835    tcg_patch32(label_ptr[0], s->code_ptr - label_ptr[0] - 4);
1836    if (label_ptr[1]) {
1837        tcg_patch32(label_ptr[1], s->code_ptr - label_ptr[1] - 4);
1838    }
1839
1840    tcg_out_ld_helper_args(s, l, &ldst_helper_param);
1841    tcg_out_branch(s, 1, qemu_ld_helpers[opc & MO_SIZE]);
1842    tcg_out_ld_helper_ret(s, l, false, &ldst_helper_param);
1843
1844    tcg_out_jmp(s, l->raddr);
1845    return true;
1846}
1847
1848/*
1849 * Generate code for the slow path for a store at the end of block
1850 */
1851static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
1852{
1853    MemOp opc = get_memop(l->oi);
1854    tcg_insn_unit **label_ptr = &l->label_ptr[0];
1855
1856    /* resolve label address */
1857    tcg_patch32(label_ptr[0], s->code_ptr - label_ptr[0] - 4);
1858    if (label_ptr[1]) {
1859        tcg_patch32(label_ptr[1], s->code_ptr - label_ptr[1] - 4);
1860    }
1861
1862    tcg_out_st_helper_args(s, l, &ldst_helper_param);
1863    tcg_out_branch(s, 1, qemu_st_helpers[opc & MO_SIZE]);
1864
1865    tcg_out_jmp(s, l->raddr);
1866    return true;
1867}
1868
1869#ifndef CONFIG_SOFTMMU
1870static HostAddress x86_guest_base = {
1871    .index = -1
1872};
1873
1874#if defined(__x86_64__) && defined(__linux__)
1875# include <asm/prctl.h>
1876# include <sys/prctl.h>
1877int arch_prctl(int code, unsigned long addr);
1878static inline int setup_guest_base_seg(void)
1879{
1880    if (arch_prctl(ARCH_SET_GS, guest_base) == 0) {
1881        return P_GS;
1882    }
1883    return 0;
1884}
1885#elif defined(__x86_64__) && \
1886      (defined (__FreeBSD__) || defined (__FreeBSD_kernel__))
1887# include <machine/sysarch.h>
1888static inline int setup_guest_base_seg(void)
1889{
1890    if (sysarch(AMD64_SET_GSBASE, &guest_base) == 0) {
1891        return P_GS;
1892    }
1893    return 0;
1894}
1895#else
1896static inline int setup_guest_base_seg(void)
1897{
1898    return 0;
1899}
1900#endif /* setup_guest_base_seg */
1901#endif /* !SOFTMMU */
1902
1903/*
1904 * For softmmu, perform the TLB load and compare.
1905 * For useronly, perform any required alignment tests.
1906 * In both cases, return a TCGLabelQemuLdst structure if the slow path
1907 * is required and fill in @h with the host address for the fast path.
1908 */
1909static TCGLabelQemuLdst *prepare_host_addr(TCGContext *s, HostAddress *h,
1910                                           TCGReg addrlo, TCGReg addrhi,
1911                                           MemOpIdx oi, bool is_ld)
1912{
1913    TCGLabelQemuLdst *ldst = NULL;
1914    MemOp opc = get_memop(oi);
1915    MemOp s_bits = opc & MO_SIZE;
1916    unsigned a_mask;
1917
1918#ifdef CONFIG_SOFTMMU
1919    h->index = TCG_REG_L0;
1920    h->ofs = 0;
1921    h->seg = 0;
1922#else
1923    *h = x86_guest_base;
1924#endif
1925    h->base = addrlo;
1926    h->aa = atom_and_align_for_opc(s, opc, MO_ATOM_IFALIGN, s_bits == MO_128);
1927    a_mask = (1 << h->aa.align) - 1;
1928
1929#ifdef CONFIG_SOFTMMU
1930    int cmp_ofs = is_ld ? offsetof(CPUTLBEntry, addr_read)
1931                        : offsetof(CPUTLBEntry, addr_write);
1932    TCGType ttype = TCG_TYPE_I32;
1933    TCGType tlbtype = TCG_TYPE_I32;
1934    int trexw = 0, hrexw = 0, tlbrexw = 0;
1935    unsigned mem_index = get_mmuidx(oi);
1936    unsigned s_mask = (1 << s_bits) - 1;
1937    int tlb_mask;
1938
1939    ldst = new_ldst_label(s);
1940    ldst->is_ld = is_ld;
1941    ldst->oi = oi;
1942    ldst->addrlo_reg = addrlo;
1943    ldst->addrhi_reg = addrhi;
1944
1945    if (TCG_TARGET_REG_BITS == 64) {
1946        ttype = s->addr_type;
1947        trexw = (ttype == TCG_TYPE_I32 ? 0 : P_REXW);
1948        if (TCG_TYPE_PTR == TCG_TYPE_I64) {
1949            hrexw = P_REXW;
1950            if (s->page_bits + s->tlb_dyn_max_bits > 32) {
1951                tlbtype = TCG_TYPE_I64;
1952                tlbrexw = P_REXW;
1953            }
1954        }
1955    }
1956
1957    tcg_out_mov(s, tlbtype, TCG_REG_L0, addrlo);
1958    tcg_out_shifti(s, SHIFT_SHR + tlbrexw, TCG_REG_L0,
1959                   s->page_bits - CPU_TLB_ENTRY_BITS);
1960
1961    tcg_out_modrm_offset(s, OPC_AND_GvEv + trexw, TCG_REG_L0, TCG_AREG0,
1962                         TLB_MASK_TABLE_OFS(mem_index) +
1963                         offsetof(CPUTLBDescFast, mask));
1964
1965    tcg_out_modrm_offset(s, OPC_ADD_GvEv + hrexw, TCG_REG_L0, TCG_AREG0,
1966                         TLB_MASK_TABLE_OFS(mem_index) +
1967                         offsetof(CPUTLBDescFast, table));
1968
1969    /*
1970     * If the required alignment is at least as large as the access, simply
1971     * copy the address and mask.  For lesser alignments, check that we don't
1972     * cross pages for the complete access.
1973     */
1974    if (a_mask >= s_mask) {
1975        tcg_out_mov(s, ttype, TCG_REG_L1, addrlo);
1976    } else {
1977        tcg_out_modrm_offset(s, OPC_LEA + trexw, TCG_REG_L1,
1978                             addrlo, s_mask - a_mask);
1979    }
1980    tlb_mask = s->page_mask | a_mask;
1981    tgen_arithi(s, ARITH_AND + trexw, TCG_REG_L1, tlb_mask, 0);
1982
1983    /* cmp 0(TCG_REG_L0), TCG_REG_L1 */
1984    tcg_out_modrm_offset(s, OPC_CMP_GvEv + trexw,
1985                         TCG_REG_L1, TCG_REG_L0, cmp_ofs);
1986
1987    /* jne slow_path */
1988    tcg_out_opc(s, OPC_JCC_long + JCC_JNE, 0, 0, 0);
1989    ldst->label_ptr[0] = s->code_ptr;
1990    s->code_ptr += 4;
1991
1992    if (TCG_TARGET_REG_BITS == 32 && s->addr_type == TCG_TYPE_I64) {
1993        /* cmp 4(TCG_REG_L0), addrhi */
1994        tcg_out_modrm_offset(s, OPC_CMP_GvEv, addrhi, TCG_REG_L0, cmp_ofs + 4);
1995
1996        /* jne slow_path */
1997        tcg_out_opc(s, OPC_JCC_long + JCC_JNE, 0, 0, 0);
1998        ldst->label_ptr[1] = s->code_ptr;
1999        s->code_ptr += 4;
2000    }
2001
2002    /* TLB Hit.  */
2003    tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_L0, TCG_REG_L0,
2004               offsetof(CPUTLBEntry, addend));
2005#else
2006    if (a_mask) {
2007        ldst = new_ldst_label(s);
2008
2009        ldst->is_ld = is_ld;
2010        ldst->oi = oi;
2011        ldst->addrlo_reg = addrlo;
2012        ldst->addrhi_reg = addrhi;
2013
2014        tcg_out_testi(s, addrlo, a_mask);
2015        /* jne slow_path */
2016        tcg_out_opc(s, OPC_JCC_long + JCC_JNE, 0, 0, 0);
2017        ldst->label_ptr[0] = s->code_ptr;
2018        s->code_ptr += 4;
2019    }
2020#endif
2021
2022    return ldst;
2023}
2024
2025static void tcg_out_qemu_ld_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
2026                                   HostAddress h, TCGType type, MemOp memop)
2027{
2028    bool use_movbe = false;
2029    int rexw = (type == TCG_TYPE_I32 ? 0 : P_REXW);
2030    int movop = OPC_MOVL_GvEv;
2031
2032    /* Do big-endian loads with movbe.  */
2033    if (memop & MO_BSWAP) {
2034        tcg_debug_assert(have_movbe);
2035        use_movbe = true;
2036        movop = OPC_MOVBE_GyMy;
2037    }
2038
2039    switch (memop & MO_SSIZE) {
2040    case MO_UB:
2041        tcg_out_modrm_sib_offset(s, OPC_MOVZBL + h.seg, datalo,
2042                                 h.base, h.index, 0, h.ofs);
2043        break;
2044    case MO_SB:
2045        tcg_out_modrm_sib_offset(s, OPC_MOVSBL + rexw + h.seg, datalo,
2046                                 h.base, h.index, 0, h.ofs);
2047        break;
2048    case MO_UW:
2049        if (use_movbe) {
2050            /* There is no extending movbe; only low 16-bits are modified.  */
2051            if (datalo != h.base && datalo != h.index) {
2052                /* XOR breaks dependency chains.  */
2053                tgen_arithr(s, ARITH_XOR, datalo, datalo);
2054                tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + P_DATA16 + h.seg,
2055                                         datalo, h.base, h.index, 0, h.ofs);
2056            } else {
2057                tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + P_DATA16 + h.seg,
2058                                         datalo, h.base, h.index, 0, h.ofs);
2059                tcg_out_ext16u(s, datalo, datalo);
2060            }
2061        } else {
2062            tcg_out_modrm_sib_offset(s, OPC_MOVZWL + h.seg, datalo,
2063                                     h.base, h.index, 0, h.ofs);
2064        }
2065        break;
2066    case MO_SW:
2067        if (use_movbe) {
2068            tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + P_DATA16 + h.seg,
2069                                     datalo, h.base, h.index, 0, h.ofs);
2070            tcg_out_ext16s(s, type, datalo, datalo);
2071        } else {
2072            tcg_out_modrm_sib_offset(s, OPC_MOVSWL + rexw + h.seg,
2073                                     datalo, h.base, h.index, 0, h.ofs);
2074        }
2075        break;
2076    case MO_UL:
2077        tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2078                                 h.base, h.index, 0, h.ofs);
2079        break;
2080#if TCG_TARGET_REG_BITS == 64
2081    case MO_SL:
2082        if (use_movbe) {
2083            tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + h.seg, datalo,
2084                                     h.base, h.index, 0, h.ofs);
2085            tcg_out_ext32s(s, datalo, datalo);
2086        } else {
2087            tcg_out_modrm_sib_offset(s, OPC_MOVSLQ + h.seg, datalo,
2088                                     h.base, h.index, 0, h.ofs);
2089        }
2090        break;
2091#endif
2092    case MO_UQ:
2093        if (TCG_TARGET_REG_BITS == 64) {
2094            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2095                                     h.base, h.index, 0, h.ofs);
2096            break;
2097        }
2098        if (use_movbe) {
2099            TCGReg t = datalo;
2100            datalo = datahi;
2101            datahi = t;
2102        }
2103        if (h.base == datalo || h.index == datalo) {
2104            tcg_out_modrm_sib_offset(s, OPC_LEA, datahi,
2105                                     h.base, h.index, 0, h.ofs);
2106            tcg_out_modrm_offset(s, movop + h.seg, datalo, datahi, 0);
2107            tcg_out_modrm_offset(s, movop + h.seg, datahi, datahi, 4);
2108        } else {
2109            tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2110                                     h.base, h.index, 0, h.ofs);
2111            tcg_out_modrm_sib_offset(s, movop + h.seg, datahi,
2112                                     h.base, h.index, 0, h.ofs + 4);
2113        }
2114        break;
2115
2116    case MO_128:
2117        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
2118
2119        /*
2120         * Without 16-byte atomicity, use integer regs.
2121         * That is where we want the data, and it allows bswaps.
2122         */
2123        if (h.aa.atom < MO_128) {
2124            if (use_movbe) {
2125                TCGReg t = datalo;
2126                datalo = datahi;
2127                datahi = t;
2128            }
2129            if (h.base == datalo || h.index == datalo) {
2130                tcg_out_modrm_sib_offset(s, OPC_LEA + P_REXW, datahi,
2131                                         h.base, h.index, 0, h.ofs);
2132                tcg_out_modrm_offset(s, movop + P_REXW + h.seg,
2133                                     datalo, datahi, 0);
2134                tcg_out_modrm_offset(s, movop + P_REXW + h.seg,
2135                                     datahi, datahi, 8);
2136            } else {
2137                tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2138                                         h.base, h.index, 0, h.ofs);
2139                tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datahi,
2140                                         h.base, h.index, 0, h.ofs + 8);
2141            }
2142            break;
2143        }
2144
2145        /*
2146         * With 16-byte atomicity, a vector load is required.
2147         * If we already have 16-byte alignment, then VMOVDQA always works.
2148         * Else if VMOVDQU has atomicity with dynamic alignment, use that.
2149         * Else use we require a runtime test for alignment for VMOVDQA;
2150         * use VMOVDQU on the unaligned nonatomic path for simplicity.
2151         */
2152        if (h.aa.align >= MO_128) {
2153            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_VxWx + h.seg,
2154                                         TCG_TMP_VEC, 0,
2155                                         h.base, h.index, 0, h.ofs);
2156        } else if (cpuinfo & CPUINFO_ATOMIC_VMOVDQU) {
2157            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_VxWx + h.seg,
2158                                         TCG_TMP_VEC, 0,
2159                                         h.base, h.index, 0, h.ofs);
2160        } else {
2161            TCGLabel *l1 = gen_new_label();
2162            TCGLabel *l2 = gen_new_label();
2163
2164            tcg_out_testi(s, h.base, 15);
2165            tcg_out_jxx(s, JCC_JNE, l1, true);
2166
2167            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_VxWx + h.seg,
2168                                         TCG_TMP_VEC, 0,
2169                                         h.base, h.index, 0, h.ofs);
2170            tcg_out_jxx(s, JCC_JMP, l2, true);
2171
2172            tcg_out_label(s, l1);
2173            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_VxWx + h.seg,
2174                                         TCG_TMP_VEC, 0,
2175                                         h.base, h.index, 0, h.ofs);
2176            tcg_out_label(s, l2);
2177        }
2178        tcg_out_vec_to_pair(s, TCG_TYPE_I64, datalo, datahi, TCG_TMP_VEC);
2179        break;
2180
2181    default:
2182        g_assert_not_reached();
2183    }
2184}
2185
2186static void tcg_out_qemu_ld(TCGContext *s, TCGReg datalo, TCGReg datahi,
2187                            TCGReg addrlo, TCGReg addrhi,
2188                            MemOpIdx oi, TCGType data_type)
2189{
2190    TCGLabelQemuLdst *ldst;
2191    HostAddress h;
2192
2193    ldst = prepare_host_addr(s, &h, addrlo, addrhi, oi, true);
2194    tcg_out_qemu_ld_direct(s, datalo, datahi, h, data_type, get_memop(oi));
2195
2196    if (ldst) {
2197        ldst->type = data_type;
2198        ldst->datalo_reg = datalo;
2199        ldst->datahi_reg = datahi;
2200        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
2201    }
2202}
2203
2204static void tcg_out_qemu_st_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
2205                                   HostAddress h, MemOp memop)
2206{
2207    bool use_movbe = false;
2208    int movop = OPC_MOVL_EvGv;
2209
2210    /*
2211     * Do big-endian stores with movbe or softmmu.
2212     * User-only without movbe will have its swapping done generically.
2213     */
2214    if (memop & MO_BSWAP) {
2215        tcg_debug_assert(have_movbe);
2216        use_movbe = true;
2217        movop = OPC_MOVBE_MyGy;
2218    }
2219
2220    switch (memop & MO_SIZE) {
2221    case MO_8:
2222        /* This is handled with constraints on INDEX_op_qemu_st8_i32. */
2223        tcg_debug_assert(TCG_TARGET_REG_BITS == 64 || datalo < 4);
2224        tcg_out_modrm_sib_offset(s, OPC_MOVB_EvGv + P_REXB_R + h.seg,
2225                                 datalo, h.base, h.index, 0, h.ofs);
2226        break;
2227    case MO_16:
2228        tcg_out_modrm_sib_offset(s, movop + P_DATA16 + h.seg, datalo,
2229                                 h.base, h.index, 0, h.ofs);
2230        break;
2231    case MO_32:
2232        tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2233                                 h.base, h.index, 0, h.ofs);
2234        break;
2235    case MO_64:
2236        if (TCG_TARGET_REG_BITS == 64) {
2237            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2238                                     h.base, h.index, 0, h.ofs);
2239        } else {
2240            if (use_movbe) {
2241                TCGReg t = datalo;
2242                datalo = datahi;
2243                datahi = t;
2244            }
2245            tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2246                                     h.base, h.index, 0, h.ofs);
2247            tcg_out_modrm_sib_offset(s, movop + h.seg, datahi,
2248                                     h.base, h.index, 0, h.ofs + 4);
2249        }
2250        break;
2251
2252    case MO_128:
2253        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
2254
2255        /*
2256         * Without 16-byte atomicity, use integer regs.
2257         * That is where we have the data, and it allows bswaps.
2258         */
2259        if (h.aa.atom < MO_128) {
2260            if (use_movbe) {
2261                TCGReg t = datalo;
2262                datalo = datahi;
2263                datahi = t;
2264            }
2265            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2266                                     h.base, h.index, 0, h.ofs);
2267            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datahi,
2268                                     h.base, h.index, 0, h.ofs + 8);
2269            break;
2270        }
2271
2272        /*
2273         * With 16-byte atomicity, a vector store is required.
2274         * If we already have 16-byte alignment, then VMOVDQA always works.
2275         * Else if VMOVDQU has atomicity with dynamic alignment, use that.
2276         * Else use we require a runtime test for alignment for VMOVDQA;
2277         * use VMOVDQU on the unaligned nonatomic path for simplicity.
2278         */
2279        tcg_out_pair_to_vec(s, TCG_TYPE_I64, TCG_TMP_VEC, datalo, datahi);
2280        if (h.aa.align >= MO_128) {
2281            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_WxVx + h.seg,
2282                                         TCG_TMP_VEC, 0,
2283                                         h.base, h.index, 0, h.ofs);
2284        } else if (cpuinfo & CPUINFO_ATOMIC_VMOVDQU) {
2285            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_WxVx + h.seg,
2286                                         TCG_TMP_VEC, 0,
2287                                         h.base, h.index, 0, h.ofs);
2288        } else {
2289            TCGLabel *l1 = gen_new_label();
2290            TCGLabel *l2 = gen_new_label();
2291
2292            tcg_out_testi(s, h.base, 15);
2293            tcg_out_jxx(s, JCC_JNE, l1, true);
2294
2295            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_WxVx + h.seg,
2296                                         TCG_TMP_VEC, 0,
2297                                         h.base, h.index, 0, h.ofs);
2298            tcg_out_jxx(s, JCC_JMP, l2, true);
2299
2300            tcg_out_label(s, l1);
2301            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_WxVx + h.seg,
2302                                         TCG_TMP_VEC, 0,
2303                                         h.base, h.index, 0, h.ofs);
2304            tcg_out_label(s, l2);
2305        }
2306        break;
2307
2308    default:
2309        g_assert_not_reached();
2310    }
2311}
2312
2313static void tcg_out_qemu_st(TCGContext *s, TCGReg datalo, TCGReg datahi,
2314                            TCGReg addrlo, TCGReg addrhi,
2315                            MemOpIdx oi, TCGType data_type)
2316{
2317    TCGLabelQemuLdst *ldst;
2318    HostAddress h;
2319
2320    ldst = prepare_host_addr(s, &h, addrlo, addrhi, oi, false);
2321    tcg_out_qemu_st_direct(s, datalo, datahi, h, get_memop(oi));
2322
2323    if (ldst) {
2324        ldst->type = data_type;
2325        ldst->datalo_reg = datalo;
2326        ldst->datahi_reg = datahi;
2327        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
2328    }
2329}
2330
2331static void tcg_out_exit_tb(TCGContext *s, uintptr_t a0)
2332{
2333    /* Reuse the zeroing that exists for goto_ptr.  */
2334    if (a0 == 0) {
2335        tcg_out_jmp(s, tcg_code_gen_epilogue);
2336    } else {
2337        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_EAX, a0);
2338        tcg_out_jmp(s, tb_ret_addr);
2339    }
2340}
2341
2342static void tcg_out_goto_tb(TCGContext *s, int which)
2343{
2344    /*
2345     * Jump displacement must be aligned for atomic patching;
2346     * see if we need to add extra nops before jump
2347     */
2348    int gap = QEMU_ALIGN_PTR_UP(s->code_ptr + 1, 4) - s->code_ptr;
2349    if (gap != 1) {
2350        tcg_out_nopn(s, gap - 1);
2351    }
2352    tcg_out8(s, OPC_JMP_long); /* jmp im */
2353    set_jmp_insn_offset(s, which);
2354    tcg_out32(s, 0);
2355    set_jmp_reset_offset(s, which);
2356}
2357
2358void tb_target_set_jmp_target(const TranslationBlock *tb, int n,
2359                              uintptr_t jmp_rx, uintptr_t jmp_rw)
2360{
2361    /* patch the branch destination */
2362    uintptr_t addr = tb->jmp_target_addr[n];
2363    qatomic_set((int32_t *)jmp_rw, addr - (jmp_rx + 4));
2364    /* no need to flush icache explicitly */
2365}
2366
2367static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
2368                              const TCGArg args[TCG_MAX_OP_ARGS],
2369                              const int const_args[TCG_MAX_OP_ARGS])
2370{
2371    TCGArg a0, a1, a2;
2372    int c, const_a2, vexop, rexw = 0;
2373
2374#if TCG_TARGET_REG_BITS == 64
2375# define OP_32_64(x) \
2376        case glue(glue(INDEX_op_, x), _i64): \
2377            rexw = P_REXW; /* FALLTHRU */    \
2378        case glue(glue(INDEX_op_, x), _i32)
2379#else
2380# define OP_32_64(x) \
2381        case glue(glue(INDEX_op_, x), _i32)
2382#endif
2383
2384    /* Hoist the loads of the most common arguments.  */
2385    a0 = args[0];
2386    a1 = args[1];
2387    a2 = args[2];
2388    const_a2 = const_args[2];
2389
2390    switch (opc) {
2391    case INDEX_op_goto_ptr:
2392        /* jmp to the given host address (could be epilogue) */
2393        tcg_out_modrm(s, OPC_GRP5, EXT5_JMPN_Ev, a0);
2394        break;
2395    case INDEX_op_br:
2396        tcg_out_jxx(s, JCC_JMP, arg_label(a0), 0);
2397        break;
2398    OP_32_64(ld8u):
2399        /* Note that we can ignore REXW for the zero-extend to 64-bit.  */
2400        tcg_out_modrm_offset(s, OPC_MOVZBL, a0, a1, a2);
2401        break;
2402    OP_32_64(ld8s):
2403        tcg_out_modrm_offset(s, OPC_MOVSBL + rexw, a0, a1, a2);
2404        break;
2405    OP_32_64(ld16u):
2406        /* Note that we can ignore REXW for the zero-extend to 64-bit.  */
2407        tcg_out_modrm_offset(s, OPC_MOVZWL, a0, a1, a2);
2408        break;
2409    OP_32_64(ld16s):
2410        tcg_out_modrm_offset(s, OPC_MOVSWL + rexw, a0, a1, a2);
2411        break;
2412#if TCG_TARGET_REG_BITS == 64
2413    case INDEX_op_ld32u_i64:
2414#endif
2415    case INDEX_op_ld_i32:
2416        tcg_out_ld(s, TCG_TYPE_I32, a0, a1, a2);
2417        break;
2418
2419    OP_32_64(st8):
2420        if (const_args[0]) {
2421            tcg_out_modrm_offset(s, OPC_MOVB_EvIz, 0, a1, a2);
2422            tcg_out8(s, a0);
2423        } else {
2424            tcg_out_modrm_offset(s, OPC_MOVB_EvGv | P_REXB_R, a0, a1, a2);
2425        }
2426        break;
2427    OP_32_64(st16):
2428        if (const_args[0]) {
2429            tcg_out_modrm_offset(s, OPC_MOVL_EvIz | P_DATA16, 0, a1, a2);
2430            tcg_out16(s, a0);
2431        } else {
2432            tcg_out_modrm_offset(s, OPC_MOVL_EvGv | P_DATA16, a0, a1, a2);
2433        }
2434        break;
2435#if TCG_TARGET_REG_BITS == 64
2436    case INDEX_op_st32_i64:
2437#endif
2438    case INDEX_op_st_i32:
2439        if (const_args[0]) {
2440            tcg_out_modrm_offset(s, OPC_MOVL_EvIz, 0, a1, a2);
2441            tcg_out32(s, a0);
2442        } else {
2443            tcg_out_st(s, TCG_TYPE_I32, a0, a1, a2);
2444        }
2445        break;
2446
2447    OP_32_64(add):
2448        /* For 3-operand addition, use LEA.  */
2449        if (a0 != a1) {
2450            TCGArg c3 = 0;
2451            if (const_a2) {
2452                c3 = a2, a2 = -1;
2453            } else if (a0 == a2) {
2454                /* Watch out for dest = src + dest, since we've removed
2455                   the matching constraint on the add.  */
2456                tgen_arithr(s, ARITH_ADD + rexw, a0, a1);
2457                break;
2458            }
2459
2460            tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, a1, a2, 0, c3);
2461            break;
2462        }
2463        c = ARITH_ADD;
2464        goto gen_arith;
2465    OP_32_64(sub):
2466        c = ARITH_SUB;
2467        goto gen_arith;
2468    OP_32_64(and):
2469        c = ARITH_AND;
2470        goto gen_arith;
2471    OP_32_64(or):
2472        c = ARITH_OR;
2473        goto gen_arith;
2474    OP_32_64(xor):
2475        c = ARITH_XOR;
2476        goto gen_arith;
2477    gen_arith:
2478        if (const_a2) {
2479            tgen_arithi(s, c + rexw, a0, a2, 0);
2480        } else {
2481            tgen_arithr(s, c + rexw, a0, a2);
2482        }
2483        break;
2484
2485    OP_32_64(andc):
2486        if (const_a2) {
2487            tcg_out_mov(s, rexw ? TCG_TYPE_I64 : TCG_TYPE_I32, a0, a1);
2488            tgen_arithi(s, ARITH_AND + rexw, a0, ~a2, 0);
2489        } else {
2490            tcg_out_vex_modrm(s, OPC_ANDN + rexw, a0, a2, a1);
2491        }
2492        break;
2493
2494    OP_32_64(mul):
2495        if (const_a2) {
2496            int32_t val;
2497            val = a2;
2498            if (val == (int8_t)val) {
2499                tcg_out_modrm(s, OPC_IMUL_GvEvIb + rexw, a0, a0);
2500                tcg_out8(s, val);
2501            } else {
2502                tcg_out_modrm(s, OPC_IMUL_GvEvIz + rexw, a0, a0);
2503                tcg_out32(s, val);
2504            }
2505        } else {
2506            tcg_out_modrm(s, OPC_IMUL_GvEv + rexw, a0, a2);
2507        }
2508        break;
2509
2510    OP_32_64(div2):
2511        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_IDIV, args[4]);
2512        break;
2513    OP_32_64(divu2):
2514        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_DIV, args[4]);
2515        break;
2516
2517    OP_32_64(shl):
2518        /* For small constant 3-operand shift, use LEA.  */
2519        if (const_a2 && a0 != a1 && (a2 - 1) < 3) {
2520            if (a2 - 1 == 0) {
2521                /* shl $1,a1,a0 -> lea (a1,a1),a0 */
2522                tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, a1, a1, 0, 0);
2523            } else {
2524                /* shl $n,a1,a0 -> lea 0(,a1,n),a0 */
2525                tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, -1, a1, a2, 0);
2526            }
2527            break;
2528        }
2529        c = SHIFT_SHL;
2530        vexop = OPC_SHLX;
2531        goto gen_shift_maybe_vex;
2532    OP_32_64(shr):
2533        c = SHIFT_SHR;
2534        vexop = OPC_SHRX;
2535        goto gen_shift_maybe_vex;
2536    OP_32_64(sar):
2537        c = SHIFT_SAR;
2538        vexop = OPC_SARX;
2539        goto gen_shift_maybe_vex;
2540    OP_32_64(rotl):
2541        c = SHIFT_ROL;
2542        goto gen_shift;
2543    OP_32_64(rotr):
2544        c = SHIFT_ROR;
2545        goto gen_shift;
2546    gen_shift_maybe_vex:
2547        if (have_bmi2) {
2548            if (!const_a2) {
2549                tcg_out_vex_modrm(s, vexop + rexw, a0, a2, a1);
2550                break;
2551            }
2552            tcg_out_mov(s, rexw ? TCG_TYPE_I64 : TCG_TYPE_I32, a0, a1);
2553        }
2554        /* FALLTHRU */
2555    gen_shift:
2556        if (const_a2) {
2557            tcg_out_shifti(s, c + rexw, a0, a2);
2558        } else {
2559            tcg_out_modrm(s, OPC_SHIFT_cl + rexw, c, a0);
2560        }
2561        break;
2562
2563    OP_32_64(ctz):
2564        tcg_out_ctz(s, rexw, args[0], args[1], args[2], const_args[2]);
2565        break;
2566    OP_32_64(clz):
2567        tcg_out_clz(s, rexw, args[0], args[1], args[2], const_args[2]);
2568        break;
2569    OP_32_64(ctpop):
2570        tcg_out_modrm(s, OPC_POPCNT + rexw, a0, a1);
2571        break;
2572
2573    case INDEX_op_brcond_i32:
2574        tcg_out_brcond32(s, a2, a0, a1, const_args[1], arg_label(args[3]), 0);
2575        break;
2576    case INDEX_op_setcond_i32:
2577        tcg_out_setcond32(s, args[3], a0, a1, a2, const_a2);
2578        break;
2579    case INDEX_op_movcond_i32:
2580        tcg_out_movcond32(s, args[5], a0, a1, a2, const_a2, args[3]);
2581        break;
2582
2583    OP_32_64(bswap16):
2584        if (a2 & TCG_BSWAP_OS) {
2585            /* Output must be sign-extended. */
2586            if (rexw) {
2587                tcg_out_bswap64(s, a0);
2588                tcg_out_shifti(s, SHIFT_SAR + rexw, a0, 48);
2589            } else {
2590                tcg_out_bswap32(s, a0);
2591                tcg_out_shifti(s, SHIFT_SAR, a0, 16);
2592            }
2593        } else if ((a2 & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
2594            /* Output must be zero-extended, but input isn't. */
2595            tcg_out_bswap32(s, a0);
2596            tcg_out_shifti(s, SHIFT_SHR, a0, 16);
2597        } else {
2598            tcg_out_rolw_8(s, a0);
2599        }
2600        break;
2601    OP_32_64(bswap32):
2602        tcg_out_bswap32(s, a0);
2603        if (rexw && (a2 & TCG_BSWAP_OS)) {
2604            tcg_out_ext32s(s, a0, a0);
2605        }
2606        break;
2607
2608    OP_32_64(neg):
2609        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NEG, a0);
2610        break;
2611    OP_32_64(not):
2612        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NOT, a0);
2613        break;
2614
2615    case INDEX_op_qemu_ld_a64_i32:
2616        if (TCG_TARGET_REG_BITS == 32) {
2617            tcg_out_qemu_ld(s, a0, -1, a1, a2, args[3], TCG_TYPE_I32);
2618            break;
2619        }
2620        /* fall through */
2621    case INDEX_op_qemu_ld_a32_i32:
2622        tcg_out_qemu_ld(s, a0, -1, a1, -1, a2, TCG_TYPE_I32);
2623        break;
2624    case INDEX_op_qemu_ld_a32_i64:
2625        if (TCG_TARGET_REG_BITS == 64) {
2626            tcg_out_qemu_ld(s, a0, -1, a1, -1, a2, TCG_TYPE_I64);
2627        } else {
2628            tcg_out_qemu_ld(s, a0, a1, a2, -1, args[3], TCG_TYPE_I64);
2629        }
2630        break;
2631    case INDEX_op_qemu_ld_a64_i64:
2632        if (TCG_TARGET_REG_BITS == 64) {
2633            tcg_out_qemu_ld(s, a0, -1, a1, -1, a2, TCG_TYPE_I64);
2634        } else {
2635            tcg_out_qemu_ld(s, a0, a1, a2, args[3], args[4], TCG_TYPE_I64);
2636        }
2637        break;
2638    case INDEX_op_qemu_ld_a32_i128:
2639    case INDEX_op_qemu_ld_a64_i128:
2640        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
2641        tcg_out_qemu_ld(s, a0, a1, a2, -1, args[3], TCG_TYPE_I128);
2642        break;
2643
2644    case INDEX_op_qemu_st_a64_i32:
2645    case INDEX_op_qemu_st8_a64_i32:
2646        if (TCG_TARGET_REG_BITS == 32) {
2647            tcg_out_qemu_st(s, a0, -1, a1, a2, args[3], TCG_TYPE_I32);
2648            break;
2649        }
2650        /* fall through */
2651    case INDEX_op_qemu_st_a32_i32:
2652    case INDEX_op_qemu_st8_a32_i32:
2653        tcg_out_qemu_st(s, a0, -1, a1, -1, a2, TCG_TYPE_I32);
2654        break;
2655    case INDEX_op_qemu_st_a32_i64:
2656        if (TCG_TARGET_REG_BITS == 64) {
2657            tcg_out_qemu_st(s, a0, -1, a1, -1, a2, TCG_TYPE_I64);
2658        } else {
2659            tcg_out_qemu_st(s, a0, a1, a2, -1, args[3], TCG_TYPE_I64);
2660        }
2661        break;
2662    case INDEX_op_qemu_st_a64_i64:
2663        if (TCG_TARGET_REG_BITS == 64) {
2664            tcg_out_qemu_st(s, a0, -1, a1, -1, a2, TCG_TYPE_I64);
2665        } else {
2666            tcg_out_qemu_st(s, a0, a1, a2, args[3], args[4], TCG_TYPE_I64);
2667        }
2668        break;
2669    case INDEX_op_qemu_st_a32_i128:
2670    case INDEX_op_qemu_st_a64_i128:
2671        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
2672        tcg_out_qemu_st(s, a0, a1, a2, -1, args[3], TCG_TYPE_I128);
2673        break;
2674
2675    OP_32_64(mulu2):
2676        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_MUL, args[3]);
2677        break;
2678    OP_32_64(muls2):
2679        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_IMUL, args[3]);
2680        break;
2681    OP_32_64(add2):
2682        if (const_args[4]) {
2683            tgen_arithi(s, ARITH_ADD + rexw, a0, args[4], 1);
2684        } else {
2685            tgen_arithr(s, ARITH_ADD + rexw, a0, args[4]);
2686        }
2687        if (const_args[5]) {
2688            tgen_arithi(s, ARITH_ADC + rexw, a1, args[5], 1);
2689        } else {
2690            tgen_arithr(s, ARITH_ADC + rexw, a1, args[5]);
2691        }
2692        break;
2693    OP_32_64(sub2):
2694        if (const_args[4]) {
2695            tgen_arithi(s, ARITH_SUB + rexw, a0, args[4], 1);
2696        } else {
2697            tgen_arithr(s, ARITH_SUB + rexw, a0, args[4]);
2698        }
2699        if (const_args[5]) {
2700            tgen_arithi(s, ARITH_SBB + rexw, a1, args[5], 1);
2701        } else {
2702            tgen_arithr(s, ARITH_SBB + rexw, a1, args[5]);
2703        }
2704        break;
2705
2706#if TCG_TARGET_REG_BITS == 32
2707    case INDEX_op_brcond2_i32:
2708        tcg_out_brcond2(s, args, const_args, 0);
2709        break;
2710    case INDEX_op_setcond2_i32:
2711        tcg_out_setcond2(s, args, const_args);
2712        break;
2713#else /* TCG_TARGET_REG_BITS == 64 */
2714    case INDEX_op_ld32s_i64:
2715        tcg_out_modrm_offset(s, OPC_MOVSLQ, a0, a1, a2);
2716        break;
2717    case INDEX_op_ld_i64:
2718        tcg_out_ld(s, TCG_TYPE_I64, a0, a1, a2);
2719        break;
2720    case INDEX_op_st_i64:
2721        if (const_args[0]) {
2722            tcg_out_modrm_offset(s, OPC_MOVL_EvIz | P_REXW, 0, a1, a2);
2723            tcg_out32(s, a0);
2724        } else {
2725            tcg_out_st(s, TCG_TYPE_I64, a0, a1, a2);
2726        }
2727        break;
2728
2729    case INDEX_op_brcond_i64:
2730        tcg_out_brcond64(s, a2, a0, a1, const_args[1], arg_label(args[3]), 0);
2731        break;
2732    case INDEX_op_setcond_i64:
2733        tcg_out_setcond64(s, args[3], a0, a1, a2, const_a2);
2734        break;
2735    case INDEX_op_movcond_i64:
2736        tcg_out_movcond64(s, args[5], a0, a1, a2, const_a2, args[3]);
2737        break;
2738
2739    case INDEX_op_bswap64_i64:
2740        tcg_out_bswap64(s, a0);
2741        break;
2742    case INDEX_op_extrh_i64_i32:
2743        tcg_out_shifti(s, SHIFT_SHR + P_REXW, a0, 32);
2744        break;
2745#endif
2746
2747    OP_32_64(deposit):
2748        if (args[3] == 0 && args[4] == 8) {
2749            /* load bits 0..7 */
2750            tcg_out_modrm(s, OPC_MOVB_EvGv | P_REXB_R | P_REXB_RM, a2, a0);
2751        } else if (args[3] == 8 && args[4] == 8) {
2752            /* load bits 8..15 */
2753            tcg_out_modrm(s, OPC_MOVB_EvGv, a2, a0 + 4);
2754        } else if (args[3] == 0 && args[4] == 16) {
2755            /* load bits 0..15 */
2756            tcg_out_modrm(s, OPC_MOVL_EvGv | P_DATA16, a2, a0);
2757        } else {
2758            g_assert_not_reached();
2759        }
2760        break;
2761
2762    case INDEX_op_extract_i64:
2763        if (a2 + args[3] == 32) {
2764            /* This is a 32-bit zero-extending right shift.  */
2765            tcg_out_mov(s, TCG_TYPE_I32, a0, a1);
2766            tcg_out_shifti(s, SHIFT_SHR, a0, a2);
2767            break;
2768        }
2769        /* FALLTHRU */
2770    case INDEX_op_extract_i32:
2771        /* On the off-chance that we can use the high-byte registers.
2772           Otherwise we emit the same ext16 + shift pattern that we
2773           would have gotten from the normal tcg-op.c expansion.  */
2774        tcg_debug_assert(a2 == 8 && args[3] == 8);
2775        if (a1 < 4 && a0 < 8) {
2776            tcg_out_modrm(s, OPC_MOVZBL, a0, a1 + 4);
2777        } else {
2778            tcg_out_ext16u(s, a0, a1);
2779            tcg_out_shifti(s, SHIFT_SHR, a0, 8);
2780        }
2781        break;
2782
2783    case INDEX_op_sextract_i32:
2784        /* We don't implement sextract_i64, as we cannot sign-extend to
2785           64-bits without using the REX prefix that explicitly excludes
2786           access to the high-byte registers.  */
2787        tcg_debug_assert(a2 == 8 && args[3] == 8);
2788        if (a1 < 4 && a0 < 8) {
2789            tcg_out_modrm(s, OPC_MOVSBL, a0, a1 + 4);
2790        } else {
2791            tcg_out_ext16s(s, TCG_TYPE_I32, a0, a1);
2792            tcg_out_shifti(s, SHIFT_SAR, a0, 8);
2793        }
2794        break;
2795
2796    OP_32_64(extract2):
2797        /* Note that SHRD outputs to the r/m operand.  */
2798        tcg_out_modrm(s, OPC_SHRD_Ib + rexw, a2, a0);
2799        tcg_out8(s, args[3]);
2800        break;
2801
2802    case INDEX_op_mb:
2803        tcg_out_mb(s, a0);
2804        break;
2805    case INDEX_op_mov_i32:  /* Always emitted via tcg_out_mov.  */
2806    case INDEX_op_mov_i64:
2807    case INDEX_op_call:     /* Always emitted via tcg_out_call.  */
2808    case INDEX_op_exit_tb:  /* Always emitted via tcg_out_exit_tb.  */
2809    case INDEX_op_goto_tb:  /* Always emitted via tcg_out_goto_tb.  */
2810    case INDEX_op_ext8s_i32:  /* Always emitted via tcg_reg_alloc_op.  */
2811    case INDEX_op_ext8s_i64:
2812    case INDEX_op_ext8u_i32:
2813    case INDEX_op_ext8u_i64:
2814    case INDEX_op_ext16s_i32:
2815    case INDEX_op_ext16s_i64:
2816    case INDEX_op_ext16u_i32:
2817    case INDEX_op_ext16u_i64:
2818    case INDEX_op_ext32s_i64:
2819    case INDEX_op_ext32u_i64:
2820    case INDEX_op_ext_i32_i64:
2821    case INDEX_op_extu_i32_i64:
2822    case INDEX_op_extrl_i64_i32:
2823    default:
2824        g_assert_not_reached();
2825    }
2826
2827#undef OP_32_64
2828}
2829
2830static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
2831                           unsigned vecl, unsigned vece,
2832                           const TCGArg args[TCG_MAX_OP_ARGS],
2833                           const int const_args[TCG_MAX_OP_ARGS])
2834{
2835    static int const add_insn[4] = {
2836        OPC_PADDB, OPC_PADDW, OPC_PADDD, OPC_PADDQ
2837    };
2838    static int const ssadd_insn[4] = {
2839        OPC_PADDSB, OPC_PADDSW, OPC_UD2, OPC_UD2
2840    };
2841    static int const usadd_insn[4] = {
2842        OPC_PADDUB, OPC_PADDUW, OPC_UD2, OPC_UD2
2843    };
2844    static int const sub_insn[4] = {
2845        OPC_PSUBB, OPC_PSUBW, OPC_PSUBD, OPC_PSUBQ
2846    };
2847    static int const sssub_insn[4] = {
2848        OPC_PSUBSB, OPC_PSUBSW, OPC_UD2, OPC_UD2
2849    };
2850    static int const ussub_insn[4] = {
2851        OPC_PSUBUB, OPC_PSUBUW, OPC_UD2, OPC_UD2
2852    };
2853    static int const mul_insn[4] = {
2854        OPC_UD2, OPC_PMULLW, OPC_PMULLD, OPC_VPMULLQ
2855    };
2856    static int const shift_imm_insn[4] = {
2857        OPC_UD2, OPC_PSHIFTW_Ib, OPC_PSHIFTD_Ib, OPC_PSHIFTQ_Ib
2858    };
2859    static int const cmpeq_insn[4] = {
2860        OPC_PCMPEQB, OPC_PCMPEQW, OPC_PCMPEQD, OPC_PCMPEQQ
2861    };
2862    static int const cmpgt_insn[4] = {
2863        OPC_PCMPGTB, OPC_PCMPGTW, OPC_PCMPGTD, OPC_PCMPGTQ
2864    };
2865    static int const punpckl_insn[4] = {
2866        OPC_PUNPCKLBW, OPC_PUNPCKLWD, OPC_PUNPCKLDQ, OPC_PUNPCKLQDQ
2867    };
2868    static int const punpckh_insn[4] = {
2869        OPC_PUNPCKHBW, OPC_PUNPCKHWD, OPC_PUNPCKHDQ, OPC_PUNPCKHQDQ
2870    };
2871    static int const packss_insn[4] = {
2872        OPC_PACKSSWB, OPC_PACKSSDW, OPC_UD2, OPC_UD2
2873    };
2874    static int const packus_insn[4] = {
2875        OPC_PACKUSWB, OPC_PACKUSDW, OPC_UD2, OPC_UD2
2876    };
2877    static int const smin_insn[4] = {
2878        OPC_PMINSB, OPC_PMINSW, OPC_PMINSD, OPC_VPMINSQ
2879    };
2880    static int const smax_insn[4] = {
2881        OPC_PMAXSB, OPC_PMAXSW, OPC_PMAXSD, OPC_VPMAXSQ
2882    };
2883    static int const umin_insn[4] = {
2884        OPC_PMINUB, OPC_PMINUW, OPC_PMINUD, OPC_VPMINUQ
2885    };
2886    static int const umax_insn[4] = {
2887        OPC_PMAXUB, OPC_PMAXUW, OPC_PMAXUD, OPC_VPMAXUQ
2888    };
2889    static int const rotlv_insn[4] = {
2890        OPC_UD2, OPC_UD2, OPC_VPROLVD, OPC_VPROLVQ
2891    };
2892    static int const rotrv_insn[4] = {
2893        OPC_UD2, OPC_UD2, OPC_VPRORVD, OPC_VPRORVQ
2894    };
2895    static int const shlv_insn[4] = {
2896        OPC_UD2, OPC_VPSLLVW, OPC_VPSLLVD, OPC_VPSLLVQ
2897    };
2898    static int const shrv_insn[4] = {
2899        OPC_UD2, OPC_VPSRLVW, OPC_VPSRLVD, OPC_VPSRLVQ
2900    };
2901    static int const sarv_insn[4] = {
2902        OPC_UD2, OPC_VPSRAVW, OPC_VPSRAVD, OPC_VPSRAVQ
2903    };
2904    static int const shls_insn[4] = {
2905        OPC_UD2, OPC_PSLLW, OPC_PSLLD, OPC_PSLLQ
2906    };
2907    static int const shrs_insn[4] = {
2908        OPC_UD2, OPC_PSRLW, OPC_PSRLD, OPC_PSRLQ
2909    };
2910    static int const sars_insn[4] = {
2911        OPC_UD2, OPC_PSRAW, OPC_PSRAD, OPC_VPSRAQ
2912    };
2913    static int const vpshldi_insn[4] = {
2914        OPC_UD2, OPC_VPSHLDW, OPC_VPSHLDD, OPC_VPSHLDQ
2915    };
2916    static int const vpshldv_insn[4] = {
2917        OPC_UD2, OPC_VPSHLDVW, OPC_VPSHLDVD, OPC_VPSHLDVQ
2918    };
2919    static int const vpshrdv_insn[4] = {
2920        OPC_UD2, OPC_VPSHRDVW, OPC_VPSHRDVD, OPC_VPSHRDVQ
2921    };
2922    static int const abs_insn[4] = {
2923        OPC_PABSB, OPC_PABSW, OPC_PABSD, OPC_VPABSQ
2924    };
2925
2926    TCGType type = vecl + TCG_TYPE_V64;
2927    int insn, sub;
2928    TCGArg a0, a1, a2, a3;
2929
2930    a0 = args[0];
2931    a1 = args[1];
2932    a2 = args[2];
2933
2934    switch (opc) {
2935    case INDEX_op_add_vec:
2936        insn = add_insn[vece];
2937        goto gen_simd;
2938    case INDEX_op_ssadd_vec:
2939        insn = ssadd_insn[vece];
2940        goto gen_simd;
2941    case INDEX_op_usadd_vec:
2942        insn = usadd_insn[vece];
2943        goto gen_simd;
2944    case INDEX_op_sub_vec:
2945        insn = sub_insn[vece];
2946        goto gen_simd;
2947    case INDEX_op_sssub_vec:
2948        insn = sssub_insn[vece];
2949        goto gen_simd;
2950    case INDEX_op_ussub_vec:
2951        insn = ussub_insn[vece];
2952        goto gen_simd;
2953    case INDEX_op_mul_vec:
2954        insn = mul_insn[vece];
2955        goto gen_simd;
2956    case INDEX_op_and_vec:
2957        insn = OPC_PAND;
2958        goto gen_simd;
2959    case INDEX_op_or_vec:
2960        insn = OPC_POR;
2961        goto gen_simd;
2962    case INDEX_op_xor_vec:
2963        insn = OPC_PXOR;
2964        goto gen_simd;
2965    case INDEX_op_smin_vec:
2966        insn = smin_insn[vece];
2967        goto gen_simd;
2968    case INDEX_op_umin_vec:
2969        insn = umin_insn[vece];
2970        goto gen_simd;
2971    case INDEX_op_smax_vec:
2972        insn = smax_insn[vece];
2973        goto gen_simd;
2974    case INDEX_op_umax_vec:
2975        insn = umax_insn[vece];
2976        goto gen_simd;
2977    case INDEX_op_shlv_vec:
2978        insn = shlv_insn[vece];
2979        goto gen_simd;
2980    case INDEX_op_shrv_vec:
2981        insn = shrv_insn[vece];
2982        goto gen_simd;
2983    case INDEX_op_sarv_vec:
2984        insn = sarv_insn[vece];
2985        goto gen_simd;
2986    case INDEX_op_rotlv_vec:
2987        insn = rotlv_insn[vece];
2988        goto gen_simd;
2989    case INDEX_op_rotrv_vec:
2990        insn = rotrv_insn[vece];
2991        goto gen_simd;
2992    case INDEX_op_shls_vec:
2993        insn = shls_insn[vece];
2994        goto gen_simd;
2995    case INDEX_op_shrs_vec:
2996        insn = shrs_insn[vece];
2997        goto gen_simd;
2998    case INDEX_op_sars_vec:
2999        insn = sars_insn[vece];
3000        goto gen_simd;
3001    case INDEX_op_x86_punpckl_vec:
3002        insn = punpckl_insn[vece];
3003        goto gen_simd;
3004    case INDEX_op_x86_punpckh_vec:
3005        insn = punpckh_insn[vece];
3006        goto gen_simd;
3007    case INDEX_op_x86_packss_vec:
3008        insn = packss_insn[vece];
3009        goto gen_simd;
3010    case INDEX_op_x86_packus_vec:
3011        insn = packus_insn[vece];
3012        goto gen_simd;
3013    case INDEX_op_x86_vpshldv_vec:
3014        insn = vpshldv_insn[vece];
3015        a1 = a2;
3016        a2 = args[3];
3017        goto gen_simd;
3018    case INDEX_op_x86_vpshrdv_vec:
3019        insn = vpshrdv_insn[vece];
3020        a1 = a2;
3021        a2 = args[3];
3022        goto gen_simd;
3023#if TCG_TARGET_REG_BITS == 32
3024    case INDEX_op_dup2_vec:
3025        /* First merge the two 32-bit inputs to a single 64-bit element. */
3026        tcg_out_vex_modrm(s, OPC_PUNPCKLDQ, a0, a1, a2);
3027        /* Then replicate the 64-bit elements across the rest of the vector. */
3028        if (type != TCG_TYPE_V64) {
3029            tcg_out_dup_vec(s, type, MO_64, a0, a0);
3030        }
3031        break;
3032#endif
3033    case INDEX_op_abs_vec:
3034        insn = abs_insn[vece];
3035        a2 = a1;
3036        a1 = 0;
3037        goto gen_simd;
3038    gen_simd:
3039        tcg_debug_assert(insn != OPC_UD2);
3040        if (type == TCG_TYPE_V256) {
3041            insn |= P_VEXL;
3042        }
3043        tcg_out_vex_modrm(s, insn, a0, a1, a2);
3044        break;
3045
3046    case INDEX_op_cmp_vec:
3047        sub = args[3];
3048        if (sub == TCG_COND_EQ) {
3049            insn = cmpeq_insn[vece];
3050        } else if (sub == TCG_COND_GT) {
3051            insn = cmpgt_insn[vece];
3052        } else {
3053            g_assert_not_reached();
3054        }
3055        goto gen_simd;
3056
3057    case INDEX_op_andc_vec:
3058        insn = OPC_PANDN;
3059        if (type == TCG_TYPE_V256) {
3060            insn |= P_VEXL;
3061        }
3062        tcg_out_vex_modrm(s, insn, a0, a2, a1);
3063        break;
3064
3065    case INDEX_op_shli_vec:
3066        insn = shift_imm_insn[vece];
3067        sub = 6;
3068        goto gen_shift;
3069    case INDEX_op_shri_vec:
3070        insn = shift_imm_insn[vece];
3071        sub = 2;
3072        goto gen_shift;
3073    case INDEX_op_sari_vec:
3074        if (vece == MO_64) {
3075            insn = OPC_PSHIFTD_Ib | P_VEXW | P_EVEX;
3076        } else {
3077            insn = shift_imm_insn[vece];
3078        }
3079        sub = 4;
3080        goto gen_shift;
3081    case INDEX_op_rotli_vec:
3082        insn = OPC_PSHIFTD_Ib | P_EVEX;  /* VPROL[DQ] */
3083        if (vece == MO_64) {
3084            insn |= P_VEXW;
3085        }
3086        sub = 1;
3087        goto gen_shift;
3088    gen_shift:
3089        tcg_debug_assert(vece != MO_8);
3090        if (type == TCG_TYPE_V256) {
3091            insn |= P_VEXL;
3092        }
3093        tcg_out_vex_modrm(s, insn, sub, a0, a1);
3094        tcg_out8(s, a2);
3095        break;
3096
3097    case INDEX_op_ld_vec:
3098        tcg_out_ld(s, type, a0, a1, a2);
3099        break;
3100    case INDEX_op_st_vec:
3101        tcg_out_st(s, type, a0, a1, a2);
3102        break;
3103    case INDEX_op_dupm_vec:
3104        tcg_out_dupm_vec(s, type, vece, a0, a1, a2);
3105        break;
3106
3107    case INDEX_op_x86_shufps_vec:
3108        insn = OPC_SHUFPS;
3109        sub = args[3];
3110        goto gen_simd_imm8;
3111    case INDEX_op_x86_blend_vec:
3112        if (vece == MO_16) {
3113            insn = OPC_PBLENDW;
3114        } else if (vece == MO_32) {
3115            insn = (have_avx2 ? OPC_VPBLENDD : OPC_BLENDPS);
3116        } else {
3117            g_assert_not_reached();
3118        }
3119        sub = args[3];
3120        goto gen_simd_imm8;
3121    case INDEX_op_x86_vperm2i128_vec:
3122        insn = OPC_VPERM2I128;
3123        sub = args[3];
3124        goto gen_simd_imm8;
3125    case INDEX_op_x86_vpshldi_vec:
3126        insn = vpshldi_insn[vece];
3127        sub = args[3];
3128        goto gen_simd_imm8;
3129
3130    case INDEX_op_not_vec:
3131        insn = OPC_VPTERNLOGQ;
3132        a2 = a1;
3133        sub = 0x33; /* !B */
3134        goto gen_simd_imm8;
3135    case INDEX_op_nor_vec:
3136        insn = OPC_VPTERNLOGQ;
3137        sub = 0x11; /* norCB */
3138        goto gen_simd_imm8;
3139    case INDEX_op_nand_vec:
3140        insn = OPC_VPTERNLOGQ;
3141        sub = 0x77; /* nandCB */
3142        goto gen_simd_imm8;
3143    case INDEX_op_eqv_vec:
3144        insn = OPC_VPTERNLOGQ;
3145        sub = 0x99; /* xnorCB */
3146        goto gen_simd_imm8;
3147    case INDEX_op_orc_vec:
3148        insn = OPC_VPTERNLOGQ;
3149        sub = 0xdd; /* orB!C */
3150        goto gen_simd_imm8;
3151
3152    case INDEX_op_bitsel_vec:
3153        insn = OPC_VPTERNLOGQ;
3154        a3 = args[3];
3155        if (a0 == a1) {
3156            a1 = a2;
3157            a2 = a3;
3158            sub = 0xca; /* A?B:C */
3159        } else if (a0 == a2) {
3160            a2 = a3;
3161            sub = 0xe2; /* B?A:C */
3162        } else {
3163            tcg_out_mov(s, type, a0, a3);
3164            sub = 0xb8; /* B?C:A */
3165        }
3166        goto gen_simd_imm8;
3167
3168    gen_simd_imm8:
3169        tcg_debug_assert(insn != OPC_UD2);
3170        if (type == TCG_TYPE_V256) {
3171            insn |= P_VEXL;
3172        }
3173        tcg_out_vex_modrm(s, insn, a0, a1, a2);
3174        tcg_out8(s, sub);
3175        break;
3176
3177    case INDEX_op_x86_vpblendvb_vec:
3178        insn = OPC_VPBLENDVB;
3179        if (type == TCG_TYPE_V256) {
3180            insn |= P_VEXL;
3181        }
3182        tcg_out_vex_modrm(s, insn, a0, a1, a2);
3183        tcg_out8(s, args[3] << 4);
3184        break;
3185
3186    case INDEX_op_x86_psrldq_vec:
3187        tcg_out_vex_modrm(s, OPC_GRP14, 3, a0, a1);
3188        tcg_out8(s, a2);
3189        break;
3190
3191    case INDEX_op_mov_vec:  /* Always emitted via tcg_out_mov.  */
3192    case INDEX_op_dup_vec:  /* Always emitted via tcg_out_dup_vec.  */
3193    default:
3194        g_assert_not_reached();
3195    }
3196}
3197
3198static TCGConstraintSetIndex tcg_target_op_def(TCGOpcode op)
3199{
3200    switch (op) {
3201    case INDEX_op_goto_ptr:
3202        return C_O0_I1(r);
3203
3204    case INDEX_op_ld8u_i32:
3205    case INDEX_op_ld8u_i64:
3206    case INDEX_op_ld8s_i32:
3207    case INDEX_op_ld8s_i64:
3208    case INDEX_op_ld16u_i32:
3209    case INDEX_op_ld16u_i64:
3210    case INDEX_op_ld16s_i32:
3211    case INDEX_op_ld16s_i64:
3212    case INDEX_op_ld_i32:
3213    case INDEX_op_ld32u_i64:
3214    case INDEX_op_ld32s_i64:
3215    case INDEX_op_ld_i64:
3216        return C_O1_I1(r, r);
3217
3218    case INDEX_op_st8_i32:
3219    case INDEX_op_st8_i64:
3220        return C_O0_I2(qi, r);
3221
3222    case INDEX_op_st16_i32:
3223    case INDEX_op_st16_i64:
3224    case INDEX_op_st_i32:
3225    case INDEX_op_st32_i64:
3226        return C_O0_I2(ri, r);
3227
3228    case INDEX_op_st_i64:
3229        return C_O0_I2(re, r);
3230
3231    case INDEX_op_add_i32:
3232    case INDEX_op_add_i64:
3233        return C_O1_I2(r, r, re);
3234
3235    case INDEX_op_sub_i32:
3236    case INDEX_op_sub_i64:
3237    case INDEX_op_mul_i32:
3238    case INDEX_op_mul_i64:
3239    case INDEX_op_or_i32:
3240    case INDEX_op_or_i64:
3241    case INDEX_op_xor_i32:
3242    case INDEX_op_xor_i64:
3243        return C_O1_I2(r, 0, re);
3244
3245    case INDEX_op_and_i32:
3246    case INDEX_op_and_i64:
3247        return C_O1_I2(r, 0, reZ);
3248
3249    case INDEX_op_andc_i32:
3250    case INDEX_op_andc_i64:
3251        return C_O1_I2(r, r, rI);
3252
3253    case INDEX_op_shl_i32:
3254    case INDEX_op_shl_i64:
3255    case INDEX_op_shr_i32:
3256    case INDEX_op_shr_i64:
3257    case INDEX_op_sar_i32:
3258    case INDEX_op_sar_i64:
3259        return have_bmi2 ? C_O1_I2(r, r, ri) : C_O1_I2(r, 0, ci);
3260
3261    case INDEX_op_rotl_i32:
3262    case INDEX_op_rotl_i64:
3263    case INDEX_op_rotr_i32:
3264    case INDEX_op_rotr_i64:
3265        return C_O1_I2(r, 0, ci);
3266
3267    case INDEX_op_brcond_i32:
3268    case INDEX_op_brcond_i64:
3269        return C_O0_I2(r, re);
3270
3271    case INDEX_op_bswap16_i32:
3272    case INDEX_op_bswap16_i64:
3273    case INDEX_op_bswap32_i32:
3274    case INDEX_op_bswap32_i64:
3275    case INDEX_op_bswap64_i64:
3276    case INDEX_op_neg_i32:
3277    case INDEX_op_neg_i64:
3278    case INDEX_op_not_i32:
3279    case INDEX_op_not_i64:
3280    case INDEX_op_extrh_i64_i32:
3281        return C_O1_I1(r, 0);
3282
3283    case INDEX_op_ext8s_i32:
3284    case INDEX_op_ext8s_i64:
3285    case INDEX_op_ext8u_i32:
3286    case INDEX_op_ext8u_i64:
3287        return C_O1_I1(r, q);
3288
3289    case INDEX_op_ext16s_i32:
3290    case INDEX_op_ext16s_i64:
3291    case INDEX_op_ext16u_i32:
3292    case INDEX_op_ext16u_i64:
3293    case INDEX_op_ext32s_i64:
3294    case INDEX_op_ext32u_i64:
3295    case INDEX_op_ext_i32_i64:
3296    case INDEX_op_extu_i32_i64:
3297    case INDEX_op_extrl_i64_i32:
3298    case INDEX_op_extract_i32:
3299    case INDEX_op_extract_i64:
3300    case INDEX_op_sextract_i32:
3301    case INDEX_op_ctpop_i32:
3302    case INDEX_op_ctpop_i64:
3303        return C_O1_I1(r, r);
3304
3305    case INDEX_op_extract2_i32:
3306    case INDEX_op_extract2_i64:
3307        return C_O1_I2(r, 0, r);
3308
3309    case INDEX_op_deposit_i32:
3310    case INDEX_op_deposit_i64:
3311        return C_O1_I2(Q, 0, Q);
3312
3313    case INDEX_op_setcond_i32:
3314    case INDEX_op_setcond_i64:
3315        return C_O1_I2(q, r, re);
3316
3317    case INDEX_op_movcond_i32:
3318    case INDEX_op_movcond_i64:
3319        return C_O1_I4(r, r, re, r, 0);
3320
3321    case INDEX_op_div2_i32:
3322    case INDEX_op_div2_i64:
3323    case INDEX_op_divu2_i32:
3324    case INDEX_op_divu2_i64:
3325        return C_O2_I3(a, d, 0, 1, r);
3326
3327    case INDEX_op_mulu2_i32:
3328    case INDEX_op_mulu2_i64:
3329    case INDEX_op_muls2_i32:
3330    case INDEX_op_muls2_i64:
3331        return C_O2_I2(a, d, a, r);
3332
3333    case INDEX_op_add2_i32:
3334    case INDEX_op_add2_i64:
3335    case INDEX_op_sub2_i32:
3336    case INDEX_op_sub2_i64:
3337        return C_O2_I4(r, r, 0, 1, re, re);
3338
3339    case INDEX_op_ctz_i32:
3340    case INDEX_op_ctz_i64:
3341        return have_bmi1 ? C_N1_I2(r, r, rW) : C_N1_I2(r, r, r);
3342
3343    case INDEX_op_clz_i32:
3344    case INDEX_op_clz_i64:
3345        return have_lzcnt ? C_N1_I2(r, r, rW) : C_N1_I2(r, r, r);
3346
3347    case INDEX_op_qemu_ld_a32_i32:
3348        return C_O1_I1(r, L);
3349    case INDEX_op_qemu_ld_a64_i32:
3350        return TCG_TARGET_REG_BITS == 64 ? C_O1_I1(r, L) : C_O1_I2(r, L, L);
3351
3352    case INDEX_op_qemu_st_a32_i32:
3353        return C_O0_I2(L, L);
3354    case INDEX_op_qemu_st_a64_i32:
3355        return TCG_TARGET_REG_BITS == 64 ? C_O0_I2(L, L) : C_O0_I3(L, L, L);
3356    case INDEX_op_qemu_st8_a32_i32:
3357        return C_O0_I2(s, L);
3358    case INDEX_op_qemu_st8_a64_i32:
3359        return TCG_TARGET_REG_BITS == 64 ? C_O0_I2(s, L) : C_O0_I3(s, L, L);
3360
3361    case INDEX_op_qemu_ld_a32_i64:
3362        return TCG_TARGET_REG_BITS == 64 ? C_O1_I1(r, L) : C_O2_I1(r, r, L);
3363    case INDEX_op_qemu_ld_a64_i64:
3364        return TCG_TARGET_REG_BITS == 64 ? C_O1_I1(r, L) : C_O2_I2(r, r, L, L);
3365
3366    case INDEX_op_qemu_st_a32_i64:
3367        return TCG_TARGET_REG_BITS == 64 ? C_O0_I2(L, L) : C_O0_I3(L, L, L);
3368    case INDEX_op_qemu_st_a64_i64:
3369        return TCG_TARGET_REG_BITS == 64 ? C_O0_I2(L, L) : C_O0_I4(L, L, L, L);
3370
3371    case INDEX_op_qemu_ld_a32_i128:
3372    case INDEX_op_qemu_ld_a64_i128:
3373        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
3374        return C_O2_I1(r, r, L);
3375    case INDEX_op_qemu_st_a32_i128:
3376    case INDEX_op_qemu_st_a64_i128:
3377        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
3378        return C_O0_I3(L, L, L);
3379
3380    case INDEX_op_brcond2_i32:
3381        return C_O0_I4(r, r, ri, ri);
3382
3383    case INDEX_op_setcond2_i32:
3384        return C_O1_I4(r, r, r, ri, ri);
3385
3386    case INDEX_op_ld_vec:
3387    case INDEX_op_dupm_vec:
3388        return C_O1_I1(x, r);
3389
3390    case INDEX_op_st_vec:
3391        return C_O0_I2(x, r);
3392
3393    case INDEX_op_add_vec:
3394    case INDEX_op_sub_vec:
3395    case INDEX_op_mul_vec:
3396    case INDEX_op_and_vec:
3397    case INDEX_op_or_vec:
3398    case INDEX_op_xor_vec:
3399    case INDEX_op_andc_vec:
3400    case INDEX_op_orc_vec:
3401    case INDEX_op_nand_vec:
3402    case INDEX_op_nor_vec:
3403    case INDEX_op_eqv_vec:
3404    case INDEX_op_ssadd_vec:
3405    case INDEX_op_usadd_vec:
3406    case INDEX_op_sssub_vec:
3407    case INDEX_op_ussub_vec:
3408    case INDEX_op_smin_vec:
3409    case INDEX_op_umin_vec:
3410    case INDEX_op_smax_vec:
3411    case INDEX_op_umax_vec:
3412    case INDEX_op_shlv_vec:
3413    case INDEX_op_shrv_vec:
3414    case INDEX_op_sarv_vec:
3415    case INDEX_op_rotlv_vec:
3416    case INDEX_op_rotrv_vec:
3417    case INDEX_op_shls_vec:
3418    case INDEX_op_shrs_vec:
3419    case INDEX_op_sars_vec:
3420    case INDEX_op_cmp_vec:
3421    case INDEX_op_x86_shufps_vec:
3422    case INDEX_op_x86_blend_vec:
3423    case INDEX_op_x86_packss_vec:
3424    case INDEX_op_x86_packus_vec:
3425    case INDEX_op_x86_vperm2i128_vec:
3426    case INDEX_op_x86_punpckl_vec:
3427    case INDEX_op_x86_punpckh_vec:
3428    case INDEX_op_x86_vpshldi_vec:
3429#if TCG_TARGET_REG_BITS == 32
3430    case INDEX_op_dup2_vec:
3431#endif
3432        return C_O1_I2(x, x, x);
3433
3434    case INDEX_op_abs_vec:
3435    case INDEX_op_dup_vec:
3436    case INDEX_op_not_vec:
3437    case INDEX_op_shli_vec:
3438    case INDEX_op_shri_vec:
3439    case INDEX_op_sari_vec:
3440    case INDEX_op_rotli_vec:
3441    case INDEX_op_x86_psrldq_vec:
3442        return C_O1_I1(x, x);
3443
3444    case INDEX_op_x86_vpshldv_vec:
3445    case INDEX_op_x86_vpshrdv_vec:
3446        return C_O1_I3(x, 0, x, x);
3447
3448    case INDEX_op_bitsel_vec:
3449    case INDEX_op_x86_vpblendvb_vec:
3450        return C_O1_I3(x, x, x, x);
3451
3452    default:
3453        g_assert_not_reached();
3454    }
3455}
3456
3457int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
3458{
3459    switch (opc) {
3460    case INDEX_op_add_vec:
3461    case INDEX_op_sub_vec:
3462    case INDEX_op_and_vec:
3463    case INDEX_op_or_vec:
3464    case INDEX_op_xor_vec:
3465    case INDEX_op_andc_vec:
3466    case INDEX_op_orc_vec:
3467    case INDEX_op_nand_vec:
3468    case INDEX_op_nor_vec:
3469    case INDEX_op_eqv_vec:
3470    case INDEX_op_not_vec:
3471    case INDEX_op_bitsel_vec:
3472        return 1;
3473    case INDEX_op_cmp_vec:
3474    case INDEX_op_cmpsel_vec:
3475        return -1;
3476
3477    case INDEX_op_rotli_vec:
3478        return have_avx512vl && vece >= MO_32 ? 1 : -1;
3479
3480    case INDEX_op_shli_vec:
3481    case INDEX_op_shri_vec:
3482        /* We must expand the operation for MO_8.  */
3483        return vece == MO_8 ? -1 : 1;
3484
3485    case INDEX_op_sari_vec:
3486        switch (vece) {
3487        case MO_8:
3488            return -1;
3489        case MO_16:
3490        case MO_32:
3491            return 1;
3492        case MO_64:
3493            if (have_avx512vl) {
3494                return 1;
3495            }
3496            /*
3497             * We can emulate this for MO_64, but it does not pay off
3498             * unless we're producing at least 4 values.
3499             */
3500            return type >= TCG_TYPE_V256 ? -1 : 0;
3501        }
3502        return 0;
3503
3504    case INDEX_op_shls_vec:
3505    case INDEX_op_shrs_vec:
3506        return vece >= MO_16;
3507    case INDEX_op_sars_vec:
3508        switch (vece) {
3509        case MO_16:
3510        case MO_32:
3511            return 1;
3512        case MO_64:
3513            return have_avx512vl;
3514        }
3515        return 0;
3516    case INDEX_op_rotls_vec:
3517        return vece >= MO_16 ? -1 : 0;
3518
3519    case INDEX_op_shlv_vec:
3520    case INDEX_op_shrv_vec:
3521        switch (vece) {
3522        case MO_16:
3523            return have_avx512bw;
3524        case MO_32:
3525        case MO_64:
3526            return have_avx2;
3527        }
3528        return 0;
3529    case INDEX_op_sarv_vec:
3530        switch (vece) {
3531        case MO_16:
3532            return have_avx512bw;
3533        case MO_32:
3534            return have_avx2;
3535        case MO_64:
3536            return have_avx512vl;
3537        }
3538        return 0;
3539    case INDEX_op_rotlv_vec:
3540    case INDEX_op_rotrv_vec:
3541        switch (vece) {
3542        case MO_16:
3543            return have_avx512vbmi2 ? -1 : 0;
3544        case MO_32:
3545        case MO_64:
3546            return have_avx512vl ? 1 : have_avx2 ? -1 : 0;
3547        }
3548        return 0;
3549
3550    case INDEX_op_mul_vec:
3551        switch (vece) {
3552        case MO_8:
3553            return -1;
3554        case MO_64:
3555            return have_avx512dq;
3556        }
3557        return 1;
3558
3559    case INDEX_op_ssadd_vec:
3560    case INDEX_op_usadd_vec:
3561    case INDEX_op_sssub_vec:
3562    case INDEX_op_ussub_vec:
3563        return vece <= MO_16;
3564    case INDEX_op_smin_vec:
3565    case INDEX_op_smax_vec:
3566    case INDEX_op_umin_vec:
3567    case INDEX_op_umax_vec:
3568    case INDEX_op_abs_vec:
3569        return vece <= MO_32 || have_avx512vl;
3570
3571    default:
3572        return 0;
3573    }
3574}
3575
3576static void expand_vec_shi(TCGType type, unsigned vece, TCGOpcode opc,
3577                           TCGv_vec v0, TCGv_vec v1, TCGArg imm)
3578{
3579    TCGv_vec t1, t2;
3580
3581    tcg_debug_assert(vece == MO_8);
3582
3583    t1 = tcg_temp_new_vec(type);
3584    t2 = tcg_temp_new_vec(type);
3585
3586    /*
3587     * Unpack to W, shift, and repack.  Tricky bits:
3588     * (1) Use punpck*bw x,x to produce DDCCBBAA,
3589     *     i.e. duplicate in other half of the 16-bit lane.
3590     * (2) For right-shift, add 8 so that the high half of the lane
3591     *     becomes zero.  For left-shift, and left-rotate, we must
3592     *     shift up and down again.
3593     * (3) Step 2 leaves high half zero such that PACKUSWB
3594     *     (pack with unsigned saturation) does not modify
3595     *     the quantity.
3596     */
3597    vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
3598              tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
3599    vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
3600              tcgv_vec_arg(t2), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
3601
3602    if (opc != INDEX_op_rotli_vec) {
3603        imm += 8;
3604    }
3605    if (opc == INDEX_op_shri_vec) {
3606        tcg_gen_shri_vec(MO_16, t1, t1, imm);
3607        tcg_gen_shri_vec(MO_16, t2, t2, imm);
3608    } else {
3609        tcg_gen_shli_vec(MO_16, t1, t1, imm);
3610        tcg_gen_shli_vec(MO_16, t2, t2, imm);
3611        tcg_gen_shri_vec(MO_16, t1, t1, 8);
3612        tcg_gen_shri_vec(MO_16, t2, t2, 8);
3613    }
3614
3615    vec_gen_3(INDEX_op_x86_packus_vec, type, MO_8,
3616              tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t2));
3617    tcg_temp_free_vec(t1);
3618    tcg_temp_free_vec(t2);
3619}
3620
3621static void expand_vec_sari(TCGType type, unsigned vece,
3622                            TCGv_vec v0, TCGv_vec v1, TCGArg imm)
3623{
3624    TCGv_vec t1, t2;
3625
3626    switch (vece) {
3627    case MO_8:
3628        /* Unpack to W, shift, and repack, as in expand_vec_shi.  */
3629        t1 = tcg_temp_new_vec(type);
3630        t2 = tcg_temp_new_vec(type);
3631        vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
3632                  tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
3633        vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
3634                  tcgv_vec_arg(t2), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
3635        tcg_gen_sari_vec(MO_16, t1, t1, imm + 8);
3636        tcg_gen_sari_vec(MO_16, t2, t2, imm + 8);
3637        vec_gen_3(INDEX_op_x86_packss_vec, type, MO_8,
3638                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t2));
3639        tcg_temp_free_vec(t1);
3640        tcg_temp_free_vec(t2);
3641        break;
3642
3643    case MO_64:
3644        t1 = tcg_temp_new_vec(type);
3645        if (imm <= 32) {
3646            /*
3647             * We can emulate a small sign extend by performing an arithmetic
3648             * 32-bit shift and overwriting the high half of a 64-bit logical
3649             * shift.  Note that the ISA says shift of 32 is valid, but TCG
3650             * does not, so we have to bound the smaller shift -- we get the
3651             * same result in the high half either way.
3652             */
3653            tcg_gen_sari_vec(MO_32, t1, v1, MIN(imm, 31));
3654            tcg_gen_shri_vec(MO_64, v0, v1, imm);
3655            vec_gen_4(INDEX_op_x86_blend_vec, type, MO_32,
3656                      tcgv_vec_arg(v0), tcgv_vec_arg(v0),
3657                      tcgv_vec_arg(t1), 0xaa);
3658        } else {
3659            /* Otherwise we will need to use a compare vs 0 to produce
3660             * the sign-extend, shift and merge.
3661             */
3662            tcg_gen_cmp_vec(TCG_COND_GT, MO_64, t1,
3663                            tcg_constant_vec(type, MO_64, 0), v1);
3664            tcg_gen_shri_vec(MO_64, v0, v1, imm);
3665            tcg_gen_shli_vec(MO_64, t1, t1, 64 - imm);
3666            tcg_gen_or_vec(MO_64, v0, v0, t1);
3667        }
3668        tcg_temp_free_vec(t1);
3669        break;
3670
3671    default:
3672        g_assert_not_reached();
3673    }
3674}
3675
3676static void expand_vec_rotli(TCGType type, unsigned vece,
3677                             TCGv_vec v0, TCGv_vec v1, TCGArg imm)
3678{
3679    TCGv_vec t;
3680
3681    if (vece == MO_8) {
3682        expand_vec_shi(type, vece, INDEX_op_rotli_vec, v0, v1, imm);
3683        return;
3684    }
3685
3686    if (have_avx512vbmi2) {
3687        vec_gen_4(INDEX_op_x86_vpshldi_vec, type, vece,
3688                  tcgv_vec_arg(v0), tcgv_vec_arg(v1), tcgv_vec_arg(v1), imm);
3689        return;
3690    }
3691
3692    t = tcg_temp_new_vec(type);
3693    tcg_gen_shli_vec(vece, t, v1, imm);
3694    tcg_gen_shri_vec(vece, v0, v1, (8 << vece) - imm);
3695    tcg_gen_or_vec(vece, v0, v0, t);
3696    tcg_temp_free_vec(t);
3697}
3698
3699static void expand_vec_rotv(TCGType type, unsigned vece, TCGv_vec v0,
3700                            TCGv_vec v1, TCGv_vec sh, bool right)
3701{
3702    TCGv_vec t;
3703
3704    if (have_avx512vbmi2) {
3705        vec_gen_4(right ? INDEX_op_x86_vpshrdv_vec : INDEX_op_x86_vpshldv_vec,
3706                  type, vece, tcgv_vec_arg(v0), tcgv_vec_arg(v1),
3707                  tcgv_vec_arg(v1), tcgv_vec_arg(sh));
3708        return;
3709    }
3710
3711    t = tcg_temp_new_vec(type);
3712    tcg_gen_dupi_vec(vece, t, 8 << vece);
3713    tcg_gen_sub_vec(vece, t, t, sh);
3714    if (right) {
3715        tcg_gen_shlv_vec(vece, t, v1, t);
3716        tcg_gen_shrv_vec(vece, v0, v1, sh);
3717    } else {
3718        tcg_gen_shrv_vec(vece, t, v1, t);
3719        tcg_gen_shlv_vec(vece, v0, v1, sh);
3720    }
3721    tcg_gen_or_vec(vece, v0, v0, t);
3722    tcg_temp_free_vec(t);
3723}
3724
3725static void expand_vec_rotls(TCGType type, unsigned vece,
3726                             TCGv_vec v0, TCGv_vec v1, TCGv_i32 lsh)
3727{
3728    TCGv_vec t = tcg_temp_new_vec(type);
3729
3730    tcg_debug_assert(vece != MO_8);
3731
3732    if (vece >= MO_32 ? have_avx512vl : have_avx512vbmi2) {
3733        tcg_gen_dup_i32_vec(vece, t, lsh);
3734        if (vece >= MO_32) {
3735            tcg_gen_rotlv_vec(vece, v0, v1, t);
3736        } else {
3737            expand_vec_rotv(type, vece, v0, v1, t, false);
3738        }
3739    } else {
3740        TCGv_i32 rsh = tcg_temp_new_i32();
3741
3742        tcg_gen_neg_i32(rsh, lsh);
3743        tcg_gen_andi_i32(rsh, rsh, (8 << vece) - 1);
3744        tcg_gen_shls_vec(vece, t, v1, lsh);
3745        tcg_gen_shrs_vec(vece, v0, v1, rsh);
3746        tcg_gen_or_vec(vece, v0, v0, t);
3747
3748        tcg_temp_free_i32(rsh);
3749    }
3750
3751    tcg_temp_free_vec(t);
3752}
3753
3754static void expand_vec_mul(TCGType type, unsigned vece,
3755                           TCGv_vec v0, TCGv_vec v1, TCGv_vec v2)
3756{
3757    TCGv_vec t1, t2, t3, t4, zero;
3758
3759    tcg_debug_assert(vece == MO_8);
3760
3761    /*
3762     * Unpack v1 bytes to words, 0 | x.
3763     * Unpack v2 bytes to words, y | 0.
3764     * This leaves the 8-bit result, x * y, with 8 bits of right padding.
3765     * Shift logical right by 8 bits to clear the high 8 bytes before
3766     * using an unsigned saturated pack.
3767     *
3768     * The difference between the V64, V128 and V256 cases is merely how
3769     * we distribute the expansion between temporaries.
3770     */
3771    switch (type) {
3772    case TCG_TYPE_V64:
3773        t1 = tcg_temp_new_vec(TCG_TYPE_V128);
3774        t2 = tcg_temp_new_vec(TCG_TYPE_V128);
3775        zero = tcg_constant_vec(TCG_TYPE_V128, MO_8, 0);
3776        vec_gen_3(INDEX_op_x86_punpckl_vec, TCG_TYPE_V128, MO_8,
3777                  tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
3778        vec_gen_3(INDEX_op_x86_punpckl_vec, TCG_TYPE_V128, MO_8,
3779                  tcgv_vec_arg(t2), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
3780        tcg_gen_mul_vec(MO_16, t1, t1, t2);
3781        tcg_gen_shri_vec(MO_16, t1, t1, 8);
3782        vec_gen_3(INDEX_op_x86_packus_vec, TCG_TYPE_V128, MO_8,
3783                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t1));
3784        tcg_temp_free_vec(t1);
3785        tcg_temp_free_vec(t2);
3786        break;
3787
3788    case TCG_TYPE_V128:
3789    case TCG_TYPE_V256:
3790        t1 = tcg_temp_new_vec(type);
3791        t2 = tcg_temp_new_vec(type);
3792        t3 = tcg_temp_new_vec(type);
3793        t4 = tcg_temp_new_vec(type);
3794        zero = tcg_constant_vec(TCG_TYPE_V128, MO_8, 0);
3795        vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
3796                  tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
3797        vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
3798                  tcgv_vec_arg(t2), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
3799        vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
3800                  tcgv_vec_arg(t3), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
3801        vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
3802                  tcgv_vec_arg(t4), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
3803        tcg_gen_mul_vec(MO_16, t1, t1, t2);
3804        tcg_gen_mul_vec(MO_16, t3, t3, t4);
3805        tcg_gen_shri_vec(MO_16, t1, t1, 8);
3806        tcg_gen_shri_vec(MO_16, t3, t3, 8);
3807        vec_gen_3(INDEX_op_x86_packus_vec, type, MO_8,
3808                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t3));
3809        tcg_temp_free_vec(t1);
3810        tcg_temp_free_vec(t2);
3811        tcg_temp_free_vec(t3);
3812        tcg_temp_free_vec(t4);
3813        break;
3814
3815    default:
3816        g_assert_not_reached();
3817    }
3818}
3819
3820static bool expand_vec_cmp_noinv(TCGType type, unsigned vece, TCGv_vec v0,
3821                                 TCGv_vec v1, TCGv_vec v2, TCGCond cond)
3822{
3823    enum {
3824        NEED_INV  = 1,
3825        NEED_SWAP = 2,
3826        NEED_BIAS = 4,
3827        NEED_UMIN = 8,
3828        NEED_UMAX = 16,
3829    };
3830    TCGv_vec t1, t2, t3;
3831    uint8_t fixup;
3832
3833    switch (cond) {
3834    case TCG_COND_EQ:
3835    case TCG_COND_GT:
3836        fixup = 0;
3837        break;
3838    case TCG_COND_NE:
3839    case TCG_COND_LE:
3840        fixup = NEED_INV;
3841        break;
3842    case TCG_COND_LT:
3843        fixup = NEED_SWAP;
3844        break;
3845    case TCG_COND_GE:
3846        fixup = NEED_SWAP | NEED_INV;
3847        break;
3848    case TCG_COND_LEU:
3849        if (tcg_can_emit_vec_op(INDEX_op_umin_vec, type, vece)) {
3850            fixup = NEED_UMIN;
3851        } else {
3852            fixup = NEED_BIAS | NEED_INV;
3853        }
3854        break;
3855    case TCG_COND_GTU:
3856        if (tcg_can_emit_vec_op(INDEX_op_umin_vec, type, vece)) {
3857            fixup = NEED_UMIN | NEED_INV;
3858        } else {
3859            fixup = NEED_BIAS;
3860        }
3861        break;
3862    case TCG_COND_GEU:
3863        if (tcg_can_emit_vec_op(INDEX_op_umax_vec, type, vece)) {
3864            fixup = NEED_UMAX;
3865        } else {
3866            fixup = NEED_BIAS | NEED_SWAP | NEED_INV;
3867        }
3868        break;
3869    case TCG_COND_LTU:
3870        if (tcg_can_emit_vec_op(INDEX_op_umax_vec, type, vece)) {
3871            fixup = NEED_UMAX | NEED_INV;
3872        } else {
3873            fixup = NEED_BIAS | NEED_SWAP;
3874        }
3875        break;
3876    default:
3877        g_assert_not_reached();
3878    }
3879
3880    if (fixup & NEED_INV) {
3881        cond = tcg_invert_cond(cond);
3882    }
3883    if (fixup & NEED_SWAP) {
3884        t1 = v1, v1 = v2, v2 = t1;
3885        cond = tcg_swap_cond(cond);
3886    }
3887
3888    t1 = t2 = NULL;
3889    if (fixup & (NEED_UMIN | NEED_UMAX)) {
3890        t1 = tcg_temp_new_vec(type);
3891        if (fixup & NEED_UMIN) {
3892            tcg_gen_umin_vec(vece, t1, v1, v2);
3893        } else {
3894            tcg_gen_umax_vec(vece, t1, v1, v2);
3895        }
3896        v2 = t1;
3897        cond = TCG_COND_EQ;
3898    } else if (fixup & NEED_BIAS) {
3899        t1 = tcg_temp_new_vec(type);
3900        t2 = tcg_temp_new_vec(type);
3901        t3 = tcg_constant_vec(type, vece, 1ull << ((8 << vece) - 1));
3902        tcg_gen_sub_vec(vece, t1, v1, t3);
3903        tcg_gen_sub_vec(vece, t2, v2, t3);
3904        v1 = t1;
3905        v2 = t2;
3906        cond = tcg_signed_cond(cond);
3907    }
3908
3909    tcg_debug_assert(cond == TCG_COND_EQ || cond == TCG_COND_GT);
3910    /* Expand directly; do not recurse.  */
3911    vec_gen_4(INDEX_op_cmp_vec, type, vece,
3912              tcgv_vec_arg(v0), tcgv_vec_arg(v1), tcgv_vec_arg(v2), cond);
3913
3914    if (t1) {
3915        tcg_temp_free_vec(t1);
3916        if (t2) {
3917            tcg_temp_free_vec(t2);
3918        }
3919    }
3920    return fixup & NEED_INV;
3921}
3922
3923static void expand_vec_cmp(TCGType type, unsigned vece, TCGv_vec v0,
3924                           TCGv_vec v1, TCGv_vec v2, TCGCond cond)
3925{
3926    if (expand_vec_cmp_noinv(type, vece, v0, v1, v2, cond)) {
3927        tcg_gen_not_vec(vece, v0, v0);
3928    }
3929}
3930
3931static void expand_vec_cmpsel(TCGType type, unsigned vece, TCGv_vec v0,
3932                              TCGv_vec c1, TCGv_vec c2,
3933                              TCGv_vec v3, TCGv_vec v4, TCGCond cond)
3934{
3935    TCGv_vec t = tcg_temp_new_vec(type);
3936
3937    if (expand_vec_cmp_noinv(type, vece, t, c1, c2, cond)) {
3938        /* Invert the sense of the compare by swapping arguments.  */
3939        TCGv_vec x;
3940        x = v3, v3 = v4, v4 = x;
3941    }
3942    vec_gen_4(INDEX_op_x86_vpblendvb_vec, type, vece,
3943              tcgv_vec_arg(v0), tcgv_vec_arg(v4),
3944              tcgv_vec_arg(v3), tcgv_vec_arg(t));
3945    tcg_temp_free_vec(t);
3946}
3947
3948void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
3949                       TCGArg a0, ...)
3950{
3951    va_list va;
3952    TCGArg a2;
3953    TCGv_vec v0, v1, v2, v3, v4;
3954
3955    va_start(va, a0);
3956    v0 = temp_tcgv_vec(arg_temp(a0));
3957    v1 = temp_tcgv_vec(arg_temp(va_arg(va, TCGArg)));
3958    a2 = va_arg(va, TCGArg);
3959
3960    switch (opc) {
3961    case INDEX_op_shli_vec:
3962    case INDEX_op_shri_vec:
3963        expand_vec_shi(type, vece, opc, v0, v1, a2);
3964        break;
3965
3966    case INDEX_op_sari_vec:
3967        expand_vec_sari(type, vece, v0, v1, a2);
3968        break;
3969
3970    case INDEX_op_rotli_vec:
3971        expand_vec_rotli(type, vece, v0, v1, a2);
3972        break;
3973
3974    case INDEX_op_rotls_vec:
3975        expand_vec_rotls(type, vece, v0, v1, temp_tcgv_i32(arg_temp(a2)));
3976        break;
3977
3978    case INDEX_op_rotlv_vec:
3979        v2 = temp_tcgv_vec(arg_temp(a2));
3980        expand_vec_rotv(type, vece, v0, v1, v2, false);
3981        break;
3982    case INDEX_op_rotrv_vec:
3983        v2 = temp_tcgv_vec(arg_temp(a2));
3984        expand_vec_rotv(type, vece, v0, v1, v2, true);
3985        break;
3986
3987    case INDEX_op_mul_vec:
3988        v2 = temp_tcgv_vec(arg_temp(a2));
3989        expand_vec_mul(type, vece, v0, v1, v2);
3990        break;
3991
3992    case INDEX_op_cmp_vec:
3993        v2 = temp_tcgv_vec(arg_temp(a2));
3994        expand_vec_cmp(type, vece, v0, v1, v2, va_arg(va, TCGArg));
3995        break;
3996
3997    case INDEX_op_cmpsel_vec:
3998        v2 = temp_tcgv_vec(arg_temp(a2));
3999        v3 = temp_tcgv_vec(arg_temp(va_arg(va, TCGArg)));
4000        v4 = temp_tcgv_vec(arg_temp(va_arg(va, TCGArg)));
4001        expand_vec_cmpsel(type, vece, v0, v1, v2, v3, v4, va_arg(va, TCGArg));
4002        break;
4003
4004    default:
4005        break;
4006    }
4007
4008    va_end(va);
4009}
4010
4011static const int tcg_target_callee_save_regs[] = {
4012#if TCG_TARGET_REG_BITS == 64
4013    TCG_REG_RBP,
4014    TCG_REG_RBX,
4015#if defined(_WIN64)
4016    TCG_REG_RDI,
4017    TCG_REG_RSI,
4018#endif
4019    TCG_REG_R12,
4020    TCG_REG_R13,
4021    TCG_REG_R14, /* Currently used for the global env. */
4022    TCG_REG_R15,
4023#else
4024    TCG_REG_EBP, /* Currently used for the global env. */
4025    TCG_REG_EBX,
4026    TCG_REG_ESI,
4027    TCG_REG_EDI,
4028#endif
4029};
4030
4031/* Compute frame size via macros, to share between tcg_target_qemu_prologue
4032   and tcg_register_jit.  */
4033
4034#define PUSH_SIZE \
4035    ((1 + ARRAY_SIZE(tcg_target_callee_save_regs)) \
4036     * (TCG_TARGET_REG_BITS / 8))
4037
4038#define FRAME_SIZE \
4039    ((PUSH_SIZE \
4040      + TCG_STATIC_CALL_ARGS_SIZE \
4041      + CPU_TEMP_BUF_NLONGS * sizeof(long) \
4042      + TCG_TARGET_STACK_ALIGN - 1) \
4043     & ~(TCG_TARGET_STACK_ALIGN - 1))
4044
4045/* Generate global QEMU prologue and epilogue code */
4046static void tcg_target_qemu_prologue(TCGContext *s)
4047{
4048    int i, stack_addend;
4049
4050    /* TB prologue */
4051
4052    /* Reserve some stack space, also for TCG temps.  */
4053    stack_addend = FRAME_SIZE - PUSH_SIZE;
4054    tcg_set_frame(s, TCG_REG_CALL_STACK, TCG_STATIC_CALL_ARGS_SIZE,
4055                  CPU_TEMP_BUF_NLONGS * sizeof(long));
4056
4057    /* Save all callee saved registers.  */
4058    for (i = 0; i < ARRAY_SIZE(tcg_target_callee_save_regs); i++) {
4059        tcg_out_push(s, tcg_target_callee_save_regs[i]);
4060    }
4061
4062#if TCG_TARGET_REG_BITS == 32
4063    tcg_out_ld(s, TCG_TYPE_PTR, TCG_AREG0, TCG_REG_ESP,
4064               (ARRAY_SIZE(tcg_target_callee_save_regs) + 1) * 4);
4065    tcg_out_addi(s, TCG_REG_ESP, -stack_addend);
4066    /* jmp *tb.  */
4067    tcg_out_modrm_offset(s, OPC_GRP5, EXT5_JMPN_Ev, TCG_REG_ESP,
4068                         (ARRAY_SIZE(tcg_target_callee_save_regs) + 2) * 4
4069                         + stack_addend);
4070#else
4071# if !defined(CONFIG_SOFTMMU)
4072    if (guest_base) {
4073        int seg = setup_guest_base_seg();
4074        if (seg != 0) {
4075            x86_guest_base.seg = seg;
4076        } else if (guest_base == (int32_t)guest_base) {
4077            x86_guest_base.ofs = guest_base;
4078        } else {
4079            /* Choose R12 because, as a base, it requires a SIB byte. */
4080            x86_guest_base.index = TCG_REG_R12;
4081            tcg_out_movi(s, TCG_TYPE_PTR, x86_guest_base.index, guest_base);
4082            tcg_regset_set_reg(s->reserved_regs, x86_guest_base.index);
4083        }
4084    }
4085# endif
4086    tcg_out_mov(s, TCG_TYPE_PTR, TCG_AREG0, tcg_target_call_iarg_regs[0]);
4087    tcg_out_addi(s, TCG_REG_ESP, -stack_addend);
4088    /* jmp *tb.  */
4089    tcg_out_modrm(s, OPC_GRP5, EXT5_JMPN_Ev, tcg_target_call_iarg_regs[1]);
4090#endif
4091
4092    /*
4093     * Return path for goto_ptr. Set return value to 0, a-la exit_tb,
4094     * and fall through to the rest of the epilogue.
4095     */
4096    tcg_code_gen_epilogue = tcg_splitwx_to_rx(s->code_ptr);
4097    tcg_out_movi(s, TCG_TYPE_REG, TCG_REG_EAX, 0);
4098
4099    /* TB epilogue */
4100    tb_ret_addr = tcg_splitwx_to_rx(s->code_ptr);
4101
4102    tcg_out_addi(s, TCG_REG_CALL_STACK, stack_addend);
4103
4104    if (have_avx2) {
4105        tcg_out_vex_opc(s, OPC_VZEROUPPER, 0, 0, 0, 0);
4106    }
4107    for (i = ARRAY_SIZE(tcg_target_callee_save_regs) - 1; i >= 0; i--) {
4108        tcg_out_pop(s, tcg_target_callee_save_regs[i]);
4109    }
4110    tcg_out_opc(s, OPC_RET, 0, 0, 0);
4111}
4112
4113static void tcg_out_nop_fill(tcg_insn_unit *p, int count)
4114{
4115    memset(p, 0x90, count);
4116}
4117
4118static void tcg_target_init(TCGContext *s)
4119{
4120    tcg_target_available_regs[TCG_TYPE_I32] = ALL_GENERAL_REGS;
4121    if (TCG_TARGET_REG_BITS == 64) {
4122        tcg_target_available_regs[TCG_TYPE_I64] = ALL_GENERAL_REGS;
4123    }
4124    if (have_avx1) {
4125        tcg_target_available_regs[TCG_TYPE_V64] = ALL_VECTOR_REGS;
4126        tcg_target_available_regs[TCG_TYPE_V128] = ALL_VECTOR_REGS;
4127    }
4128    if (have_avx2) {
4129        tcg_target_available_regs[TCG_TYPE_V256] = ALL_VECTOR_REGS;
4130    }
4131
4132    tcg_target_call_clobber_regs = ALL_VECTOR_REGS;
4133    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_EAX);
4134    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_EDX);
4135    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_ECX);
4136    if (TCG_TARGET_REG_BITS == 64) {
4137#if !defined(_WIN64)
4138        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_RDI);
4139        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_RSI);
4140#endif
4141        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R8);
4142        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R9);
4143        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R10);
4144        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R11);
4145    }
4146
4147    s->reserved_regs = 0;
4148    tcg_regset_set_reg(s->reserved_regs, TCG_REG_CALL_STACK);
4149    tcg_regset_set_reg(s->reserved_regs, TCG_TMP_VEC);
4150#ifdef _WIN64
4151    /* These are call saved, and we don't save them, so don't use them. */
4152    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM6);
4153    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM7);
4154    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM8);
4155    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM9);
4156    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM10);
4157    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM11);
4158    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM12);
4159    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM13);
4160    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM14);
4161    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM15);
4162#endif
4163}
4164
4165typedef struct {
4166    DebugFrameHeader h;
4167    uint8_t fde_def_cfa[4];
4168    uint8_t fde_reg_ofs[14];
4169} DebugFrame;
4170
4171/* We're expecting a 2 byte uleb128 encoded value.  */
4172QEMU_BUILD_BUG_ON(FRAME_SIZE >= (1 << 14));
4173
4174#if !defined(__ELF__)
4175    /* Host machine without ELF. */
4176#elif TCG_TARGET_REG_BITS == 64
4177#define ELF_HOST_MACHINE EM_X86_64
4178static const DebugFrame debug_frame = {
4179    .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
4180    .h.cie.id = -1,
4181    .h.cie.version = 1,
4182    .h.cie.code_align = 1,
4183    .h.cie.data_align = 0x78,             /* sleb128 -8 */
4184    .h.cie.return_column = 16,
4185
4186    /* Total FDE size does not include the "len" member.  */
4187    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
4188
4189    .fde_def_cfa = {
4190        12, 7,                          /* DW_CFA_def_cfa %rsp, ... */
4191        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
4192        (FRAME_SIZE >> 7)
4193    },
4194    .fde_reg_ofs = {
4195        0x90, 1,                        /* DW_CFA_offset, %rip, -8 */
4196        /* The following ordering must match tcg_target_callee_save_regs.  */
4197        0x86, 2,                        /* DW_CFA_offset, %rbp, -16 */
4198        0x83, 3,                        /* DW_CFA_offset, %rbx, -24 */
4199        0x8c, 4,                        /* DW_CFA_offset, %r12, -32 */
4200        0x8d, 5,                        /* DW_CFA_offset, %r13, -40 */
4201        0x8e, 6,                        /* DW_CFA_offset, %r14, -48 */
4202        0x8f, 7,                        /* DW_CFA_offset, %r15, -56 */
4203    }
4204};
4205#else
4206#define ELF_HOST_MACHINE EM_386
4207static const DebugFrame debug_frame = {
4208    .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
4209    .h.cie.id = -1,
4210    .h.cie.version = 1,
4211    .h.cie.code_align = 1,
4212    .h.cie.data_align = 0x7c,             /* sleb128 -4 */
4213    .h.cie.return_column = 8,
4214
4215    /* Total FDE size does not include the "len" member.  */
4216    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
4217
4218    .fde_def_cfa = {
4219        12, 4,                          /* DW_CFA_def_cfa %esp, ... */
4220        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
4221        (FRAME_SIZE >> 7)
4222    },
4223    .fde_reg_ofs = {
4224        0x88, 1,                        /* DW_CFA_offset, %eip, -4 */
4225        /* The following ordering must match tcg_target_callee_save_regs.  */
4226        0x85, 2,                        /* DW_CFA_offset, %ebp, -8 */
4227        0x83, 3,                        /* DW_CFA_offset, %ebx, -12 */
4228        0x86, 4,                        /* DW_CFA_offset, %esi, -16 */
4229        0x87, 5,                        /* DW_CFA_offset, %edi, -20 */
4230    }
4231};
4232#endif
4233
4234#if defined(ELF_HOST_MACHINE)
4235void tcg_register_jit(const void *buf, size_t buf_size)
4236{
4237    tcg_register_jit_int(buf, buf_size, &debug_frame, sizeof(debug_frame));
4238}
4239#endif
4240