xref: /openbmc/qemu/tcg/i386/tcg-target.c.inc (revision 5f6d4f79)
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#if TCG_TARGET_REG_BITS == 64
148# define ALL_GENERAL_REGS      0x0000ffffu
149# define ALL_VECTOR_REGS       0xffff0000u
150# define ALL_BYTEL_REGS        ALL_GENERAL_REGS
151#else
152# define ALL_GENERAL_REGS      0x000000ffu
153# define ALL_VECTOR_REGS       0x00ff0000u
154# define ALL_BYTEL_REGS        0x0000000fu
155#endif
156#ifdef CONFIG_SOFTMMU
157# define SOFTMMU_RESERVE_REGS  ((1 << TCG_REG_L0) | (1 << TCG_REG_L1))
158#else
159# define SOFTMMU_RESERVE_REGS  0
160#endif
161
162/* For 64-bit, we always know that CMOV is available.  */
163#if TCG_TARGET_REG_BITS == 64
164# define have_cmov      true
165#else
166# define have_cmov      (cpuinfo & CPUINFO_CMOV)
167#endif
168#define have_bmi2       (cpuinfo & CPUINFO_BMI2)
169#define have_lzcnt      (cpuinfo & CPUINFO_LZCNT)
170
171static const tcg_insn_unit *tb_ret_addr;
172
173static bool patch_reloc(tcg_insn_unit *code_ptr, int type,
174                        intptr_t value, intptr_t addend)
175{
176    value += addend;
177    switch(type) {
178    case R_386_PC32:
179        value -= (uintptr_t)tcg_splitwx_to_rx(code_ptr);
180        if (value != (int32_t)value) {
181            return false;
182        }
183        /* FALLTHRU */
184    case R_386_32:
185        tcg_patch32(code_ptr, value);
186        break;
187    case R_386_PC8:
188        value -= (uintptr_t)tcg_splitwx_to_rx(code_ptr);
189        if (value != (int8_t)value) {
190            return false;
191        }
192        tcg_patch8(code_ptr, value);
193        break;
194    default:
195        g_assert_not_reached();
196    }
197    return true;
198}
199
200/* test if a constant matches the constraint */
201static bool tcg_target_const_match(int64_t val, TCGType type, int ct, int vece)
202{
203    if (ct & TCG_CT_CONST) {
204        return 1;
205    }
206    if (type == TCG_TYPE_I32) {
207        if (ct & (TCG_CT_CONST_S32 | TCG_CT_CONST_U32 | TCG_CT_CONST_I32)) {
208            return 1;
209        }
210    } else {
211        if ((ct & TCG_CT_CONST_S32) && val == (int32_t)val) {
212            return 1;
213        }
214        if ((ct & TCG_CT_CONST_U32) && val == (uint32_t)val) {
215            return 1;
216        }
217        if ((ct & TCG_CT_CONST_I32) && ~val == (int32_t)~val) {
218            return 1;
219        }
220    }
221    if ((ct & TCG_CT_CONST_WSZ) && val == (type == TCG_TYPE_I32 ? 32 : 64)) {
222        return 1;
223    }
224    return 0;
225}
226
227# define LOWREGMASK(x)	((x) & 7)
228
229#define P_EXT		0x100		/* 0x0f opcode prefix */
230#define P_EXT38         0x200           /* 0x0f 0x38 opcode prefix */
231#define P_DATA16        0x400           /* 0x66 opcode prefix */
232#define P_VEXW          0x1000          /* Set VEX.W = 1 */
233#if TCG_TARGET_REG_BITS == 64
234# define P_REXW         P_VEXW          /* Set REX.W = 1; match VEXW */
235# define P_REXB_R       0x2000          /* REG field as byte register */
236# define P_REXB_RM      0x4000          /* R/M field as byte register */
237# define P_GS           0x8000          /* gs segment override */
238#else
239# define P_REXW		0
240# define P_REXB_R	0
241# define P_REXB_RM	0
242# define P_GS           0
243#endif
244#define P_EXT3A         0x10000         /* 0x0f 0x3a opcode prefix */
245#define P_SIMDF3        0x20000         /* 0xf3 opcode prefix */
246#define P_SIMDF2        0x40000         /* 0xf2 opcode prefix */
247#define P_VEXL          0x80000         /* Set VEX.L = 1 */
248#define P_EVEX          0x100000        /* Requires EVEX encoding */
249
250#define OPC_ARITH_EvIz	(0x81)
251#define OPC_ARITH_EvIb	(0x83)
252#define OPC_ARITH_GvEv	(0x03)		/* ... plus (ARITH_FOO << 3) */
253#define OPC_ANDN        (0xf2 | P_EXT38)
254#define OPC_ADD_GvEv	(OPC_ARITH_GvEv | (ARITH_ADD << 3))
255#define OPC_AND_GvEv    (OPC_ARITH_GvEv | (ARITH_AND << 3))
256#define OPC_BLENDPS     (0x0c | P_EXT3A | P_DATA16)
257#define OPC_BSF         (0xbc | P_EXT)
258#define OPC_BSR         (0xbd | P_EXT)
259#define OPC_BSWAP	(0xc8 | P_EXT)
260#define OPC_CALL_Jz	(0xe8)
261#define OPC_CMOVCC      (0x40 | P_EXT)  /* ... plus condition code */
262#define OPC_CMP_GvEv	(OPC_ARITH_GvEv | (ARITH_CMP << 3))
263#define OPC_DEC_r32	(0x48)
264#define OPC_IMUL_GvEv	(0xaf | P_EXT)
265#define OPC_IMUL_GvEvIb	(0x6b)
266#define OPC_IMUL_GvEvIz	(0x69)
267#define OPC_INC_r32	(0x40)
268#define OPC_JCC_long	(0x80 | P_EXT)	/* ... plus condition code */
269#define OPC_JCC_short	(0x70)		/* ... plus condition code */
270#define OPC_JMP_long	(0xe9)
271#define OPC_JMP_short	(0xeb)
272#define OPC_LEA         (0x8d)
273#define OPC_LZCNT       (0xbd | P_EXT | P_SIMDF3)
274#define OPC_MOVB_EvGv	(0x88)		/* stores, more or less */
275#define OPC_MOVL_EvGv	(0x89)		/* stores, more or less */
276#define OPC_MOVL_GvEv	(0x8b)		/* loads, more or less */
277#define OPC_MOVB_EvIz   (0xc6)
278#define OPC_MOVL_EvIz	(0xc7)
279#define OPC_MOVB_Ib     (0xb0)
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    if (opc & P_GS) {
599        tcg_out8(s, 0x65);
600    }
601    /* Use the two byte form if possible, which cannot encode
602       VEX.W, VEX.B, VEX.X, or an m-mmmm field other than P_EXT.  */
603    if ((opc & (P_EXT | P_EXT38 | P_EXT3A | P_VEXW)) == P_EXT
604        && ((rm | index) & 8) == 0) {
605        /* Two byte VEX prefix.  */
606        tcg_out8(s, 0xc5);
607
608        tmp = (r & 8 ? 0 : 0x80);              /* VEX.R */
609    } else {
610        /* Three byte VEX prefix.  */
611        tcg_out8(s, 0xc4);
612
613        /* VEX.m-mmmm */
614        if (opc & P_EXT3A) {
615            tmp = 3;
616        } else if (opc & P_EXT38) {
617            tmp = 2;
618        } else if (opc & P_EXT) {
619            tmp = 1;
620        } else {
621            g_assert_not_reached();
622        }
623        tmp |= (r & 8 ? 0 : 0x80);             /* VEX.R */
624        tmp |= (index & 8 ? 0 : 0x40);         /* VEX.X */
625        tmp |= (rm & 8 ? 0 : 0x20);            /* VEX.B */
626        tcg_out8(s, tmp);
627
628        tmp = (opc & P_VEXW ? 0x80 : 0);       /* VEX.W */
629    }
630
631    tmp |= (opc & P_VEXL ? 0x04 : 0);      /* VEX.L */
632    /* VEX.pp */
633    if (opc & P_DATA16) {
634        tmp |= 1;                          /* 0x66 */
635    } else if (opc & P_SIMDF3) {
636        tmp |= 2;                          /* 0xf3 */
637    } else if (opc & P_SIMDF2) {
638        tmp |= 3;                          /* 0xf2 */
639    }
640    tmp |= (~v & 15) << 3;                 /* VEX.vvvv */
641    tcg_out8(s, tmp);
642    tcg_out8(s, opc);
643}
644
645static void tcg_out_evex_opc(TCGContext *s, int opc, int r, int v,
646                             int rm, int index)
647{
648    /* The entire 4-byte evex prefix; with R' and V' set. */
649    uint32_t p = 0x08041062;
650    int mm, pp;
651
652    tcg_debug_assert(have_avx512vl);
653
654    /* EVEX.mm */
655    if (opc & P_EXT3A) {
656        mm = 3;
657    } else if (opc & P_EXT38) {
658        mm = 2;
659    } else if (opc & P_EXT) {
660        mm = 1;
661    } else {
662        g_assert_not_reached();
663    }
664
665    /* EVEX.pp */
666    if (opc & P_DATA16) {
667        pp = 1;                          /* 0x66 */
668    } else if (opc & P_SIMDF3) {
669        pp = 2;                          /* 0xf3 */
670    } else if (opc & P_SIMDF2) {
671        pp = 3;                          /* 0xf2 */
672    } else {
673        pp = 0;
674    }
675
676    p = deposit32(p, 8, 2, mm);
677    p = deposit32(p, 13, 1, (rm & 8) == 0);             /* EVEX.RXB.B */
678    p = deposit32(p, 14, 1, (index & 8) == 0);          /* EVEX.RXB.X */
679    p = deposit32(p, 15, 1, (r & 8) == 0);              /* EVEX.RXB.R */
680    p = deposit32(p, 16, 2, pp);
681    p = deposit32(p, 19, 4, ~v);
682    p = deposit32(p, 23, 1, (opc & P_VEXW) != 0);
683    p = deposit32(p, 29, 2, (opc & P_VEXL) != 0);
684
685    tcg_out32(s, p);
686    tcg_out8(s, opc);
687}
688
689static void tcg_out_vex_modrm(TCGContext *s, int opc, int r, int v, int rm)
690{
691    if (opc & P_EVEX) {
692        tcg_out_evex_opc(s, opc, r, v, rm, 0);
693    } else {
694        tcg_out_vex_opc(s, opc, r, v, rm, 0);
695    }
696    tcg_out8(s, 0xc0 | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
697}
698
699/* Output an opcode with a full "rm + (index<<shift) + offset" address mode.
700   We handle either RM and INDEX missing with a negative value.  In 64-bit
701   mode for absolute addresses, ~RM is the size of the immediate operand
702   that will follow the instruction.  */
703
704static void tcg_out_sib_offset(TCGContext *s, int r, int rm, int index,
705                               int shift, intptr_t offset)
706{
707    int mod, len;
708
709    if (index < 0 && rm < 0) {
710        if (TCG_TARGET_REG_BITS == 64) {
711            /* Try for a rip-relative addressing mode.  This has replaced
712               the 32-bit-mode absolute addressing encoding.  */
713            intptr_t pc = (intptr_t)s->code_ptr + 5 + ~rm;
714            intptr_t disp = offset - pc;
715            if (disp == (int32_t)disp) {
716                tcg_out8(s, (LOWREGMASK(r) << 3) | 5);
717                tcg_out32(s, disp);
718                return;
719            }
720
721            /* Try for an absolute address encoding.  This requires the
722               use of the MODRM+SIB encoding and is therefore larger than
723               rip-relative addressing.  */
724            if (offset == (int32_t)offset) {
725                tcg_out8(s, (LOWREGMASK(r) << 3) | 4);
726                tcg_out8(s, (4 << 3) | 5);
727                tcg_out32(s, offset);
728                return;
729            }
730
731            /* ??? The memory isn't directly addressable.  */
732            g_assert_not_reached();
733        } else {
734            /* Absolute address.  */
735            tcg_out8(s, (r << 3) | 5);
736            tcg_out32(s, offset);
737            return;
738        }
739    }
740
741    /* Find the length of the immediate addend.  Note that the encoding
742       that would be used for (%ebp) indicates absolute addressing.  */
743    if (rm < 0) {
744        mod = 0, len = 4, rm = 5;
745    } else if (offset == 0 && LOWREGMASK(rm) != TCG_REG_EBP) {
746        mod = 0, len = 0;
747    } else if (offset == (int8_t)offset) {
748        mod = 0x40, len = 1;
749    } else {
750        mod = 0x80, len = 4;
751    }
752
753    /* Use a single byte MODRM format if possible.  Note that the encoding
754       that would be used for %esp is the escape to the two byte form.  */
755    if (index < 0 && LOWREGMASK(rm) != TCG_REG_ESP) {
756        /* Single byte MODRM format.  */
757        tcg_out8(s, mod | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
758    } else {
759        /* Two byte MODRM+SIB format.  */
760
761        /* Note that the encoding that would place %esp into the index
762           field indicates no index register.  In 64-bit mode, the REX.X
763           bit counts, so %r12 can be used as the index.  */
764        if (index < 0) {
765            index = 4;
766        } else {
767            tcg_debug_assert(index != TCG_REG_ESP);
768        }
769
770        tcg_out8(s, mod | (LOWREGMASK(r) << 3) | 4);
771        tcg_out8(s, (shift << 6) | (LOWREGMASK(index) << 3) | LOWREGMASK(rm));
772    }
773
774    if (len == 1) {
775        tcg_out8(s, offset);
776    } else if (len == 4) {
777        tcg_out32(s, offset);
778    }
779}
780
781static void tcg_out_modrm_sib_offset(TCGContext *s, int opc, int r, int rm,
782                                     int index, int shift, intptr_t offset)
783{
784    tcg_out_opc(s, opc, r, rm < 0 ? 0 : rm, index < 0 ? 0 : index);
785    tcg_out_sib_offset(s, r, rm, index, shift, offset);
786}
787
788static void tcg_out_vex_modrm_sib_offset(TCGContext *s, int opc, int r, int v,
789                                         int rm, int index, int shift,
790                                         intptr_t offset)
791{
792    tcg_out_vex_opc(s, opc, r, v, rm < 0 ? 0 : rm, index < 0 ? 0 : index);
793    tcg_out_sib_offset(s, r, rm, index, shift, offset);
794}
795
796/* A simplification of the above with no index or shift.  */
797static inline void tcg_out_modrm_offset(TCGContext *s, int opc, int r,
798                                        int rm, intptr_t offset)
799{
800    tcg_out_modrm_sib_offset(s, opc, r, rm, -1, 0, offset);
801}
802
803static inline void tcg_out_vex_modrm_offset(TCGContext *s, int opc, int r,
804                                            int v, int rm, intptr_t offset)
805{
806    tcg_out_vex_modrm_sib_offset(s, opc, r, v, rm, -1, 0, offset);
807}
808
809/* Output an opcode with an expected reference to the constant pool.  */
810static inline void tcg_out_modrm_pool(TCGContext *s, int opc, int r)
811{
812    tcg_out_opc(s, opc, r, 0, 0);
813    /* Absolute for 32-bit, pc-relative for 64-bit.  */
814    tcg_out8(s, LOWREGMASK(r) << 3 | 5);
815    tcg_out32(s, 0);
816}
817
818/* Output an opcode with an expected reference to the constant pool.  */
819static inline void tcg_out_vex_modrm_pool(TCGContext *s, int opc, int r)
820{
821    tcg_out_vex_opc(s, opc, r, 0, 0, 0);
822    /* Absolute for 32-bit, pc-relative for 64-bit.  */
823    tcg_out8(s, LOWREGMASK(r) << 3 | 5);
824    tcg_out32(s, 0);
825}
826
827/* Generate dest op= src.  Uses the same ARITH_* codes as tgen_arithi.  */
828static inline void tgen_arithr(TCGContext *s, int subop, int dest, int src)
829{
830    /* Propagate an opcode prefix, such as P_REXW.  */
831    int ext = subop & ~0x7;
832    subop &= 0x7;
833
834    tcg_out_modrm(s, OPC_ARITH_GvEv + (subop << 3) + ext, dest, src);
835}
836
837static bool tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
838{
839    int rexw = 0;
840
841    if (arg == ret) {
842        return true;
843    }
844    switch (type) {
845    case TCG_TYPE_I64:
846        rexw = P_REXW;
847        /* fallthru */
848    case TCG_TYPE_I32:
849        if (ret < 16) {
850            if (arg < 16) {
851                tcg_out_modrm(s, OPC_MOVL_GvEv + rexw, ret, arg);
852            } else {
853                tcg_out_vex_modrm(s, OPC_MOVD_EyVy + rexw, arg, 0, ret);
854            }
855        } else {
856            if (arg < 16) {
857                tcg_out_vex_modrm(s, OPC_MOVD_VyEy + rexw, ret, 0, arg);
858            } else {
859                tcg_out_vex_modrm(s, OPC_MOVQ_VqWq, ret, 0, arg);
860            }
861        }
862        break;
863
864    case TCG_TYPE_V64:
865        tcg_debug_assert(ret >= 16 && arg >= 16);
866        tcg_out_vex_modrm(s, OPC_MOVQ_VqWq, ret, 0, arg);
867        break;
868    case TCG_TYPE_V128:
869        tcg_debug_assert(ret >= 16 && arg >= 16);
870        tcg_out_vex_modrm(s, OPC_MOVDQA_VxWx, ret, 0, arg);
871        break;
872    case TCG_TYPE_V256:
873        tcg_debug_assert(ret >= 16 && arg >= 16);
874        tcg_out_vex_modrm(s, OPC_MOVDQA_VxWx | P_VEXL, ret, 0, arg);
875        break;
876
877    default:
878        g_assert_not_reached();
879    }
880    return true;
881}
882
883static const int avx2_dup_insn[4] = {
884    OPC_VPBROADCASTB, OPC_VPBROADCASTW,
885    OPC_VPBROADCASTD, OPC_VPBROADCASTQ,
886};
887
888static bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece,
889                            TCGReg r, TCGReg a)
890{
891    if (have_avx2) {
892        int vex_l = (type == TCG_TYPE_V256 ? P_VEXL : 0);
893        tcg_out_vex_modrm(s, avx2_dup_insn[vece] + vex_l, r, 0, a);
894    } else {
895        switch (vece) {
896        case MO_8:
897            /* ??? With zero in a register, use PSHUFB.  */
898            tcg_out_vex_modrm(s, OPC_PUNPCKLBW, r, a, a);
899            a = r;
900            /* FALLTHRU */
901        case MO_16:
902            tcg_out_vex_modrm(s, OPC_PUNPCKLWD, r, a, a);
903            a = r;
904            /* FALLTHRU */
905        case MO_32:
906            tcg_out_vex_modrm(s, OPC_PSHUFD, r, 0, a);
907            /* imm8 operand: all output lanes selected from input lane 0.  */
908            tcg_out8(s, 0);
909            break;
910        case MO_64:
911            tcg_out_vex_modrm(s, OPC_PUNPCKLQDQ, r, a, a);
912            break;
913        default:
914            g_assert_not_reached();
915        }
916    }
917    return true;
918}
919
920static bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece,
921                             TCGReg r, TCGReg base, intptr_t offset)
922{
923    if (have_avx2) {
924        int vex_l = (type == TCG_TYPE_V256 ? P_VEXL : 0);
925        tcg_out_vex_modrm_offset(s, avx2_dup_insn[vece] + vex_l,
926                                 r, 0, base, offset);
927    } else {
928        switch (vece) {
929        case MO_64:
930            tcg_out_vex_modrm_offset(s, OPC_MOVDDUP, r, 0, base, offset);
931            break;
932        case MO_32:
933            tcg_out_vex_modrm_offset(s, OPC_VBROADCASTSS, r, 0, base, offset);
934            break;
935        case MO_16:
936            tcg_out_vex_modrm_offset(s, OPC_VPINSRW, r, r, base, offset);
937            tcg_out8(s, 0); /* imm8 */
938            tcg_out_dup_vec(s, type, vece, r, r);
939            break;
940        case MO_8:
941            tcg_out_vex_modrm_offset(s, OPC_VPINSRB, r, r, base, offset);
942            tcg_out8(s, 0); /* imm8 */
943            tcg_out_dup_vec(s, type, vece, r, r);
944            break;
945        default:
946            g_assert_not_reached();
947        }
948    }
949    return true;
950}
951
952static void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece,
953                             TCGReg ret, int64_t arg)
954{
955    int vex_l = (type == TCG_TYPE_V256 ? P_VEXL : 0);
956
957    if (arg == 0) {
958        tcg_out_vex_modrm(s, OPC_PXOR, ret, ret, ret);
959        return;
960    }
961    if (arg == -1) {
962        tcg_out_vex_modrm(s, OPC_PCMPEQB + vex_l, ret, ret, ret);
963        return;
964    }
965
966    if (TCG_TARGET_REG_BITS == 32 && vece < MO_64) {
967        if (have_avx2) {
968            tcg_out_vex_modrm_pool(s, OPC_VPBROADCASTD + vex_l, ret);
969        } else {
970            tcg_out_vex_modrm_pool(s, OPC_VBROADCASTSS, ret);
971        }
972        new_pool_label(s, arg, R_386_32, s->code_ptr - 4, 0);
973    } else {
974        if (type == TCG_TYPE_V64) {
975            tcg_out_vex_modrm_pool(s, OPC_MOVQ_VqWq, ret);
976        } else if (have_avx2) {
977            tcg_out_vex_modrm_pool(s, OPC_VPBROADCASTQ + vex_l, ret);
978        } else {
979            tcg_out_vex_modrm_pool(s, OPC_MOVDDUP, ret);
980        }
981        if (TCG_TARGET_REG_BITS == 64) {
982            new_pool_label(s, arg, R_386_PC32, s->code_ptr - 4, -4);
983        } else {
984            new_pool_l2(s, R_386_32, s->code_ptr - 4, 0, arg, arg >> 32);
985        }
986    }
987}
988
989static void tcg_out_movi_vec(TCGContext *s, TCGType type,
990                             TCGReg ret, tcg_target_long arg)
991{
992    if (arg == 0) {
993        tcg_out_vex_modrm(s, OPC_PXOR, ret, ret, ret);
994        return;
995    }
996    if (arg == -1) {
997        tcg_out_vex_modrm(s, OPC_PCMPEQB, ret, ret, ret);
998        return;
999    }
1000
1001    int rexw = (type == TCG_TYPE_I32 ? 0 : P_REXW);
1002    tcg_out_vex_modrm_pool(s, OPC_MOVD_VyEy + rexw, ret);
1003    if (TCG_TARGET_REG_BITS == 64) {
1004        new_pool_label(s, arg, R_386_PC32, s->code_ptr - 4, -4);
1005    } else {
1006        new_pool_label(s, arg, R_386_32, s->code_ptr - 4, 0);
1007    }
1008}
1009
1010static void tcg_out_movi_int(TCGContext *s, TCGType type,
1011                             TCGReg ret, tcg_target_long arg)
1012{
1013    tcg_target_long diff;
1014
1015    if (arg == 0) {
1016        tgen_arithr(s, ARITH_XOR, ret, ret);
1017        return;
1018    }
1019    if (arg == (uint32_t)arg || type == TCG_TYPE_I32) {
1020        tcg_out_opc(s, OPC_MOVL_Iv + LOWREGMASK(ret), 0, ret, 0);
1021        tcg_out32(s, arg);
1022        return;
1023    }
1024    if (arg == (int32_t)arg) {
1025        tcg_out_modrm(s, OPC_MOVL_EvIz + P_REXW, 0, ret);
1026        tcg_out32(s, arg);
1027        return;
1028    }
1029
1030    /* Try a 7 byte pc-relative lea before the 10 byte movq.  */
1031    diff = tcg_pcrel_diff(s, (const void *)arg) - 7;
1032    if (diff == (int32_t)diff) {
1033        tcg_out_opc(s, OPC_LEA | P_REXW, ret, 0, 0);
1034        tcg_out8(s, (LOWREGMASK(ret) << 3) | 5);
1035        tcg_out32(s, diff);
1036        return;
1037    }
1038
1039    tcg_out_opc(s, OPC_MOVL_Iv + P_REXW + LOWREGMASK(ret), 0, ret, 0);
1040    tcg_out64(s, arg);
1041}
1042
1043static void tcg_out_movi(TCGContext *s, TCGType type,
1044                         TCGReg ret, tcg_target_long arg)
1045{
1046    switch (type) {
1047    case TCG_TYPE_I32:
1048#if TCG_TARGET_REG_BITS == 64
1049    case TCG_TYPE_I64:
1050#endif
1051        if (ret < 16) {
1052            tcg_out_movi_int(s, type, ret, arg);
1053        } else {
1054            tcg_out_movi_vec(s, type, ret, arg);
1055        }
1056        break;
1057    default:
1058        g_assert_not_reached();
1059    }
1060}
1061
1062static bool tcg_out_xchg(TCGContext *s, TCGType type, TCGReg r1, TCGReg r2)
1063{
1064    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1065    tcg_out_modrm(s, OPC_XCHG_EvGv + rexw, r1, r2);
1066    return true;
1067}
1068
1069static void tcg_out_addi_ptr(TCGContext *s, TCGReg rd, TCGReg rs,
1070                             tcg_target_long imm)
1071{
1072    /* This function is only used for passing structs by reference. */
1073    tcg_debug_assert(imm == (int32_t)imm);
1074    tcg_out_modrm_offset(s, OPC_LEA | P_REXW, rd, rs, imm);
1075}
1076
1077static inline void tcg_out_pushi(TCGContext *s, tcg_target_long val)
1078{
1079    if (val == (int8_t)val) {
1080        tcg_out_opc(s, OPC_PUSH_Ib, 0, 0, 0);
1081        tcg_out8(s, val);
1082    } else if (val == (int32_t)val) {
1083        tcg_out_opc(s, OPC_PUSH_Iv, 0, 0, 0);
1084        tcg_out32(s, val);
1085    } else {
1086        g_assert_not_reached();
1087    }
1088}
1089
1090static inline void tcg_out_mb(TCGContext *s, TCGArg a0)
1091{
1092    /* Given the strength of x86 memory ordering, we only need care for
1093       store-load ordering.  Experimentally, "lock orl $0,0(%esp)" is
1094       faster than "mfence", so don't bother with the sse insn.  */
1095    if (a0 & TCG_MO_ST_LD) {
1096        tcg_out8(s, 0xf0);
1097        tcg_out_modrm_offset(s, OPC_ARITH_EvIb, ARITH_OR, TCG_REG_ESP, 0);
1098        tcg_out8(s, 0);
1099    }
1100}
1101
1102static inline void tcg_out_push(TCGContext *s, int reg)
1103{
1104    tcg_out_opc(s, OPC_PUSH_r32 + LOWREGMASK(reg), 0, reg, 0);
1105}
1106
1107static inline void tcg_out_pop(TCGContext *s, int reg)
1108{
1109    tcg_out_opc(s, OPC_POP_r32 + LOWREGMASK(reg), 0, reg, 0);
1110}
1111
1112static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg ret,
1113                       TCGReg arg1, intptr_t arg2)
1114{
1115    switch (type) {
1116    case TCG_TYPE_I32:
1117        if (ret < 16) {
1118            tcg_out_modrm_offset(s, OPC_MOVL_GvEv, ret, arg1, arg2);
1119        } else {
1120            tcg_out_vex_modrm_offset(s, OPC_MOVD_VyEy, ret, 0, arg1, arg2);
1121        }
1122        break;
1123    case TCG_TYPE_I64:
1124        if (ret < 16) {
1125            tcg_out_modrm_offset(s, OPC_MOVL_GvEv | P_REXW, ret, arg1, arg2);
1126            break;
1127        }
1128        /* FALLTHRU */
1129    case TCG_TYPE_V64:
1130        /* There is no instruction that can validate 8-byte alignment.  */
1131        tcg_debug_assert(ret >= 16);
1132        tcg_out_vex_modrm_offset(s, OPC_MOVQ_VqWq, ret, 0, arg1, arg2);
1133        break;
1134    case TCG_TYPE_V128:
1135        /*
1136         * The gvec infrastructure is asserts that v128 vector loads
1137         * and stores use a 16-byte aligned offset.  Validate that the
1138         * final pointer is aligned by using an insn that will SIGSEGV.
1139         */
1140        tcg_debug_assert(ret >= 16);
1141        tcg_out_vex_modrm_offset(s, OPC_MOVDQA_VxWx, ret, 0, arg1, arg2);
1142        break;
1143    case TCG_TYPE_V256:
1144        /*
1145         * The gvec infrastructure only requires 16-byte alignment,
1146         * so here we must use an unaligned load.
1147         */
1148        tcg_debug_assert(ret >= 16);
1149        tcg_out_vex_modrm_offset(s, OPC_MOVDQU_VxWx | P_VEXL,
1150                                 ret, 0, arg1, arg2);
1151        break;
1152    default:
1153        g_assert_not_reached();
1154    }
1155}
1156
1157static void tcg_out_st(TCGContext *s, TCGType type, TCGReg arg,
1158                       TCGReg arg1, intptr_t arg2)
1159{
1160    switch (type) {
1161    case TCG_TYPE_I32:
1162        if (arg < 16) {
1163            tcg_out_modrm_offset(s, OPC_MOVL_EvGv, arg, arg1, arg2);
1164        } else {
1165            tcg_out_vex_modrm_offset(s, OPC_MOVD_EyVy, arg, 0, arg1, arg2);
1166        }
1167        break;
1168    case TCG_TYPE_I64:
1169        if (arg < 16) {
1170            tcg_out_modrm_offset(s, OPC_MOVL_EvGv | P_REXW, arg, arg1, arg2);
1171            break;
1172        }
1173        /* FALLTHRU */
1174    case TCG_TYPE_V64:
1175        /* There is no instruction that can validate 8-byte alignment.  */
1176        tcg_debug_assert(arg >= 16);
1177        tcg_out_vex_modrm_offset(s, OPC_MOVQ_WqVq, arg, 0, arg1, arg2);
1178        break;
1179    case TCG_TYPE_V128:
1180        /*
1181         * The gvec infrastructure is asserts that v128 vector loads
1182         * and stores use a 16-byte aligned offset.  Validate that the
1183         * final pointer is aligned by using an insn that will SIGSEGV.
1184         *
1185         * This specific instance is also used by TCG_CALL_RET_BY_VEC,
1186         * for _WIN64, which must have SSE2 but may not have AVX.
1187         */
1188        tcg_debug_assert(arg >= 16);
1189        if (have_avx1) {
1190            tcg_out_vex_modrm_offset(s, OPC_MOVDQA_WxVx, arg, 0, arg1, arg2);
1191        } else {
1192            tcg_out_modrm_offset(s, OPC_MOVDQA_WxVx, arg, arg1, arg2);
1193        }
1194        break;
1195    case TCG_TYPE_V256:
1196        /*
1197         * The gvec infrastructure only requires 16-byte alignment,
1198         * so here we must use an unaligned store.
1199         */
1200        tcg_debug_assert(arg >= 16);
1201        tcg_out_vex_modrm_offset(s, OPC_MOVDQU_WxVx | P_VEXL,
1202                                 arg, 0, arg1, arg2);
1203        break;
1204    default:
1205        g_assert_not_reached();
1206    }
1207}
1208
1209static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val,
1210                        TCGReg base, intptr_t ofs)
1211{
1212    int rexw = 0;
1213    if (TCG_TARGET_REG_BITS == 64 && type == TCG_TYPE_I64) {
1214        if (val != (int32_t)val) {
1215            return false;
1216        }
1217        rexw = P_REXW;
1218    } else if (type != TCG_TYPE_I32) {
1219        return false;
1220    }
1221    tcg_out_modrm_offset(s, OPC_MOVL_EvIz | rexw, 0, base, ofs);
1222    tcg_out32(s, val);
1223    return true;
1224}
1225
1226static void tcg_out_shifti(TCGContext *s, int subopc, int reg, int count)
1227{
1228    /* Propagate an opcode prefix, such as P_DATA16.  */
1229    int ext = subopc & ~0x7;
1230    subopc &= 0x7;
1231
1232    if (count == 1) {
1233        tcg_out_modrm(s, OPC_SHIFT_1 + ext, subopc, reg);
1234    } else {
1235        tcg_out_modrm(s, OPC_SHIFT_Ib + ext, subopc, reg);
1236        tcg_out8(s, count);
1237    }
1238}
1239
1240static inline void tcg_out_bswap32(TCGContext *s, int reg)
1241{
1242    tcg_out_opc(s, OPC_BSWAP + LOWREGMASK(reg), 0, reg, 0);
1243}
1244
1245static inline void tcg_out_rolw_8(TCGContext *s, int reg)
1246{
1247    tcg_out_shifti(s, SHIFT_ROL + P_DATA16, reg, 8);
1248}
1249
1250static void tcg_out_ext8u(TCGContext *s, TCGReg dest, TCGReg src)
1251{
1252    /* movzbl */
1253    tcg_debug_assert(src < 4 || TCG_TARGET_REG_BITS == 64);
1254    tcg_out_modrm(s, OPC_MOVZBL + P_REXB_RM, dest, src);
1255}
1256
1257static void tcg_out_ext8s(TCGContext *s, TCGType type, TCGReg dest, TCGReg src)
1258{
1259    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1260    /* movsbl */
1261    tcg_debug_assert(src < 4 || TCG_TARGET_REG_BITS == 64);
1262    tcg_out_modrm(s, OPC_MOVSBL + P_REXB_RM + rexw, dest, src);
1263}
1264
1265static void tcg_out_ext16u(TCGContext *s, TCGReg dest, TCGReg src)
1266{
1267    /* movzwl */
1268    tcg_out_modrm(s, OPC_MOVZWL, dest, src);
1269}
1270
1271static void tcg_out_ext16s(TCGContext *s, TCGType type, TCGReg dest, TCGReg src)
1272{
1273    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1274    /* movsw[lq] */
1275    tcg_out_modrm(s, OPC_MOVSWL + rexw, dest, src);
1276}
1277
1278static void tcg_out_ext32u(TCGContext *s, TCGReg dest, TCGReg src)
1279{
1280    /* 32-bit mov zero extends.  */
1281    tcg_out_modrm(s, OPC_MOVL_GvEv, dest, src);
1282}
1283
1284static void tcg_out_ext32s(TCGContext *s, TCGReg dest, TCGReg src)
1285{
1286    tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
1287    tcg_out_modrm(s, OPC_MOVSLQ, dest, src);
1288}
1289
1290static void tcg_out_exts_i32_i64(TCGContext *s, TCGReg dest, TCGReg src)
1291{
1292    tcg_out_ext32s(s, dest, src);
1293}
1294
1295static void tcg_out_extu_i32_i64(TCGContext *s, TCGReg dest, TCGReg src)
1296{
1297    if (dest != src) {
1298        tcg_out_ext32u(s, dest, src);
1299    }
1300}
1301
1302static void tcg_out_extrl_i64_i32(TCGContext *s, TCGReg dest, TCGReg src)
1303{
1304    tcg_out_ext32u(s, dest, src);
1305}
1306
1307static inline void tcg_out_bswap64(TCGContext *s, int reg)
1308{
1309    tcg_out_opc(s, OPC_BSWAP + P_REXW + LOWREGMASK(reg), 0, reg, 0);
1310}
1311
1312static void tgen_arithi(TCGContext *s, int c, int r0,
1313                        tcg_target_long val, int cf)
1314{
1315    int rexw = 0;
1316
1317    if (TCG_TARGET_REG_BITS == 64) {
1318        rexw = c & -8;
1319        c &= 7;
1320    }
1321
1322    /* ??? While INC is 2 bytes shorter than ADDL $1, they also induce
1323       partial flags update stalls on Pentium4 and are not recommended
1324       by current Intel optimization manuals.  */
1325    if (!cf && (c == ARITH_ADD || c == ARITH_SUB) && (val == 1 || val == -1)) {
1326        int is_inc = (c == ARITH_ADD) ^ (val < 0);
1327        if (TCG_TARGET_REG_BITS == 64) {
1328            /* The single-byte increment encodings are re-tasked as the
1329               REX prefixes.  Use the MODRM encoding.  */
1330            tcg_out_modrm(s, OPC_GRP5 + rexw,
1331                          (is_inc ? EXT5_INC_Ev : EXT5_DEC_Ev), r0);
1332        } else {
1333            tcg_out8(s, (is_inc ? OPC_INC_r32 : OPC_DEC_r32) + r0);
1334        }
1335        return;
1336    }
1337
1338    if (c == ARITH_AND) {
1339        if (TCG_TARGET_REG_BITS == 64) {
1340            if (val == 0xffffffffu) {
1341                tcg_out_ext32u(s, r0, r0);
1342                return;
1343            }
1344            if (val == (uint32_t)val) {
1345                /* AND with no high bits set can use a 32-bit operation.  */
1346                rexw = 0;
1347            }
1348        }
1349        if (val == 0xffu && (r0 < 4 || TCG_TARGET_REG_BITS == 64)) {
1350            tcg_out_ext8u(s, r0, r0);
1351            return;
1352        }
1353        if (val == 0xffffu) {
1354            tcg_out_ext16u(s, r0, r0);
1355            return;
1356        }
1357    }
1358
1359    if (val == (int8_t)val) {
1360        tcg_out_modrm(s, OPC_ARITH_EvIb + rexw, c, r0);
1361        tcg_out8(s, val);
1362        return;
1363    }
1364    if (rexw == 0 || val == (int32_t)val) {
1365        tcg_out_modrm(s, OPC_ARITH_EvIz + rexw, c, r0);
1366        tcg_out32(s, val);
1367        return;
1368    }
1369
1370    g_assert_not_reached();
1371}
1372
1373static void tcg_out_addi(TCGContext *s, int reg, tcg_target_long val)
1374{
1375    if (val != 0) {
1376        tgen_arithi(s, ARITH_ADD + P_REXW, reg, val, 0);
1377    }
1378}
1379
1380/* Set SMALL to force a short forward branch.  */
1381static void tcg_out_jxx(TCGContext *s, int opc, TCGLabel *l, bool small)
1382{
1383    int32_t val, val1;
1384
1385    if (l->has_value) {
1386        val = tcg_pcrel_diff(s, l->u.value_ptr);
1387        val1 = val - 2;
1388        if ((int8_t)val1 == val1) {
1389            if (opc == -1) {
1390                tcg_out8(s, OPC_JMP_short);
1391            } else {
1392                tcg_out8(s, OPC_JCC_short + opc);
1393            }
1394            tcg_out8(s, val1);
1395        } else {
1396            tcg_debug_assert(!small);
1397            if (opc == -1) {
1398                tcg_out8(s, OPC_JMP_long);
1399                tcg_out32(s, val - 5);
1400            } else {
1401                tcg_out_opc(s, OPC_JCC_long + opc, 0, 0, 0);
1402                tcg_out32(s, val - 6);
1403            }
1404        }
1405    } else if (small) {
1406        if (opc == -1) {
1407            tcg_out8(s, OPC_JMP_short);
1408        } else {
1409            tcg_out8(s, OPC_JCC_short + opc);
1410        }
1411        tcg_out_reloc(s, s->code_ptr, R_386_PC8, l, -1);
1412        s->code_ptr += 1;
1413    } else {
1414        if (opc == -1) {
1415            tcg_out8(s, OPC_JMP_long);
1416        } else {
1417            tcg_out_opc(s, OPC_JCC_long + opc, 0, 0, 0);
1418        }
1419        tcg_out_reloc(s, s->code_ptr, R_386_PC32, l, -4);
1420        s->code_ptr += 4;
1421    }
1422}
1423
1424static void tcg_out_cmp(TCGContext *s, TCGArg arg1, TCGArg arg2,
1425                        int const_arg2, int rexw)
1426{
1427    if (const_arg2) {
1428        if (arg2 == 0) {
1429            /* test r, r */
1430            tcg_out_modrm(s, OPC_TESTL + rexw, arg1, arg1);
1431        } else {
1432            tgen_arithi(s, ARITH_CMP + rexw, arg1, arg2, 0);
1433        }
1434    } else {
1435        tgen_arithr(s, ARITH_CMP + rexw, arg1, arg2);
1436    }
1437}
1438
1439static void tcg_out_brcond(TCGContext *s, int rexw, TCGCond cond,
1440                           TCGArg arg1, TCGArg arg2, int const_arg2,
1441                           TCGLabel *label, bool small)
1442{
1443    tcg_out_cmp(s, arg1, arg2, const_arg2, rexw);
1444    tcg_out_jxx(s, tcg_cond_to_jcc[cond], label, small);
1445}
1446
1447#if TCG_TARGET_REG_BITS == 32
1448static void tcg_out_brcond2(TCGContext *s, const TCGArg *args,
1449                            const int *const_args, bool small)
1450{
1451    TCGLabel *label_next = gen_new_label();
1452    TCGLabel *label_this = arg_label(args[5]);
1453
1454    switch(args[4]) {
1455    case TCG_COND_EQ:
1456        tcg_out_brcond(s, 0, TCG_COND_NE, args[0], args[2], const_args[2],
1457                       label_next, 1);
1458        tcg_out_brcond(s, 0, TCG_COND_EQ, args[1], args[3], const_args[3],
1459                       label_this, small);
1460        break;
1461    case TCG_COND_NE:
1462        tcg_out_brcond(s, 0, TCG_COND_NE, args[0], args[2], const_args[2],
1463                       label_this, small);
1464        tcg_out_brcond(s, 0, TCG_COND_NE, args[1], args[3], const_args[3],
1465                       label_this, small);
1466        break;
1467    case TCG_COND_LT:
1468        tcg_out_brcond(s, 0, TCG_COND_LT, args[1], args[3], const_args[3],
1469                       label_this, small);
1470        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1471        tcg_out_brcond(s, 0, TCG_COND_LTU, args[0], args[2], const_args[2],
1472                       label_this, small);
1473        break;
1474    case TCG_COND_LE:
1475        tcg_out_brcond(s, 0, 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_brcond(s, 0, TCG_COND_LEU, args[0], args[2], const_args[2],
1479                       label_this, small);
1480        break;
1481    case TCG_COND_GT:
1482        tcg_out_brcond(s, 0, TCG_COND_GT, args[1], args[3], const_args[3],
1483                       label_this, small);
1484        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1485        tcg_out_brcond(s, 0, TCG_COND_GTU, args[0], args[2], const_args[2],
1486                       label_this, small);
1487        break;
1488    case TCG_COND_GE:
1489        tcg_out_brcond(s, 0, 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_brcond(s, 0, TCG_COND_GEU, args[0], args[2], const_args[2],
1493                       label_this, small);
1494        break;
1495    case TCG_COND_LTU:
1496        tcg_out_brcond(s, 0, TCG_COND_LTU, args[1], args[3], const_args[3],
1497                       label_this, small);
1498        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1499        tcg_out_brcond(s, 0, TCG_COND_LTU, args[0], args[2], const_args[2],
1500                       label_this, small);
1501        break;
1502    case TCG_COND_LEU:
1503        tcg_out_brcond(s, 0, 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_brcond(s, 0, TCG_COND_LEU, args[0], args[2], const_args[2],
1507                       label_this, small);
1508        break;
1509    case TCG_COND_GTU:
1510        tcg_out_brcond(s, 0, TCG_COND_GTU, args[1], args[3], const_args[3],
1511                       label_this, small);
1512        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1513        tcg_out_brcond(s, 0, TCG_COND_GTU, args[0], args[2], const_args[2],
1514                       label_this, small);
1515        break;
1516    case TCG_COND_GEU:
1517        tcg_out_brcond(s, 0, 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_brcond(s, 0, TCG_COND_GEU, args[0], args[2], const_args[2],
1521                       label_this, small);
1522        break;
1523    default:
1524        g_assert_not_reached();
1525    }
1526    tcg_out_label(s, label_next);
1527}
1528#endif
1529
1530static void tcg_out_setcond(TCGContext *s, int rexw, TCGCond cond,
1531                            TCGArg dest, TCGArg arg1, TCGArg arg2,
1532                            int const_arg2, bool neg)
1533{
1534    bool inv = false;
1535    bool cleared;
1536
1537    switch (cond) {
1538    case TCG_COND_NE:
1539        inv = true;
1540        /* fall through */
1541    case TCG_COND_EQ:
1542        /* If arg2 is 0, convert to LTU/GEU vs 1. */
1543        if (const_arg2 && arg2 == 0) {
1544            arg2 = 1;
1545            goto do_ltu;
1546        }
1547        break;
1548
1549    case TCG_COND_LEU:
1550        inv = true;
1551        /* fall through */
1552    case TCG_COND_GTU:
1553        /* If arg2 is a register, swap for LTU/GEU. */
1554        if (!const_arg2) {
1555            TCGReg t = arg1;
1556            arg1 = arg2;
1557            arg2 = t;
1558            goto do_ltu;
1559        }
1560        break;
1561
1562    case TCG_COND_GEU:
1563        inv = true;
1564        /* fall through */
1565    case TCG_COND_LTU:
1566    do_ltu:
1567        /*
1568         * Relying on the carry bit, use SBB to produce -1 if LTU, 0 if GEU.
1569         * We can then use NEG or INC to produce the desired result.
1570         * This is always smaller than the SETCC expansion.
1571         */
1572        tcg_out_cmp(s, arg1, arg2, const_arg2, rexw);
1573
1574        /* X - X - C = -C = (C ? -1 : 0) */
1575        tgen_arithr(s, ARITH_SBB + (neg ? rexw : 0), dest, dest);
1576        if (inv && neg) {
1577            /* ~(C ? -1 : 0) = (C ? 0 : -1) */
1578            tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NOT, dest);
1579        } else if (inv) {
1580            /* (C ? -1 : 0) + 1 = (C ? 0 : 1) */
1581            tgen_arithi(s, ARITH_ADD, dest, 1, 0);
1582        } else if (!neg) {
1583            /* -(C ? -1 : 0) = (C ? 1 : 0) */
1584            tcg_out_modrm(s, OPC_GRP3_Ev, EXT3_NEG, dest);
1585        }
1586        return;
1587
1588    case TCG_COND_GE:
1589        inv = true;
1590        /* fall through */
1591    case TCG_COND_LT:
1592        /* If arg2 is 0, extract the sign bit. */
1593        if (const_arg2 && arg2 == 0) {
1594            tcg_out_mov(s, rexw ? TCG_TYPE_I64 : TCG_TYPE_I32, dest, arg1);
1595            if (inv) {
1596                tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NOT, dest);
1597            }
1598            tcg_out_shifti(s, (neg ? SHIFT_SAR : SHIFT_SHR) + rexw,
1599                           dest, rexw ? 63 : 31);
1600            return;
1601        }
1602        break;
1603
1604    default:
1605        break;
1606    }
1607
1608    /*
1609     * If dest does not overlap the inputs, clearing it first is preferred.
1610     * The XOR breaks any false dependency for the low-byte write to dest,
1611     * and is also one byte smaller than MOVZBL.
1612     */
1613    cleared = false;
1614    if (dest != arg1 && (const_arg2 || dest != arg2)) {
1615        tgen_arithr(s, ARITH_XOR, dest, dest);
1616        cleared = true;
1617    }
1618
1619    tcg_out_cmp(s, arg1, arg2, const_arg2, rexw);
1620    tcg_out_modrm(s, OPC_SETCC | tcg_cond_to_jcc[cond], 0, dest);
1621
1622    if (!cleared) {
1623        tcg_out_ext8u(s, dest, dest);
1624    }
1625    if (neg) {
1626        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NEG, dest);
1627    }
1628}
1629
1630#if TCG_TARGET_REG_BITS == 32
1631static void tcg_out_setcond2(TCGContext *s, const TCGArg *args,
1632                             const int *const_args)
1633{
1634    TCGArg new_args[6];
1635    TCGLabel *label_true, *label_over;
1636
1637    memcpy(new_args, args+1, 5*sizeof(TCGArg));
1638
1639    if (args[0] == args[1] || args[0] == args[2]
1640        || (!const_args[3] && args[0] == args[3])
1641        || (!const_args[4] && args[0] == args[4])) {
1642        /* When the destination overlaps with one of the argument
1643           registers, don't do anything tricky.  */
1644        label_true = gen_new_label();
1645        label_over = gen_new_label();
1646
1647        new_args[5] = label_arg(label_true);
1648        tcg_out_brcond2(s, new_args, const_args+1, 1);
1649
1650        tcg_out_movi(s, TCG_TYPE_I32, args[0], 0);
1651        tcg_out_jxx(s, JCC_JMP, label_over, 1);
1652        tcg_out_label(s, label_true);
1653
1654        tcg_out_movi(s, TCG_TYPE_I32, args[0], 1);
1655        tcg_out_label(s, label_over);
1656    } else {
1657        /* When the destination does not overlap one of the arguments,
1658           clear the destination first, jump if cond false, and emit an
1659           increment in the true case.  This results in smaller code.  */
1660
1661        tcg_out_movi(s, TCG_TYPE_I32, args[0], 0);
1662
1663        label_over = gen_new_label();
1664        new_args[4] = tcg_invert_cond(new_args[4]);
1665        new_args[5] = label_arg(label_over);
1666        tcg_out_brcond2(s, new_args, const_args+1, 1);
1667
1668        tgen_arithi(s, ARITH_ADD, args[0], 1, 0);
1669        tcg_out_label(s, label_over);
1670    }
1671}
1672#endif
1673
1674static void tcg_out_cmov(TCGContext *s, TCGCond cond, int rexw,
1675                         TCGReg dest, TCGReg v1)
1676{
1677    if (have_cmov) {
1678        tcg_out_modrm(s, OPC_CMOVCC | tcg_cond_to_jcc[cond] | rexw, dest, v1);
1679    } else {
1680        TCGLabel *over = gen_new_label();
1681        tcg_out_jxx(s, tcg_cond_to_jcc[tcg_invert_cond(cond)], over, 1);
1682        tcg_out_mov(s, TCG_TYPE_I32, dest, v1);
1683        tcg_out_label(s, over);
1684    }
1685}
1686
1687static void tcg_out_movcond(TCGContext *s, int rexw, TCGCond cond,
1688                            TCGReg dest, TCGReg c1, TCGArg c2, int const_c2,
1689                            TCGReg v1)
1690{
1691    tcg_out_cmp(s, c1, c2, const_c2, rexw);
1692    tcg_out_cmov(s, cond, rexw, dest, v1);
1693}
1694
1695static void tcg_out_ctz(TCGContext *s, int rexw, TCGReg dest, TCGReg arg1,
1696                        TCGArg arg2, bool const_a2)
1697{
1698    if (have_bmi1) {
1699        tcg_out_modrm(s, OPC_TZCNT + rexw, dest, arg1);
1700        if (const_a2) {
1701            tcg_debug_assert(arg2 == (rexw ? 64 : 32));
1702        } else {
1703            tcg_debug_assert(dest != arg2);
1704            tcg_out_cmov(s, TCG_COND_LTU, rexw, dest, arg2);
1705        }
1706    } else {
1707        tcg_debug_assert(dest != arg2);
1708        tcg_out_modrm(s, OPC_BSF + rexw, dest, arg1);
1709        tcg_out_cmov(s, TCG_COND_EQ, rexw, dest, arg2);
1710    }
1711}
1712
1713static void tcg_out_clz(TCGContext *s, int rexw, TCGReg dest, TCGReg arg1,
1714                        TCGArg arg2, bool const_a2)
1715{
1716    if (have_lzcnt) {
1717        tcg_out_modrm(s, OPC_LZCNT + rexw, dest, arg1);
1718        if (const_a2) {
1719            tcg_debug_assert(arg2 == (rexw ? 64 : 32));
1720        } else {
1721            tcg_debug_assert(dest != arg2);
1722            tcg_out_cmov(s, TCG_COND_LTU, rexw, dest, arg2);
1723        }
1724    } else {
1725        tcg_debug_assert(!const_a2);
1726        tcg_debug_assert(dest != arg1);
1727        tcg_debug_assert(dest != arg2);
1728
1729        /* Recall that the output of BSR is the index not the count.  */
1730        tcg_out_modrm(s, OPC_BSR + rexw, dest, arg1);
1731        tgen_arithi(s, ARITH_XOR + rexw, dest, rexw ? 63 : 31, 0);
1732
1733        /* Since we have destroyed the flags from BSR, we have to re-test.  */
1734        tcg_out_cmp(s, arg1, 0, 1, rexw);
1735        tcg_out_cmov(s, TCG_COND_EQ, rexw, dest, arg2);
1736    }
1737}
1738
1739static void tcg_out_branch(TCGContext *s, int call, const tcg_insn_unit *dest)
1740{
1741    intptr_t disp = tcg_pcrel_diff(s, dest) - 5;
1742
1743    if (disp == (int32_t)disp) {
1744        tcg_out_opc(s, call ? OPC_CALL_Jz : OPC_JMP_long, 0, 0, 0);
1745        tcg_out32(s, disp);
1746    } else {
1747        /* rip-relative addressing into the constant pool.
1748           This is 6 + 8 = 14 bytes, as compared to using an
1749           immediate load 10 + 6 = 16 bytes, plus we may
1750           be able to re-use the pool constant for more calls.  */
1751        tcg_out_opc(s, OPC_GRP5, 0, 0, 0);
1752        tcg_out8(s, (call ? EXT5_CALLN_Ev : EXT5_JMPN_Ev) << 3 | 5);
1753        new_pool_label(s, (uintptr_t)dest, R_386_PC32, s->code_ptr, -4);
1754        tcg_out32(s, 0);
1755    }
1756}
1757
1758static void tcg_out_call(TCGContext *s, const tcg_insn_unit *dest,
1759                         const TCGHelperInfo *info)
1760{
1761    tcg_out_branch(s, 1, dest);
1762
1763#ifndef _WIN32
1764    if (TCG_TARGET_REG_BITS == 32 && info->out_kind == TCG_CALL_RET_BY_REF) {
1765        /*
1766         * The sysv i386 abi for struct return places a reference as the
1767         * first argument of the stack, and pops that argument with the
1768         * return statement.  Since we want to retain the aligned stack
1769         * pointer for the callee, we do not want to actually push that
1770         * argument before the call but rely on the normal store to the
1771         * stack slot.  But we do need to compensate for the pop in order
1772         * to reset our correct stack pointer value.
1773         * Pushing a garbage value back onto the stack is quickest.
1774         */
1775        tcg_out_push(s, TCG_REG_EAX);
1776    }
1777#endif
1778}
1779
1780static void tcg_out_jmp(TCGContext *s, const tcg_insn_unit *dest)
1781{
1782    tcg_out_branch(s, 0, dest);
1783}
1784
1785static void tcg_out_nopn(TCGContext *s, int n)
1786{
1787    int i;
1788    /* Emit 1 or 2 operand size prefixes for the standard one byte nop,
1789     * "xchg %eax,%eax", forming "xchg %ax,%ax". All cores accept the
1790     * duplicate prefix, and all of the interesting recent cores can
1791     * decode and discard the duplicates in a single cycle.
1792     */
1793    tcg_debug_assert(n >= 1);
1794    for (i = 1; i < n; ++i) {
1795        tcg_out8(s, 0x66);
1796    }
1797    tcg_out8(s, 0x90);
1798}
1799
1800/* Test register R vs immediate bits I, setting Z flag for EQ/NE. */
1801static void __attribute__((unused))
1802tcg_out_testi(TCGContext *s, TCGReg r, uint32_t i)
1803{
1804    /*
1805     * This is used for testing alignment, so we can usually use testb.
1806     * For i686, we have to use testl for %esi/%edi.
1807     */
1808    if (i <= 0xff && (TCG_TARGET_REG_BITS == 64 || r < 4)) {
1809        tcg_out_modrm(s, OPC_GRP3_Eb | P_REXB_RM, EXT3_TESTi, r);
1810        tcg_out8(s, i);
1811    } else {
1812        tcg_out_modrm(s, OPC_GRP3_Ev, EXT3_TESTi, r);
1813        tcg_out32(s, i);
1814    }
1815}
1816
1817typedef struct {
1818    TCGReg base;
1819    int index;
1820    int ofs;
1821    int seg;
1822    TCGAtomAlign aa;
1823} HostAddress;
1824
1825bool tcg_target_has_memory_bswap(MemOp memop)
1826{
1827    TCGAtomAlign aa;
1828
1829    if (!have_movbe) {
1830        return false;
1831    }
1832    if ((memop & MO_SIZE) < MO_128) {
1833        return true;
1834    }
1835
1836    /*
1837     * Reject 16-byte memop with 16-byte atomicity, i.e. VMOVDQA,
1838     * but do allow a pair of 64-bit operations, i.e. MOVBEQ.
1839     */
1840    aa = atom_and_align_for_opc(tcg_ctx, memop, MO_ATOM_IFALIGN, true);
1841    return aa.atom < MO_128;
1842}
1843
1844/*
1845 * Because i686 has no register parameters and because x86_64 has xchg
1846 * to handle addr/data register overlap, we have placed all input arguments
1847 * before we need might need a scratch reg.
1848 *
1849 * Even then, a scratch is only needed for l->raddr.  Rather than expose
1850 * a general-purpose scratch when we don't actually know it's available,
1851 * use the ra_gen hook to load into RAX if needed.
1852 */
1853#if TCG_TARGET_REG_BITS == 64
1854static TCGReg ldst_ra_gen(TCGContext *s, const TCGLabelQemuLdst *l, int arg)
1855{
1856    if (arg < 0) {
1857        arg = TCG_REG_RAX;
1858    }
1859    tcg_out_movi(s, TCG_TYPE_PTR, arg, (uintptr_t)l->raddr);
1860    return arg;
1861}
1862static const TCGLdstHelperParam ldst_helper_param = {
1863    .ra_gen = ldst_ra_gen
1864};
1865#else
1866static const TCGLdstHelperParam ldst_helper_param = { };
1867#endif
1868
1869static void tcg_out_vec_to_pair(TCGContext *s, TCGType type,
1870                                TCGReg l, TCGReg h, TCGReg v)
1871{
1872    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1873
1874    /* vpmov{d,q} %v, %l */
1875    tcg_out_vex_modrm(s, OPC_MOVD_EyVy + rexw, v, 0, l);
1876    /* vpextr{d,q} $1, %v, %h */
1877    tcg_out_vex_modrm(s, OPC_PEXTRD + rexw, v, 0, h);
1878    tcg_out8(s, 1);
1879}
1880
1881static void tcg_out_pair_to_vec(TCGContext *s, TCGType type,
1882                                TCGReg v, TCGReg l, TCGReg h)
1883{
1884    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1885
1886    /* vmov{d,q} %l, %v */
1887    tcg_out_vex_modrm(s, OPC_MOVD_VyEy + rexw, v, 0, l);
1888    /* vpinsr{d,q} $1, %h, %v, %v */
1889    tcg_out_vex_modrm(s, OPC_PINSRD + rexw, v, v, h);
1890    tcg_out8(s, 1);
1891}
1892
1893/*
1894 * Generate code for the slow path for a load at the end of block
1895 */
1896static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
1897{
1898    MemOp opc = get_memop(l->oi);
1899    tcg_insn_unit **label_ptr = &l->label_ptr[0];
1900
1901    /* resolve label address */
1902    tcg_patch32(label_ptr[0], s->code_ptr - label_ptr[0] - 4);
1903    if (label_ptr[1]) {
1904        tcg_patch32(label_ptr[1], s->code_ptr - label_ptr[1] - 4);
1905    }
1906
1907    tcg_out_ld_helper_args(s, l, &ldst_helper_param);
1908    tcg_out_branch(s, 1, qemu_ld_helpers[opc & MO_SIZE]);
1909    tcg_out_ld_helper_ret(s, l, false, &ldst_helper_param);
1910
1911    tcg_out_jmp(s, l->raddr);
1912    return true;
1913}
1914
1915/*
1916 * Generate code for the slow path for a store at the end of block
1917 */
1918static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
1919{
1920    MemOp opc = get_memop(l->oi);
1921    tcg_insn_unit **label_ptr = &l->label_ptr[0];
1922
1923    /* resolve label address */
1924    tcg_patch32(label_ptr[0], s->code_ptr - label_ptr[0] - 4);
1925    if (label_ptr[1]) {
1926        tcg_patch32(label_ptr[1], s->code_ptr - label_ptr[1] - 4);
1927    }
1928
1929    tcg_out_st_helper_args(s, l, &ldst_helper_param);
1930    tcg_out_branch(s, 1, qemu_st_helpers[opc & MO_SIZE]);
1931
1932    tcg_out_jmp(s, l->raddr);
1933    return true;
1934}
1935
1936#ifndef CONFIG_SOFTMMU
1937static HostAddress x86_guest_base = {
1938    .index = -1
1939};
1940
1941#if defined(__x86_64__) && defined(__linux__)
1942# include <asm/prctl.h>
1943# include <sys/prctl.h>
1944int arch_prctl(int code, unsigned long addr);
1945static inline int setup_guest_base_seg(void)
1946{
1947    if (arch_prctl(ARCH_SET_GS, guest_base) == 0) {
1948        return P_GS;
1949    }
1950    return 0;
1951}
1952#elif defined(__x86_64__) && \
1953      (defined (__FreeBSD__) || defined (__FreeBSD_kernel__))
1954# include <machine/sysarch.h>
1955static inline int setup_guest_base_seg(void)
1956{
1957    if (sysarch(AMD64_SET_GSBASE, &guest_base) == 0) {
1958        return P_GS;
1959    }
1960    return 0;
1961}
1962#else
1963static inline int setup_guest_base_seg(void)
1964{
1965    return 0;
1966}
1967#endif /* setup_guest_base_seg */
1968#endif /* !SOFTMMU */
1969
1970#define MIN_TLB_MASK_TABLE_OFS  INT_MIN
1971
1972/*
1973 * For softmmu, perform the TLB load and compare.
1974 * For useronly, perform any required alignment tests.
1975 * In both cases, return a TCGLabelQemuLdst structure if the slow path
1976 * is required and fill in @h with the host address for the fast path.
1977 */
1978static TCGLabelQemuLdst *prepare_host_addr(TCGContext *s, HostAddress *h,
1979                                           TCGReg addrlo, TCGReg addrhi,
1980                                           MemOpIdx oi, bool is_ld)
1981{
1982    TCGLabelQemuLdst *ldst = NULL;
1983    MemOp opc = get_memop(oi);
1984    MemOp s_bits = opc & MO_SIZE;
1985    unsigned a_mask;
1986
1987#ifdef CONFIG_SOFTMMU
1988    h->index = TCG_REG_L0;
1989    h->ofs = 0;
1990    h->seg = 0;
1991#else
1992    *h = x86_guest_base;
1993#endif
1994    h->base = addrlo;
1995    h->aa = atom_and_align_for_opc(s, opc, MO_ATOM_IFALIGN, s_bits == MO_128);
1996    a_mask = (1 << h->aa.align) - 1;
1997
1998#ifdef CONFIG_SOFTMMU
1999    int cmp_ofs = is_ld ? offsetof(CPUTLBEntry, addr_read)
2000                        : offsetof(CPUTLBEntry, addr_write);
2001    TCGType ttype = TCG_TYPE_I32;
2002    TCGType tlbtype = TCG_TYPE_I32;
2003    int trexw = 0, hrexw = 0, tlbrexw = 0;
2004    unsigned mem_index = get_mmuidx(oi);
2005    unsigned s_mask = (1 << s_bits) - 1;
2006    int fast_ofs = tlb_mask_table_ofs(s, mem_index);
2007    int tlb_mask;
2008
2009    ldst = new_ldst_label(s);
2010    ldst->is_ld = is_ld;
2011    ldst->oi = oi;
2012    ldst->addrlo_reg = addrlo;
2013    ldst->addrhi_reg = addrhi;
2014
2015    if (TCG_TARGET_REG_BITS == 64) {
2016        ttype = s->addr_type;
2017        trexw = (ttype == TCG_TYPE_I32 ? 0 : P_REXW);
2018        if (TCG_TYPE_PTR == TCG_TYPE_I64) {
2019            hrexw = P_REXW;
2020            if (s->page_bits + s->tlb_dyn_max_bits > 32) {
2021                tlbtype = TCG_TYPE_I64;
2022                tlbrexw = P_REXW;
2023            }
2024        }
2025    }
2026
2027    tcg_out_mov(s, tlbtype, TCG_REG_L0, addrlo);
2028    tcg_out_shifti(s, SHIFT_SHR + tlbrexw, TCG_REG_L0,
2029                   s->page_bits - CPU_TLB_ENTRY_BITS);
2030
2031    tcg_out_modrm_offset(s, OPC_AND_GvEv + trexw, TCG_REG_L0, TCG_AREG0,
2032                         fast_ofs + offsetof(CPUTLBDescFast, mask));
2033
2034    tcg_out_modrm_offset(s, OPC_ADD_GvEv + hrexw, TCG_REG_L0, TCG_AREG0,
2035                         fast_ofs + offsetof(CPUTLBDescFast, table));
2036
2037    /*
2038     * If the required alignment is at least as large as the access, simply
2039     * copy the address and mask.  For lesser alignments, check that we don't
2040     * cross pages for the complete access.
2041     */
2042    if (a_mask >= s_mask) {
2043        tcg_out_mov(s, ttype, TCG_REG_L1, addrlo);
2044    } else {
2045        tcg_out_modrm_offset(s, OPC_LEA + trexw, TCG_REG_L1,
2046                             addrlo, s_mask - a_mask);
2047    }
2048    tlb_mask = s->page_mask | a_mask;
2049    tgen_arithi(s, ARITH_AND + trexw, TCG_REG_L1, tlb_mask, 0);
2050
2051    /* cmp 0(TCG_REG_L0), TCG_REG_L1 */
2052    tcg_out_modrm_offset(s, OPC_CMP_GvEv + trexw,
2053                         TCG_REG_L1, TCG_REG_L0, cmp_ofs);
2054
2055    /* jne slow_path */
2056    tcg_out_opc(s, OPC_JCC_long + JCC_JNE, 0, 0, 0);
2057    ldst->label_ptr[0] = s->code_ptr;
2058    s->code_ptr += 4;
2059
2060    if (TCG_TARGET_REG_BITS == 32 && s->addr_type == TCG_TYPE_I64) {
2061        /* cmp 4(TCG_REG_L0), addrhi */
2062        tcg_out_modrm_offset(s, OPC_CMP_GvEv, addrhi, TCG_REG_L0, cmp_ofs + 4);
2063
2064        /* jne slow_path */
2065        tcg_out_opc(s, OPC_JCC_long + JCC_JNE, 0, 0, 0);
2066        ldst->label_ptr[1] = s->code_ptr;
2067        s->code_ptr += 4;
2068    }
2069
2070    /* TLB Hit.  */
2071    tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_L0, TCG_REG_L0,
2072               offsetof(CPUTLBEntry, addend));
2073#else
2074    if (a_mask) {
2075        ldst = new_ldst_label(s);
2076
2077        ldst->is_ld = is_ld;
2078        ldst->oi = oi;
2079        ldst->addrlo_reg = addrlo;
2080        ldst->addrhi_reg = addrhi;
2081
2082        tcg_out_testi(s, addrlo, a_mask);
2083        /* jne slow_path */
2084        tcg_out_opc(s, OPC_JCC_long + JCC_JNE, 0, 0, 0);
2085        ldst->label_ptr[0] = s->code_ptr;
2086        s->code_ptr += 4;
2087    }
2088#endif
2089
2090    return ldst;
2091}
2092
2093static void tcg_out_qemu_ld_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
2094                                   HostAddress h, TCGType type, MemOp memop)
2095{
2096    bool use_movbe = false;
2097    int rexw = (type == TCG_TYPE_I32 ? 0 : P_REXW);
2098    int movop = OPC_MOVL_GvEv;
2099
2100    /* Do big-endian loads with movbe.  */
2101    if (memop & MO_BSWAP) {
2102        tcg_debug_assert(have_movbe);
2103        use_movbe = true;
2104        movop = OPC_MOVBE_GyMy;
2105    }
2106
2107    switch (memop & MO_SSIZE) {
2108    case MO_UB:
2109        tcg_out_modrm_sib_offset(s, OPC_MOVZBL + h.seg, datalo,
2110                                 h.base, h.index, 0, h.ofs);
2111        break;
2112    case MO_SB:
2113        tcg_out_modrm_sib_offset(s, OPC_MOVSBL + rexw + h.seg, datalo,
2114                                 h.base, h.index, 0, h.ofs);
2115        break;
2116    case MO_UW:
2117        if (use_movbe) {
2118            /* There is no extending movbe; only low 16-bits are modified.  */
2119            if (datalo != h.base && datalo != h.index) {
2120                /* XOR breaks dependency chains.  */
2121                tgen_arithr(s, ARITH_XOR, datalo, datalo);
2122                tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + P_DATA16 + h.seg,
2123                                         datalo, h.base, h.index, 0, h.ofs);
2124            } else {
2125                tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + P_DATA16 + h.seg,
2126                                         datalo, h.base, h.index, 0, h.ofs);
2127                tcg_out_ext16u(s, datalo, datalo);
2128            }
2129        } else {
2130            tcg_out_modrm_sib_offset(s, OPC_MOVZWL + h.seg, datalo,
2131                                     h.base, h.index, 0, h.ofs);
2132        }
2133        break;
2134    case MO_SW:
2135        if (use_movbe) {
2136            tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + P_DATA16 + h.seg,
2137                                     datalo, h.base, h.index, 0, h.ofs);
2138            tcg_out_ext16s(s, type, datalo, datalo);
2139        } else {
2140            tcg_out_modrm_sib_offset(s, OPC_MOVSWL + rexw + h.seg,
2141                                     datalo, h.base, h.index, 0, h.ofs);
2142        }
2143        break;
2144    case MO_UL:
2145        tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2146                                 h.base, h.index, 0, h.ofs);
2147        break;
2148#if TCG_TARGET_REG_BITS == 64
2149    case MO_SL:
2150        if (use_movbe) {
2151            tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + h.seg, datalo,
2152                                     h.base, h.index, 0, h.ofs);
2153            tcg_out_ext32s(s, datalo, datalo);
2154        } else {
2155            tcg_out_modrm_sib_offset(s, OPC_MOVSLQ + h.seg, datalo,
2156                                     h.base, h.index, 0, h.ofs);
2157        }
2158        break;
2159#endif
2160    case MO_UQ:
2161        if (TCG_TARGET_REG_BITS == 64) {
2162            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2163                                     h.base, h.index, 0, h.ofs);
2164            break;
2165        }
2166        if (use_movbe) {
2167            TCGReg t = datalo;
2168            datalo = datahi;
2169            datahi = t;
2170        }
2171        if (h.base == datalo || h.index == datalo) {
2172            tcg_out_modrm_sib_offset(s, OPC_LEA, datahi,
2173                                     h.base, h.index, 0, h.ofs);
2174            tcg_out_modrm_offset(s, movop + h.seg, datalo, datahi, 0);
2175            tcg_out_modrm_offset(s, movop + h.seg, datahi, datahi, 4);
2176        } else {
2177            tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2178                                     h.base, h.index, 0, h.ofs);
2179            tcg_out_modrm_sib_offset(s, movop + h.seg, datahi,
2180                                     h.base, h.index, 0, h.ofs + 4);
2181        }
2182        break;
2183
2184    case MO_128:
2185        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
2186
2187        /*
2188         * Without 16-byte atomicity, use integer regs.
2189         * That is where we want the data, and it allows bswaps.
2190         */
2191        if (h.aa.atom < MO_128) {
2192            if (use_movbe) {
2193                TCGReg t = datalo;
2194                datalo = datahi;
2195                datahi = t;
2196            }
2197            if (h.base == datalo || h.index == datalo) {
2198                tcg_out_modrm_sib_offset(s, OPC_LEA + P_REXW, datahi,
2199                                         h.base, h.index, 0, h.ofs);
2200                tcg_out_modrm_offset(s, movop + P_REXW + h.seg,
2201                                     datalo, datahi, 0);
2202                tcg_out_modrm_offset(s, movop + P_REXW + h.seg,
2203                                     datahi, datahi, 8);
2204            } else {
2205                tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2206                                         h.base, h.index, 0, h.ofs);
2207                tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datahi,
2208                                         h.base, h.index, 0, h.ofs + 8);
2209            }
2210            break;
2211        }
2212
2213        /*
2214         * With 16-byte atomicity, a vector load is required.
2215         * If we already have 16-byte alignment, then VMOVDQA always works.
2216         * Else if VMOVDQU has atomicity with dynamic alignment, use that.
2217         * Else use we require a runtime test for alignment for VMOVDQA;
2218         * use VMOVDQU on the unaligned nonatomic path for simplicity.
2219         */
2220        if (h.aa.align >= MO_128) {
2221            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_VxWx + h.seg,
2222                                         TCG_TMP_VEC, 0,
2223                                         h.base, h.index, 0, h.ofs);
2224        } else if (cpuinfo & CPUINFO_ATOMIC_VMOVDQU) {
2225            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_VxWx + h.seg,
2226                                         TCG_TMP_VEC, 0,
2227                                         h.base, h.index, 0, h.ofs);
2228        } else {
2229            TCGLabel *l1 = gen_new_label();
2230            TCGLabel *l2 = gen_new_label();
2231
2232            tcg_out_testi(s, h.base, 15);
2233            tcg_out_jxx(s, JCC_JNE, l1, true);
2234
2235            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_VxWx + h.seg,
2236                                         TCG_TMP_VEC, 0,
2237                                         h.base, h.index, 0, h.ofs);
2238            tcg_out_jxx(s, JCC_JMP, l2, true);
2239
2240            tcg_out_label(s, l1);
2241            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_VxWx + h.seg,
2242                                         TCG_TMP_VEC, 0,
2243                                         h.base, h.index, 0, h.ofs);
2244            tcg_out_label(s, l2);
2245        }
2246        tcg_out_vec_to_pair(s, TCG_TYPE_I64, datalo, datahi, TCG_TMP_VEC);
2247        break;
2248
2249    default:
2250        g_assert_not_reached();
2251    }
2252}
2253
2254static void tcg_out_qemu_ld(TCGContext *s, TCGReg datalo, TCGReg datahi,
2255                            TCGReg addrlo, TCGReg addrhi,
2256                            MemOpIdx oi, TCGType data_type)
2257{
2258    TCGLabelQemuLdst *ldst;
2259    HostAddress h;
2260
2261    ldst = prepare_host_addr(s, &h, addrlo, addrhi, oi, true);
2262    tcg_out_qemu_ld_direct(s, datalo, datahi, h, data_type, get_memop(oi));
2263
2264    if (ldst) {
2265        ldst->type = data_type;
2266        ldst->datalo_reg = datalo;
2267        ldst->datahi_reg = datahi;
2268        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
2269    }
2270}
2271
2272static void tcg_out_qemu_st_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
2273                                   HostAddress h, MemOp memop)
2274{
2275    bool use_movbe = false;
2276    int movop = OPC_MOVL_EvGv;
2277
2278    /*
2279     * Do big-endian stores with movbe or softmmu.
2280     * User-only without movbe will have its swapping done generically.
2281     */
2282    if (memop & MO_BSWAP) {
2283        tcg_debug_assert(have_movbe);
2284        use_movbe = true;
2285        movop = OPC_MOVBE_MyGy;
2286    }
2287
2288    switch (memop & MO_SIZE) {
2289    case MO_8:
2290        /* This is handled with constraints on INDEX_op_qemu_st8_i32. */
2291        tcg_debug_assert(TCG_TARGET_REG_BITS == 64 || datalo < 4);
2292        tcg_out_modrm_sib_offset(s, OPC_MOVB_EvGv + P_REXB_R + h.seg,
2293                                 datalo, h.base, h.index, 0, h.ofs);
2294        break;
2295    case MO_16:
2296        tcg_out_modrm_sib_offset(s, movop + P_DATA16 + h.seg, datalo,
2297                                 h.base, h.index, 0, h.ofs);
2298        break;
2299    case MO_32:
2300        tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2301                                 h.base, h.index, 0, h.ofs);
2302        break;
2303    case MO_64:
2304        if (TCG_TARGET_REG_BITS == 64) {
2305            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2306                                     h.base, h.index, 0, h.ofs);
2307        } else {
2308            if (use_movbe) {
2309                TCGReg t = datalo;
2310                datalo = datahi;
2311                datahi = t;
2312            }
2313            tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2314                                     h.base, h.index, 0, h.ofs);
2315            tcg_out_modrm_sib_offset(s, movop + h.seg, datahi,
2316                                     h.base, h.index, 0, h.ofs + 4);
2317        }
2318        break;
2319
2320    case MO_128:
2321        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
2322
2323        /*
2324         * Without 16-byte atomicity, use integer regs.
2325         * That is where we have the data, and it allows bswaps.
2326         */
2327        if (h.aa.atom < MO_128) {
2328            if (use_movbe) {
2329                TCGReg t = datalo;
2330                datalo = datahi;
2331                datahi = t;
2332            }
2333            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2334                                     h.base, h.index, 0, h.ofs);
2335            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datahi,
2336                                     h.base, h.index, 0, h.ofs + 8);
2337            break;
2338        }
2339
2340        /*
2341         * With 16-byte atomicity, a vector store is required.
2342         * If we already have 16-byte alignment, then VMOVDQA always works.
2343         * Else if VMOVDQU has atomicity with dynamic alignment, use that.
2344         * Else use we require a runtime test for alignment for VMOVDQA;
2345         * use VMOVDQU on the unaligned nonatomic path for simplicity.
2346         */
2347        tcg_out_pair_to_vec(s, TCG_TYPE_I64, TCG_TMP_VEC, datalo, datahi);
2348        if (h.aa.align >= MO_128) {
2349            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_WxVx + h.seg,
2350                                         TCG_TMP_VEC, 0,
2351                                         h.base, h.index, 0, h.ofs);
2352        } else if (cpuinfo & CPUINFO_ATOMIC_VMOVDQU) {
2353            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_WxVx + h.seg,
2354                                         TCG_TMP_VEC, 0,
2355                                         h.base, h.index, 0, h.ofs);
2356        } else {
2357            TCGLabel *l1 = gen_new_label();
2358            TCGLabel *l2 = gen_new_label();
2359
2360            tcg_out_testi(s, h.base, 15);
2361            tcg_out_jxx(s, JCC_JNE, l1, true);
2362
2363            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_WxVx + h.seg,
2364                                         TCG_TMP_VEC, 0,
2365                                         h.base, h.index, 0, h.ofs);
2366            tcg_out_jxx(s, JCC_JMP, l2, true);
2367
2368            tcg_out_label(s, l1);
2369            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_WxVx + h.seg,
2370                                         TCG_TMP_VEC, 0,
2371                                         h.base, h.index, 0, h.ofs);
2372            tcg_out_label(s, l2);
2373        }
2374        break;
2375
2376    default:
2377        g_assert_not_reached();
2378    }
2379}
2380
2381static void tcg_out_qemu_st(TCGContext *s, TCGReg datalo, TCGReg datahi,
2382                            TCGReg addrlo, TCGReg addrhi,
2383                            MemOpIdx oi, TCGType data_type)
2384{
2385    TCGLabelQemuLdst *ldst;
2386    HostAddress h;
2387
2388    ldst = prepare_host_addr(s, &h, addrlo, addrhi, oi, false);
2389    tcg_out_qemu_st_direct(s, datalo, datahi, h, get_memop(oi));
2390
2391    if (ldst) {
2392        ldst->type = data_type;
2393        ldst->datalo_reg = datalo;
2394        ldst->datahi_reg = datahi;
2395        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
2396    }
2397}
2398
2399static void tcg_out_exit_tb(TCGContext *s, uintptr_t a0)
2400{
2401    /* Reuse the zeroing that exists for goto_ptr.  */
2402    if (a0 == 0) {
2403        tcg_out_jmp(s, tcg_code_gen_epilogue);
2404    } else {
2405        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_EAX, a0);
2406        tcg_out_jmp(s, tb_ret_addr);
2407    }
2408}
2409
2410static void tcg_out_goto_tb(TCGContext *s, int which)
2411{
2412    /*
2413     * Jump displacement must be aligned for atomic patching;
2414     * see if we need to add extra nops before jump
2415     */
2416    int gap = QEMU_ALIGN_PTR_UP(s->code_ptr + 1, 4) - s->code_ptr;
2417    if (gap != 1) {
2418        tcg_out_nopn(s, gap - 1);
2419    }
2420    tcg_out8(s, OPC_JMP_long); /* jmp im */
2421    set_jmp_insn_offset(s, which);
2422    tcg_out32(s, 0);
2423    set_jmp_reset_offset(s, which);
2424}
2425
2426void tb_target_set_jmp_target(const TranslationBlock *tb, int n,
2427                              uintptr_t jmp_rx, uintptr_t jmp_rw)
2428{
2429    /* patch the branch destination */
2430    uintptr_t addr = tb->jmp_target_addr[n];
2431    qatomic_set((int32_t *)jmp_rw, addr - (jmp_rx + 4));
2432    /* no need to flush icache explicitly */
2433}
2434
2435static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
2436                              const TCGArg args[TCG_MAX_OP_ARGS],
2437                              const int const_args[TCG_MAX_OP_ARGS])
2438{
2439    TCGArg a0, a1, a2;
2440    int c, const_a2, vexop, rexw = 0;
2441
2442#if TCG_TARGET_REG_BITS == 64
2443# define OP_32_64(x) \
2444        case glue(glue(INDEX_op_, x), _i64): \
2445            rexw = P_REXW; /* FALLTHRU */    \
2446        case glue(glue(INDEX_op_, x), _i32)
2447#else
2448# define OP_32_64(x) \
2449        case glue(glue(INDEX_op_, x), _i32)
2450#endif
2451
2452    /* Hoist the loads of the most common arguments.  */
2453    a0 = args[0];
2454    a1 = args[1];
2455    a2 = args[2];
2456    const_a2 = const_args[2];
2457
2458    switch (opc) {
2459    case INDEX_op_goto_ptr:
2460        /* jmp to the given host address (could be epilogue) */
2461        tcg_out_modrm(s, OPC_GRP5, EXT5_JMPN_Ev, a0);
2462        break;
2463    case INDEX_op_br:
2464        tcg_out_jxx(s, JCC_JMP, arg_label(a0), 0);
2465        break;
2466    OP_32_64(ld8u):
2467        /* Note that we can ignore REXW for the zero-extend to 64-bit.  */
2468        tcg_out_modrm_offset(s, OPC_MOVZBL, a0, a1, a2);
2469        break;
2470    OP_32_64(ld8s):
2471        tcg_out_modrm_offset(s, OPC_MOVSBL + rexw, a0, a1, a2);
2472        break;
2473    OP_32_64(ld16u):
2474        /* Note that we can ignore REXW for the zero-extend to 64-bit.  */
2475        tcg_out_modrm_offset(s, OPC_MOVZWL, a0, a1, a2);
2476        break;
2477    OP_32_64(ld16s):
2478        tcg_out_modrm_offset(s, OPC_MOVSWL + rexw, a0, a1, a2);
2479        break;
2480#if TCG_TARGET_REG_BITS == 64
2481    case INDEX_op_ld32u_i64:
2482#endif
2483    case INDEX_op_ld_i32:
2484        tcg_out_ld(s, TCG_TYPE_I32, a0, a1, a2);
2485        break;
2486
2487    OP_32_64(st8):
2488        if (const_args[0]) {
2489            tcg_out_modrm_offset(s, OPC_MOVB_EvIz, 0, a1, a2);
2490            tcg_out8(s, a0);
2491        } else {
2492            tcg_out_modrm_offset(s, OPC_MOVB_EvGv | P_REXB_R, a0, a1, a2);
2493        }
2494        break;
2495    OP_32_64(st16):
2496        if (const_args[0]) {
2497            tcg_out_modrm_offset(s, OPC_MOVL_EvIz | P_DATA16, 0, a1, a2);
2498            tcg_out16(s, a0);
2499        } else {
2500            tcg_out_modrm_offset(s, OPC_MOVL_EvGv | P_DATA16, a0, a1, a2);
2501        }
2502        break;
2503#if TCG_TARGET_REG_BITS == 64
2504    case INDEX_op_st32_i64:
2505#endif
2506    case INDEX_op_st_i32:
2507        if (const_args[0]) {
2508            tcg_out_modrm_offset(s, OPC_MOVL_EvIz, 0, a1, a2);
2509            tcg_out32(s, a0);
2510        } else {
2511            tcg_out_st(s, TCG_TYPE_I32, a0, a1, a2);
2512        }
2513        break;
2514
2515    OP_32_64(add):
2516        /* For 3-operand addition, use LEA.  */
2517        if (a0 != a1) {
2518            TCGArg c3 = 0;
2519            if (const_a2) {
2520                c3 = a2, a2 = -1;
2521            } else if (a0 == a2) {
2522                /* Watch out for dest = src + dest, since we've removed
2523                   the matching constraint on the add.  */
2524                tgen_arithr(s, ARITH_ADD + rexw, a0, a1);
2525                break;
2526            }
2527
2528            tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, a1, a2, 0, c3);
2529            break;
2530        }
2531        c = ARITH_ADD;
2532        goto gen_arith;
2533    OP_32_64(sub):
2534        c = ARITH_SUB;
2535        goto gen_arith;
2536    OP_32_64(and):
2537        c = ARITH_AND;
2538        goto gen_arith;
2539    OP_32_64(or):
2540        c = ARITH_OR;
2541        goto gen_arith;
2542    OP_32_64(xor):
2543        c = ARITH_XOR;
2544        goto gen_arith;
2545    gen_arith:
2546        if (const_a2) {
2547            tgen_arithi(s, c + rexw, a0, a2, 0);
2548        } else {
2549            tgen_arithr(s, c + rexw, a0, a2);
2550        }
2551        break;
2552
2553    OP_32_64(andc):
2554        if (const_a2) {
2555            tcg_out_mov(s, rexw ? TCG_TYPE_I64 : TCG_TYPE_I32, a0, a1);
2556            tgen_arithi(s, ARITH_AND + rexw, a0, ~a2, 0);
2557        } else {
2558            tcg_out_vex_modrm(s, OPC_ANDN + rexw, a0, a2, a1);
2559        }
2560        break;
2561
2562    OP_32_64(mul):
2563        if (const_a2) {
2564            int32_t val;
2565            val = a2;
2566            if (val == (int8_t)val) {
2567                tcg_out_modrm(s, OPC_IMUL_GvEvIb + rexw, a0, a0);
2568                tcg_out8(s, val);
2569            } else {
2570                tcg_out_modrm(s, OPC_IMUL_GvEvIz + rexw, a0, a0);
2571                tcg_out32(s, val);
2572            }
2573        } else {
2574            tcg_out_modrm(s, OPC_IMUL_GvEv + rexw, a0, a2);
2575        }
2576        break;
2577
2578    OP_32_64(div2):
2579        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_IDIV, args[4]);
2580        break;
2581    OP_32_64(divu2):
2582        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_DIV, args[4]);
2583        break;
2584
2585    OP_32_64(shl):
2586        /* For small constant 3-operand shift, use LEA.  */
2587        if (const_a2 && a0 != a1 && (a2 - 1) < 3) {
2588            if (a2 - 1 == 0) {
2589                /* shl $1,a1,a0 -> lea (a1,a1),a0 */
2590                tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, a1, a1, 0, 0);
2591            } else {
2592                /* shl $n,a1,a0 -> lea 0(,a1,n),a0 */
2593                tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, -1, a1, a2, 0);
2594            }
2595            break;
2596        }
2597        c = SHIFT_SHL;
2598        vexop = OPC_SHLX;
2599        goto gen_shift_maybe_vex;
2600    OP_32_64(shr):
2601        c = SHIFT_SHR;
2602        vexop = OPC_SHRX;
2603        goto gen_shift_maybe_vex;
2604    OP_32_64(sar):
2605        c = SHIFT_SAR;
2606        vexop = OPC_SARX;
2607        goto gen_shift_maybe_vex;
2608    OP_32_64(rotl):
2609        c = SHIFT_ROL;
2610        goto gen_shift;
2611    OP_32_64(rotr):
2612        c = SHIFT_ROR;
2613        goto gen_shift;
2614    gen_shift_maybe_vex:
2615        if (have_bmi2) {
2616            if (!const_a2) {
2617                tcg_out_vex_modrm(s, vexop + rexw, a0, a2, a1);
2618                break;
2619            }
2620            tcg_out_mov(s, rexw ? TCG_TYPE_I64 : TCG_TYPE_I32, a0, a1);
2621        }
2622        /* FALLTHRU */
2623    gen_shift:
2624        if (const_a2) {
2625            tcg_out_shifti(s, c + rexw, a0, a2);
2626        } else {
2627            tcg_out_modrm(s, OPC_SHIFT_cl + rexw, c, a0);
2628        }
2629        break;
2630
2631    OP_32_64(ctz):
2632        tcg_out_ctz(s, rexw, args[0], args[1], args[2], const_args[2]);
2633        break;
2634    OP_32_64(clz):
2635        tcg_out_clz(s, rexw, args[0], args[1], args[2], const_args[2]);
2636        break;
2637    OP_32_64(ctpop):
2638        tcg_out_modrm(s, OPC_POPCNT + rexw, a0, a1);
2639        break;
2640
2641    OP_32_64(brcond):
2642        tcg_out_brcond(s, rexw, a2, a0, a1, const_args[1],
2643                       arg_label(args[3]), 0);
2644        break;
2645    OP_32_64(setcond):
2646        tcg_out_setcond(s, rexw, args[3], a0, a1, a2, const_a2, false);
2647        break;
2648    OP_32_64(negsetcond):
2649        tcg_out_setcond(s, rexw, args[3], a0, a1, a2, const_a2, true);
2650        break;
2651    OP_32_64(movcond):
2652        tcg_out_movcond(s, rexw, args[5], a0, a1, a2, const_a2, args[3]);
2653        break;
2654
2655    OP_32_64(bswap16):
2656        if (a2 & TCG_BSWAP_OS) {
2657            /* Output must be sign-extended. */
2658            if (rexw) {
2659                tcg_out_bswap64(s, a0);
2660                tcg_out_shifti(s, SHIFT_SAR + rexw, a0, 48);
2661            } else {
2662                tcg_out_bswap32(s, a0);
2663                tcg_out_shifti(s, SHIFT_SAR, a0, 16);
2664            }
2665        } else if ((a2 & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
2666            /* Output must be zero-extended, but input isn't. */
2667            tcg_out_bswap32(s, a0);
2668            tcg_out_shifti(s, SHIFT_SHR, a0, 16);
2669        } else {
2670            tcg_out_rolw_8(s, a0);
2671        }
2672        break;
2673    OP_32_64(bswap32):
2674        tcg_out_bswap32(s, a0);
2675        if (rexw && (a2 & TCG_BSWAP_OS)) {
2676            tcg_out_ext32s(s, a0, a0);
2677        }
2678        break;
2679
2680    OP_32_64(neg):
2681        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NEG, a0);
2682        break;
2683    OP_32_64(not):
2684        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NOT, a0);
2685        break;
2686
2687    case INDEX_op_qemu_ld_a64_i32:
2688        if (TCG_TARGET_REG_BITS == 32) {
2689            tcg_out_qemu_ld(s, a0, -1, a1, a2, args[3], TCG_TYPE_I32);
2690            break;
2691        }
2692        /* fall through */
2693    case INDEX_op_qemu_ld_a32_i32:
2694        tcg_out_qemu_ld(s, a0, -1, a1, -1, a2, TCG_TYPE_I32);
2695        break;
2696    case INDEX_op_qemu_ld_a32_i64:
2697        if (TCG_TARGET_REG_BITS == 64) {
2698            tcg_out_qemu_ld(s, a0, -1, a1, -1, a2, TCG_TYPE_I64);
2699        } else {
2700            tcg_out_qemu_ld(s, a0, a1, a2, -1, args[3], TCG_TYPE_I64);
2701        }
2702        break;
2703    case INDEX_op_qemu_ld_a64_i64:
2704        if (TCG_TARGET_REG_BITS == 64) {
2705            tcg_out_qemu_ld(s, a0, -1, a1, -1, a2, TCG_TYPE_I64);
2706        } else {
2707            tcg_out_qemu_ld(s, a0, a1, a2, args[3], args[4], TCG_TYPE_I64);
2708        }
2709        break;
2710    case INDEX_op_qemu_ld_a32_i128:
2711    case INDEX_op_qemu_ld_a64_i128:
2712        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
2713        tcg_out_qemu_ld(s, a0, a1, a2, -1, args[3], TCG_TYPE_I128);
2714        break;
2715
2716    case INDEX_op_qemu_st_a64_i32:
2717    case INDEX_op_qemu_st8_a64_i32:
2718        if (TCG_TARGET_REG_BITS == 32) {
2719            tcg_out_qemu_st(s, a0, -1, a1, a2, args[3], TCG_TYPE_I32);
2720            break;
2721        }
2722        /* fall through */
2723    case INDEX_op_qemu_st_a32_i32:
2724    case INDEX_op_qemu_st8_a32_i32:
2725        tcg_out_qemu_st(s, a0, -1, a1, -1, a2, TCG_TYPE_I32);
2726        break;
2727    case INDEX_op_qemu_st_a32_i64:
2728        if (TCG_TARGET_REG_BITS == 64) {
2729            tcg_out_qemu_st(s, a0, -1, a1, -1, a2, TCG_TYPE_I64);
2730        } else {
2731            tcg_out_qemu_st(s, a0, a1, a2, -1, args[3], TCG_TYPE_I64);
2732        }
2733        break;
2734    case INDEX_op_qemu_st_a64_i64:
2735        if (TCG_TARGET_REG_BITS == 64) {
2736            tcg_out_qemu_st(s, a0, -1, a1, -1, a2, TCG_TYPE_I64);
2737        } else {
2738            tcg_out_qemu_st(s, a0, a1, a2, args[3], args[4], TCG_TYPE_I64);
2739        }
2740        break;
2741    case INDEX_op_qemu_st_a32_i128:
2742    case INDEX_op_qemu_st_a64_i128:
2743        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
2744        tcg_out_qemu_st(s, a0, a1, a2, -1, args[3], TCG_TYPE_I128);
2745        break;
2746
2747    OP_32_64(mulu2):
2748        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_MUL, args[3]);
2749        break;
2750    OP_32_64(muls2):
2751        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_IMUL, args[3]);
2752        break;
2753    OP_32_64(add2):
2754        if (const_args[4]) {
2755            tgen_arithi(s, ARITH_ADD + rexw, a0, args[4], 1);
2756        } else {
2757            tgen_arithr(s, ARITH_ADD + rexw, a0, args[4]);
2758        }
2759        if (const_args[5]) {
2760            tgen_arithi(s, ARITH_ADC + rexw, a1, args[5], 1);
2761        } else {
2762            tgen_arithr(s, ARITH_ADC + rexw, a1, args[5]);
2763        }
2764        break;
2765    OP_32_64(sub2):
2766        if (const_args[4]) {
2767            tgen_arithi(s, ARITH_SUB + rexw, a0, args[4], 1);
2768        } else {
2769            tgen_arithr(s, ARITH_SUB + rexw, a0, args[4]);
2770        }
2771        if (const_args[5]) {
2772            tgen_arithi(s, ARITH_SBB + rexw, a1, args[5], 1);
2773        } else {
2774            tgen_arithr(s, ARITH_SBB + rexw, a1, args[5]);
2775        }
2776        break;
2777
2778#if TCG_TARGET_REG_BITS == 32
2779    case INDEX_op_brcond2_i32:
2780        tcg_out_brcond2(s, args, const_args, 0);
2781        break;
2782    case INDEX_op_setcond2_i32:
2783        tcg_out_setcond2(s, args, const_args);
2784        break;
2785#else /* TCG_TARGET_REG_BITS == 64 */
2786    case INDEX_op_ld32s_i64:
2787        tcg_out_modrm_offset(s, OPC_MOVSLQ, a0, a1, a2);
2788        break;
2789    case INDEX_op_ld_i64:
2790        tcg_out_ld(s, TCG_TYPE_I64, a0, a1, a2);
2791        break;
2792    case INDEX_op_st_i64:
2793        if (const_args[0]) {
2794            tcg_out_modrm_offset(s, OPC_MOVL_EvIz | P_REXW, 0, a1, a2);
2795            tcg_out32(s, a0);
2796        } else {
2797            tcg_out_st(s, TCG_TYPE_I64, a0, a1, a2);
2798        }
2799        break;
2800
2801    case INDEX_op_bswap64_i64:
2802        tcg_out_bswap64(s, a0);
2803        break;
2804    case INDEX_op_extrh_i64_i32:
2805        tcg_out_shifti(s, SHIFT_SHR + P_REXW, a0, 32);
2806        break;
2807#endif
2808
2809    OP_32_64(deposit):
2810        if (args[3] == 0 && args[4] == 8) {
2811            /* load bits 0..7 */
2812            if (const_a2) {
2813                tcg_out_opc(s, OPC_MOVB_Ib | P_REXB_RM | LOWREGMASK(a0),
2814                            0, a0, 0);
2815                tcg_out8(s, a2);
2816            } else {
2817                tcg_out_modrm(s, OPC_MOVB_EvGv | P_REXB_R | P_REXB_RM, a2, a0);
2818            }
2819        } else if (TCG_TARGET_REG_BITS == 32 && args[3] == 8 && args[4] == 8) {
2820            /* load bits 8..15 */
2821            if (const_a2) {
2822                tcg_out8(s, OPC_MOVB_Ib + a0 + 4);
2823                tcg_out8(s, a2);
2824            } else {
2825                tcg_out_modrm(s, OPC_MOVB_EvGv, a2, a0 + 4);
2826            }
2827        } else if (args[3] == 0 && args[4] == 16) {
2828            /* load bits 0..15 */
2829            if (const_a2) {
2830                tcg_out_opc(s, OPC_MOVL_Iv | P_DATA16 | LOWREGMASK(a0),
2831                            0, a0, 0);
2832                tcg_out16(s, a2);
2833            } else {
2834                tcg_out_modrm(s, OPC_MOVL_EvGv | P_DATA16, a2, a0);
2835            }
2836        } else {
2837            g_assert_not_reached();
2838        }
2839        break;
2840
2841    case INDEX_op_extract_i64:
2842        if (a2 + args[3] == 32) {
2843            /* This is a 32-bit zero-extending right shift.  */
2844            tcg_out_mov(s, TCG_TYPE_I32, a0, a1);
2845            tcg_out_shifti(s, SHIFT_SHR, a0, a2);
2846            break;
2847        }
2848        /* FALLTHRU */
2849    case INDEX_op_extract_i32:
2850        /* On the off-chance that we can use the high-byte registers.
2851           Otherwise we emit the same ext16 + shift pattern that we
2852           would have gotten from the normal tcg-op.c expansion.  */
2853        tcg_debug_assert(a2 == 8 && args[3] == 8);
2854        if (a1 < 4 && a0 < 8) {
2855            tcg_out_modrm(s, OPC_MOVZBL, a0, a1 + 4);
2856        } else {
2857            tcg_out_ext16u(s, a0, a1);
2858            tcg_out_shifti(s, SHIFT_SHR, a0, 8);
2859        }
2860        break;
2861
2862    case INDEX_op_sextract_i32:
2863        /* We don't implement sextract_i64, as we cannot sign-extend to
2864           64-bits without using the REX prefix that explicitly excludes
2865           access to the high-byte registers.  */
2866        tcg_debug_assert(a2 == 8 && args[3] == 8);
2867        if (a1 < 4 && a0 < 8) {
2868            tcg_out_modrm(s, OPC_MOVSBL, a0, a1 + 4);
2869        } else {
2870            tcg_out_ext16s(s, TCG_TYPE_I32, a0, a1);
2871            tcg_out_shifti(s, SHIFT_SAR, a0, 8);
2872        }
2873        break;
2874
2875    OP_32_64(extract2):
2876        /* Note that SHRD outputs to the r/m operand.  */
2877        tcg_out_modrm(s, OPC_SHRD_Ib + rexw, a2, a0);
2878        tcg_out8(s, args[3]);
2879        break;
2880
2881    case INDEX_op_mb:
2882        tcg_out_mb(s, a0);
2883        break;
2884    case INDEX_op_mov_i32:  /* Always emitted via tcg_out_mov.  */
2885    case INDEX_op_mov_i64:
2886    case INDEX_op_call:     /* Always emitted via tcg_out_call.  */
2887    case INDEX_op_exit_tb:  /* Always emitted via tcg_out_exit_tb.  */
2888    case INDEX_op_goto_tb:  /* Always emitted via tcg_out_goto_tb.  */
2889    case INDEX_op_ext8s_i32:  /* Always emitted via tcg_reg_alloc_op.  */
2890    case INDEX_op_ext8s_i64:
2891    case INDEX_op_ext8u_i32:
2892    case INDEX_op_ext8u_i64:
2893    case INDEX_op_ext16s_i32:
2894    case INDEX_op_ext16s_i64:
2895    case INDEX_op_ext16u_i32:
2896    case INDEX_op_ext16u_i64:
2897    case INDEX_op_ext32s_i64:
2898    case INDEX_op_ext32u_i64:
2899    case INDEX_op_ext_i32_i64:
2900    case INDEX_op_extu_i32_i64:
2901    case INDEX_op_extrl_i64_i32:
2902    default:
2903        g_assert_not_reached();
2904    }
2905
2906#undef OP_32_64
2907}
2908
2909static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
2910                           unsigned vecl, unsigned vece,
2911                           const TCGArg args[TCG_MAX_OP_ARGS],
2912                           const int const_args[TCG_MAX_OP_ARGS])
2913{
2914    static int const add_insn[4] = {
2915        OPC_PADDB, OPC_PADDW, OPC_PADDD, OPC_PADDQ
2916    };
2917    static int const ssadd_insn[4] = {
2918        OPC_PADDSB, OPC_PADDSW, OPC_UD2, OPC_UD2
2919    };
2920    static int const usadd_insn[4] = {
2921        OPC_PADDUB, OPC_PADDUW, OPC_UD2, OPC_UD2
2922    };
2923    static int const sub_insn[4] = {
2924        OPC_PSUBB, OPC_PSUBW, OPC_PSUBD, OPC_PSUBQ
2925    };
2926    static int const sssub_insn[4] = {
2927        OPC_PSUBSB, OPC_PSUBSW, OPC_UD2, OPC_UD2
2928    };
2929    static int const ussub_insn[4] = {
2930        OPC_PSUBUB, OPC_PSUBUW, OPC_UD2, OPC_UD2
2931    };
2932    static int const mul_insn[4] = {
2933        OPC_UD2, OPC_PMULLW, OPC_PMULLD, OPC_VPMULLQ
2934    };
2935    static int const shift_imm_insn[4] = {
2936        OPC_UD2, OPC_PSHIFTW_Ib, OPC_PSHIFTD_Ib, OPC_PSHIFTQ_Ib
2937    };
2938    static int const cmpeq_insn[4] = {
2939        OPC_PCMPEQB, OPC_PCMPEQW, OPC_PCMPEQD, OPC_PCMPEQQ
2940    };
2941    static int const cmpgt_insn[4] = {
2942        OPC_PCMPGTB, OPC_PCMPGTW, OPC_PCMPGTD, OPC_PCMPGTQ
2943    };
2944    static int const punpckl_insn[4] = {
2945        OPC_PUNPCKLBW, OPC_PUNPCKLWD, OPC_PUNPCKLDQ, OPC_PUNPCKLQDQ
2946    };
2947    static int const punpckh_insn[4] = {
2948        OPC_PUNPCKHBW, OPC_PUNPCKHWD, OPC_PUNPCKHDQ, OPC_PUNPCKHQDQ
2949    };
2950    static int const packss_insn[4] = {
2951        OPC_PACKSSWB, OPC_PACKSSDW, OPC_UD2, OPC_UD2
2952    };
2953    static int const packus_insn[4] = {
2954        OPC_PACKUSWB, OPC_PACKUSDW, OPC_UD2, OPC_UD2
2955    };
2956    static int const smin_insn[4] = {
2957        OPC_PMINSB, OPC_PMINSW, OPC_PMINSD, OPC_VPMINSQ
2958    };
2959    static int const smax_insn[4] = {
2960        OPC_PMAXSB, OPC_PMAXSW, OPC_PMAXSD, OPC_VPMAXSQ
2961    };
2962    static int const umin_insn[4] = {
2963        OPC_PMINUB, OPC_PMINUW, OPC_PMINUD, OPC_VPMINUQ
2964    };
2965    static int const umax_insn[4] = {
2966        OPC_PMAXUB, OPC_PMAXUW, OPC_PMAXUD, OPC_VPMAXUQ
2967    };
2968    static int const rotlv_insn[4] = {
2969        OPC_UD2, OPC_UD2, OPC_VPROLVD, OPC_VPROLVQ
2970    };
2971    static int const rotrv_insn[4] = {
2972        OPC_UD2, OPC_UD2, OPC_VPRORVD, OPC_VPRORVQ
2973    };
2974    static int const shlv_insn[4] = {
2975        OPC_UD2, OPC_VPSLLVW, OPC_VPSLLVD, OPC_VPSLLVQ
2976    };
2977    static int const shrv_insn[4] = {
2978        OPC_UD2, OPC_VPSRLVW, OPC_VPSRLVD, OPC_VPSRLVQ
2979    };
2980    static int const sarv_insn[4] = {
2981        OPC_UD2, OPC_VPSRAVW, OPC_VPSRAVD, OPC_VPSRAVQ
2982    };
2983    static int const shls_insn[4] = {
2984        OPC_UD2, OPC_PSLLW, OPC_PSLLD, OPC_PSLLQ
2985    };
2986    static int const shrs_insn[4] = {
2987        OPC_UD2, OPC_PSRLW, OPC_PSRLD, OPC_PSRLQ
2988    };
2989    static int const sars_insn[4] = {
2990        OPC_UD2, OPC_PSRAW, OPC_PSRAD, OPC_VPSRAQ
2991    };
2992    static int const vpshldi_insn[4] = {
2993        OPC_UD2, OPC_VPSHLDW, OPC_VPSHLDD, OPC_VPSHLDQ
2994    };
2995    static int const vpshldv_insn[4] = {
2996        OPC_UD2, OPC_VPSHLDVW, OPC_VPSHLDVD, OPC_VPSHLDVQ
2997    };
2998    static int const vpshrdv_insn[4] = {
2999        OPC_UD2, OPC_VPSHRDVW, OPC_VPSHRDVD, OPC_VPSHRDVQ
3000    };
3001    static int const abs_insn[4] = {
3002        OPC_PABSB, OPC_PABSW, OPC_PABSD, OPC_VPABSQ
3003    };
3004
3005    TCGType type = vecl + TCG_TYPE_V64;
3006    int insn, sub;
3007    TCGArg a0, a1, a2, a3;
3008
3009    a0 = args[0];
3010    a1 = args[1];
3011    a2 = args[2];
3012
3013    switch (opc) {
3014    case INDEX_op_add_vec:
3015        insn = add_insn[vece];
3016        goto gen_simd;
3017    case INDEX_op_ssadd_vec:
3018        insn = ssadd_insn[vece];
3019        goto gen_simd;
3020    case INDEX_op_usadd_vec:
3021        insn = usadd_insn[vece];
3022        goto gen_simd;
3023    case INDEX_op_sub_vec:
3024        insn = sub_insn[vece];
3025        goto gen_simd;
3026    case INDEX_op_sssub_vec:
3027        insn = sssub_insn[vece];
3028        goto gen_simd;
3029    case INDEX_op_ussub_vec:
3030        insn = ussub_insn[vece];
3031        goto gen_simd;
3032    case INDEX_op_mul_vec:
3033        insn = mul_insn[vece];
3034        goto gen_simd;
3035    case INDEX_op_and_vec:
3036        insn = OPC_PAND;
3037        goto gen_simd;
3038    case INDEX_op_or_vec:
3039        insn = OPC_POR;
3040        goto gen_simd;
3041    case INDEX_op_xor_vec:
3042        insn = OPC_PXOR;
3043        goto gen_simd;
3044    case INDEX_op_smin_vec:
3045        insn = smin_insn[vece];
3046        goto gen_simd;
3047    case INDEX_op_umin_vec:
3048        insn = umin_insn[vece];
3049        goto gen_simd;
3050    case INDEX_op_smax_vec:
3051        insn = smax_insn[vece];
3052        goto gen_simd;
3053    case INDEX_op_umax_vec:
3054        insn = umax_insn[vece];
3055        goto gen_simd;
3056    case INDEX_op_shlv_vec:
3057        insn = shlv_insn[vece];
3058        goto gen_simd;
3059    case INDEX_op_shrv_vec:
3060        insn = shrv_insn[vece];
3061        goto gen_simd;
3062    case INDEX_op_sarv_vec:
3063        insn = sarv_insn[vece];
3064        goto gen_simd;
3065    case INDEX_op_rotlv_vec:
3066        insn = rotlv_insn[vece];
3067        goto gen_simd;
3068    case INDEX_op_rotrv_vec:
3069        insn = rotrv_insn[vece];
3070        goto gen_simd;
3071    case INDEX_op_shls_vec:
3072        insn = shls_insn[vece];
3073        goto gen_simd;
3074    case INDEX_op_shrs_vec:
3075        insn = shrs_insn[vece];
3076        goto gen_simd;
3077    case INDEX_op_sars_vec:
3078        insn = sars_insn[vece];
3079        goto gen_simd;
3080    case INDEX_op_x86_punpckl_vec:
3081        insn = punpckl_insn[vece];
3082        goto gen_simd;
3083    case INDEX_op_x86_punpckh_vec:
3084        insn = punpckh_insn[vece];
3085        goto gen_simd;
3086    case INDEX_op_x86_packss_vec:
3087        insn = packss_insn[vece];
3088        goto gen_simd;
3089    case INDEX_op_x86_packus_vec:
3090        insn = packus_insn[vece];
3091        goto gen_simd;
3092    case INDEX_op_x86_vpshldv_vec:
3093        insn = vpshldv_insn[vece];
3094        a1 = a2;
3095        a2 = args[3];
3096        goto gen_simd;
3097    case INDEX_op_x86_vpshrdv_vec:
3098        insn = vpshrdv_insn[vece];
3099        a1 = a2;
3100        a2 = args[3];
3101        goto gen_simd;
3102#if TCG_TARGET_REG_BITS == 32
3103    case INDEX_op_dup2_vec:
3104        /* First merge the two 32-bit inputs to a single 64-bit element. */
3105        tcg_out_vex_modrm(s, OPC_PUNPCKLDQ, a0, a1, a2);
3106        /* Then replicate the 64-bit elements across the rest of the vector. */
3107        if (type != TCG_TYPE_V64) {
3108            tcg_out_dup_vec(s, type, MO_64, a0, a0);
3109        }
3110        break;
3111#endif
3112    case INDEX_op_abs_vec:
3113        insn = abs_insn[vece];
3114        a2 = a1;
3115        a1 = 0;
3116        goto gen_simd;
3117    gen_simd:
3118        tcg_debug_assert(insn != OPC_UD2);
3119        if (type == TCG_TYPE_V256) {
3120            insn |= P_VEXL;
3121        }
3122        tcg_out_vex_modrm(s, insn, a0, a1, a2);
3123        break;
3124
3125    case INDEX_op_cmp_vec:
3126        sub = args[3];
3127        if (sub == TCG_COND_EQ) {
3128            insn = cmpeq_insn[vece];
3129        } else if (sub == TCG_COND_GT) {
3130            insn = cmpgt_insn[vece];
3131        } else {
3132            g_assert_not_reached();
3133        }
3134        goto gen_simd;
3135
3136    case INDEX_op_andc_vec:
3137        insn = OPC_PANDN;
3138        if (type == TCG_TYPE_V256) {
3139            insn |= P_VEXL;
3140        }
3141        tcg_out_vex_modrm(s, insn, a0, a2, a1);
3142        break;
3143
3144    case INDEX_op_shli_vec:
3145        insn = shift_imm_insn[vece];
3146        sub = 6;
3147        goto gen_shift;
3148    case INDEX_op_shri_vec:
3149        insn = shift_imm_insn[vece];
3150        sub = 2;
3151        goto gen_shift;
3152    case INDEX_op_sari_vec:
3153        if (vece == MO_64) {
3154            insn = OPC_PSHIFTD_Ib | P_VEXW | P_EVEX;
3155        } else {
3156            insn = shift_imm_insn[vece];
3157        }
3158        sub = 4;
3159        goto gen_shift;
3160    case INDEX_op_rotli_vec:
3161        insn = OPC_PSHIFTD_Ib | P_EVEX;  /* VPROL[DQ] */
3162        if (vece == MO_64) {
3163            insn |= P_VEXW;
3164        }
3165        sub = 1;
3166        goto gen_shift;
3167    gen_shift:
3168        tcg_debug_assert(vece != MO_8);
3169        if (type == TCG_TYPE_V256) {
3170            insn |= P_VEXL;
3171        }
3172        tcg_out_vex_modrm(s, insn, sub, a0, a1);
3173        tcg_out8(s, a2);
3174        break;
3175
3176    case INDEX_op_ld_vec:
3177        tcg_out_ld(s, type, a0, a1, a2);
3178        break;
3179    case INDEX_op_st_vec:
3180        tcg_out_st(s, type, a0, a1, a2);
3181        break;
3182    case INDEX_op_dupm_vec:
3183        tcg_out_dupm_vec(s, type, vece, a0, a1, a2);
3184        break;
3185
3186    case INDEX_op_x86_shufps_vec:
3187        insn = OPC_SHUFPS;
3188        sub = args[3];
3189        goto gen_simd_imm8;
3190    case INDEX_op_x86_blend_vec:
3191        if (vece == MO_16) {
3192            insn = OPC_PBLENDW;
3193        } else if (vece == MO_32) {
3194            insn = (have_avx2 ? OPC_VPBLENDD : OPC_BLENDPS);
3195        } else {
3196            g_assert_not_reached();
3197        }
3198        sub = args[3];
3199        goto gen_simd_imm8;
3200    case INDEX_op_x86_vperm2i128_vec:
3201        insn = OPC_VPERM2I128;
3202        sub = args[3];
3203        goto gen_simd_imm8;
3204    case INDEX_op_x86_vpshldi_vec:
3205        insn = vpshldi_insn[vece];
3206        sub = args[3];
3207        goto gen_simd_imm8;
3208
3209    case INDEX_op_not_vec:
3210        insn = OPC_VPTERNLOGQ;
3211        a2 = a1;
3212        sub = 0x33; /* !B */
3213        goto gen_simd_imm8;
3214    case INDEX_op_nor_vec:
3215        insn = OPC_VPTERNLOGQ;
3216        sub = 0x11; /* norCB */
3217        goto gen_simd_imm8;
3218    case INDEX_op_nand_vec:
3219        insn = OPC_VPTERNLOGQ;
3220        sub = 0x77; /* nandCB */
3221        goto gen_simd_imm8;
3222    case INDEX_op_eqv_vec:
3223        insn = OPC_VPTERNLOGQ;
3224        sub = 0x99; /* xnorCB */
3225        goto gen_simd_imm8;
3226    case INDEX_op_orc_vec:
3227        insn = OPC_VPTERNLOGQ;
3228        sub = 0xdd; /* orB!C */
3229        goto gen_simd_imm8;
3230
3231    case INDEX_op_bitsel_vec:
3232        insn = OPC_VPTERNLOGQ;
3233        a3 = args[3];
3234        if (a0 == a1) {
3235            a1 = a2;
3236            a2 = a3;
3237            sub = 0xca; /* A?B:C */
3238        } else if (a0 == a2) {
3239            a2 = a3;
3240            sub = 0xe2; /* B?A:C */
3241        } else {
3242            tcg_out_mov(s, type, a0, a3);
3243            sub = 0xb8; /* B?C:A */
3244        }
3245        goto gen_simd_imm8;
3246
3247    gen_simd_imm8:
3248        tcg_debug_assert(insn != OPC_UD2);
3249        if (type == TCG_TYPE_V256) {
3250            insn |= P_VEXL;
3251        }
3252        tcg_out_vex_modrm(s, insn, a0, a1, a2);
3253        tcg_out8(s, sub);
3254        break;
3255
3256    case INDEX_op_x86_vpblendvb_vec:
3257        insn = OPC_VPBLENDVB;
3258        if (type == TCG_TYPE_V256) {
3259            insn |= P_VEXL;
3260        }
3261        tcg_out_vex_modrm(s, insn, a0, a1, a2);
3262        tcg_out8(s, args[3] << 4);
3263        break;
3264
3265    case INDEX_op_x86_psrldq_vec:
3266        tcg_out_vex_modrm(s, OPC_GRP14, 3, a0, a1);
3267        tcg_out8(s, a2);
3268        break;
3269
3270    case INDEX_op_mov_vec:  /* Always emitted via tcg_out_mov.  */
3271    case INDEX_op_dup_vec:  /* Always emitted via tcg_out_dup_vec.  */
3272    default:
3273        g_assert_not_reached();
3274    }
3275}
3276
3277static TCGConstraintSetIndex tcg_target_op_def(TCGOpcode op)
3278{
3279    switch (op) {
3280    case INDEX_op_goto_ptr:
3281        return C_O0_I1(r);
3282
3283    case INDEX_op_ld8u_i32:
3284    case INDEX_op_ld8u_i64:
3285    case INDEX_op_ld8s_i32:
3286    case INDEX_op_ld8s_i64:
3287    case INDEX_op_ld16u_i32:
3288    case INDEX_op_ld16u_i64:
3289    case INDEX_op_ld16s_i32:
3290    case INDEX_op_ld16s_i64:
3291    case INDEX_op_ld_i32:
3292    case INDEX_op_ld32u_i64:
3293    case INDEX_op_ld32s_i64:
3294    case INDEX_op_ld_i64:
3295        return C_O1_I1(r, r);
3296
3297    case INDEX_op_st8_i32:
3298    case INDEX_op_st8_i64:
3299        return C_O0_I2(qi, r);
3300
3301    case INDEX_op_st16_i32:
3302    case INDEX_op_st16_i64:
3303    case INDEX_op_st_i32:
3304    case INDEX_op_st32_i64:
3305        return C_O0_I2(ri, r);
3306
3307    case INDEX_op_st_i64:
3308        return C_O0_I2(re, r);
3309
3310    case INDEX_op_add_i32:
3311    case INDEX_op_add_i64:
3312        return C_O1_I2(r, r, re);
3313
3314    case INDEX_op_sub_i32:
3315    case INDEX_op_sub_i64:
3316    case INDEX_op_mul_i32:
3317    case INDEX_op_mul_i64:
3318    case INDEX_op_or_i32:
3319    case INDEX_op_or_i64:
3320    case INDEX_op_xor_i32:
3321    case INDEX_op_xor_i64:
3322        return C_O1_I2(r, 0, re);
3323
3324    case INDEX_op_and_i32:
3325    case INDEX_op_and_i64:
3326        return C_O1_I2(r, 0, reZ);
3327
3328    case INDEX_op_andc_i32:
3329    case INDEX_op_andc_i64:
3330        return C_O1_I2(r, r, rI);
3331
3332    case INDEX_op_shl_i32:
3333    case INDEX_op_shl_i64:
3334    case INDEX_op_shr_i32:
3335    case INDEX_op_shr_i64:
3336    case INDEX_op_sar_i32:
3337    case INDEX_op_sar_i64:
3338        return have_bmi2 ? C_O1_I2(r, r, ri) : C_O1_I2(r, 0, ci);
3339
3340    case INDEX_op_rotl_i32:
3341    case INDEX_op_rotl_i64:
3342    case INDEX_op_rotr_i32:
3343    case INDEX_op_rotr_i64:
3344        return C_O1_I2(r, 0, ci);
3345
3346    case INDEX_op_brcond_i32:
3347    case INDEX_op_brcond_i64:
3348        return C_O0_I2(r, re);
3349
3350    case INDEX_op_bswap16_i32:
3351    case INDEX_op_bswap16_i64:
3352    case INDEX_op_bswap32_i32:
3353    case INDEX_op_bswap32_i64:
3354    case INDEX_op_bswap64_i64:
3355    case INDEX_op_neg_i32:
3356    case INDEX_op_neg_i64:
3357    case INDEX_op_not_i32:
3358    case INDEX_op_not_i64:
3359    case INDEX_op_extrh_i64_i32:
3360        return C_O1_I1(r, 0);
3361
3362    case INDEX_op_ext8s_i32:
3363    case INDEX_op_ext8s_i64:
3364    case INDEX_op_ext8u_i32:
3365    case INDEX_op_ext8u_i64:
3366        return C_O1_I1(r, q);
3367
3368    case INDEX_op_ext16s_i32:
3369    case INDEX_op_ext16s_i64:
3370    case INDEX_op_ext16u_i32:
3371    case INDEX_op_ext16u_i64:
3372    case INDEX_op_ext32s_i64:
3373    case INDEX_op_ext32u_i64:
3374    case INDEX_op_ext_i32_i64:
3375    case INDEX_op_extu_i32_i64:
3376    case INDEX_op_extrl_i64_i32:
3377    case INDEX_op_extract_i32:
3378    case INDEX_op_extract_i64:
3379    case INDEX_op_sextract_i32:
3380    case INDEX_op_ctpop_i32:
3381    case INDEX_op_ctpop_i64:
3382        return C_O1_I1(r, r);
3383
3384    case INDEX_op_extract2_i32:
3385    case INDEX_op_extract2_i64:
3386        return C_O1_I2(r, 0, r);
3387
3388    case INDEX_op_deposit_i32:
3389    case INDEX_op_deposit_i64:
3390        return C_O1_I2(q, 0, qi);
3391
3392    case INDEX_op_setcond_i32:
3393    case INDEX_op_setcond_i64:
3394    case INDEX_op_negsetcond_i32:
3395    case INDEX_op_negsetcond_i64:
3396        return C_O1_I2(q, r, re);
3397
3398    case INDEX_op_movcond_i32:
3399    case INDEX_op_movcond_i64:
3400        return C_O1_I4(r, r, re, r, 0);
3401
3402    case INDEX_op_div2_i32:
3403    case INDEX_op_div2_i64:
3404    case INDEX_op_divu2_i32:
3405    case INDEX_op_divu2_i64:
3406        return C_O2_I3(a, d, 0, 1, r);
3407
3408    case INDEX_op_mulu2_i32:
3409    case INDEX_op_mulu2_i64:
3410    case INDEX_op_muls2_i32:
3411    case INDEX_op_muls2_i64:
3412        return C_O2_I2(a, d, a, r);
3413
3414    case INDEX_op_add2_i32:
3415    case INDEX_op_add2_i64:
3416    case INDEX_op_sub2_i32:
3417    case INDEX_op_sub2_i64:
3418        return C_N1_O1_I4(r, r, 0, 1, re, re);
3419
3420    case INDEX_op_ctz_i32:
3421    case INDEX_op_ctz_i64:
3422        return have_bmi1 ? C_N1_I2(r, r, rW) : C_N1_I2(r, r, r);
3423
3424    case INDEX_op_clz_i32:
3425    case INDEX_op_clz_i64:
3426        return have_lzcnt ? C_N1_I2(r, r, rW) : C_N1_I2(r, r, r);
3427
3428    case INDEX_op_qemu_ld_a32_i32:
3429        return C_O1_I1(r, L);
3430    case INDEX_op_qemu_ld_a64_i32:
3431        return TCG_TARGET_REG_BITS == 64 ? C_O1_I1(r, L) : C_O1_I2(r, L, L);
3432
3433    case INDEX_op_qemu_st_a32_i32:
3434        return C_O0_I2(L, L);
3435    case INDEX_op_qemu_st_a64_i32:
3436        return TCG_TARGET_REG_BITS == 64 ? C_O0_I2(L, L) : C_O0_I3(L, L, L);
3437    case INDEX_op_qemu_st8_a32_i32:
3438        return C_O0_I2(s, L);
3439    case INDEX_op_qemu_st8_a64_i32:
3440        return TCG_TARGET_REG_BITS == 64 ? C_O0_I2(s, L) : C_O0_I3(s, L, L);
3441
3442    case INDEX_op_qemu_ld_a32_i64:
3443        return TCG_TARGET_REG_BITS == 64 ? C_O1_I1(r, L) : C_O2_I1(r, r, L);
3444    case INDEX_op_qemu_ld_a64_i64:
3445        return TCG_TARGET_REG_BITS == 64 ? C_O1_I1(r, L) : C_O2_I2(r, r, L, L);
3446
3447    case INDEX_op_qemu_st_a32_i64:
3448        return TCG_TARGET_REG_BITS == 64 ? C_O0_I2(L, L) : C_O0_I3(L, L, L);
3449    case INDEX_op_qemu_st_a64_i64:
3450        return TCG_TARGET_REG_BITS == 64 ? C_O0_I2(L, L) : C_O0_I4(L, L, L, L);
3451
3452    case INDEX_op_qemu_ld_a32_i128:
3453    case INDEX_op_qemu_ld_a64_i128:
3454        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
3455        return C_O2_I1(r, r, L);
3456    case INDEX_op_qemu_st_a32_i128:
3457    case INDEX_op_qemu_st_a64_i128:
3458        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
3459        return C_O0_I3(L, L, L);
3460
3461    case INDEX_op_brcond2_i32:
3462        return C_O0_I4(r, r, ri, ri);
3463
3464    case INDEX_op_setcond2_i32:
3465        return C_O1_I4(r, r, r, ri, ri);
3466
3467    case INDEX_op_ld_vec:
3468    case INDEX_op_dupm_vec:
3469        return C_O1_I1(x, r);
3470
3471    case INDEX_op_st_vec:
3472        return C_O0_I2(x, r);
3473
3474    case INDEX_op_add_vec:
3475    case INDEX_op_sub_vec:
3476    case INDEX_op_mul_vec:
3477    case INDEX_op_and_vec:
3478    case INDEX_op_or_vec:
3479    case INDEX_op_xor_vec:
3480    case INDEX_op_andc_vec:
3481    case INDEX_op_orc_vec:
3482    case INDEX_op_nand_vec:
3483    case INDEX_op_nor_vec:
3484    case INDEX_op_eqv_vec:
3485    case INDEX_op_ssadd_vec:
3486    case INDEX_op_usadd_vec:
3487    case INDEX_op_sssub_vec:
3488    case INDEX_op_ussub_vec:
3489    case INDEX_op_smin_vec:
3490    case INDEX_op_umin_vec:
3491    case INDEX_op_smax_vec:
3492    case INDEX_op_umax_vec:
3493    case INDEX_op_shlv_vec:
3494    case INDEX_op_shrv_vec:
3495    case INDEX_op_sarv_vec:
3496    case INDEX_op_rotlv_vec:
3497    case INDEX_op_rotrv_vec:
3498    case INDEX_op_shls_vec:
3499    case INDEX_op_shrs_vec:
3500    case INDEX_op_sars_vec:
3501    case INDEX_op_cmp_vec:
3502    case INDEX_op_x86_shufps_vec:
3503    case INDEX_op_x86_blend_vec:
3504    case INDEX_op_x86_packss_vec:
3505    case INDEX_op_x86_packus_vec:
3506    case INDEX_op_x86_vperm2i128_vec:
3507    case INDEX_op_x86_punpckl_vec:
3508    case INDEX_op_x86_punpckh_vec:
3509    case INDEX_op_x86_vpshldi_vec:
3510#if TCG_TARGET_REG_BITS == 32
3511    case INDEX_op_dup2_vec:
3512#endif
3513        return C_O1_I2(x, x, x);
3514
3515    case INDEX_op_abs_vec:
3516    case INDEX_op_dup_vec:
3517    case INDEX_op_not_vec:
3518    case INDEX_op_shli_vec:
3519    case INDEX_op_shri_vec:
3520    case INDEX_op_sari_vec:
3521    case INDEX_op_rotli_vec:
3522    case INDEX_op_x86_psrldq_vec:
3523        return C_O1_I1(x, x);
3524
3525    case INDEX_op_x86_vpshldv_vec:
3526    case INDEX_op_x86_vpshrdv_vec:
3527        return C_O1_I3(x, 0, x, x);
3528
3529    case INDEX_op_bitsel_vec:
3530    case INDEX_op_x86_vpblendvb_vec:
3531        return C_O1_I3(x, x, x, x);
3532
3533    default:
3534        g_assert_not_reached();
3535    }
3536}
3537
3538int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
3539{
3540    switch (opc) {
3541    case INDEX_op_add_vec:
3542    case INDEX_op_sub_vec:
3543    case INDEX_op_and_vec:
3544    case INDEX_op_or_vec:
3545    case INDEX_op_xor_vec:
3546    case INDEX_op_andc_vec:
3547    case INDEX_op_orc_vec:
3548    case INDEX_op_nand_vec:
3549    case INDEX_op_nor_vec:
3550    case INDEX_op_eqv_vec:
3551    case INDEX_op_not_vec:
3552    case INDEX_op_bitsel_vec:
3553        return 1;
3554    case INDEX_op_cmp_vec:
3555    case INDEX_op_cmpsel_vec:
3556        return -1;
3557
3558    case INDEX_op_rotli_vec:
3559        return have_avx512vl && vece >= MO_32 ? 1 : -1;
3560
3561    case INDEX_op_shli_vec:
3562    case INDEX_op_shri_vec:
3563        /* We must expand the operation for MO_8.  */
3564        return vece == MO_8 ? -1 : 1;
3565
3566    case INDEX_op_sari_vec:
3567        switch (vece) {
3568        case MO_8:
3569            return -1;
3570        case MO_16:
3571        case MO_32:
3572            return 1;
3573        case MO_64:
3574            if (have_avx512vl) {
3575                return 1;
3576            }
3577            /*
3578             * We can emulate this for MO_64, but it does not pay off
3579             * unless we're producing at least 4 values.
3580             */
3581            return type >= TCG_TYPE_V256 ? -1 : 0;
3582        }
3583        return 0;
3584
3585    case INDEX_op_shls_vec:
3586    case INDEX_op_shrs_vec:
3587        return vece >= MO_16;
3588    case INDEX_op_sars_vec:
3589        switch (vece) {
3590        case MO_16:
3591        case MO_32:
3592            return 1;
3593        case MO_64:
3594            return have_avx512vl;
3595        }
3596        return 0;
3597    case INDEX_op_rotls_vec:
3598        return vece >= MO_16 ? -1 : 0;
3599
3600    case INDEX_op_shlv_vec:
3601    case INDEX_op_shrv_vec:
3602        switch (vece) {
3603        case MO_16:
3604            return have_avx512bw;
3605        case MO_32:
3606        case MO_64:
3607            return have_avx2;
3608        }
3609        return 0;
3610    case INDEX_op_sarv_vec:
3611        switch (vece) {
3612        case MO_16:
3613            return have_avx512bw;
3614        case MO_32:
3615            return have_avx2;
3616        case MO_64:
3617            return have_avx512vl;
3618        }
3619        return 0;
3620    case INDEX_op_rotlv_vec:
3621    case INDEX_op_rotrv_vec:
3622        switch (vece) {
3623        case MO_16:
3624            return have_avx512vbmi2 ? -1 : 0;
3625        case MO_32:
3626        case MO_64:
3627            return have_avx512vl ? 1 : have_avx2 ? -1 : 0;
3628        }
3629        return 0;
3630
3631    case INDEX_op_mul_vec:
3632        switch (vece) {
3633        case MO_8:
3634            return -1;
3635        case MO_64:
3636            return have_avx512dq;
3637        }
3638        return 1;
3639
3640    case INDEX_op_ssadd_vec:
3641    case INDEX_op_usadd_vec:
3642    case INDEX_op_sssub_vec:
3643    case INDEX_op_ussub_vec:
3644        return vece <= MO_16;
3645    case INDEX_op_smin_vec:
3646    case INDEX_op_smax_vec:
3647    case INDEX_op_umin_vec:
3648    case INDEX_op_umax_vec:
3649    case INDEX_op_abs_vec:
3650        return vece <= MO_32 || have_avx512vl;
3651
3652    default:
3653        return 0;
3654    }
3655}
3656
3657static void expand_vec_shi(TCGType type, unsigned vece, TCGOpcode opc,
3658                           TCGv_vec v0, TCGv_vec v1, TCGArg imm)
3659{
3660    TCGv_vec t1, t2;
3661
3662    tcg_debug_assert(vece == MO_8);
3663
3664    t1 = tcg_temp_new_vec(type);
3665    t2 = tcg_temp_new_vec(type);
3666
3667    /*
3668     * Unpack to W, shift, and repack.  Tricky bits:
3669     * (1) Use punpck*bw x,x to produce DDCCBBAA,
3670     *     i.e. duplicate in other half of the 16-bit lane.
3671     * (2) For right-shift, add 8 so that the high half of the lane
3672     *     becomes zero.  For left-shift, and left-rotate, we must
3673     *     shift up and down again.
3674     * (3) Step 2 leaves high half zero such that PACKUSWB
3675     *     (pack with unsigned saturation) does not modify
3676     *     the quantity.
3677     */
3678    vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
3679              tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
3680    vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
3681              tcgv_vec_arg(t2), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
3682
3683    if (opc != INDEX_op_rotli_vec) {
3684        imm += 8;
3685    }
3686    if (opc == INDEX_op_shri_vec) {
3687        tcg_gen_shri_vec(MO_16, t1, t1, imm);
3688        tcg_gen_shri_vec(MO_16, t2, t2, imm);
3689    } else {
3690        tcg_gen_shli_vec(MO_16, t1, t1, imm);
3691        tcg_gen_shli_vec(MO_16, t2, t2, imm);
3692        tcg_gen_shri_vec(MO_16, t1, t1, 8);
3693        tcg_gen_shri_vec(MO_16, t2, t2, 8);
3694    }
3695
3696    vec_gen_3(INDEX_op_x86_packus_vec, type, MO_8,
3697              tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t2));
3698    tcg_temp_free_vec(t1);
3699    tcg_temp_free_vec(t2);
3700}
3701
3702static void expand_vec_sari(TCGType type, unsigned vece,
3703                            TCGv_vec v0, TCGv_vec v1, TCGArg imm)
3704{
3705    TCGv_vec t1, t2;
3706
3707    switch (vece) {
3708    case MO_8:
3709        /* Unpack to W, shift, and repack, as in expand_vec_shi.  */
3710        t1 = tcg_temp_new_vec(type);
3711        t2 = tcg_temp_new_vec(type);
3712        vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
3713                  tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
3714        vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
3715                  tcgv_vec_arg(t2), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
3716        tcg_gen_sari_vec(MO_16, t1, t1, imm + 8);
3717        tcg_gen_sari_vec(MO_16, t2, t2, imm + 8);
3718        vec_gen_3(INDEX_op_x86_packss_vec, type, MO_8,
3719                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t2));
3720        tcg_temp_free_vec(t1);
3721        tcg_temp_free_vec(t2);
3722        break;
3723
3724    case MO_64:
3725        t1 = tcg_temp_new_vec(type);
3726        if (imm <= 32) {
3727            /*
3728             * We can emulate a small sign extend by performing an arithmetic
3729             * 32-bit shift and overwriting the high half of a 64-bit logical
3730             * shift.  Note that the ISA says shift of 32 is valid, but TCG
3731             * does not, so we have to bound the smaller shift -- we get the
3732             * same result in the high half either way.
3733             */
3734            tcg_gen_sari_vec(MO_32, t1, v1, MIN(imm, 31));
3735            tcg_gen_shri_vec(MO_64, v0, v1, imm);
3736            vec_gen_4(INDEX_op_x86_blend_vec, type, MO_32,
3737                      tcgv_vec_arg(v0), tcgv_vec_arg(v0),
3738                      tcgv_vec_arg(t1), 0xaa);
3739        } else {
3740            /* Otherwise we will need to use a compare vs 0 to produce
3741             * the sign-extend, shift and merge.
3742             */
3743            tcg_gen_cmp_vec(TCG_COND_GT, MO_64, t1,
3744                            tcg_constant_vec(type, MO_64, 0), v1);
3745            tcg_gen_shri_vec(MO_64, v0, v1, imm);
3746            tcg_gen_shli_vec(MO_64, t1, t1, 64 - imm);
3747            tcg_gen_or_vec(MO_64, v0, v0, t1);
3748        }
3749        tcg_temp_free_vec(t1);
3750        break;
3751
3752    default:
3753        g_assert_not_reached();
3754    }
3755}
3756
3757static void expand_vec_rotli(TCGType type, unsigned vece,
3758                             TCGv_vec v0, TCGv_vec v1, TCGArg imm)
3759{
3760    TCGv_vec t;
3761
3762    if (vece == MO_8) {
3763        expand_vec_shi(type, vece, INDEX_op_rotli_vec, v0, v1, imm);
3764        return;
3765    }
3766
3767    if (have_avx512vbmi2) {
3768        vec_gen_4(INDEX_op_x86_vpshldi_vec, type, vece,
3769                  tcgv_vec_arg(v0), tcgv_vec_arg(v1), tcgv_vec_arg(v1), imm);
3770        return;
3771    }
3772
3773    t = tcg_temp_new_vec(type);
3774    tcg_gen_shli_vec(vece, t, v1, imm);
3775    tcg_gen_shri_vec(vece, v0, v1, (8 << vece) - imm);
3776    tcg_gen_or_vec(vece, v0, v0, t);
3777    tcg_temp_free_vec(t);
3778}
3779
3780static void expand_vec_rotv(TCGType type, unsigned vece, TCGv_vec v0,
3781                            TCGv_vec v1, TCGv_vec sh, bool right)
3782{
3783    TCGv_vec t;
3784
3785    if (have_avx512vbmi2) {
3786        vec_gen_4(right ? INDEX_op_x86_vpshrdv_vec : INDEX_op_x86_vpshldv_vec,
3787                  type, vece, tcgv_vec_arg(v0), tcgv_vec_arg(v1),
3788                  tcgv_vec_arg(v1), tcgv_vec_arg(sh));
3789        return;
3790    }
3791
3792    t = tcg_temp_new_vec(type);
3793    tcg_gen_dupi_vec(vece, t, 8 << vece);
3794    tcg_gen_sub_vec(vece, t, t, sh);
3795    if (right) {
3796        tcg_gen_shlv_vec(vece, t, v1, t);
3797        tcg_gen_shrv_vec(vece, v0, v1, sh);
3798    } else {
3799        tcg_gen_shrv_vec(vece, t, v1, t);
3800        tcg_gen_shlv_vec(vece, v0, v1, sh);
3801    }
3802    tcg_gen_or_vec(vece, v0, v0, t);
3803    tcg_temp_free_vec(t);
3804}
3805
3806static void expand_vec_rotls(TCGType type, unsigned vece,
3807                             TCGv_vec v0, TCGv_vec v1, TCGv_i32 lsh)
3808{
3809    TCGv_vec t = tcg_temp_new_vec(type);
3810
3811    tcg_debug_assert(vece != MO_8);
3812
3813    if (vece >= MO_32 ? have_avx512vl : have_avx512vbmi2) {
3814        tcg_gen_dup_i32_vec(vece, t, lsh);
3815        if (vece >= MO_32) {
3816            tcg_gen_rotlv_vec(vece, v0, v1, t);
3817        } else {
3818            expand_vec_rotv(type, vece, v0, v1, t, false);
3819        }
3820    } else {
3821        TCGv_i32 rsh = tcg_temp_new_i32();
3822
3823        tcg_gen_neg_i32(rsh, lsh);
3824        tcg_gen_andi_i32(rsh, rsh, (8 << vece) - 1);
3825        tcg_gen_shls_vec(vece, t, v1, lsh);
3826        tcg_gen_shrs_vec(vece, v0, v1, rsh);
3827        tcg_gen_or_vec(vece, v0, v0, t);
3828
3829        tcg_temp_free_i32(rsh);
3830    }
3831
3832    tcg_temp_free_vec(t);
3833}
3834
3835static void expand_vec_mul(TCGType type, unsigned vece,
3836                           TCGv_vec v0, TCGv_vec v1, TCGv_vec v2)
3837{
3838    TCGv_vec t1, t2, t3, t4, zero;
3839
3840    tcg_debug_assert(vece == MO_8);
3841
3842    /*
3843     * Unpack v1 bytes to words, 0 | x.
3844     * Unpack v2 bytes to words, y | 0.
3845     * This leaves the 8-bit result, x * y, with 8 bits of right padding.
3846     * Shift logical right by 8 bits to clear the high 8 bytes before
3847     * using an unsigned saturated pack.
3848     *
3849     * The difference between the V64, V128 and V256 cases is merely how
3850     * we distribute the expansion between temporaries.
3851     */
3852    switch (type) {
3853    case TCG_TYPE_V64:
3854        t1 = tcg_temp_new_vec(TCG_TYPE_V128);
3855        t2 = tcg_temp_new_vec(TCG_TYPE_V128);
3856        zero = tcg_constant_vec(TCG_TYPE_V128, MO_8, 0);
3857        vec_gen_3(INDEX_op_x86_punpckl_vec, TCG_TYPE_V128, MO_8,
3858                  tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
3859        vec_gen_3(INDEX_op_x86_punpckl_vec, TCG_TYPE_V128, MO_8,
3860                  tcgv_vec_arg(t2), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
3861        tcg_gen_mul_vec(MO_16, t1, t1, t2);
3862        tcg_gen_shri_vec(MO_16, t1, t1, 8);
3863        vec_gen_3(INDEX_op_x86_packus_vec, TCG_TYPE_V128, MO_8,
3864                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t1));
3865        tcg_temp_free_vec(t1);
3866        tcg_temp_free_vec(t2);
3867        break;
3868
3869    case TCG_TYPE_V128:
3870    case TCG_TYPE_V256:
3871        t1 = tcg_temp_new_vec(type);
3872        t2 = tcg_temp_new_vec(type);
3873        t3 = tcg_temp_new_vec(type);
3874        t4 = tcg_temp_new_vec(type);
3875        zero = tcg_constant_vec(TCG_TYPE_V128, MO_8, 0);
3876        vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
3877                  tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
3878        vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
3879                  tcgv_vec_arg(t2), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
3880        vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
3881                  tcgv_vec_arg(t3), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
3882        vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
3883                  tcgv_vec_arg(t4), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
3884        tcg_gen_mul_vec(MO_16, t1, t1, t2);
3885        tcg_gen_mul_vec(MO_16, t3, t3, t4);
3886        tcg_gen_shri_vec(MO_16, t1, t1, 8);
3887        tcg_gen_shri_vec(MO_16, t3, t3, 8);
3888        vec_gen_3(INDEX_op_x86_packus_vec, type, MO_8,
3889                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t3));
3890        tcg_temp_free_vec(t1);
3891        tcg_temp_free_vec(t2);
3892        tcg_temp_free_vec(t3);
3893        tcg_temp_free_vec(t4);
3894        break;
3895
3896    default:
3897        g_assert_not_reached();
3898    }
3899}
3900
3901static bool expand_vec_cmp_noinv(TCGType type, unsigned vece, TCGv_vec v0,
3902                                 TCGv_vec v1, TCGv_vec v2, TCGCond cond)
3903{
3904    enum {
3905        NEED_INV  = 1,
3906        NEED_SWAP = 2,
3907        NEED_BIAS = 4,
3908        NEED_UMIN = 8,
3909        NEED_UMAX = 16,
3910    };
3911    TCGv_vec t1, t2, t3;
3912    uint8_t fixup;
3913
3914    switch (cond) {
3915    case TCG_COND_EQ:
3916    case TCG_COND_GT:
3917        fixup = 0;
3918        break;
3919    case TCG_COND_NE:
3920    case TCG_COND_LE:
3921        fixup = NEED_INV;
3922        break;
3923    case TCG_COND_LT:
3924        fixup = NEED_SWAP;
3925        break;
3926    case TCG_COND_GE:
3927        fixup = NEED_SWAP | NEED_INV;
3928        break;
3929    case TCG_COND_LEU:
3930        if (tcg_can_emit_vec_op(INDEX_op_umin_vec, type, vece)) {
3931            fixup = NEED_UMIN;
3932        } else {
3933            fixup = NEED_BIAS | NEED_INV;
3934        }
3935        break;
3936    case TCG_COND_GTU:
3937        if (tcg_can_emit_vec_op(INDEX_op_umin_vec, type, vece)) {
3938            fixup = NEED_UMIN | NEED_INV;
3939        } else {
3940            fixup = NEED_BIAS;
3941        }
3942        break;
3943    case TCG_COND_GEU:
3944        if (tcg_can_emit_vec_op(INDEX_op_umax_vec, type, vece)) {
3945            fixup = NEED_UMAX;
3946        } else {
3947            fixup = NEED_BIAS | NEED_SWAP | NEED_INV;
3948        }
3949        break;
3950    case TCG_COND_LTU:
3951        if (tcg_can_emit_vec_op(INDEX_op_umax_vec, type, vece)) {
3952            fixup = NEED_UMAX | NEED_INV;
3953        } else {
3954            fixup = NEED_BIAS | NEED_SWAP;
3955        }
3956        break;
3957    default:
3958        g_assert_not_reached();
3959    }
3960
3961    if (fixup & NEED_INV) {
3962        cond = tcg_invert_cond(cond);
3963    }
3964    if (fixup & NEED_SWAP) {
3965        t1 = v1, v1 = v2, v2 = t1;
3966        cond = tcg_swap_cond(cond);
3967    }
3968
3969    t1 = t2 = NULL;
3970    if (fixup & (NEED_UMIN | NEED_UMAX)) {
3971        t1 = tcg_temp_new_vec(type);
3972        if (fixup & NEED_UMIN) {
3973            tcg_gen_umin_vec(vece, t1, v1, v2);
3974        } else {
3975            tcg_gen_umax_vec(vece, t1, v1, v2);
3976        }
3977        v2 = t1;
3978        cond = TCG_COND_EQ;
3979    } else if (fixup & NEED_BIAS) {
3980        t1 = tcg_temp_new_vec(type);
3981        t2 = tcg_temp_new_vec(type);
3982        t3 = tcg_constant_vec(type, vece, 1ull << ((8 << vece) - 1));
3983        tcg_gen_sub_vec(vece, t1, v1, t3);
3984        tcg_gen_sub_vec(vece, t2, v2, t3);
3985        v1 = t1;
3986        v2 = t2;
3987        cond = tcg_signed_cond(cond);
3988    }
3989
3990    tcg_debug_assert(cond == TCG_COND_EQ || cond == TCG_COND_GT);
3991    /* Expand directly; do not recurse.  */
3992    vec_gen_4(INDEX_op_cmp_vec, type, vece,
3993              tcgv_vec_arg(v0), tcgv_vec_arg(v1), tcgv_vec_arg(v2), cond);
3994
3995    if (t1) {
3996        tcg_temp_free_vec(t1);
3997        if (t2) {
3998            tcg_temp_free_vec(t2);
3999        }
4000    }
4001    return fixup & NEED_INV;
4002}
4003
4004static void expand_vec_cmp(TCGType type, unsigned vece, TCGv_vec v0,
4005                           TCGv_vec v1, TCGv_vec v2, TCGCond cond)
4006{
4007    if (expand_vec_cmp_noinv(type, vece, v0, v1, v2, cond)) {
4008        tcg_gen_not_vec(vece, v0, v0);
4009    }
4010}
4011
4012static void expand_vec_cmpsel(TCGType type, unsigned vece, TCGv_vec v0,
4013                              TCGv_vec c1, TCGv_vec c2,
4014                              TCGv_vec v3, TCGv_vec v4, TCGCond cond)
4015{
4016    TCGv_vec t = tcg_temp_new_vec(type);
4017
4018    if (expand_vec_cmp_noinv(type, vece, t, c1, c2, cond)) {
4019        /* Invert the sense of the compare by swapping arguments.  */
4020        TCGv_vec x;
4021        x = v3, v3 = v4, v4 = x;
4022    }
4023    vec_gen_4(INDEX_op_x86_vpblendvb_vec, type, vece,
4024              tcgv_vec_arg(v0), tcgv_vec_arg(v4),
4025              tcgv_vec_arg(v3), tcgv_vec_arg(t));
4026    tcg_temp_free_vec(t);
4027}
4028
4029void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
4030                       TCGArg a0, ...)
4031{
4032    va_list va;
4033    TCGArg a2;
4034    TCGv_vec v0, v1, v2, v3, v4;
4035
4036    va_start(va, a0);
4037    v0 = temp_tcgv_vec(arg_temp(a0));
4038    v1 = temp_tcgv_vec(arg_temp(va_arg(va, TCGArg)));
4039    a2 = va_arg(va, TCGArg);
4040
4041    switch (opc) {
4042    case INDEX_op_shli_vec:
4043    case INDEX_op_shri_vec:
4044        expand_vec_shi(type, vece, opc, v0, v1, a2);
4045        break;
4046
4047    case INDEX_op_sari_vec:
4048        expand_vec_sari(type, vece, v0, v1, a2);
4049        break;
4050
4051    case INDEX_op_rotli_vec:
4052        expand_vec_rotli(type, vece, v0, v1, a2);
4053        break;
4054
4055    case INDEX_op_rotls_vec:
4056        expand_vec_rotls(type, vece, v0, v1, temp_tcgv_i32(arg_temp(a2)));
4057        break;
4058
4059    case INDEX_op_rotlv_vec:
4060        v2 = temp_tcgv_vec(arg_temp(a2));
4061        expand_vec_rotv(type, vece, v0, v1, v2, false);
4062        break;
4063    case INDEX_op_rotrv_vec:
4064        v2 = temp_tcgv_vec(arg_temp(a2));
4065        expand_vec_rotv(type, vece, v0, v1, v2, true);
4066        break;
4067
4068    case INDEX_op_mul_vec:
4069        v2 = temp_tcgv_vec(arg_temp(a2));
4070        expand_vec_mul(type, vece, v0, v1, v2);
4071        break;
4072
4073    case INDEX_op_cmp_vec:
4074        v2 = temp_tcgv_vec(arg_temp(a2));
4075        expand_vec_cmp(type, vece, v0, v1, v2, va_arg(va, TCGArg));
4076        break;
4077
4078    case INDEX_op_cmpsel_vec:
4079        v2 = temp_tcgv_vec(arg_temp(a2));
4080        v3 = temp_tcgv_vec(arg_temp(va_arg(va, TCGArg)));
4081        v4 = temp_tcgv_vec(arg_temp(va_arg(va, TCGArg)));
4082        expand_vec_cmpsel(type, vece, v0, v1, v2, v3, v4, va_arg(va, TCGArg));
4083        break;
4084
4085    default:
4086        break;
4087    }
4088
4089    va_end(va);
4090}
4091
4092static const int tcg_target_callee_save_regs[] = {
4093#if TCG_TARGET_REG_BITS == 64
4094    TCG_REG_RBP,
4095    TCG_REG_RBX,
4096#if defined(_WIN64)
4097    TCG_REG_RDI,
4098    TCG_REG_RSI,
4099#endif
4100    TCG_REG_R12,
4101    TCG_REG_R13,
4102    TCG_REG_R14, /* Currently used for the global env. */
4103    TCG_REG_R15,
4104#else
4105    TCG_REG_EBP, /* Currently used for the global env. */
4106    TCG_REG_EBX,
4107    TCG_REG_ESI,
4108    TCG_REG_EDI,
4109#endif
4110};
4111
4112/* Compute frame size via macros, to share between tcg_target_qemu_prologue
4113   and tcg_register_jit.  */
4114
4115#define PUSH_SIZE \
4116    ((1 + ARRAY_SIZE(tcg_target_callee_save_regs)) \
4117     * (TCG_TARGET_REG_BITS / 8))
4118
4119#define FRAME_SIZE \
4120    ((PUSH_SIZE \
4121      + TCG_STATIC_CALL_ARGS_SIZE \
4122      + CPU_TEMP_BUF_NLONGS * sizeof(long) \
4123      + TCG_TARGET_STACK_ALIGN - 1) \
4124     & ~(TCG_TARGET_STACK_ALIGN - 1))
4125
4126/* Generate global QEMU prologue and epilogue code */
4127static void tcg_target_qemu_prologue(TCGContext *s)
4128{
4129    int i, stack_addend;
4130
4131    /* TB prologue */
4132
4133    /* Reserve some stack space, also for TCG temps.  */
4134    stack_addend = FRAME_SIZE - PUSH_SIZE;
4135    tcg_set_frame(s, TCG_REG_CALL_STACK, TCG_STATIC_CALL_ARGS_SIZE,
4136                  CPU_TEMP_BUF_NLONGS * sizeof(long));
4137
4138    /* Save all callee saved registers.  */
4139    for (i = 0; i < ARRAY_SIZE(tcg_target_callee_save_regs); i++) {
4140        tcg_out_push(s, tcg_target_callee_save_regs[i]);
4141    }
4142
4143#if TCG_TARGET_REG_BITS == 32
4144    tcg_out_ld(s, TCG_TYPE_PTR, TCG_AREG0, TCG_REG_ESP,
4145               (ARRAY_SIZE(tcg_target_callee_save_regs) + 1) * 4);
4146    tcg_out_addi(s, TCG_REG_ESP, -stack_addend);
4147    /* jmp *tb.  */
4148    tcg_out_modrm_offset(s, OPC_GRP5, EXT5_JMPN_Ev, TCG_REG_ESP,
4149                         (ARRAY_SIZE(tcg_target_callee_save_regs) + 2) * 4
4150                         + stack_addend);
4151#else
4152# if !defined(CONFIG_SOFTMMU)
4153    if (guest_base) {
4154        int seg = setup_guest_base_seg();
4155        if (seg != 0) {
4156            x86_guest_base.seg = seg;
4157        } else if (guest_base == (int32_t)guest_base) {
4158            x86_guest_base.ofs = guest_base;
4159        } else {
4160            /* Choose R12 because, as a base, it requires a SIB byte. */
4161            x86_guest_base.index = TCG_REG_R12;
4162            tcg_out_movi(s, TCG_TYPE_PTR, x86_guest_base.index, guest_base);
4163            tcg_regset_set_reg(s->reserved_regs, x86_guest_base.index);
4164        }
4165    }
4166# endif
4167    tcg_out_mov(s, TCG_TYPE_PTR, TCG_AREG0, tcg_target_call_iarg_regs[0]);
4168    tcg_out_addi(s, TCG_REG_ESP, -stack_addend);
4169    /* jmp *tb.  */
4170    tcg_out_modrm(s, OPC_GRP5, EXT5_JMPN_Ev, tcg_target_call_iarg_regs[1]);
4171#endif
4172
4173    /*
4174     * Return path for goto_ptr. Set return value to 0, a-la exit_tb,
4175     * and fall through to the rest of the epilogue.
4176     */
4177    tcg_code_gen_epilogue = tcg_splitwx_to_rx(s->code_ptr);
4178    tcg_out_movi(s, TCG_TYPE_REG, TCG_REG_EAX, 0);
4179
4180    /* TB epilogue */
4181    tb_ret_addr = tcg_splitwx_to_rx(s->code_ptr);
4182
4183    tcg_out_addi(s, TCG_REG_CALL_STACK, stack_addend);
4184
4185    if (have_avx2) {
4186        tcg_out_vex_opc(s, OPC_VZEROUPPER, 0, 0, 0, 0);
4187    }
4188    for (i = ARRAY_SIZE(tcg_target_callee_save_regs) - 1; i >= 0; i--) {
4189        tcg_out_pop(s, tcg_target_callee_save_regs[i]);
4190    }
4191    tcg_out_opc(s, OPC_RET, 0, 0, 0);
4192}
4193
4194static void tcg_out_tb_start(TCGContext *s)
4195{
4196    /* nothing to do */
4197}
4198
4199static void tcg_out_nop_fill(tcg_insn_unit *p, int count)
4200{
4201    memset(p, 0x90, count);
4202}
4203
4204static void tcg_target_init(TCGContext *s)
4205{
4206    tcg_target_available_regs[TCG_TYPE_I32] = ALL_GENERAL_REGS;
4207    if (TCG_TARGET_REG_BITS == 64) {
4208        tcg_target_available_regs[TCG_TYPE_I64] = ALL_GENERAL_REGS;
4209    }
4210    if (have_avx1) {
4211        tcg_target_available_regs[TCG_TYPE_V64] = ALL_VECTOR_REGS;
4212        tcg_target_available_regs[TCG_TYPE_V128] = ALL_VECTOR_REGS;
4213    }
4214    if (have_avx2) {
4215        tcg_target_available_regs[TCG_TYPE_V256] = ALL_VECTOR_REGS;
4216    }
4217
4218    tcg_target_call_clobber_regs = ALL_VECTOR_REGS;
4219    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_EAX);
4220    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_EDX);
4221    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_ECX);
4222    if (TCG_TARGET_REG_BITS == 64) {
4223#if !defined(_WIN64)
4224        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_RDI);
4225        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_RSI);
4226#endif
4227        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R8);
4228        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R9);
4229        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R10);
4230        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R11);
4231    }
4232
4233    s->reserved_regs = 0;
4234    tcg_regset_set_reg(s->reserved_regs, TCG_REG_CALL_STACK);
4235    tcg_regset_set_reg(s->reserved_regs, TCG_TMP_VEC);
4236#ifdef _WIN64
4237    /* These are call saved, and we don't save them, so don't use them. */
4238    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM6);
4239    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM7);
4240    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM8);
4241    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM9);
4242    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM10);
4243    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM11);
4244    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM12);
4245    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM13);
4246    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM14);
4247    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM15);
4248#endif
4249}
4250
4251typedef struct {
4252    DebugFrameHeader h;
4253    uint8_t fde_def_cfa[4];
4254    uint8_t fde_reg_ofs[14];
4255} DebugFrame;
4256
4257/* We're expecting a 2 byte uleb128 encoded value.  */
4258QEMU_BUILD_BUG_ON(FRAME_SIZE >= (1 << 14));
4259
4260#if !defined(__ELF__)
4261    /* Host machine without ELF. */
4262#elif TCG_TARGET_REG_BITS == 64
4263#define ELF_HOST_MACHINE EM_X86_64
4264static const DebugFrame debug_frame = {
4265    .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
4266    .h.cie.id = -1,
4267    .h.cie.version = 1,
4268    .h.cie.code_align = 1,
4269    .h.cie.data_align = 0x78,             /* sleb128 -8 */
4270    .h.cie.return_column = 16,
4271
4272    /* Total FDE size does not include the "len" member.  */
4273    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
4274
4275    .fde_def_cfa = {
4276        12, 7,                          /* DW_CFA_def_cfa %rsp, ... */
4277        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
4278        (FRAME_SIZE >> 7)
4279    },
4280    .fde_reg_ofs = {
4281        0x90, 1,                        /* DW_CFA_offset, %rip, -8 */
4282        /* The following ordering must match tcg_target_callee_save_regs.  */
4283        0x86, 2,                        /* DW_CFA_offset, %rbp, -16 */
4284        0x83, 3,                        /* DW_CFA_offset, %rbx, -24 */
4285        0x8c, 4,                        /* DW_CFA_offset, %r12, -32 */
4286        0x8d, 5,                        /* DW_CFA_offset, %r13, -40 */
4287        0x8e, 6,                        /* DW_CFA_offset, %r14, -48 */
4288        0x8f, 7,                        /* DW_CFA_offset, %r15, -56 */
4289    }
4290};
4291#else
4292#define ELF_HOST_MACHINE EM_386
4293static const DebugFrame debug_frame = {
4294    .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
4295    .h.cie.id = -1,
4296    .h.cie.version = 1,
4297    .h.cie.code_align = 1,
4298    .h.cie.data_align = 0x7c,             /* sleb128 -4 */
4299    .h.cie.return_column = 8,
4300
4301    /* Total FDE size does not include the "len" member.  */
4302    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
4303
4304    .fde_def_cfa = {
4305        12, 4,                          /* DW_CFA_def_cfa %esp, ... */
4306        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
4307        (FRAME_SIZE >> 7)
4308    },
4309    .fde_reg_ofs = {
4310        0x88, 1,                        /* DW_CFA_offset, %eip, -4 */
4311        /* The following ordering must match tcg_target_callee_save_regs.  */
4312        0x85, 2,                        /* DW_CFA_offset, %ebp, -8 */
4313        0x83, 3,                        /* DW_CFA_offset, %ebx, -12 */
4314        0x86, 4,                        /* DW_CFA_offset, %esi, -16 */
4315        0x87, 5,                        /* DW_CFA_offset, %edi, -20 */
4316    }
4317};
4318#endif
4319
4320#if defined(ELF_HOST_MACHINE)
4321void tcg_register_jit(const void *buf, size_t buf_size)
4322{
4323    tcg_register_jit_int(buf, buf_size, &debug_frame, sizeof(debug_frame));
4324}
4325#endif
4326