xref: /openbmc/qemu/tcg/i386/tcg-target.c.inc (revision d31f1185fb029b44c439a6961a6cb087df6567d9)
1/*
2 * Tiny Code Generator for QEMU
3 *
4 * Copyright (c) 2008 Fabrice Bellard
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to deal
8 * in the Software without restriction, including without limitation the rights
9 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
10 * copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
22 * THE SOFTWARE.
23 */
24
25/* Used for function call generation. */
26#define TCG_TARGET_STACK_ALIGN 16
27#if defined(_WIN64)
28#define TCG_TARGET_CALL_STACK_OFFSET 32
29#else
30#define TCG_TARGET_CALL_STACK_OFFSET 0
31#endif
32#define TCG_TARGET_CALL_ARG_I32      TCG_CALL_ARG_NORMAL
33#define TCG_TARGET_CALL_ARG_I64      TCG_CALL_ARG_NORMAL
34#if defined(_WIN64)
35# define TCG_TARGET_CALL_ARG_I128    TCG_CALL_ARG_BY_REF
36# define TCG_TARGET_CALL_RET_I128    TCG_CALL_RET_BY_VEC
37#elif TCG_TARGET_REG_BITS == 64
38# define TCG_TARGET_CALL_ARG_I128    TCG_CALL_ARG_NORMAL
39# define TCG_TARGET_CALL_RET_I128    TCG_CALL_RET_NORMAL
40#else
41# define TCG_TARGET_CALL_ARG_I128    TCG_CALL_ARG_NORMAL
42# define TCG_TARGET_CALL_RET_I128    TCG_CALL_RET_BY_REF
43#endif
44
45#ifdef CONFIG_DEBUG_TCG
46static const char * const tcg_target_reg_names[TCG_TARGET_NB_REGS] = {
47#if TCG_TARGET_REG_BITS == 64
48    "%rax", "%rcx", "%rdx", "%rbx", "%rsp", "%rbp", "%rsi", "%rdi",
49#else
50    "%eax", "%ecx", "%edx", "%ebx", "%esp", "%ebp", "%esi", "%edi",
51#endif
52    "%r8",  "%r9",  "%r10", "%r11", "%r12", "%r13", "%r14", "%r15",
53    "%xmm0", "%xmm1", "%xmm2", "%xmm3", "%xmm4", "%xmm5", "%xmm6", "%xmm7",
54#if TCG_TARGET_REG_BITS == 64
55    "%xmm8", "%xmm9", "%xmm10", "%xmm11",
56    "%xmm12", "%xmm13", "%xmm14", "%xmm15",
57#endif
58};
59#endif
60
61static const int tcg_target_reg_alloc_order[] = {
62#if TCG_TARGET_REG_BITS == 64
63    TCG_REG_RBP,
64    TCG_REG_RBX,
65    TCG_REG_R12,
66    TCG_REG_R13,
67    TCG_REG_R14,
68    TCG_REG_R15,
69    TCG_REG_R10,
70    TCG_REG_R11,
71    TCG_REG_R9,
72    TCG_REG_R8,
73    TCG_REG_RCX,
74    TCG_REG_RDX,
75    TCG_REG_RSI,
76    TCG_REG_RDI,
77    TCG_REG_RAX,
78#else
79    TCG_REG_EBX,
80    TCG_REG_ESI,
81    TCG_REG_EDI,
82    TCG_REG_EBP,
83    TCG_REG_ECX,
84    TCG_REG_EDX,
85    TCG_REG_EAX,
86#endif
87    TCG_REG_XMM0,
88    TCG_REG_XMM1,
89    TCG_REG_XMM2,
90    TCG_REG_XMM3,
91    TCG_REG_XMM4,
92    TCG_REG_XMM5,
93#ifndef _WIN64
94    /* The Win64 ABI has xmm6-xmm15 as caller-saves, and we do not save
95       any of them.  Therefore only allow xmm0-xmm5 to be allocated.  */
96    TCG_REG_XMM6,
97    TCG_REG_XMM7,
98#if TCG_TARGET_REG_BITS == 64
99    TCG_REG_XMM8,
100    TCG_REG_XMM9,
101    TCG_REG_XMM10,
102    TCG_REG_XMM11,
103    TCG_REG_XMM12,
104    TCG_REG_XMM13,
105    TCG_REG_XMM14,
106    TCG_REG_XMM15,
107#endif
108#endif
109};
110
111#define TCG_TMP_VEC  TCG_REG_XMM5
112
113static const int tcg_target_call_iarg_regs[] = {
114#if TCG_TARGET_REG_BITS == 64
115#if defined(_WIN64)
116    TCG_REG_RCX,
117    TCG_REG_RDX,
118#else
119    TCG_REG_RDI,
120    TCG_REG_RSI,
121    TCG_REG_RDX,
122    TCG_REG_RCX,
123#endif
124    TCG_REG_R8,
125    TCG_REG_R9,
126#else
127    /* 32 bit mode uses stack based calling convention (GCC default). */
128#endif
129};
130
131static TCGReg tcg_target_call_oarg_reg(TCGCallReturnKind kind, int slot)
132{
133    switch (kind) {
134    case TCG_CALL_RET_NORMAL:
135        tcg_debug_assert(slot >= 0 && slot <= 1);
136        return slot ? TCG_REG_EDX : TCG_REG_EAX;
137#ifdef _WIN64
138    case TCG_CALL_RET_BY_VEC:
139        tcg_debug_assert(slot == 0);
140        return TCG_REG_XMM0;
141#endif
142    default:
143        g_assert_not_reached();
144    }
145}
146
147/* Constants we accept.  */
148#define TCG_CT_CONST_S32 0x100
149#define TCG_CT_CONST_U32 0x200
150#define TCG_CT_CONST_I32 0x400
151#define TCG_CT_CONST_WSZ 0x800
152#define TCG_CT_CONST_TST 0x1000
153#define TCG_CT_CONST_ZERO 0x2000
154
155/* Registers used with L constraint, which are the first argument
156   registers on x86_64, and two random call clobbered registers on
157   i386. */
158#if TCG_TARGET_REG_BITS == 64
159# define TCG_REG_L0 tcg_target_call_iarg_regs[0]
160# define TCG_REG_L1 tcg_target_call_iarg_regs[1]
161#else
162# define TCG_REG_L0 TCG_REG_EAX
163# define TCG_REG_L1 TCG_REG_EDX
164#endif
165
166#if TCG_TARGET_REG_BITS == 64
167# define ALL_GENERAL_REGS      0x0000ffffu
168# define ALL_VECTOR_REGS       0xffff0000u
169# define ALL_BYTEL_REGS        ALL_GENERAL_REGS
170#else
171# define ALL_GENERAL_REGS      0x000000ffu
172# define ALL_VECTOR_REGS       0x00ff0000u
173# define ALL_BYTEL_REGS        0x0000000fu
174#endif
175#define SOFTMMU_RESERVE_REGS \
176    (tcg_use_softmmu ? (1 << TCG_REG_L0) | (1 << TCG_REG_L1) : 0)
177
178#define have_bmi2       (cpuinfo & CPUINFO_BMI2)
179#define have_lzcnt      (cpuinfo & CPUINFO_LZCNT)
180
181static const tcg_insn_unit *tb_ret_addr;
182
183static bool patch_reloc(tcg_insn_unit *code_ptr, int type,
184                        intptr_t value, intptr_t addend)
185{
186    value += addend;
187    switch(type) {
188    case R_386_PC32:
189        value -= (uintptr_t)tcg_splitwx_to_rx(code_ptr);
190        if (value != (int32_t)value) {
191            return false;
192        }
193        /* FALLTHRU */
194    case R_386_32:
195        tcg_patch32(code_ptr, value);
196        break;
197    case R_386_PC8:
198        value -= (uintptr_t)tcg_splitwx_to_rx(code_ptr);
199        if (value != (int8_t)value) {
200            return false;
201        }
202        tcg_patch8(code_ptr, value);
203        break;
204    default:
205        g_assert_not_reached();
206    }
207    return true;
208}
209
210/* test if a constant matches the constraint */
211static bool tcg_target_const_match(int64_t val, int ct,
212                                   TCGType type, TCGCond cond, int vece)
213{
214    if (ct & TCG_CT_CONST) {
215        return 1;
216    }
217    if (type == TCG_TYPE_I32) {
218        if (ct & (TCG_CT_CONST_S32 | TCG_CT_CONST_U32 |
219                  TCG_CT_CONST_I32 | TCG_CT_CONST_TST)) {
220            return 1;
221        }
222    } else {
223        if ((ct & TCG_CT_CONST_S32) && val == (int32_t)val) {
224            return 1;
225        }
226        if ((ct & TCG_CT_CONST_U32) && val == (uint32_t)val) {
227            return 1;
228        }
229        if ((ct & TCG_CT_CONST_I32) && ~val == (int32_t)~val) {
230            return 1;
231        }
232        /*
233         * This will be used in combination with TCG_CT_CONST_S32,
234         * so "normal" TESTQ is already matched.  Also accept:
235         *    TESTQ -> TESTL   (uint32_t)
236         *    TESTQ -> BT      (is_power_of_2)
237         */
238        if ((ct & TCG_CT_CONST_TST)
239            && is_tst_cond(cond)
240            && (val == (uint32_t)val || is_power_of_2(val))) {
241            return 1;
242        }
243    }
244    if ((ct & TCG_CT_CONST_WSZ) && val == (type == TCG_TYPE_I32 ? 32 : 64)) {
245        return 1;
246    }
247    if ((ct & TCG_CT_CONST_ZERO) && val == 0) {
248        return 1;
249    }
250    return 0;
251}
252
253# define LOWREGMASK(x)	((x) & 7)
254
255#define P_EXT		0x100		/* 0x0f opcode prefix */
256#define P_EXT38         0x200           /* 0x0f 0x38 opcode prefix */
257#define P_DATA16        0x400           /* 0x66 opcode prefix */
258#define P_VEXW          0x1000          /* Set VEX.W = 1 */
259#if TCG_TARGET_REG_BITS == 64
260# define P_REXW         P_VEXW          /* Set REX.W = 1; match VEXW */
261# define P_REXB_R       0x2000          /* REG field as byte register */
262# define P_REXB_RM      0x4000          /* R/M field as byte register */
263# define P_GS           0x8000          /* gs segment override */
264#else
265# define P_REXW		0
266# define P_REXB_R	0
267# define P_REXB_RM	0
268# define P_GS           0
269#endif
270#define P_EXT3A         0x10000         /* 0x0f 0x3a opcode prefix */
271#define P_SIMDF3        0x20000         /* 0xf3 opcode prefix */
272#define P_SIMDF2        0x40000         /* 0xf2 opcode prefix */
273#define P_VEXL          0x80000         /* Set VEX.L = 1 */
274#define P_EVEX          0x100000        /* Requires EVEX encoding */
275
276#define OPC_ARITH_EbIb	(0x80)
277#define OPC_ARITH_EvIz	(0x81)
278#define OPC_ARITH_EvIb	(0x83)
279#define OPC_ARITH_GvEv	(0x03)		/* ... plus (ARITH_FOO << 3) */
280#define OPC_ANDN        (0xf2 | P_EXT38)
281#define OPC_ADD_GvEv	(OPC_ARITH_GvEv | (ARITH_ADD << 3))
282#define OPC_AND_GvEv    (OPC_ARITH_GvEv | (ARITH_AND << 3))
283#define OPC_BLENDPS     (0x0c | P_EXT3A | P_DATA16)
284#define OPC_BSF         (0xbc | P_EXT)
285#define OPC_BSR         (0xbd | P_EXT)
286#define OPC_BSWAP	(0xc8 | P_EXT)
287#define OPC_CALL_Jz	(0xe8)
288#define OPC_CMOVCC      (0x40 | P_EXT)  /* ... plus condition code */
289#define OPC_CMP_GvEv	(OPC_ARITH_GvEv | (ARITH_CMP << 3))
290#define OPC_DEC_r32	(0x48)
291#define OPC_IMUL_GvEv	(0xaf | P_EXT)
292#define OPC_IMUL_GvEvIb	(0x6b)
293#define OPC_IMUL_GvEvIz	(0x69)
294#define OPC_INC_r32	(0x40)
295#define OPC_JCC_long	(0x80 | P_EXT)	/* ... plus condition code */
296#define OPC_JCC_short	(0x70)		/* ... plus condition code */
297#define OPC_JMP_long	(0xe9)
298#define OPC_JMP_short	(0xeb)
299#define OPC_LEA         (0x8d)
300#define OPC_LZCNT       (0xbd | P_EXT | P_SIMDF3)
301#define OPC_MOVB_EvGv	(0x88)		/* stores, more or less */
302#define OPC_MOVL_EvGv	(0x89)		/* stores, more or less */
303#define OPC_MOVL_GvEv	(0x8b)		/* loads, more or less */
304#define OPC_MOVB_EvIz   (0xc6)
305#define OPC_MOVL_EvIz	(0xc7)
306#define OPC_MOVB_Ib     (0xb0)
307#define OPC_MOVL_Iv     (0xb8)
308#define OPC_MOVBE_GyMy  (0xf0 | P_EXT38)
309#define OPC_MOVBE_MyGy  (0xf1 | P_EXT38)
310#define OPC_MOVD_VyEy   (0x6e | P_EXT | P_DATA16)
311#define OPC_MOVD_EyVy   (0x7e | P_EXT | P_DATA16)
312#define OPC_MOVDDUP     (0x12 | P_EXT | P_SIMDF2)
313#define OPC_MOVDQA_VxWx (0x6f | P_EXT | P_DATA16)
314#define OPC_MOVDQA_WxVx (0x7f | P_EXT | P_DATA16)
315#define OPC_MOVDQU_VxWx (0x6f | P_EXT | P_SIMDF3)
316#define OPC_MOVDQU_WxVx (0x7f | P_EXT | P_SIMDF3)
317#define OPC_MOVQ_VqWq   (0x7e | P_EXT | P_SIMDF3)
318#define OPC_MOVQ_WqVq   (0xd6 | P_EXT | P_DATA16)
319#define OPC_MOVSBL	(0xbe | P_EXT)
320#define OPC_MOVSWL	(0xbf | P_EXT)
321#define OPC_MOVSLQ	(0x63 | P_REXW)
322#define OPC_MOVZBL	(0xb6 | P_EXT)
323#define OPC_MOVZWL	(0xb7 | P_EXT)
324#define OPC_PABSB       (0x1c | P_EXT38 | P_DATA16)
325#define OPC_PABSW       (0x1d | P_EXT38 | P_DATA16)
326#define OPC_PABSD       (0x1e | P_EXT38 | P_DATA16)
327#define OPC_VPABSQ      (0x1f | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
328#define OPC_PACKSSDW    (0x6b | P_EXT | P_DATA16)
329#define OPC_PACKSSWB    (0x63 | P_EXT | P_DATA16)
330#define OPC_PACKUSDW    (0x2b | P_EXT38 | P_DATA16)
331#define OPC_PACKUSWB    (0x67 | P_EXT | P_DATA16)
332#define OPC_PADDB       (0xfc | P_EXT | P_DATA16)
333#define OPC_PADDW       (0xfd | P_EXT | P_DATA16)
334#define OPC_PADDD       (0xfe | P_EXT | P_DATA16)
335#define OPC_PADDQ       (0xd4 | P_EXT | P_DATA16)
336#define OPC_PADDSB      (0xec | P_EXT | P_DATA16)
337#define OPC_PADDSW      (0xed | P_EXT | P_DATA16)
338#define OPC_PADDUB      (0xdc | P_EXT | P_DATA16)
339#define OPC_PADDUW      (0xdd | P_EXT | P_DATA16)
340#define OPC_PAND        (0xdb | P_EXT | P_DATA16)
341#define OPC_PANDN       (0xdf | P_EXT | P_DATA16)
342#define OPC_PBLENDW     (0x0e | P_EXT3A | P_DATA16)
343#define OPC_PCMPEQB     (0x74 | P_EXT | P_DATA16)
344#define OPC_PCMPEQW     (0x75 | P_EXT | P_DATA16)
345#define OPC_PCMPEQD     (0x76 | P_EXT | P_DATA16)
346#define OPC_PCMPEQQ     (0x29 | P_EXT38 | P_DATA16)
347#define OPC_PCMPGTB     (0x64 | P_EXT | P_DATA16)
348#define OPC_PCMPGTW     (0x65 | P_EXT | P_DATA16)
349#define OPC_PCMPGTD     (0x66 | P_EXT | P_DATA16)
350#define OPC_PCMPGTQ     (0x37 | P_EXT38 | P_DATA16)
351#define OPC_PEXTRD      (0x16 | P_EXT3A | P_DATA16)
352#define OPC_PINSRD      (0x22 | P_EXT3A | P_DATA16)
353#define OPC_PMAXSB      (0x3c | P_EXT38 | P_DATA16)
354#define OPC_PMAXSW      (0xee | P_EXT | P_DATA16)
355#define OPC_PMAXSD      (0x3d | P_EXT38 | P_DATA16)
356#define OPC_VPMAXSQ     (0x3d | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
357#define OPC_PMAXUB      (0xde | P_EXT | P_DATA16)
358#define OPC_PMAXUW      (0x3e | P_EXT38 | P_DATA16)
359#define OPC_PMAXUD      (0x3f | P_EXT38 | P_DATA16)
360#define OPC_VPMAXUQ     (0x3f | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
361#define OPC_PMINSB      (0x38 | P_EXT38 | P_DATA16)
362#define OPC_PMINSW      (0xea | P_EXT | P_DATA16)
363#define OPC_PMINSD      (0x39 | P_EXT38 | P_DATA16)
364#define OPC_VPMINSQ     (0x39 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
365#define OPC_PMINUB      (0xda | P_EXT | P_DATA16)
366#define OPC_PMINUW      (0x3a | P_EXT38 | P_DATA16)
367#define OPC_PMINUD      (0x3b | P_EXT38 | P_DATA16)
368#define OPC_VPMINUQ     (0x3b | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
369#define OPC_PMOVSXBW    (0x20 | P_EXT38 | P_DATA16)
370#define OPC_PMOVSXWD    (0x23 | P_EXT38 | P_DATA16)
371#define OPC_PMOVSXDQ    (0x25 | P_EXT38 | P_DATA16)
372#define OPC_PMOVZXBW    (0x30 | P_EXT38 | P_DATA16)
373#define OPC_PMOVZXWD    (0x33 | P_EXT38 | P_DATA16)
374#define OPC_PMOVZXDQ    (0x35 | P_EXT38 | P_DATA16)
375#define OPC_PMULLW      (0xd5 | P_EXT | P_DATA16)
376#define OPC_PMULLD      (0x40 | P_EXT38 | P_DATA16)
377#define OPC_VPMULLQ     (0x40 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
378#define OPC_POR         (0xeb | P_EXT | P_DATA16)
379#define OPC_PSHUFB      (0x00 | P_EXT38 | P_DATA16)
380#define OPC_PSHUFD      (0x70 | P_EXT | P_DATA16)
381#define OPC_PSHUFLW     (0x70 | P_EXT | P_SIMDF2)
382#define OPC_PSHUFHW     (0x70 | P_EXT | P_SIMDF3)
383#define OPC_PSHIFTW_Ib  (0x71 | P_EXT | P_DATA16) /* /2 /6 /4 */
384#define OPC_PSHIFTD_Ib  (0x72 | P_EXT | P_DATA16) /* /1 /2 /6 /4 */
385#define OPC_PSHIFTQ_Ib  (0x73 | P_EXT | P_DATA16) /* /2 /6 /4 */
386#define OPC_PSLLW       (0xf1 | P_EXT | P_DATA16)
387#define OPC_PSLLD       (0xf2 | P_EXT | P_DATA16)
388#define OPC_PSLLQ       (0xf3 | P_EXT | P_DATA16)
389#define OPC_PSRAW       (0xe1 | P_EXT | P_DATA16)
390#define OPC_PSRAD       (0xe2 | P_EXT | P_DATA16)
391#define OPC_VPSRAQ      (0xe2 | P_EXT | P_DATA16 | P_VEXW | P_EVEX)
392#define OPC_PSRLW       (0xd1 | P_EXT | P_DATA16)
393#define OPC_PSRLD       (0xd2 | P_EXT | P_DATA16)
394#define OPC_PSRLQ       (0xd3 | P_EXT | P_DATA16)
395#define OPC_PSUBB       (0xf8 | P_EXT | P_DATA16)
396#define OPC_PSUBW       (0xf9 | P_EXT | P_DATA16)
397#define OPC_PSUBD       (0xfa | P_EXT | P_DATA16)
398#define OPC_PSUBQ       (0xfb | P_EXT | P_DATA16)
399#define OPC_PSUBSB      (0xe8 | P_EXT | P_DATA16)
400#define OPC_PSUBSW      (0xe9 | P_EXT | P_DATA16)
401#define OPC_PSUBUB      (0xd8 | P_EXT | P_DATA16)
402#define OPC_PSUBUW      (0xd9 | P_EXT | P_DATA16)
403#define OPC_PUNPCKLBW   (0x60 | P_EXT | P_DATA16)
404#define OPC_PUNPCKLWD   (0x61 | P_EXT | P_DATA16)
405#define OPC_PUNPCKLDQ   (0x62 | P_EXT | P_DATA16)
406#define OPC_PUNPCKLQDQ  (0x6c | P_EXT | P_DATA16)
407#define OPC_PUNPCKHBW   (0x68 | P_EXT | P_DATA16)
408#define OPC_PUNPCKHWD   (0x69 | P_EXT | P_DATA16)
409#define OPC_PUNPCKHDQ   (0x6a | P_EXT | P_DATA16)
410#define OPC_PUNPCKHQDQ  (0x6d | P_EXT | P_DATA16)
411#define OPC_PXOR        (0xef | P_EXT | P_DATA16)
412#define OPC_POP_r32	(0x58)
413#define OPC_POPCNT      (0xb8 | P_EXT | P_SIMDF3)
414#define OPC_PUSH_r32	(0x50)
415#define OPC_PUSH_Iv	(0x68)
416#define OPC_PUSH_Ib	(0x6a)
417#define OPC_RET		(0xc3)
418#define OPC_SETCC	(0x90 | P_EXT | P_REXB_RM) /* ... plus cc */
419#define OPC_SHIFT_1	(0xd1)
420#define OPC_SHIFT_Ib	(0xc1)
421#define OPC_SHIFT_cl	(0xd3)
422#define OPC_SARX        (0xf7 | P_EXT38 | P_SIMDF3)
423#define OPC_SHUFPS      (0xc6 | P_EXT)
424#define OPC_SHLX        (0xf7 | P_EXT38 | P_DATA16)
425#define OPC_SHRX        (0xf7 | P_EXT38 | P_SIMDF2)
426#define OPC_SHRD_Ib     (0xac | P_EXT)
427#define OPC_TESTB	(0x84)
428#define OPC_TESTL	(0x85)
429#define OPC_TZCNT       (0xbc | P_EXT | P_SIMDF3)
430#define OPC_UD2         (0x0b | P_EXT)
431#define OPC_VPBLENDD    (0x02 | P_EXT3A | P_DATA16)
432#define OPC_VPBLENDVB   (0x4c | P_EXT3A | P_DATA16)
433#define OPC_VPBLENDMB   (0x66 | P_EXT38 | P_DATA16 | P_EVEX)
434#define OPC_VPBLENDMW   (0x66 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
435#define OPC_VPBLENDMD   (0x64 | P_EXT38 | P_DATA16 | P_EVEX)
436#define OPC_VPBLENDMQ   (0x64 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
437#define OPC_VPCMPB      (0x3f | P_EXT3A | P_DATA16 | P_EVEX)
438#define OPC_VPCMPUB     (0x3e | P_EXT3A | P_DATA16 | P_EVEX)
439#define OPC_VPCMPW      (0x3f | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
440#define OPC_VPCMPUW     (0x3e | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
441#define OPC_VPCMPD      (0x1f | P_EXT3A | P_DATA16 | P_EVEX)
442#define OPC_VPCMPUD     (0x1e | P_EXT3A | P_DATA16 | P_EVEX)
443#define OPC_VPCMPQ      (0x1f | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
444#define OPC_VPCMPUQ     (0x1e | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
445#define OPC_VPINSRB     (0x20 | P_EXT3A | P_DATA16)
446#define OPC_VPINSRW     (0xc4 | P_EXT | P_DATA16)
447#define OPC_VBROADCASTSS (0x18 | P_EXT38 | P_DATA16)
448#define OPC_VBROADCASTSD (0x19 | P_EXT38 | P_DATA16)
449#define OPC_VPBROADCASTB (0x78 | P_EXT38 | P_DATA16)
450#define OPC_VPBROADCASTW (0x79 | P_EXT38 | P_DATA16)
451#define OPC_VPBROADCASTD (0x58 | P_EXT38 | P_DATA16)
452#define OPC_VPBROADCASTQ (0x59 | P_EXT38 | P_DATA16)
453#define OPC_VPMOVM2B    (0x28 | P_EXT38 | P_SIMDF3 | P_EVEX)
454#define OPC_VPMOVM2W    (0x28 | P_EXT38 | P_SIMDF3 | P_VEXW | P_EVEX)
455#define OPC_VPMOVM2D    (0x38 | P_EXT38 | P_SIMDF3 | P_EVEX)
456#define OPC_VPMOVM2Q    (0x38 | P_EXT38 | P_SIMDF3 | P_VEXW | P_EVEX)
457#define OPC_VPERMQ      (0x00 | P_EXT3A | P_DATA16 | P_VEXW)
458#define OPC_VPERM2I128  (0x46 | P_EXT3A | P_DATA16 | P_VEXL)
459#define OPC_VPROLVD     (0x15 | P_EXT38 | P_DATA16 | P_EVEX)
460#define OPC_VPROLVQ     (0x15 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
461#define OPC_VPRORVD     (0x14 | P_EXT38 | P_DATA16 | P_EVEX)
462#define OPC_VPRORVQ     (0x14 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
463#define OPC_VPSHLDW     (0x70 | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
464#define OPC_VPSHLDD     (0x71 | P_EXT3A | P_DATA16 | P_EVEX)
465#define OPC_VPSHLDQ     (0x71 | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
466#define OPC_VPSHLDVW    (0x70 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
467#define OPC_VPSHLDVD    (0x71 | P_EXT38 | P_DATA16 | P_EVEX)
468#define OPC_VPSHLDVQ    (0x71 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
469#define OPC_VPSHRDVW    (0x72 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
470#define OPC_VPSHRDVD    (0x73 | P_EXT38 | P_DATA16 | P_EVEX)
471#define OPC_VPSHRDVQ    (0x73 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
472#define OPC_VPSLLVW     (0x12 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
473#define OPC_VPSLLVD     (0x47 | P_EXT38 | P_DATA16)
474#define OPC_VPSLLVQ     (0x47 | P_EXT38 | P_DATA16 | P_VEXW)
475#define OPC_VPSRAVW     (0x11 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
476#define OPC_VPSRAVD     (0x46 | P_EXT38 | P_DATA16)
477#define OPC_VPSRAVQ     (0x46 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
478#define OPC_VPSRLVW     (0x10 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
479#define OPC_VPSRLVD     (0x45 | P_EXT38 | P_DATA16)
480#define OPC_VPSRLVQ     (0x45 | P_EXT38 | P_DATA16 | P_VEXW)
481#define OPC_VPTERNLOGQ  (0x25 | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
482#define OPC_VPTESTMB    (0x26 | P_EXT38 | P_DATA16 | P_EVEX)
483#define OPC_VPTESTMW    (0x26 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
484#define OPC_VPTESTMD    (0x27 | P_EXT38 | P_DATA16 | P_EVEX)
485#define OPC_VPTESTMQ    (0x27 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
486#define OPC_VPTESTNMB   (0x26 | P_EXT38 | P_SIMDF3 | P_EVEX)
487#define OPC_VPTESTNMW   (0x26 | P_EXT38 | P_SIMDF3 | P_VEXW | P_EVEX)
488#define OPC_VPTESTNMD   (0x27 | P_EXT38 | P_SIMDF3 | P_EVEX)
489#define OPC_VPTESTNMQ   (0x27 | P_EXT38 | P_SIMDF3 | P_VEXW | P_EVEX)
490#define OPC_VZEROUPPER  (0x77 | P_EXT)
491#define OPC_XCHG_ax_r32	(0x90)
492#define OPC_XCHG_EvGv   (0x87)
493
494#define OPC_GRP3_Eb     (0xf6)
495#define OPC_GRP3_Ev     (0xf7)
496#define OPC_GRP5        (0xff)
497#define OPC_GRP14       (0x73 | P_EXT | P_DATA16)
498#define OPC_GRPBT       (0xba | P_EXT)
499
500#define OPC_GRPBT_BT    4
501#define OPC_GRPBT_BTS   5
502#define OPC_GRPBT_BTR   6
503#define OPC_GRPBT_BTC   7
504
505/* Group 1 opcode extensions for 0x80-0x83.
506   These are also used as modifiers for OPC_ARITH.  */
507#define ARITH_ADD 0
508#define ARITH_OR  1
509#define ARITH_ADC 2
510#define ARITH_SBB 3
511#define ARITH_AND 4
512#define ARITH_SUB 5
513#define ARITH_XOR 6
514#define ARITH_CMP 7
515
516/* Group 2 opcode extensions for 0xc0, 0xc1, 0xd0-0xd3.  */
517#define SHIFT_ROL 0
518#define SHIFT_ROR 1
519#define SHIFT_SHL 4
520#define SHIFT_SHR 5
521#define SHIFT_SAR 7
522
523/* Group 3 opcode extensions for 0xf6, 0xf7.  To be used with OPC_GRP3.  */
524#define EXT3_TESTi 0
525#define EXT3_NOT   2
526#define EXT3_NEG   3
527#define EXT3_MUL   4
528#define EXT3_IMUL  5
529#define EXT3_DIV   6
530#define EXT3_IDIV  7
531
532/* Group 5 opcode extensions for 0xff.  To be used with OPC_GRP5.  */
533#define EXT5_INC_Ev	0
534#define EXT5_DEC_Ev	1
535#define EXT5_CALLN_Ev	2
536#define EXT5_JMPN_Ev	4
537
538/* Condition codes to be added to OPC_JCC_{long,short}.  */
539#define JCC_JMP (-1)
540#define JCC_JO  0x0
541#define JCC_JNO 0x1
542#define JCC_JB  0x2
543#define JCC_JAE 0x3
544#define JCC_JE  0x4
545#define JCC_JNE 0x5
546#define JCC_JBE 0x6
547#define JCC_JA  0x7
548#define JCC_JS  0x8
549#define JCC_JNS 0x9
550#define JCC_JP  0xa
551#define JCC_JNP 0xb
552#define JCC_JL  0xc
553#define JCC_JGE 0xd
554#define JCC_JLE 0xe
555#define JCC_JG  0xf
556
557static const uint8_t tcg_cond_to_jcc[] = {
558    [TCG_COND_EQ] = JCC_JE,
559    [TCG_COND_NE] = JCC_JNE,
560    [TCG_COND_LT] = JCC_JL,
561    [TCG_COND_GE] = JCC_JGE,
562    [TCG_COND_LE] = JCC_JLE,
563    [TCG_COND_GT] = JCC_JG,
564    [TCG_COND_LTU] = JCC_JB,
565    [TCG_COND_GEU] = JCC_JAE,
566    [TCG_COND_LEU] = JCC_JBE,
567    [TCG_COND_GTU] = JCC_JA,
568    [TCG_COND_TSTEQ] = JCC_JE,
569    [TCG_COND_TSTNE] = JCC_JNE,
570};
571
572#if TCG_TARGET_REG_BITS == 64
573static void tcg_out_opc(TCGContext *s, int opc, int r, int rm, int x)
574{
575    int rex;
576
577    if (opc & P_GS) {
578        tcg_out8(s, 0x65);
579    }
580    if (opc & P_DATA16) {
581        /* We should never be asking for both 16 and 64-bit operation.  */
582        tcg_debug_assert((opc & P_REXW) == 0);
583        tcg_out8(s, 0x66);
584    }
585    if (opc & P_SIMDF3) {
586        tcg_out8(s, 0xf3);
587    } else if (opc & P_SIMDF2) {
588        tcg_out8(s, 0xf2);
589    }
590
591    rex = 0;
592    rex |= (opc & P_REXW) ? 0x8 : 0x0;  /* REX.W */
593    rex |= (r & 8) >> 1;                /* REX.R */
594    rex |= (x & 8) >> 2;                /* REX.X */
595    rex |= (rm & 8) >> 3;               /* REX.B */
596
597    /* P_REXB_{R,RM} indicates that the given register is the low byte.
598       For %[abcd]l we need no REX prefix, but for %{si,di,bp,sp}l we do,
599       as otherwise the encoding indicates %[abcd]h.  Note that the values
600       that are ORed in merely indicate that the REX byte must be present;
601       those bits get discarded in output.  */
602    rex |= opc & (r >= 4 ? P_REXB_R : 0);
603    rex |= opc & (rm >= 4 ? P_REXB_RM : 0);
604
605    if (rex) {
606        tcg_out8(s, (uint8_t)(rex | 0x40));
607    }
608
609    if (opc & (P_EXT | P_EXT38 | P_EXT3A)) {
610        tcg_out8(s, 0x0f);
611        if (opc & P_EXT38) {
612            tcg_out8(s, 0x38);
613        } else if (opc & P_EXT3A) {
614            tcg_out8(s, 0x3a);
615        }
616    }
617
618    tcg_out8(s, opc);
619}
620#else
621static void tcg_out_opc(TCGContext *s, int opc)
622{
623    if (opc & P_DATA16) {
624        tcg_out8(s, 0x66);
625    }
626    if (opc & P_SIMDF3) {
627        tcg_out8(s, 0xf3);
628    } else if (opc & P_SIMDF2) {
629        tcg_out8(s, 0xf2);
630    }
631    if (opc & (P_EXT | P_EXT38 | P_EXT3A)) {
632        tcg_out8(s, 0x0f);
633        if (opc & P_EXT38) {
634            tcg_out8(s, 0x38);
635        } else if (opc & P_EXT3A) {
636            tcg_out8(s, 0x3a);
637        }
638    }
639    tcg_out8(s, opc);
640}
641/* Discard the register arguments to tcg_out_opc early, so as not to penalize
642   the 32-bit compilation paths.  This method works with all versions of gcc,
643   whereas relying on optimization may not be able to exclude them.  */
644#define tcg_out_opc(s, opc, r, rm, x)  (tcg_out_opc)(s, opc)
645#endif
646
647static void tcg_out_modrm(TCGContext *s, int opc, int r, int rm)
648{
649    tcg_out_opc(s, opc, r, rm, 0);
650    tcg_out8(s, 0xc0 | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
651}
652
653static void tcg_out_vex_opc(TCGContext *s, int opc, int r, int v,
654                            int rm, int index)
655{
656    int tmp;
657
658    if (opc & P_GS) {
659        tcg_out8(s, 0x65);
660    }
661    /* Use the two byte form if possible, which cannot encode
662       VEX.W, VEX.B, VEX.X, or an m-mmmm field other than P_EXT.  */
663    if ((opc & (P_EXT | P_EXT38 | P_EXT3A | P_VEXW)) == P_EXT
664        && ((rm | index) & 8) == 0) {
665        /* Two byte VEX prefix.  */
666        tcg_out8(s, 0xc5);
667
668        tmp = (r & 8 ? 0 : 0x80);              /* VEX.R */
669    } else {
670        /* Three byte VEX prefix.  */
671        tcg_out8(s, 0xc4);
672
673        /* VEX.m-mmmm */
674        if (opc & P_EXT3A) {
675            tmp = 3;
676        } else if (opc & P_EXT38) {
677            tmp = 2;
678        } else if (opc & P_EXT) {
679            tmp = 1;
680        } else {
681            g_assert_not_reached();
682        }
683        tmp |= (r & 8 ? 0 : 0x80);             /* VEX.R */
684        tmp |= (index & 8 ? 0 : 0x40);         /* VEX.X */
685        tmp |= (rm & 8 ? 0 : 0x20);            /* VEX.B */
686        tcg_out8(s, tmp);
687
688        tmp = (opc & P_VEXW ? 0x80 : 0);       /* VEX.W */
689    }
690
691    tmp |= (opc & P_VEXL ? 0x04 : 0);      /* VEX.L */
692    /* VEX.pp */
693    if (opc & P_DATA16) {
694        tmp |= 1;                          /* 0x66 */
695    } else if (opc & P_SIMDF3) {
696        tmp |= 2;                          /* 0xf3 */
697    } else if (opc & P_SIMDF2) {
698        tmp |= 3;                          /* 0xf2 */
699    }
700    tmp |= (~v & 15) << 3;                 /* VEX.vvvv */
701    tcg_out8(s, tmp);
702    tcg_out8(s, opc);
703}
704
705static void tcg_out_evex_opc(TCGContext *s, int opc, int r, int v,
706                             int rm, int index, int aaa, bool z)
707{
708    /* The entire 4-byte evex prefix; with R' and V' set. */
709    uint32_t p = 0x08041062;
710    int mm, pp;
711
712    tcg_debug_assert(have_avx512vl);
713
714    /* EVEX.mm */
715    if (opc & P_EXT3A) {
716        mm = 3;
717    } else if (opc & P_EXT38) {
718        mm = 2;
719    } else if (opc & P_EXT) {
720        mm = 1;
721    } else {
722        g_assert_not_reached();
723    }
724
725    /* EVEX.pp */
726    if (opc & P_DATA16) {
727        pp = 1;                          /* 0x66 */
728    } else if (opc & P_SIMDF3) {
729        pp = 2;                          /* 0xf3 */
730    } else if (opc & P_SIMDF2) {
731        pp = 3;                          /* 0xf2 */
732    } else {
733        pp = 0;
734    }
735
736    p = deposit32(p, 8, 2, mm);
737    p = deposit32(p, 13, 1, (rm & 8) == 0);             /* EVEX.RXB.B */
738    p = deposit32(p, 14, 1, (index & 8) == 0);          /* EVEX.RXB.X */
739    p = deposit32(p, 15, 1, (r & 8) == 0);              /* EVEX.RXB.R */
740    p = deposit32(p, 16, 2, pp);
741    p = deposit32(p, 19, 4, ~v);
742    p = deposit32(p, 23, 1, (opc & P_VEXW) != 0);
743    p = deposit32(p, 24, 3, aaa);
744    p = deposit32(p, 29, 2, (opc & P_VEXL) != 0);
745    p = deposit32(p, 31, 1, z);
746
747    tcg_out32(s, p);
748    tcg_out8(s, opc);
749}
750
751static void tcg_out_vex_modrm(TCGContext *s, int opc, int r, int v, int rm)
752{
753    if (opc & P_EVEX) {
754        tcg_out_evex_opc(s, opc, r, v, rm, 0, 0, false);
755    } else {
756        tcg_out_vex_opc(s, opc, r, v, rm, 0);
757    }
758    tcg_out8(s, 0xc0 | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
759}
760
761static void tcg_out_vex_modrm_type(TCGContext *s, int opc,
762                                   int r, int v, int rm, TCGType type)
763{
764    if (type == TCG_TYPE_V256) {
765        opc |= P_VEXL;
766    }
767    tcg_out_vex_modrm(s, opc, r, v, rm);
768}
769
770static void tcg_out_evex_modrm_type(TCGContext *s, int opc, int r, int v,
771                                    int rm, int aaa, bool z, TCGType type)
772{
773    if (type == TCG_TYPE_V256) {
774        opc |= P_VEXL;
775    }
776    tcg_out_evex_opc(s, opc, r, v, rm, 0, aaa, z);
777    tcg_out8(s, 0xc0 | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
778}
779
780/* Output an opcode with a full "rm + (index<<shift) + offset" address mode.
781   We handle either RM and INDEX missing with a negative value.  In 64-bit
782   mode for absolute addresses, ~RM is the size of the immediate operand
783   that will follow the instruction.  */
784
785static void tcg_out_sib_offset(TCGContext *s, int r, int rm, int index,
786                               int shift, intptr_t offset)
787{
788    int mod, len;
789
790    if (index < 0 && rm < 0) {
791        if (TCG_TARGET_REG_BITS == 64) {
792            /* Try for a rip-relative addressing mode.  This has replaced
793               the 32-bit-mode absolute addressing encoding.  */
794            intptr_t pc = (intptr_t)s->code_ptr + 5 + ~rm;
795            intptr_t disp = offset - pc;
796            if (disp == (int32_t)disp) {
797                tcg_out8(s, (LOWREGMASK(r) << 3) | 5);
798                tcg_out32(s, disp);
799                return;
800            }
801
802            /* Try for an absolute address encoding.  This requires the
803               use of the MODRM+SIB encoding and is therefore larger than
804               rip-relative addressing.  */
805            if (offset == (int32_t)offset) {
806                tcg_out8(s, (LOWREGMASK(r) << 3) | 4);
807                tcg_out8(s, (4 << 3) | 5);
808                tcg_out32(s, offset);
809                return;
810            }
811
812            /* ??? The memory isn't directly addressable.  */
813            g_assert_not_reached();
814        } else {
815            /* Absolute address.  */
816            tcg_out8(s, (r << 3) | 5);
817            tcg_out32(s, offset);
818            return;
819        }
820    }
821
822    /* Find the length of the immediate addend.  Note that the encoding
823       that would be used for (%ebp) indicates absolute addressing.  */
824    if (rm < 0) {
825        mod = 0, len = 4, rm = 5;
826    } else if (offset == 0 && LOWREGMASK(rm) != TCG_REG_EBP) {
827        mod = 0, len = 0;
828    } else if (offset == (int8_t)offset) {
829        mod = 0x40, len = 1;
830    } else {
831        mod = 0x80, len = 4;
832    }
833
834    /* Use a single byte MODRM format if possible.  Note that the encoding
835       that would be used for %esp is the escape to the two byte form.  */
836    if (index < 0 && LOWREGMASK(rm) != TCG_REG_ESP) {
837        /* Single byte MODRM format.  */
838        tcg_out8(s, mod | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
839    } else {
840        /* Two byte MODRM+SIB format.  */
841
842        /* Note that the encoding that would place %esp into the index
843           field indicates no index register.  In 64-bit mode, the REX.X
844           bit counts, so %r12 can be used as the index.  */
845        if (index < 0) {
846            index = 4;
847        } else {
848            tcg_debug_assert(index != TCG_REG_ESP);
849        }
850
851        tcg_out8(s, mod | (LOWREGMASK(r) << 3) | 4);
852        tcg_out8(s, (shift << 6) | (LOWREGMASK(index) << 3) | LOWREGMASK(rm));
853    }
854
855    if (len == 1) {
856        tcg_out8(s, offset);
857    } else if (len == 4) {
858        tcg_out32(s, offset);
859    }
860}
861
862static void tcg_out_modrm_sib_offset(TCGContext *s, int opc, int r, int rm,
863                                     int index, int shift, intptr_t offset)
864{
865    tcg_out_opc(s, opc, r, rm < 0 ? 0 : rm, index < 0 ? 0 : index);
866    tcg_out_sib_offset(s, r, rm, index, shift, offset);
867}
868
869static void tcg_out_vex_modrm_sib_offset(TCGContext *s, int opc, int r, int v,
870                                         int rm, int index, int shift,
871                                         intptr_t offset)
872{
873    tcg_out_vex_opc(s, opc, r, v, rm < 0 ? 0 : rm, index < 0 ? 0 : index);
874    tcg_out_sib_offset(s, r, rm, index, shift, offset);
875}
876
877/* A simplification of the above with no index or shift.  */
878static inline void tcg_out_modrm_offset(TCGContext *s, int opc, int r,
879                                        int rm, intptr_t offset)
880{
881    tcg_out_modrm_sib_offset(s, opc, r, rm, -1, 0, offset);
882}
883
884static inline void tcg_out_vex_modrm_offset(TCGContext *s, int opc, int r,
885                                            int v, int rm, intptr_t offset)
886{
887    tcg_out_vex_modrm_sib_offset(s, opc, r, v, rm, -1, 0, offset);
888}
889
890/* Output an opcode with an expected reference to the constant pool.  */
891static inline void tcg_out_modrm_pool(TCGContext *s, int opc, int r)
892{
893    tcg_out_opc(s, opc, r, 0, 0);
894    /* Absolute for 32-bit, pc-relative for 64-bit.  */
895    tcg_out8(s, LOWREGMASK(r) << 3 | 5);
896    tcg_out32(s, 0);
897}
898
899/* Output an opcode with an expected reference to the constant pool.  */
900static inline void tcg_out_vex_modrm_pool(TCGContext *s, int opc, int r)
901{
902    tcg_out_vex_opc(s, opc, r, 0, 0, 0);
903    /* Absolute for 32-bit, pc-relative for 64-bit.  */
904    tcg_out8(s, LOWREGMASK(r) << 3 | 5);
905    tcg_out32(s, 0);
906}
907
908/* Generate dest op= src.  Uses the same ARITH_* codes as tgen_arithi.  */
909static inline void tgen_arithr(TCGContext *s, int subop, int dest, int src)
910{
911    /* Propagate an opcode prefix, such as P_REXW.  */
912    int ext = subop & ~0x7;
913    subop &= 0x7;
914
915    tcg_out_modrm(s, OPC_ARITH_GvEv + (subop << 3) + ext, dest, src);
916}
917
918static bool tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
919{
920    int rexw = 0;
921
922    if (arg == ret) {
923        return true;
924    }
925    switch (type) {
926    case TCG_TYPE_I64:
927        rexw = P_REXW;
928        /* fallthru */
929    case TCG_TYPE_I32:
930        if (ret < 16) {
931            if (arg < 16) {
932                tcg_out_modrm(s, OPC_MOVL_GvEv + rexw, ret, arg);
933            } else {
934                tcg_out_vex_modrm(s, OPC_MOVD_EyVy + rexw, arg, 0, ret);
935            }
936        } else {
937            if (arg < 16) {
938                tcg_out_vex_modrm(s, OPC_MOVD_VyEy + rexw, ret, 0, arg);
939            } else {
940                tcg_out_vex_modrm(s, OPC_MOVQ_VqWq, ret, 0, arg);
941            }
942        }
943        break;
944
945    case TCG_TYPE_V64:
946        tcg_debug_assert(ret >= 16 && arg >= 16);
947        tcg_out_vex_modrm(s, OPC_MOVQ_VqWq, ret, 0, arg);
948        break;
949    case TCG_TYPE_V128:
950        tcg_debug_assert(ret >= 16 && arg >= 16);
951        tcg_out_vex_modrm(s, OPC_MOVDQA_VxWx, ret, 0, arg);
952        break;
953    case TCG_TYPE_V256:
954        tcg_debug_assert(ret >= 16 && arg >= 16);
955        tcg_out_vex_modrm(s, OPC_MOVDQA_VxWx | P_VEXL, ret, 0, arg);
956        break;
957
958    default:
959        g_assert_not_reached();
960    }
961    return true;
962}
963
964static const int avx2_dup_insn[4] = {
965    OPC_VPBROADCASTB, OPC_VPBROADCASTW,
966    OPC_VPBROADCASTD, OPC_VPBROADCASTQ,
967};
968
969static bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece,
970                            TCGReg r, TCGReg a)
971{
972    if (have_avx2) {
973        tcg_out_vex_modrm_type(s, avx2_dup_insn[vece], r, 0, a, type);
974    } else {
975        switch (vece) {
976        case MO_8:
977            /* ??? With zero in a register, use PSHUFB.  */
978            tcg_out_vex_modrm(s, OPC_PUNPCKLBW, r, a, a);
979            a = r;
980            /* FALLTHRU */
981        case MO_16:
982            tcg_out_vex_modrm(s, OPC_PUNPCKLWD, r, a, a);
983            a = r;
984            /* FALLTHRU */
985        case MO_32:
986            tcg_out_vex_modrm(s, OPC_PSHUFD, r, 0, a);
987            /* imm8 operand: all output lanes selected from input lane 0.  */
988            tcg_out8(s, 0);
989            break;
990        case MO_64:
991            tcg_out_vex_modrm(s, OPC_PUNPCKLQDQ, r, a, a);
992            break;
993        default:
994            g_assert_not_reached();
995        }
996    }
997    return true;
998}
999
1000static bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece,
1001                             TCGReg r, TCGReg base, intptr_t offset)
1002{
1003    if (have_avx2) {
1004        int vex_l = (type == TCG_TYPE_V256 ? P_VEXL : 0);
1005        tcg_out_vex_modrm_offset(s, avx2_dup_insn[vece] + vex_l,
1006                                 r, 0, base, offset);
1007    } else {
1008        switch (vece) {
1009        case MO_64:
1010            tcg_out_vex_modrm_offset(s, OPC_MOVDDUP, r, 0, base, offset);
1011            break;
1012        case MO_32:
1013            tcg_out_vex_modrm_offset(s, OPC_VBROADCASTSS, r, 0, base, offset);
1014            break;
1015        case MO_16:
1016            tcg_out_vex_modrm_offset(s, OPC_VPINSRW, r, r, base, offset);
1017            tcg_out8(s, 0); /* imm8 */
1018            tcg_out_dup_vec(s, type, vece, r, r);
1019            break;
1020        case MO_8:
1021            tcg_out_vex_modrm_offset(s, OPC_VPINSRB, r, r, base, offset);
1022            tcg_out8(s, 0); /* imm8 */
1023            tcg_out_dup_vec(s, type, vece, r, r);
1024            break;
1025        default:
1026            g_assert_not_reached();
1027        }
1028    }
1029    return true;
1030}
1031
1032static void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece,
1033                             TCGReg ret, int64_t arg)
1034{
1035    int vex_l = (type == TCG_TYPE_V256 ? P_VEXL : 0);
1036
1037    if (arg == 0) {
1038        tcg_out_vex_modrm(s, OPC_PXOR, ret, ret, ret);
1039        return;
1040    }
1041    if (arg == -1) {
1042        tcg_out_vex_modrm(s, OPC_PCMPEQB + vex_l, ret, ret, ret);
1043        return;
1044    }
1045
1046    if (TCG_TARGET_REG_BITS == 32 && vece < MO_64) {
1047        if (have_avx2) {
1048            tcg_out_vex_modrm_pool(s, OPC_VPBROADCASTD + vex_l, ret);
1049        } else {
1050            tcg_out_vex_modrm_pool(s, OPC_VBROADCASTSS, ret);
1051        }
1052        new_pool_label(s, arg, R_386_32, s->code_ptr - 4, 0);
1053    } else {
1054        if (type == TCG_TYPE_V64) {
1055            tcg_out_vex_modrm_pool(s, OPC_MOVQ_VqWq, ret);
1056        } else if (have_avx2) {
1057            tcg_out_vex_modrm_pool(s, OPC_VPBROADCASTQ + vex_l, ret);
1058        } else {
1059            tcg_out_vex_modrm_pool(s, OPC_MOVDDUP, ret);
1060        }
1061        if (TCG_TARGET_REG_BITS == 64) {
1062            new_pool_label(s, arg, R_386_PC32, s->code_ptr - 4, -4);
1063        } else {
1064            new_pool_l2(s, R_386_32, s->code_ptr - 4, 0, arg, arg >> 32);
1065        }
1066    }
1067}
1068
1069static void tcg_out_movi_vec(TCGContext *s, TCGType type,
1070                             TCGReg ret, tcg_target_long arg)
1071{
1072    if (arg == 0) {
1073        tcg_out_vex_modrm(s, OPC_PXOR, ret, ret, ret);
1074        return;
1075    }
1076    if (arg == -1) {
1077        tcg_out_vex_modrm(s, OPC_PCMPEQB, ret, ret, ret);
1078        return;
1079    }
1080
1081    int rexw = (type == TCG_TYPE_I32 ? 0 : P_REXW);
1082    tcg_out_vex_modrm_pool(s, OPC_MOVD_VyEy + rexw, ret);
1083    if (TCG_TARGET_REG_BITS == 64) {
1084        new_pool_label(s, arg, R_386_PC32, s->code_ptr - 4, -4);
1085    } else {
1086        new_pool_label(s, arg, R_386_32, s->code_ptr - 4, 0);
1087    }
1088}
1089
1090static void tcg_out_movi_int(TCGContext *s, TCGType type,
1091                             TCGReg ret, tcg_target_long arg)
1092{
1093    tcg_target_long diff;
1094
1095    if (arg == 0) {
1096        tgen_arithr(s, ARITH_XOR, ret, ret);
1097        return;
1098    }
1099    if (arg == (uint32_t)arg || type == TCG_TYPE_I32) {
1100        tcg_out_opc(s, OPC_MOVL_Iv + LOWREGMASK(ret), 0, ret, 0);
1101        tcg_out32(s, arg);
1102        return;
1103    }
1104    if (arg == (int32_t)arg) {
1105        tcg_out_modrm(s, OPC_MOVL_EvIz + P_REXW, 0, ret);
1106        tcg_out32(s, arg);
1107        return;
1108    }
1109
1110    /* Try a 7 byte pc-relative lea before the 10 byte movq.  */
1111    diff = tcg_pcrel_diff(s, (const void *)arg) - 7;
1112    if (diff == (int32_t)diff) {
1113        tcg_out_opc(s, OPC_LEA | P_REXW, ret, 0, 0);
1114        tcg_out8(s, (LOWREGMASK(ret) << 3) | 5);
1115        tcg_out32(s, diff);
1116        return;
1117    }
1118
1119    tcg_out_opc(s, OPC_MOVL_Iv + P_REXW + LOWREGMASK(ret), 0, ret, 0);
1120    tcg_out64(s, arg);
1121}
1122
1123static void tcg_out_movi(TCGContext *s, TCGType type,
1124                         TCGReg ret, tcg_target_long arg)
1125{
1126    switch (type) {
1127    case TCG_TYPE_I32:
1128#if TCG_TARGET_REG_BITS == 64
1129    case TCG_TYPE_I64:
1130#endif
1131        if (ret < 16) {
1132            tcg_out_movi_int(s, type, ret, arg);
1133        } else {
1134            tcg_out_movi_vec(s, type, ret, arg);
1135        }
1136        break;
1137    default:
1138        g_assert_not_reached();
1139    }
1140}
1141
1142static bool tcg_out_xchg(TCGContext *s, TCGType type, TCGReg r1, TCGReg r2)
1143{
1144    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1145    tcg_out_modrm(s, OPC_XCHG_EvGv + rexw, r1, r2);
1146    return true;
1147}
1148
1149static void tcg_out_addi_ptr(TCGContext *s, TCGReg rd, TCGReg rs,
1150                             tcg_target_long imm)
1151{
1152    /* This function is only used for passing structs by reference. */
1153    tcg_debug_assert(imm == (int32_t)imm);
1154    tcg_out_modrm_offset(s, OPC_LEA | P_REXW, rd, rs, imm);
1155}
1156
1157static inline void tcg_out_pushi(TCGContext *s, tcg_target_long val)
1158{
1159    if (val == (int8_t)val) {
1160        tcg_out_opc(s, OPC_PUSH_Ib, 0, 0, 0);
1161        tcg_out8(s, val);
1162    } else if (val == (int32_t)val) {
1163        tcg_out_opc(s, OPC_PUSH_Iv, 0, 0, 0);
1164        tcg_out32(s, val);
1165    } else {
1166        g_assert_not_reached();
1167    }
1168}
1169
1170static inline void tcg_out_mb(TCGContext *s, TCGArg a0)
1171{
1172    /* Given the strength of x86 memory ordering, we only need care for
1173       store-load ordering.  Experimentally, "lock orl $0,0(%esp)" is
1174       faster than "mfence", so don't bother with the sse insn.  */
1175    if (a0 & TCG_MO_ST_LD) {
1176        tcg_out8(s, 0xf0);
1177        tcg_out_modrm_offset(s, OPC_ARITH_EvIb, ARITH_OR, TCG_REG_ESP, 0);
1178        tcg_out8(s, 0);
1179    }
1180}
1181
1182static inline void tcg_out_push(TCGContext *s, int reg)
1183{
1184    tcg_out_opc(s, OPC_PUSH_r32 + LOWREGMASK(reg), 0, reg, 0);
1185}
1186
1187static inline void tcg_out_pop(TCGContext *s, int reg)
1188{
1189    tcg_out_opc(s, OPC_POP_r32 + LOWREGMASK(reg), 0, reg, 0);
1190}
1191
1192static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg ret,
1193                       TCGReg arg1, intptr_t arg2)
1194{
1195    switch (type) {
1196    case TCG_TYPE_I32:
1197        if (ret < 16) {
1198            tcg_out_modrm_offset(s, OPC_MOVL_GvEv, ret, arg1, arg2);
1199        } else {
1200            tcg_out_vex_modrm_offset(s, OPC_MOVD_VyEy, ret, 0, arg1, arg2);
1201        }
1202        break;
1203    case TCG_TYPE_I64:
1204        if (ret < 16) {
1205            tcg_out_modrm_offset(s, OPC_MOVL_GvEv | P_REXW, ret, arg1, arg2);
1206            break;
1207        }
1208        /* FALLTHRU */
1209    case TCG_TYPE_V64:
1210        /* There is no instruction that can validate 8-byte alignment.  */
1211        tcg_debug_assert(ret >= 16);
1212        tcg_out_vex_modrm_offset(s, OPC_MOVQ_VqWq, ret, 0, arg1, arg2);
1213        break;
1214    case TCG_TYPE_V128:
1215        /*
1216         * The gvec infrastructure is asserts that v128 vector loads
1217         * and stores use a 16-byte aligned offset.  Validate that the
1218         * final pointer is aligned by using an insn that will SIGSEGV.
1219         */
1220        tcg_debug_assert(ret >= 16);
1221        tcg_out_vex_modrm_offset(s, OPC_MOVDQA_VxWx, ret, 0, arg1, arg2);
1222        break;
1223    case TCG_TYPE_V256:
1224        /*
1225         * The gvec infrastructure only requires 16-byte alignment,
1226         * so here we must use an unaligned load.
1227         */
1228        tcg_debug_assert(ret >= 16);
1229        tcg_out_vex_modrm_offset(s, OPC_MOVDQU_VxWx | P_VEXL,
1230                                 ret, 0, arg1, arg2);
1231        break;
1232    default:
1233        g_assert_not_reached();
1234    }
1235}
1236
1237static void tcg_out_st(TCGContext *s, TCGType type, TCGReg arg,
1238                       TCGReg arg1, intptr_t arg2)
1239{
1240    switch (type) {
1241    case TCG_TYPE_I32:
1242        if (arg < 16) {
1243            tcg_out_modrm_offset(s, OPC_MOVL_EvGv, arg, arg1, arg2);
1244        } else {
1245            tcg_out_vex_modrm_offset(s, OPC_MOVD_EyVy, arg, 0, arg1, arg2);
1246        }
1247        break;
1248    case TCG_TYPE_I64:
1249        if (arg < 16) {
1250            tcg_out_modrm_offset(s, OPC_MOVL_EvGv | P_REXW, arg, arg1, arg2);
1251            break;
1252        }
1253        /* FALLTHRU */
1254    case TCG_TYPE_V64:
1255        /* There is no instruction that can validate 8-byte alignment.  */
1256        tcg_debug_assert(arg >= 16);
1257        tcg_out_vex_modrm_offset(s, OPC_MOVQ_WqVq, arg, 0, arg1, arg2);
1258        break;
1259    case TCG_TYPE_V128:
1260        /*
1261         * The gvec infrastructure is asserts that v128 vector loads
1262         * and stores use a 16-byte aligned offset.  Validate that the
1263         * final pointer is aligned by using an insn that will SIGSEGV.
1264         *
1265         * This specific instance is also used by TCG_CALL_RET_BY_VEC,
1266         * for _WIN64, which must have SSE2 but may not have AVX.
1267         */
1268        tcg_debug_assert(arg >= 16);
1269        if (have_avx1) {
1270            tcg_out_vex_modrm_offset(s, OPC_MOVDQA_WxVx, arg, 0, arg1, arg2);
1271        } else {
1272            tcg_out_modrm_offset(s, OPC_MOVDQA_WxVx, arg, arg1, arg2);
1273        }
1274        break;
1275    case TCG_TYPE_V256:
1276        /*
1277         * The gvec infrastructure only requires 16-byte alignment,
1278         * so here we must use an unaligned store.
1279         */
1280        tcg_debug_assert(arg >= 16);
1281        tcg_out_vex_modrm_offset(s, OPC_MOVDQU_WxVx | P_VEXL,
1282                                 arg, 0, arg1, arg2);
1283        break;
1284    default:
1285        g_assert_not_reached();
1286    }
1287}
1288
1289static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val,
1290                        TCGReg base, intptr_t ofs)
1291{
1292    int rexw = 0;
1293    if (TCG_TARGET_REG_BITS == 64 && type == TCG_TYPE_I64) {
1294        if (val != (int32_t)val) {
1295            return false;
1296        }
1297        rexw = P_REXW;
1298    } else if (type != TCG_TYPE_I32) {
1299        return false;
1300    }
1301    tcg_out_modrm_offset(s, OPC_MOVL_EvIz | rexw, 0, base, ofs);
1302    tcg_out32(s, val);
1303    return true;
1304}
1305
1306static void tcg_out_shifti(TCGContext *s, int subopc, int reg, int count)
1307{
1308    /* Propagate an opcode prefix, such as P_DATA16.  */
1309    int ext = subopc & ~0x7;
1310    subopc &= 0x7;
1311
1312    if (count == 1) {
1313        tcg_out_modrm(s, OPC_SHIFT_1 + ext, subopc, reg);
1314    } else {
1315        tcg_out_modrm(s, OPC_SHIFT_Ib + ext, subopc, reg);
1316        tcg_out8(s, count);
1317    }
1318}
1319
1320static inline void tcg_out_bswap32(TCGContext *s, int reg)
1321{
1322    tcg_out_opc(s, OPC_BSWAP + LOWREGMASK(reg), 0, reg, 0);
1323}
1324
1325static inline void tcg_out_rolw_8(TCGContext *s, int reg)
1326{
1327    tcg_out_shifti(s, SHIFT_ROL + P_DATA16, reg, 8);
1328}
1329
1330static void tcg_out_ext8u(TCGContext *s, TCGReg dest, TCGReg src)
1331{
1332    if (TCG_TARGET_REG_BITS == 32 && src >= 4) {
1333        tcg_out_mov(s, TCG_TYPE_I32, dest, src);
1334        if (dest >= 4) {
1335            tcg_out_modrm(s, OPC_ARITH_EvIz, ARITH_AND, dest);
1336            tcg_out32(s, 0xff);
1337            return;
1338        }
1339        src = dest;
1340    }
1341    tcg_out_modrm(s, OPC_MOVZBL + P_REXB_RM, dest, src);
1342}
1343
1344static void tcg_out_ext8s(TCGContext *s, TCGType type, TCGReg dest, TCGReg src)
1345{
1346    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1347
1348    if (TCG_TARGET_REG_BITS == 32 && src >= 4) {
1349        tcg_out_mov(s, TCG_TYPE_I32, dest, src);
1350        if (dest >= 4) {
1351            tcg_out_shifti(s, SHIFT_SHL, dest, 24);
1352            tcg_out_shifti(s, SHIFT_SAR, dest, 24);
1353            return;
1354        }
1355        src = dest;
1356    }
1357    tcg_out_modrm(s, OPC_MOVSBL + P_REXB_RM + rexw, dest, src);
1358}
1359
1360static void tcg_out_ext16u(TCGContext *s, TCGReg dest, TCGReg src)
1361{
1362    /* movzwl */
1363    tcg_out_modrm(s, OPC_MOVZWL, dest, src);
1364}
1365
1366static void tcg_out_ext16s(TCGContext *s, TCGType type, TCGReg dest, TCGReg src)
1367{
1368    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1369    /* movsw[lq] */
1370    tcg_out_modrm(s, OPC_MOVSWL + rexw, dest, src);
1371}
1372
1373static void tcg_out_ext32u(TCGContext *s, TCGReg dest, TCGReg src)
1374{
1375    /* 32-bit mov zero extends.  */
1376    tcg_out_modrm(s, OPC_MOVL_GvEv, dest, src);
1377}
1378
1379static void tcg_out_ext32s(TCGContext *s, TCGReg dest, TCGReg src)
1380{
1381    tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
1382    tcg_out_modrm(s, OPC_MOVSLQ, dest, src);
1383}
1384
1385static void tcg_out_exts_i32_i64(TCGContext *s, TCGReg dest, TCGReg src)
1386{
1387    tcg_out_ext32s(s, dest, src);
1388}
1389
1390static void tcg_out_extu_i32_i64(TCGContext *s, TCGReg dest, TCGReg src)
1391{
1392    if (dest != src) {
1393        tcg_out_ext32u(s, dest, src);
1394    }
1395}
1396
1397static void tcg_out_extrl_i64_i32(TCGContext *s, TCGReg dest, TCGReg src)
1398{
1399    tcg_out_ext32u(s, dest, src);
1400}
1401
1402static inline void tcg_out_bswap64(TCGContext *s, int reg)
1403{
1404    tcg_out_opc(s, OPC_BSWAP + P_REXW + LOWREGMASK(reg), 0, reg, 0);
1405}
1406
1407static void tgen_arithi(TCGContext *s, int c, int r0,
1408                        tcg_target_long val, int cf)
1409{
1410    int rexw = 0;
1411
1412    if (TCG_TARGET_REG_BITS == 64) {
1413        rexw = c & -8;
1414        c &= 7;
1415    }
1416
1417    switch (c) {
1418    case ARITH_ADD:
1419    case ARITH_SUB:
1420        if (!cf) {
1421            /*
1422             * ??? While INC is 2 bytes shorter than ADDL $1, they also induce
1423             * partial flags update stalls on Pentium4 and are not recommended
1424             * by current Intel optimization manuals.
1425             */
1426            if (val == 1 || val == -1) {
1427                int is_inc = (c == ARITH_ADD) ^ (val < 0);
1428                if (TCG_TARGET_REG_BITS == 64) {
1429                    /*
1430                     * The single-byte increment encodings are re-tasked
1431                     * as the REX prefixes.  Use the MODRM encoding.
1432                     */
1433                    tcg_out_modrm(s, OPC_GRP5 + rexw,
1434                                  (is_inc ? EXT5_INC_Ev : EXT5_DEC_Ev), r0);
1435                } else {
1436                    tcg_out8(s, (is_inc ? OPC_INC_r32 : OPC_DEC_r32) + r0);
1437                }
1438                return;
1439            }
1440            if (val == 128) {
1441                /*
1442                 * Facilitate using an 8-bit immediate.  Carry is inverted
1443                 * by this transformation, so do it only if cf == 0.
1444                 */
1445                c ^= ARITH_ADD ^ ARITH_SUB;
1446                val = -128;
1447            }
1448        }
1449        break;
1450
1451    case ARITH_AND:
1452        if (TCG_TARGET_REG_BITS == 64) {
1453            if (val == 0xffffffffu) {
1454                tcg_out_ext32u(s, r0, r0);
1455                return;
1456            }
1457            if (val == (uint32_t)val) {
1458                /* AND with no high bits set can use a 32-bit operation.  */
1459                rexw = 0;
1460            }
1461        }
1462        if (val == 0xffu && (r0 < 4 || TCG_TARGET_REG_BITS == 64)) {
1463            tcg_out_ext8u(s, r0, r0);
1464            return;
1465        }
1466        if (val == 0xffffu) {
1467            tcg_out_ext16u(s, r0, r0);
1468            return;
1469        }
1470        break;
1471
1472    case ARITH_OR:
1473    case ARITH_XOR:
1474        if (val >= 0x80 && val <= 0xff
1475            && (r0 < 4 || TCG_TARGET_REG_BITS == 64)) {
1476            tcg_out_modrm(s, OPC_ARITH_EbIb + P_REXB_RM, c, r0);
1477            tcg_out8(s, val);
1478            return;
1479        }
1480        break;
1481    }
1482
1483    if (val == (int8_t)val) {
1484        tcg_out_modrm(s, OPC_ARITH_EvIb + rexw, c, r0);
1485        tcg_out8(s, val);
1486        return;
1487    }
1488    if (rexw == 0 || val == (int32_t)val) {
1489        tcg_out_modrm(s, OPC_ARITH_EvIz + rexw, c, r0);
1490        tcg_out32(s, val);
1491        return;
1492    }
1493
1494    g_assert_not_reached();
1495}
1496
1497static void tcg_out_addi(TCGContext *s, int reg, tcg_target_long val)
1498{
1499    if (val != 0) {
1500        tgen_arithi(s, ARITH_ADD + P_REXW, reg, val, 0);
1501    }
1502}
1503
1504/* Set SMALL to force a short forward branch.  */
1505static void tcg_out_jxx(TCGContext *s, int opc, TCGLabel *l, bool small)
1506{
1507    int32_t val, val1;
1508
1509    if (l->has_value) {
1510        val = tcg_pcrel_diff(s, l->u.value_ptr);
1511        val1 = val - 2;
1512        if ((int8_t)val1 == val1) {
1513            if (opc == -1) {
1514                tcg_out8(s, OPC_JMP_short);
1515            } else {
1516                tcg_out8(s, OPC_JCC_short + opc);
1517            }
1518            tcg_out8(s, val1);
1519        } else {
1520            tcg_debug_assert(!small);
1521            if (opc == -1) {
1522                tcg_out8(s, OPC_JMP_long);
1523                tcg_out32(s, val - 5);
1524            } else {
1525                tcg_out_opc(s, OPC_JCC_long + opc, 0, 0, 0);
1526                tcg_out32(s, val - 6);
1527            }
1528        }
1529    } else if (small) {
1530        if (opc == -1) {
1531            tcg_out8(s, OPC_JMP_short);
1532        } else {
1533            tcg_out8(s, OPC_JCC_short + opc);
1534        }
1535        tcg_out_reloc(s, s->code_ptr, R_386_PC8, l, -1);
1536        s->code_ptr += 1;
1537    } else {
1538        if (opc == -1) {
1539            tcg_out8(s, OPC_JMP_long);
1540        } else {
1541            tcg_out_opc(s, OPC_JCC_long + opc, 0, 0, 0);
1542        }
1543        tcg_out_reloc(s, s->code_ptr, R_386_PC32, l, -4);
1544        s->code_ptr += 4;
1545    }
1546}
1547
1548static int tcg_out_cmp(TCGContext *s, TCGCond cond, TCGArg arg1,
1549                       TCGArg arg2, int const_arg2, int rexw)
1550{
1551    int jz, js;
1552
1553    if (!is_tst_cond(cond)) {
1554        if (!const_arg2) {
1555            tgen_arithr(s, ARITH_CMP + rexw, arg1, arg2);
1556        } else if (arg2 == 0) {
1557            tcg_out_modrm(s, OPC_TESTL + rexw, arg1, arg1);
1558        } else {
1559            tcg_debug_assert(!rexw || arg2 == (int32_t)arg2);
1560            tgen_arithi(s, ARITH_CMP + rexw, arg1, arg2, 0);
1561        }
1562        return tcg_cond_to_jcc[cond];
1563    }
1564
1565    jz = tcg_cond_to_jcc[cond];
1566    js = (cond == TCG_COND_TSTNE ? JCC_JS : JCC_JNS);
1567
1568    if (!const_arg2) {
1569        tcg_out_modrm(s, OPC_TESTL + rexw, arg1, arg2);
1570        return jz;
1571    }
1572
1573    if (arg2 <= 0xff && (TCG_TARGET_REG_BITS == 64 || arg1 < 4)) {
1574        if (arg2 == 0x80) {
1575            tcg_out_modrm(s, OPC_TESTB | P_REXB_R, arg1, arg1);
1576            return js;
1577        }
1578        if (arg2 == 0xff) {
1579            tcg_out_modrm(s, OPC_TESTB | P_REXB_R, arg1, arg1);
1580            return jz;
1581        }
1582        tcg_out_modrm(s, OPC_GRP3_Eb | P_REXB_RM, EXT3_TESTi, arg1);
1583        tcg_out8(s, arg2);
1584        return jz;
1585    }
1586
1587    if ((arg2 & ~0xff00) == 0 && arg1 < 4) {
1588        if (arg2 == 0x8000) {
1589            tcg_out_modrm(s, OPC_TESTB, arg1 + 4, arg1 + 4);
1590            return js;
1591        }
1592        if (arg2 == 0xff00) {
1593            tcg_out_modrm(s, OPC_TESTB, arg1 + 4, arg1 + 4);
1594            return jz;
1595        }
1596        tcg_out_modrm(s, OPC_GRP3_Eb, EXT3_TESTi, arg1 + 4);
1597        tcg_out8(s, arg2 >> 8);
1598        return jz;
1599    }
1600
1601    if (arg2 == 0xffff) {
1602        tcg_out_modrm(s, OPC_TESTL | P_DATA16, arg1, arg1);
1603        return jz;
1604    }
1605    if (arg2 == 0xffffffffu) {
1606        tcg_out_modrm(s, OPC_TESTL, arg1, arg1);
1607        return jz;
1608    }
1609
1610    if (is_power_of_2(rexw ? arg2 : (uint32_t)arg2)) {
1611        int jc = (cond == TCG_COND_TSTNE ? JCC_JB : JCC_JAE);
1612        int sh = ctz64(arg2);
1613
1614        rexw = (sh & 32 ? P_REXW : 0);
1615        if ((sh & 31) == 31) {
1616            tcg_out_modrm(s, OPC_TESTL | rexw, arg1, arg1);
1617            return js;
1618        } else {
1619            tcg_out_modrm(s, OPC_GRPBT | rexw, OPC_GRPBT_BT, arg1);
1620            tcg_out8(s, sh);
1621            return jc;
1622        }
1623    }
1624
1625    if (rexw) {
1626        if (arg2 == (uint32_t)arg2) {
1627            rexw = 0;
1628        } else {
1629            tcg_debug_assert(arg2 == (int32_t)arg2);
1630        }
1631    }
1632    tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_TESTi, arg1);
1633    tcg_out32(s, arg2);
1634    return jz;
1635}
1636
1637static void tcg_out_brcond(TCGContext *s, int rexw, TCGCond cond,
1638                           TCGArg arg1, TCGArg arg2, int const_arg2,
1639                           TCGLabel *label, bool small)
1640{
1641    int jcc = tcg_out_cmp(s, cond, arg1, arg2, const_arg2, rexw);
1642    tcg_out_jxx(s, jcc, label, small);
1643}
1644
1645#if TCG_TARGET_REG_BITS == 32
1646static void tcg_out_brcond2(TCGContext *s, const TCGArg *args,
1647                            const int *const_args, bool small)
1648{
1649    TCGLabel *label_next = gen_new_label();
1650    TCGLabel *label_this = arg_label(args[5]);
1651    TCGCond cond = args[4];
1652
1653    switch (cond) {
1654    case TCG_COND_EQ:
1655    case TCG_COND_TSTEQ:
1656        tcg_out_brcond(s, 0, tcg_invert_cond(cond),
1657                       args[0], args[2], const_args[2], label_next, 1);
1658        tcg_out_brcond(s, 0, cond, args[1], args[3], const_args[3],
1659                       label_this, small);
1660        break;
1661    case TCG_COND_NE:
1662    case TCG_COND_TSTNE:
1663        tcg_out_brcond(s, 0, cond, args[0], args[2], const_args[2],
1664                       label_this, small);
1665        tcg_out_brcond(s, 0, cond, args[1], args[3], const_args[3],
1666                       label_this, small);
1667        break;
1668    case TCG_COND_LT:
1669        tcg_out_brcond(s, 0, TCG_COND_LT, args[1], args[3], const_args[3],
1670                       label_this, small);
1671        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1672        tcg_out_brcond(s, 0, TCG_COND_LTU, args[0], args[2], const_args[2],
1673                       label_this, small);
1674        break;
1675    case TCG_COND_LE:
1676        tcg_out_brcond(s, 0, TCG_COND_LT, args[1], args[3], const_args[3],
1677                       label_this, small);
1678        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1679        tcg_out_brcond(s, 0, TCG_COND_LEU, args[0], args[2], const_args[2],
1680                       label_this, small);
1681        break;
1682    case TCG_COND_GT:
1683        tcg_out_brcond(s, 0, TCG_COND_GT, args[1], args[3], const_args[3],
1684                       label_this, small);
1685        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1686        tcg_out_brcond(s, 0, TCG_COND_GTU, args[0], args[2], const_args[2],
1687                       label_this, small);
1688        break;
1689    case TCG_COND_GE:
1690        tcg_out_brcond(s, 0, TCG_COND_GT, args[1], args[3], const_args[3],
1691                       label_this, small);
1692        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1693        tcg_out_brcond(s, 0, TCG_COND_GEU, args[0], args[2], const_args[2],
1694                       label_this, small);
1695        break;
1696    case TCG_COND_LTU:
1697        tcg_out_brcond(s, 0, TCG_COND_LTU, args[1], args[3], const_args[3],
1698                       label_this, small);
1699        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1700        tcg_out_brcond(s, 0, TCG_COND_LTU, args[0], args[2], const_args[2],
1701                       label_this, small);
1702        break;
1703    case TCG_COND_LEU:
1704        tcg_out_brcond(s, 0, TCG_COND_LTU, args[1], args[3], const_args[3],
1705                       label_this, small);
1706        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1707        tcg_out_brcond(s, 0, TCG_COND_LEU, args[0], args[2], const_args[2],
1708                       label_this, small);
1709        break;
1710    case TCG_COND_GTU:
1711        tcg_out_brcond(s, 0, TCG_COND_GTU, args[1], args[3], const_args[3],
1712                       label_this, small);
1713        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1714        tcg_out_brcond(s, 0, TCG_COND_GTU, args[0], args[2], const_args[2],
1715                       label_this, small);
1716        break;
1717    case TCG_COND_GEU:
1718        tcg_out_brcond(s, 0, TCG_COND_GTU, args[1], args[3], const_args[3],
1719                       label_this, small);
1720        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1721        tcg_out_brcond(s, 0, TCG_COND_GEU, args[0], args[2], const_args[2],
1722                       label_this, small);
1723        break;
1724    default:
1725        g_assert_not_reached();
1726    }
1727    tcg_out_label(s, label_next);
1728}
1729#endif
1730
1731static void tcg_out_setcond(TCGContext *s, int rexw, TCGCond cond,
1732                            TCGArg dest, TCGArg arg1, TCGArg arg2,
1733                            int const_arg2, bool neg)
1734{
1735    int cmp_rexw = rexw;
1736    bool inv = false;
1737    bool cleared;
1738    int jcc;
1739
1740    switch (cond) {
1741    case TCG_COND_NE:
1742        inv = true;
1743        /* fall through */
1744    case TCG_COND_EQ:
1745        /* If arg2 is 0, convert to LTU/GEU vs 1. */
1746        if (const_arg2 && arg2 == 0) {
1747            arg2 = 1;
1748            goto do_ltu;
1749        }
1750        break;
1751
1752    case TCG_COND_TSTNE:
1753        inv = true;
1754        /* fall through */
1755    case TCG_COND_TSTEQ:
1756        /* If arg2 is -1, convert to LTU/GEU vs 1. */
1757        if (const_arg2 && arg2 == 0xffffffffu) {
1758            arg2 = 1;
1759            cmp_rexw = 0;
1760            goto do_ltu;
1761        }
1762        break;
1763
1764    case TCG_COND_LEU:
1765        inv = true;
1766        /* fall through */
1767    case TCG_COND_GTU:
1768        /* If arg2 is a register, swap for LTU/GEU. */
1769        if (!const_arg2) {
1770            TCGReg t = arg1;
1771            arg1 = arg2;
1772            arg2 = t;
1773            goto do_ltu;
1774        }
1775        break;
1776
1777    case TCG_COND_GEU:
1778        inv = true;
1779        /* fall through */
1780    case TCG_COND_LTU:
1781    do_ltu:
1782        /*
1783         * Relying on the carry bit, use SBB to produce -1 if LTU, 0 if GEU.
1784         * We can then use NEG or INC to produce the desired result.
1785         * This is always smaller than the SETCC expansion.
1786         */
1787        tcg_out_cmp(s, TCG_COND_LTU, arg1, arg2, const_arg2, cmp_rexw);
1788
1789        /* X - X - C = -C = (C ? -1 : 0) */
1790        tgen_arithr(s, ARITH_SBB + (neg ? rexw : 0), dest, dest);
1791        if (inv && neg) {
1792            /* ~(C ? -1 : 0) = (C ? 0 : -1) */
1793            tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NOT, dest);
1794        } else if (inv) {
1795            /* (C ? -1 : 0) + 1 = (C ? 0 : 1) */
1796            tgen_arithi(s, ARITH_ADD, dest, 1, 0);
1797        } else if (!neg) {
1798            /* -(C ? -1 : 0) = (C ? 1 : 0) */
1799            tcg_out_modrm(s, OPC_GRP3_Ev, EXT3_NEG, dest);
1800        }
1801        return;
1802
1803    case TCG_COND_GE:
1804        inv = true;
1805        /* fall through */
1806    case TCG_COND_LT:
1807        /* If arg2 is 0, extract the sign bit. */
1808        if (const_arg2 && arg2 == 0) {
1809            tcg_out_mov(s, rexw ? TCG_TYPE_I64 : TCG_TYPE_I32, dest, arg1);
1810            if (inv) {
1811                tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NOT, dest);
1812            }
1813            tcg_out_shifti(s, (neg ? SHIFT_SAR : SHIFT_SHR) + rexw,
1814                           dest, rexw ? 63 : 31);
1815            return;
1816        }
1817        break;
1818
1819    default:
1820        break;
1821    }
1822
1823    /*
1824     * If dest does not overlap the inputs, clearing it first is preferred.
1825     * The XOR breaks any false dependency for the low-byte write to dest,
1826     * and is also one byte smaller than MOVZBL.
1827     */
1828    cleared = false;
1829    if (dest != arg1 && (const_arg2 || dest != arg2)) {
1830        tgen_arithr(s, ARITH_XOR, dest, dest);
1831        cleared = true;
1832    }
1833
1834    jcc = tcg_out_cmp(s, cond, arg1, arg2, const_arg2, cmp_rexw);
1835    tcg_out_modrm(s, OPC_SETCC | jcc, 0, dest);
1836
1837    if (!cleared) {
1838        tcg_out_ext8u(s, dest, dest);
1839    }
1840    if (neg) {
1841        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NEG, dest);
1842    }
1843}
1844
1845#if TCG_TARGET_REG_BITS == 32
1846static void tcg_out_setcond2(TCGContext *s, const TCGArg *args,
1847                             const int *const_args)
1848{
1849    TCGArg new_args[6];
1850    TCGLabel *label_true, *label_over;
1851
1852    memcpy(new_args, args+1, 5*sizeof(TCGArg));
1853
1854    if (args[0] == args[1] || args[0] == args[2]
1855        || (!const_args[3] && args[0] == args[3])
1856        || (!const_args[4] && args[0] == args[4])) {
1857        /* When the destination overlaps with one of the argument
1858           registers, don't do anything tricky.  */
1859        label_true = gen_new_label();
1860        label_over = gen_new_label();
1861
1862        new_args[5] = label_arg(label_true);
1863        tcg_out_brcond2(s, new_args, const_args+1, 1);
1864
1865        tcg_out_movi(s, TCG_TYPE_I32, args[0], 0);
1866        tcg_out_jxx(s, JCC_JMP, label_over, 1);
1867        tcg_out_label(s, label_true);
1868
1869        tcg_out_movi(s, TCG_TYPE_I32, args[0], 1);
1870        tcg_out_label(s, label_over);
1871    } else {
1872        /* When the destination does not overlap one of the arguments,
1873           clear the destination first, jump if cond false, and emit an
1874           increment in the true case.  This results in smaller code.  */
1875
1876        tcg_out_movi(s, TCG_TYPE_I32, args[0], 0);
1877
1878        label_over = gen_new_label();
1879        new_args[4] = tcg_invert_cond(new_args[4]);
1880        new_args[5] = label_arg(label_over);
1881        tcg_out_brcond2(s, new_args, const_args+1, 1);
1882
1883        tgen_arithi(s, ARITH_ADD, args[0], 1, 0);
1884        tcg_out_label(s, label_over);
1885    }
1886}
1887#endif
1888
1889static void tcg_out_cmov(TCGContext *s, int jcc, int rexw,
1890                         TCGReg dest, TCGReg v1)
1891{
1892    tcg_out_modrm(s, OPC_CMOVCC | jcc | rexw, dest, v1);
1893}
1894
1895static void tcg_out_movcond(TCGContext *s, int rexw, TCGCond cond,
1896                            TCGReg dest, TCGReg c1, TCGArg c2, int const_c2,
1897                            TCGReg v1)
1898{
1899    int jcc = tcg_out_cmp(s, cond, c1, c2, const_c2, rexw);
1900    tcg_out_cmov(s, jcc, rexw, dest, v1);
1901}
1902
1903static void tcg_out_ctz(TCGContext *s, int rexw, TCGReg dest, TCGReg arg1,
1904                        TCGArg arg2, bool const_a2)
1905{
1906    if (have_bmi1) {
1907        tcg_out_modrm(s, OPC_TZCNT + rexw, dest, arg1);
1908        if (const_a2) {
1909            tcg_debug_assert(arg2 == (rexw ? 64 : 32));
1910        } else {
1911            tcg_debug_assert(dest != arg2);
1912            tcg_out_cmov(s, JCC_JB, rexw, dest, arg2);
1913        }
1914    } else {
1915        tcg_debug_assert(dest != arg2);
1916        tcg_out_modrm(s, OPC_BSF + rexw, dest, arg1);
1917        tcg_out_cmov(s, JCC_JE, rexw, dest, arg2);
1918    }
1919}
1920
1921static void tcg_out_clz(TCGContext *s, int rexw, TCGReg dest, TCGReg arg1,
1922                        TCGArg arg2, bool const_a2)
1923{
1924    if (have_lzcnt) {
1925        tcg_out_modrm(s, OPC_LZCNT + rexw, dest, arg1);
1926        if (const_a2) {
1927            tcg_debug_assert(arg2 == (rexw ? 64 : 32));
1928        } else {
1929            tcg_debug_assert(dest != arg2);
1930            tcg_out_cmov(s, JCC_JB, rexw, dest, arg2);
1931        }
1932    } else {
1933        tcg_debug_assert(!const_a2);
1934        tcg_debug_assert(dest != arg1);
1935        tcg_debug_assert(dest != arg2);
1936
1937        /* Recall that the output of BSR is the index not the count.  */
1938        tcg_out_modrm(s, OPC_BSR + rexw, dest, arg1);
1939        tgen_arithi(s, ARITH_XOR + rexw, dest, rexw ? 63 : 31, 0);
1940
1941        /* Since we have destroyed the flags from BSR, we have to re-test.  */
1942        int jcc = tcg_out_cmp(s, TCG_COND_EQ, arg1, 0, 1, rexw);
1943        tcg_out_cmov(s, jcc, rexw, dest, arg2);
1944    }
1945}
1946
1947static void tcg_out_branch(TCGContext *s, int call, const tcg_insn_unit *dest)
1948{
1949    intptr_t disp = tcg_pcrel_diff(s, dest) - 5;
1950
1951    if (disp == (int32_t)disp) {
1952        tcg_out_opc(s, call ? OPC_CALL_Jz : OPC_JMP_long, 0, 0, 0);
1953        tcg_out32(s, disp);
1954    } else {
1955        /* rip-relative addressing into the constant pool.
1956           This is 6 + 8 = 14 bytes, as compared to using an
1957           immediate load 10 + 6 = 16 bytes, plus we may
1958           be able to re-use the pool constant for more calls.  */
1959        tcg_out_opc(s, OPC_GRP5, 0, 0, 0);
1960        tcg_out8(s, (call ? EXT5_CALLN_Ev : EXT5_JMPN_Ev) << 3 | 5);
1961        new_pool_label(s, (uintptr_t)dest, R_386_PC32, s->code_ptr, -4);
1962        tcg_out32(s, 0);
1963    }
1964}
1965
1966static void tcg_out_call(TCGContext *s, const tcg_insn_unit *dest,
1967                         const TCGHelperInfo *info)
1968{
1969    tcg_out_branch(s, 1, dest);
1970
1971#ifndef _WIN32
1972    if (TCG_TARGET_REG_BITS == 32 && info->out_kind == TCG_CALL_RET_BY_REF) {
1973        /*
1974         * The sysv i386 abi for struct return places a reference as the
1975         * first argument of the stack, and pops that argument with the
1976         * return statement.  Since we want to retain the aligned stack
1977         * pointer for the callee, we do not want to actually push that
1978         * argument before the call but rely on the normal store to the
1979         * stack slot.  But we do need to compensate for the pop in order
1980         * to reset our correct stack pointer value.
1981         * Pushing a garbage value back onto the stack is quickest.
1982         */
1983        tcg_out_push(s, TCG_REG_EAX);
1984    }
1985#endif
1986}
1987
1988static void tcg_out_jmp(TCGContext *s, const tcg_insn_unit *dest)
1989{
1990    tcg_out_branch(s, 0, dest);
1991}
1992
1993static void tcg_out_nopn(TCGContext *s, int n)
1994{
1995    int i;
1996    /* Emit 1 or 2 operand size prefixes for the standard one byte nop,
1997     * "xchg %eax,%eax", forming "xchg %ax,%ax". All cores accept the
1998     * duplicate prefix, and all of the interesting recent cores can
1999     * decode and discard the duplicates in a single cycle.
2000     */
2001    tcg_debug_assert(n >= 1);
2002    for (i = 1; i < n; ++i) {
2003        tcg_out8(s, 0x66);
2004    }
2005    tcg_out8(s, 0x90);
2006}
2007
2008typedef struct {
2009    TCGReg base;
2010    int index;
2011    int ofs;
2012    int seg;
2013    TCGAtomAlign aa;
2014} HostAddress;
2015
2016bool tcg_target_has_memory_bswap(MemOp memop)
2017{
2018    TCGAtomAlign aa;
2019
2020    if (!have_movbe) {
2021        return false;
2022    }
2023    if ((memop & MO_SIZE) < MO_128) {
2024        return true;
2025    }
2026
2027    /*
2028     * Reject 16-byte memop with 16-byte atomicity, i.e. VMOVDQA,
2029     * but do allow a pair of 64-bit operations, i.e. MOVBEQ.
2030     */
2031    aa = atom_and_align_for_opc(tcg_ctx, memop, MO_ATOM_IFALIGN, true);
2032    return aa.atom < MO_128;
2033}
2034
2035/*
2036 * Because i686 has no register parameters and because x86_64 has xchg
2037 * to handle addr/data register overlap, we have placed all input arguments
2038 * before we need might need a scratch reg.
2039 *
2040 * Even then, a scratch is only needed for l->raddr.  Rather than expose
2041 * a general-purpose scratch when we don't actually know it's available,
2042 * use the ra_gen hook to load into RAX if needed.
2043 */
2044#if TCG_TARGET_REG_BITS == 64
2045static TCGReg ldst_ra_gen(TCGContext *s, const TCGLabelQemuLdst *l, int arg)
2046{
2047    if (arg < 0) {
2048        arg = TCG_REG_RAX;
2049    }
2050    tcg_out_movi(s, TCG_TYPE_PTR, arg, (uintptr_t)l->raddr);
2051    return arg;
2052}
2053static const TCGLdstHelperParam ldst_helper_param = {
2054    .ra_gen = ldst_ra_gen
2055};
2056#else
2057static const TCGLdstHelperParam ldst_helper_param = { };
2058#endif
2059
2060static void tcg_out_vec_to_pair(TCGContext *s, TCGType type,
2061                                TCGReg l, TCGReg h, TCGReg v)
2062{
2063    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2064
2065    /* vpmov{d,q} %v, %l */
2066    tcg_out_vex_modrm(s, OPC_MOVD_EyVy + rexw, v, 0, l);
2067    /* vpextr{d,q} $1, %v, %h */
2068    tcg_out_vex_modrm(s, OPC_PEXTRD + rexw, v, 0, h);
2069    tcg_out8(s, 1);
2070}
2071
2072static void tcg_out_pair_to_vec(TCGContext *s, TCGType type,
2073                                TCGReg v, TCGReg l, TCGReg h)
2074{
2075    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2076
2077    /* vmov{d,q} %l, %v */
2078    tcg_out_vex_modrm(s, OPC_MOVD_VyEy + rexw, v, 0, l);
2079    /* vpinsr{d,q} $1, %h, %v, %v */
2080    tcg_out_vex_modrm(s, OPC_PINSRD + rexw, v, v, h);
2081    tcg_out8(s, 1);
2082}
2083
2084/*
2085 * Generate code for the slow path for a load at the end of block
2086 */
2087static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
2088{
2089    MemOp opc = get_memop(l->oi);
2090    tcg_insn_unit **label_ptr = &l->label_ptr[0];
2091
2092    /* resolve label address */
2093    tcg_patch32(label_ptr[0], s->code_ptr - label_ptr[0] - 4);
2094    if (label_ptr[1]) {
2095        tcg_patch32(label_ptr[1], s->code_ptr - label_ptr[1] - 4);
2096    }
2097
2098    tcg_out_ld_helper_args(s, l, &ldst_helper_param);
2099    tcg_out_branch(s, 1, qemu_ld_helpers[opc & MO_SIZE]);
2100    tcg_out_ld_helper_ret(s, l, false, &ldst_helper_param);
2101
2102    tcg_out_jmp(s, l->raddr);
2103    return true;
2104}
2105
2106/*
2107 * Generate code for the slow path for a store at the end of block
2108 */
2109static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
2110{
2111    MemOp opc = get_memop(l->oi);
2112    tcg_insn_unit **label_ptr = &l->label_ptr[0];
2113
2114    /* resolve label address */
2115    tcg_patch32(label_ptr[0], s->code_ptr - label_ptr[0] - 4);
2116    if (label_ptr[1]) {
2117        tcg_patch32(label_ptr[1], s->code_ptr - label_ptr[1] - 4);
2118    }
2119
2120    tcg_out_st_helper_args(s, l, &ldst_helper_param);
2121    tcg_out_branch(s, 1, qemu_st_helpers[opc & MO_SIZE]);
2122
2123    tcg_out_jmp(s, l->raddr);
2124    return true;
2125}
2126
2127#ifdef CONFIG_USER_ONLY
2128static HostAddress x86_guest_base = {
2129    .index = -1
2130};
2131
2132#if defined(__x86_64__) && defined(__linux__)
2133# include <asm/prctl.h>
2134# include <sys/prctl.h>
2135int arch_prctl(int code, unsigned long addr);
2136static inline int setup_guest_base_seg(void)
2137{
2138    if (arch_prctl(ARCH_SET_GS, guest_base) == 0) {
2139        return P_GS;
2140    }
2141    return 0;
2142}
2143#define setup_guest_base_seg  setup_guest_base_seg
2144#elif defined(__x86_64__) && \
2145      (defined (__FreeBSD__) || defined (__FreeBSD_kernel__))
2146# include <machine/sysarch.h>
2147static inline int setup_guest_base_seg(void)
2148{
2149    if (sysarch(AMD64_SET_GSBASE, &guest_base) == 0) {
2150        return P_GS;
2151    }
2152    return 0;
2153}
2154#define setup_guest_base_seg  setup_guest_base_seg
2155#endif
2156#else
2157# define x86_guest_base (*(HostAddress *)({ qemu_build_not_reached(); NULL; }))
2158#endif /* CONFIG_USER_ONLY */
2159#ifndef setup_guest_base_seg
2160# define setup_guest_base_seg()  0
2161#endif
2162
2163#define MIN_TLB_MASK_TABLE_OFS  INT_MIN
2164
2165/*
2166 * For softmmu, perform the TLB load and compare.
2167 * For useronly, perform any required alignment tests.
2168 * In both cases, return a TCGLabelQemuLdst structure if the slow path
2169 * is required and fill in @h with the host address for the fast path.
2170 */
2171static TCGLabelQemuLdst *prepare_host_addr(TCGContext *s, HostAddress *h,
2172                                           TCGReg addrlo, TCGReg addrhi,
2173                                           MemOpIdx oi, bool is_ld)
2174{
2175    TCGLabelQemuLdst *ldst = NULL;
2176    MemOp opc = get_memop(oi);
2177    MemOp s_bits = opc & MO_SIZE;
2178    unsigned a_mask;
2179
2180    if (tcg_use_softmmu) {
2181        h->index = TCG_REG_L0;
2182        h->ofs = 0;
2183        h->seg = 0;
2184    } else {
2185        *h = x86_guest_base;
2186    }
2187    h->base = addrlo;
2188    h->aa = atom_and_align_for_opc(s, opc, MO_ATOM_IFALIGN, s_bits == MO_128);
2189    a_mask = (1 << h->aa.align) - 1;
2190
2191    if (tcg_use_softmmu) {
2192        int cmp_ofs = is_ld ? offsetof(CPUTLBEntry, addr_read)
2193                            : offsetof(CPUTLBEntry, addr_write);
2194        TCGType ttype = TCG_TYPE_I32;
2195        TCGType tlbtype = TCG_TYPE_I32;
2196        int trexw = 0, hrexw = 0, tlbrexw = 0;
2197        unsigned mem_index = get_mmuidx(oi);
2198        unsigned s_mask = (1 << s_bits) - 1;
2199        int fast_ofs = tlb_mask_table_ofs(s, mem_index);
2200        int tlb_mask;
2201
2202        ldst = new_ldst_label(s);
2203        ldst->is_ld = is_ld;
2204        ldst->oi = oi;
2205        ldst->addrlo_reg = addrlo;
2206        ldst->addrhi_reg = addrhi;
2207
2208        if (TCG_TARGET_REG_BITS == 64) {
2209            ttype = s->addr_type;
2210            trexw = (ttype == TCG_TYPE_I32 ? 0 : P_REXW);
2211            if (TCG_TYPE_PTR == TCG_TYPE_I64) {
2212                hrexw = P_REXW;
2213                if (s->page_bits + s->tlb_dyn_max_bits > 32) {
2214                    tlbtype = TCG_TYPE_I64;
2215                    tlbrexw = P_REXW;
2216                }
2217            }
2218        }
2219
2220        tcg_out_mov(s, tlbtype, TCG_REG_L0, addrlo);
2221        tcg_out_shifti(s, SHIFT_SHR + tlbrexw, TCG_REG_L0,
2222                       s->page_bits - CPU_TLB_ENTRY_BITS);
2223
2224        tcg_out_modrm_offset(s, OPC_AND_GvEv + trexw, TCG_REG_L0, TCG_AREG0,
2225                             fast_ofs + offsetof(CPUTLBDescFast, mask));
2226
2227        tcg_out_modrm_offset(s, OPC_ADD_GvEv + hrexw, TCG_REG_L0, TCG_AREG0,
2228                             fast_ofs + offsetof(CPUTLBDescFast, table));
2229
2230        /*
2231         * If the required alignment is at least as large as the access,
2232         * simply copy the address and mask.  For lesser alignments,
2233         * check that we don't cross pages for the complete access.
2234         */
2235        if (a_mask >= s_mask) {
2236            tcg_out_mov(s, ttype, TCG_REG_L1, addrlo);
2237        } else {
2238            tcg_out_modrm_offset(s, OPC_LEA + trexw, TCG_REG_L1,
2239                                 addrlo, s_mask - a_mask);
2240        }
2241        tlb_mask = s->page_mask | a_mask;
2242        tgen_arithi(s, ARITH_AND + trexw, TCG_REG_L1, tlb_mask, 0);
2243
2244        /* cmp 0(TCG_REG_L0), TCG_REG_L1 */
2245        tcg_out_modrm_offset(s, OPC_CMP_GvEv + trexw,
2246                             TCG_REG_L1, TCG_REG_L0, cmp_ofs);
2247
2248        /* jne slow_path */
2249        tcg_out_opc(s, OPC_JCC_long + JCC_JNE, 0, 0, 0);
2250        ldst->label_ptr[0] = s->code_ptr;
2251        s->code_ptr += 4;
2252
2253        if (TCG_TARGET_REG_BITS == 32 && s->addr_type == TCG_TYPE_I64) {
2254            /* cmp 4(TCG_REG_L0), addrhi */
2255            tcg_out_modrm_offset(s, OPC_CMP_GvEv, addrhi,
2256                                 TCG_REG_L0, cmp_ofs + 4);
2257
2258            /* jne slow_path */
2259            tcg_out_opc(s, OPC_JCC_long + JCC_JNE, 0, 0, 0);
2260            ldst->label_ptr[1] = s->code_ptr;
2261            s->code_ptr += 4;
2262        }
2263
2264        /* TLB Hit.  */
2265        tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_L0, TCG_REG_L0,
2266                   offsetof(CPUTLBEntry, addend));
2267    } else if (a_mask) {
2268        int jcc;
2269
2270        ldst = new_ldst_label(s);
2271        ldst->is_ld = is_ld;
2272        ldst->oi = oi;
2273        ldst->addrlo_reg = addrlo;
2274        ldst->addrhi_reg = addrhi;
2275
2276        /* jne slow_path */
2277        jcc = tcg_out_cmp(s, TCG_COND_TSTNE, addrlo, a_mask, true, false);
2278        tcg_out_opc(s, OPC_JCC_long + jcc, 0, 0, 0);
2279        ldst->label_ptr[0] = s->code_ptr;
2280        s->code_ptr += 4;
2281    }
2282
2283    return ldst;
2284}
2285
2286static void tcg_out_qemu_ld_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
2287                                   HostAddress h, TCGType type, MemOp memop)
2288{
2289    bool use_movbe = false;
2290    int rexw = (type == TCG_TYPE_I32 ? 0 : P_REXW);
2291    int movop = OPC_MOVL_GvEv;
2292
2293    /* Do big-endian loads with movbe.  */
2294    if (memop & MO_BSWAP) {
2295        tcg_debug_assert(have_movbe);
2296        use_movbe = true;
2297        movop = OPC_MOVBE_GyMy;
2298    }
2299
2300    switch (memop & MO_SSIZE) {
2301    case MO_UB:
2302        tcg_out_modrm_sib_offset(s, OPC_MOVZBL + h.seg, datalo,
2303                                 h.base, h.index, 0, h.ofs);
2304        break;
2305    case MO_SB:
2306        tcg_out_modrm_sib_offset(s, OPC_MOVSBL + rexw + h.seg, datalo,
2307                                 h.base, h.index, 0, h.ofs);
2308        break;
2309    case MO_UW:
2310        if (use_movbe) {
2311            /* There is no extending movbe; only low 16-bits are modified.  */
2312            if (datalo != h.base && datalo != h.index) {
2313                /* XOR breaks dependency chains.  */
2314                tgen_arithr(s, ARITH_XOR, datalo, datalo);
2315                tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + P_DATA16 + h.seg,
2316                                         datalo, h.base, h.index, 0, h.ofs);
2317            } else {
2318                tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + P_DATA16 + h.seg,
2319                                         datalo, h.base, h.index, 0, h.ofs);
2320                tcg_out_ext16u(s, datalo, datalo);
2321            }
2322        } else {
2323            tcg_out_modrm_sib_offset(s, OPC_MOVZWL + h.seg, datalo,
2324                                     h.base, h.index, 0, h.ofs);
2325        }
2326        break;
2327    case MO_SW:
2328        if (use_movbe) {
2329            tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + P_DATA16 + h.seg,
2330                                     datalo, h.base, h.index, 0, h.ofs);
2331            tcg_out_ext16s(s, type, datalo, datalo);
2332        } else {
2333            tcg_out_modrm_sib_offset(s, OPC_MOVSWL + rexw + h.seg,
2334                                     datalo, h.base, h.index, 0, h.ofs);
2335        }
2336        break;
2337    case MO_UL:
2338        tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2339                                 h.base, h.index, 0, h.ofs);
2340        break;
2341#if TCG_TARGET_REG_BITS == 64
2342    case MO_SL:
2343        if (use_movbe) {
2344            tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + h.seg, datalo,
2345                                     h.base, h.index, 0, h.ofs);
2346            tcg_out_ext32s(s, datalo, datalo);
2347        } else {
2348            tcg_out_modrm_sib_offset(s, OPC_MOVSLQ + h.seg, datalo,
2349                                     h.base, h.index, 0, h.ofs);
2350        }
2351        break;
2352#endif
2353    case MO_UQ:
2354        if (TCG_TARGET_REG_BITS == 64) {
2355            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2356                                     h.base, h.index, 0, h.ofs);
2357            break;
2358        }
2359        if (use_movbe) {
2360            TCGReg t = datalo;
2361            datalo = datahi;
2362            datahi = t;
2363        }
2364        if (h.base == datalo || h.index == datalo) {
2365            tcg_out_modrm_sib_offset(s, OPC_LEA, datahi,
2366                                     h.base, h.index, 0, h.ofs);
2367            tcg_out_modrm_offset(s, movop + h.seg, datalo, datahi, 0);
2368            tcg_out_modrm_offset(s, movop + h.seg, datahi, datahi, 4);
2369        } else {
2370            tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2371                                     h.base, h.index, 0, h.ofs);
2372            tcg_out_modrm_sib_offset(s, movop + h.seg, datahi,
2373                                     h.base, h.index, 0, h.ofs + 4);
2374        }
2375        break;
2376
2377    case MO_128:
2378        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
2379
2380        /*
2381         * Without 16-byte atomicity, use integer regs.
2382         * That is where we want the data, and it allows bswaps.
2383         */
2384        if (h.aa.atom < MO_128) {
2385            if (use_movbe) {
2386                TCGReg t = datalo;
2387                datalo = datahi;
2388                datahi = t;
2389            }
2390            if (h.base == datalo || h.index == datalo) {
2391                tcg_out_modrm_sib_offset(s, OPC_LEA + P_REXW, datahi,
2392                                         h.base, h.index, 0, h.ofs);
2393                tcg_out_modrm_offset(s, movop + P_REXW + h.seg,
2394                                     datalo, datahi, 0);
2395                tcg_out_modrm_offset(s, movop + P_REXW + h.seg,
2396                                     datahi, datahi, 8);
2397            } else {
2398                tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2399                                         h.base, h.index, 0, h.ofs);
2400                tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datahi,
2401                                         h.base, h.index, 0, h.ofs + 8);
2402            }
2403            break;
2404        }
2405
2406        /*
2407         * With 16-byte atomicity, a vector load is required.
2408         * If we already have 16-byte alignment, then VMOVDQA always works.
2409         * Else if VMOVDQU has atomicity with dynamic alignment, use that.
2410         * Else use we require a runtime test for alignment for VMOVDQA;
2411         * use VMOVDQU on the unaligned nonatomic path for simplicity.
2412         */
2413        if (h.aa.align >= MO_128) {
2414            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_VxWx + h.seg,
2415                                         TCG_TMP_VEC, 0,
2416                                         h.base, h.index, 0, h.ofs);
2417        } else if (cpuinfo & CPUINFO_ATOMIC_VMOVDQU) {
2418            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_VxWx + h.seg,
2419                                         TCG_TMP_VEC, 0,
2420                                         h.base, h.index, 0, h.ofs);
2421        } else {
2422            TCGLabel *l1 = gen_new_label();
2423            TCGLabel *l2 = gen_new_label();
2424            int jcc;
2425
2426            jcc = tcg_out_cmp(s, TCG_COND_TSTNE, h.base, 15, true, false);
2427            tcg_out_jxx(s, jcc, l1, true);
2428
2429            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_VxWx + h.seg,
2430                                         TCG_TMP_VEC, 0,
2431                                         h.base, h.index, 0, h.ofs);
2432            tcg_out_jxx(s, JCC_JMP, l2, true);
2433
2434            tcg_out_label(s, l1);
2435            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_VxWx + h.seg,
2436                                         TCG_TMP_VEC, 0,
2437                                         h.base, h.index, 0, h.ofs);
2438            tcg_out_label(s, l2);
2439        }
2440        tcg_out_vec_to_pair(s, TCG_TYPE_I64, datalo, datahi, TCG_TMP_VEC);
2441        break;
2442
2443    default:
2444        g_assert_not_reached();
2445    }
2446}
2447
2448static void tcg_out_qemu_ld(TCGContext *s, TCGReg datalo, TCGReg datahi,
2449                            TCGReg addrlo, TCGReg addrhi,
2450                            MemOpIdx oi, TCGType data_type)
2451{
2452    TCGLabelQemuLdst *ldst;
2453    HostAddress h;
2454
2455    ldst = prepare_host_addr(s, &h, addrlo, addrhi, oi, true);
2456    tcg_out_qemu_ld_direct(s, datalo, datahi, h, data_type, get_memop(oi));
2457
2458    if (ldst) {
2459        ldst->type = data_type;
2460        ldst->datalo_reg = datalo;
2461        ldst->datahi_reg = datahi;
2462        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
2463    }
2464}
2465
2466static void tcg_out_qemu_st_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
2467                                   HostAddress h, MemOp memop)
2468{
2469    bool use_movbe = false;
2470    int movop = OPC_MOVL_EvGv;
2471
2472    /*
2473     * Do big-endian stores with movbe or system-mode.
2474     * User-only without movbe will have its swapping done generically.
2475     */
2476    if (memop & MO_BSWAP) {
2477        tcg_debug_assert(have_movbe);
2478        use_movbe = true;
2479        movop = OPC_MOVBE_MyGy;
2480    }
2481
2482    switch (memop & MO_SIZE) {
2483    case MO_8:
2484        /* This is handled with constraints on INDEX_op_qemu_st8_i32. */
2485        tcg_debug_assert(TCG_TARGET_REG_BITS == 64 || datalo < 4);
2486        tcg_out_modrm_sib_offset(s, OPC_MOVB_EvGv + P_REXB_R + h.seg,
2487                                 datalo, h.base, h.index, 0, h.ofs);
2488        break;
2489    case MO_16:
2490        tcg_out_modrm_sib_offset(s, movop + P_DATA16 + h.seg, datalo,
2491                                 h.base, h.index, 0, h.ofs);
2492        break;
2493    case MO_32:
2494        tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2495                                 h.base, h.index, 0, h.ofs);
2496        break;
2497    case MO_64:
2498        if (TCG_TARGET_REG_BITS == 64) {
2499            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2500                                     h.base, h.index, 0, h.ofs);
2501        } else {
2502            if (use_movbe) {
2503                TCGReg t = datalo;
2504                datalo = datahi;
2505                datahi = t;
2506            }
2507            tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2508                                     h.base, h.index, 0, h.ofs);
2509            tcg_out_modrm_sib_offset(s, movop + h.seg, datahi,
2510                                     h.base, h.index, 0, h.ofs + 4);
2511        }
2512        break;
2513
2514    case MO_128:
2515        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
2516
2517        /*
2518         * Without 16-byte atomicity, use integer regs.
2519         * That is where we have the data, and it allows bswaps.
2520         */
2521        if (h.aa.atom < MO_128) {
2522            if (use_movbe) {
2523                TCGReg t = datalo;
2524                datalo = datahi;
2525                datahi = t;
2526            }
2527            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2528                                     h.base, h.index, 0, h.ofs);
2529            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datahi,
2530                                     h.base, h.index, 0, h.ofs + 8);
2531            break;
2532        }
2533
2534        /*
2535         * With 16-byte atomicity, a vector store is required.
2536         * If we already have 16-byte alignment, then VMOVDQA always works.
2537         * Else if VMOVDQU has atomicity with dynamic alignment, use that.
2538         * Else use we require a runtime test for alignment for VMOVDQA;
2539         * use VMOVDQU on the unaligned nonatomic path for simplicity.
2540         */
2541        tcg_out_pair_to_vec(s, TCG_TYPE_I64, TCG_TMP_VEC, datalo, datahi);
2542        if (h.aa.align >= MO_128) {
2543            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_WxVx + h.seg,
2544                                         TCG_TMP_VEC, 0,
2545                                         h.base, h.index, 0, h.ofs);
2546        } else if (cpuinfo & CPUINFO_ATOMIC_VMOVDQU) {
2547            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_WxVx + h.seg,
2548                                         TCG_TMP_VEC, 0,
2549                                         h.base, h.index, 0, h.ofs);
2550        } else {
2551            TCGLabel *l1 = gen_new_label();
2552            TCGLabel *l2 = gen_new_label();
2553            int jcc;
2554
2555            jcc = tcg_out_cmp(s, TCG_COND_TSTNE, h.base, 15, true, false);
2556            tcg_out_jxx(s, jcc, l1, true);
2557
2558            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_WxVx + h.seg,
2559                                         TCG_TMP_VEC, 0,
2560                                         h.base, h.index, 0, h.ofs);
2561            tcg_out_jxx(s, JCC_JMP, l2, true);
2562
2563            tcg_out_label(s, l1);
2564            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_WxVx + h.seg,
2565                                         TCG_TMP_VEC, 0,
2566                                         h.base, h.index, 0, h.ofs);
2567            tcg_out_label(s, l2);
2568        }
2569        break;
2570
2571    default:
2572        g_assert_not_reached();
2573    }
2574}
2575
2576static void tcg_out_qemu_st(TCGContext *s, TCGReg datalo, TCGReg datahi,
2577                            TCGReg addrlo, TCGReg addrhi,
2578                            MemOpIdx oi, TCGType data_type)
2579{
2580    TCGLabelQemuLdst *ldst;
2581    HostAddress h;
2582
2583    ldst = prepare_host_addr(s, &h, addrlo, addrhi, oi, false);
2584    tcg_out_qemu_st_direct(s, datalo, datahi, h, get_memop(oi));
2585
2586    if (ldst) {
2587        ldst->type = data_type;
2588        ldst->datalo_reg = datalo;
2589        ldst->datahi_reg = datahi;
2590        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
2591    }
2592}
2593
2594static void tcg_out_exit_tb(TCGContext *s, uintptr_t a0)
2595{
2596    /* Reuse the zeroing that exists for goto_ptr.  */
2597    if (a0 == 0) {
2598        tcg_out_jmp(s, tcg_code_gen_epilogue);
2599    } else {
2600        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_EAX, a0);
2601        tcg_out_jmp(s, tb_ret_addr);
2602    }
2603}
2604
2605static void tcg_out_goto_tb(TCGContext *s, int which)
2606{
2607    /*
2608     * Jump displacement must be aligned for atomic patching;
2609     * see if we need to add extra nops before jump
2610     */
2611    int gap = QEMU_ALIGN_PTR_UP(s->code_ptr + 1, 4) - s->code_ptr;
2612    if (gap != 1) {
2613        tcg_out_nopn(s, gap - 1);
2614    }
2615    tcg_out8(s, OPC_JMP_long); /* jmp im */
2616    set_jmp_insn_offset(s, which);
2617    tcg_out32(s, 0);
2618    set_jmp_reset_offset(s, which);
2619}
2620
2621void tb_target_set_jmp_target(const TranslationBlock *tb, int n,
2622                              uintptr_t jmp_rx, uintptr_t jmp_rw)
2623{
2624    /* patch the branch destination */
2625    uintptr_t addr = tb->jmp_target_addr[n];
2626    qatomic_set((int32_t *)jmp_rw, addr - (jmp_rx + 4));
2627    /* no need to flush icache explicitly */
2628}
2629
2630static void tcg_out_op(TCGContext *s, TCGOpcode opc, TCGType type,
2631                       const TCGArg args[TCG_MAX_OP_ARGS],
2632                       const int const_args[TCG_MAX_OP_ARGS])
2633{
2634    TCGArg a0, a1, a2;
2635    int c, const_a2, vexop, rexw;
2636
2637#if TCG_TARGET_REG_BITS == 64
2638# define OP_32_64(x) \
2639        case glue(glue(INDEX_op_, x), _i64): \
2640        case glue(glue(INDEX_op_, x), _i32)
2641#else
2642# define OP_32_64(x) \
2643        case glue(glue(INDEX_op_, x), _i32)
2644#endif
2645
2646    /* Hoist the loads of the most common arguments.  */
2647    a0 = args[0];
2648    a1 = args[1];
2649    a2 = args[2];
2650    const_a2 = const_args[2];
2651    rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2652
2653    switch (opc) {
2654    case INDEX_op_goto_ptr:
2655        /* jmp to the given host address (could be epilogue) */
2656        tcg_out_modrm(s, OPC_GRP5, EXT5_JMPN_Ev, a0);
2657        break;
2658    case INDEX_op_br:
2659        tcg_out_jxx(s, JCC_JMP, arg_label(a0), 0);
2660        break;
2661    OP_32_64(ld8u):
2662        /* Note that we can ignore REXW for the zero-extend to 64-bit.  */
2663        tcg_out_modrm_offset(s, OPC_MOVZBL, a0, a1, a2);
2664        break;
2665    OP_32_64(ld8s):
2666        tcg_out_modrm_offset(s, OPC_MOVSBL + rexw, a0, a1, a2);
2667        break;
2668    OP_32_64(ld16u):
2669        /* Note that we can ignore REXW for the zero-extend to 64-bit.  */
2670        tcg_out_modrm_offset(s, OPC_MOVZWL, a0, a1, a2);
2671        break;
2672    OP_32_64(ld16s):
2673        tcg_out_modrm_offset(s, OPC_MOVSWL + rexw, a0, a1, a2);
2674        break;
2675#if TCG_TARGET_REG_BITS == 64
2676    case INDEX_op_ld32u_i64:
2677#endif
2678    case INDEX_op_ld_i32:
2679        tcg_out_ld(s, TCG_TYPE_I32, a0, a1, a2);
2680        break;
2681
2682    OP_32_64(st8):
2683        if (const_args[0]) {
2684            tcg_out_modrm_offset(s, OPC_MOVB_EvIz, 0, a1, a2);
2685            tcg_out8(s, a0);
2686        } else {
2687            tcg_out_modrm_offset(s, OPC_MOVB_EvGv | P_REXB_R, a0, a1, a2);
2688        }
2689        break;
2690    OP_32_64(st16):
2691        if (const_args[0]) {
2692            tcg_out_modrm_offset(s, OPC_MOVL_EvIz | P_DATA16, 0, a1, a2);
2693            tcg_out16(s, a0);
2694        } else {
2695            tcg_out_modrm_offset(s, OPC_MOVL_EvGv | P_DATA16, a0, a1, a2);
2696        }
2697        break;
2698#if TCG_TARGET_REG_BITS == 64
2699    case INDEX_op_st32_i64:
2700#endif
2701    case INDEX_op_st_i32:
2702        if (const_args[0]) {
2703            tcg_out_modrm_offset(s, OPC_MOVL_EvIz, 0, a1, a2);
2704            tcg_out32(s, a0);
2705        } else {
2706            tcg_out_st(s, TCG_TYPE_I32, a0, a1, a2);
2707        }
2708        break;
2709
2710    OP_32_64(add):
2711        /* For 3-operand addition, use LEA.  */
2712        if (a0 != a1) {
2713            TCGArg c3 = 0;
2714            if (const_a2) {
2715                c3 = a2, a2 = -1;
2716            } else if (a0 == a2) {
2717                /* Watch out for dest = src + dest, since we've removed
2718                   the matching constraint on the add.  */
2719                tgen_arithr(s, ARITH_ADD + rexw, a0, a1);
2720                break;
2721            }
2722
2723            tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, a1, a2, 0, c3);
2724            break;
2725        }
2726        c = ARITH_ADD;
2727        goto gen_arith;
2728    OP_32_64(sub):
2729        c = ARITH_SUB;
2730        goto gen_arith;
2731    OP_32_64(and):
2732        c = ARITH_AND;
2733        goto gen_arith;
2734    OP_32_64(or):
2735        c = ARITH_OR;
2736        goto gen_arith;
2737    OP_32_64(xor):
2738        c = ARITH_XOR;
2739        goto gen_arith;
2740    gen_arith:
2741        if (const_a2) {
2742            tgen_arithi(s, c + rexw, a0, a2, 0);
2743        } else {
2744            tgen_arithr(s, c + rexw, a0, a2);
2745        }
2746        break;
2747
2748    OP_32_64(andc):
2749        if (const_a2) {
2750            tcg_out_mov(s, rexw ? TCG_TYPE_I64 : TCG_TYPE_I32, a0, a1);
2751            tgen_arithi(s, ARITH_AND + rexw, a0, ~a2, 0);
2752        } else {
2753            tcg_out_vex_modrm(s, OPC_ANDN + rexw, a0, a2, a1);
2754        }
2755        break;
2756
2757    OP_32_64(mul):
2758        if (const_a2) {
2759            int32_t val;
2760            val = a2;
2761            if (val == (int8_t)val) {
2762                tcg_out_modrm(s, OPC_IMUL_GvEvIb + rexw, a0, a0);
2763                tcg_out8(s, val);
2764            } else {
2765                tcg_out_modrm(s, OPC_IMUL_GvEvIz + rexw, a0, a0);
2766                tcg_out32(s, val);
2767            }
2768        } else {
2769            tcg_out_modrm(s, OPC_IMUL_GvEv + rexw, a0, a2);
2770        }
2771        break;
2772
2773    OP_32_64(div2):
2774        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_IDIV, args[4]);
2775        break;
2776    OP_32_64(divu2):
2777        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_DIV, args[4]);
2778        break;
2779
2780    OP_32_64(shl):
2781        /* For small constant 3-operand shift, use LEA.  */
2782        if (const_a2 && a0 != a1 && (a2 - 1) < 3) {
2783            if (a2 - 1 == 0) {
2784                /* shl $1,a1,a0 -> lea (a1,a1),a0 */
2785                tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, a1, a1, 0, 0);
2786            } else {
2787                /* shl $n,a1,a0 -> lea 0(,a1,n),a0 */
2788                tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, -1, a1, a2, 0);
2789            }
2790            break;
2791        }
2792        c = SHIFT_SHL;
2793        vexop = OPC_SHLX;
2794        goto gen_shift_maybe_vex;
2795    OP_32_64(shr):
2796        c = SHIFT_SHR;
2797        vexop = OPC_SHRX;
2798        goto gen_shift_maybe_vex;
2799    OP_32_64(sar):
2800        c = SHIFT_SAR;
2801        vexop = OPC_SARX;
2802        goto gen_shift_maybe_vex;
2803    OP_32_64(rotl):
2804        c = SHIFT_ROL;
2805        goto gen_shift;
2806    OP_32_64(rotr):
2807        c = SHIFT_ROR;
2808        goto gen_shift;
2809    gen_shift_maybe_vex:
2810        if (have_bmi2) {
2811            if (!const_a2) {
2812                tcg_out_vex_modrm(s, vexop + rexw, a0, a2, a1);
2813                break;
2814            }
2815            tcg_out_mov(s, rexw ? TCG_TYPE_I64 : TCG_TYPE_I32, a0, a1);
2816        }
2817        /* FALLTHRU */
2818    gen_shift:
2819        if (const_a2) {
2820            tcg_out_shifti(s, c + rexw, a0, a2);
2821        } else {
2822            tcg_out_modrm(s, OPC_SHIFT_cl + rexw, c, a0);
2823        }
2824        break;
2825
2826    OP_32_64(ctz):
2827        tcg_out_ctz(s, rexw, args[0], args[1], args[2], const_args[2]);
2828        break;
2829    OP_32_64(clz):
2830        tcg_out_clz(s, rexw, args[0], args[1], args[2], const_args[2]);
2831        break;
2832    OP_32_64(ctpop):
2833        tcg_out_modrm(s, OPC_POPCNT + rexw, a0, a1);
2834        break;
2835
2836    OP_32_64(brcond):
2837        tcg_out_brcond(s, rexw, a2, a0, a1, const_args[1],
2838                       arg_label(args[3]), 0);
2839        break;
2840    OP_32_64(setcond):
2841        tcg_out_setcond(s, rexw, args[3], a0, a1, a2, const_a2, false);
2842        break;
2843    OP_32_64(negsetcond):
2844        tcg_out_setcond(s, rexw, args[3], a0, a1, a2, const_a2, true);
2845        break;
2846    OP_32_64(movcond):
2847        tcg_out_movcond(s, rexw, args[5], a0, a1, a2, const_a2, args[3]);
2848        break;
2849
2850    OP_32_64(bswap16):
2851        if (a2 & TCG_BSWAP_OS) {
2852            /* Output must be sign-extended. */
2853            if (rexw) {
2854                tcg_out_bswap64(s, a0);
2855                tcg_out_shifti(s, SHIFT_SAR + rexw, a0, 48);
2856            } else {
2857                tcg_out_bswap32(s, a0);
2858                tcg_out_shifti(s, SHIFT_SAR, a0, 16);
2859            }
2860        } else if ((a2 & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
2861            /* Output must be zero-extended, but input isn't. */
2862            tcg_out_bswap32(s, a0);
2863            tcg_out_shifti(s, SHIFT_SHR, a0, 16);
2864        } else {
2865            tcg_out_rolw_8(s, a0);
2866        }
2867        break;
2868    OP_32_64(bswap32):
2869        tcg_out_bswap32(s, a0);
2870        if (rexw && (a2 & TCG_BSWAP_OS)) {
2871            tcg_out_ext32s(s, a0, a0);
2872        }
2873        break;
2874
2875    OP_32_64(neg):
2876        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NEG, a0);
2877        break;
2878    OP_32_64(not):
2879        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NOT, a0);
2880        break;
2881
2882    case INDEX_op_qemu_ld_a64_i32:
2883        if (TCG_TARGET_REG_BITS == 32) {
2884            tcg_out_qemu_ld(s, a0, -1, a1, a2, args[3], TCG_TYPE_I32);
2885            break;
2886        }
2887        /* fall through */
2888    case INDEX_op_qemu_ld_a32_i32:
2889        tcg_out_qemu_ld(s, a0, -1, a1, -1, a2, TCG_TYPE_I32);
2890        break;
2891    case INDEX_op_qemu_ld_a32_i64:
2892        if (TCG_TARGET_REG_BITS == 64) {
2893            tcg_out_qemu_ld(s, a0, -1, a1, -1, a2, TCG_TYPE_I64);
2894        } else {
2895            tcg_out_qemu_ld(s, a0, a1, a2, -1, args[3], TCG_TYPE_I64);
2896        }
2897        break;
2898    case INDEX_op_qemu_ld_a64_i64:
2899        if (TCG_TARGET_REG_BITS == 64) {
2900            tcg_out_qemu_ld(s, a0, -1, a1, -1, a2, TCG_TYPE_I64);
2901        } else {
2902            tcg_out_qemu_ld(s, a0, a1, a2, args[3], args[4], TCG_TYPE_I64);
2903        }
2904        break;
2905    case INDEX_op_qemu_ld_a32_i128:
2906    case INDEX_op_qemu_ld_a64_i128:
2907        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
2908        tcg_out_qemu_ld(s, a0, a1, a2, -1, args[3], TCG_TYPE_I128);
2909        break;
2910
2911    case INDEX_op_qemu_st_a64_i32:
2912    case INDEX_op_qemu_st8_a64_i32:
2913        if (TCG_TARGET_REG_BITS == 32) {
2914            tcg_out_qemu_st(s, a0, -1, a1, a2, args[3], TCG_TYPE_I32);
2915            break;
2916        }
2917        /* fall through */
2918    case INDEX_op_qemu_st_a32_i32:
2919    case INDEX_op_qemu_st8_a32_i32:
2920        tcg_out_qemu_st(s, a0, -1, a1, -1, a2, TCG_TYPE_I32);
2921        break;
2922    case INDEX_op_qemu_st_a32_i64:
2923        if (TCG_TARGET_REG_BITS == 64) {
2924            tcg_out_qemu_st(s, a0, -1, a1, -1, a2, TCG_TYPE_I64);
2925        } else {
2926            tcg_out_qemu_st(s, a0, a1, a2, -1, args[3], TCG_TYPE_I64);
2927        }
2928        break;
2929    case INDEX_op_qemu_st_a64_i64:
2930        if (TCG_TARGET_REG_BITS == 64) {
2931            tcg_out_qemu_st(s, a0, -1, a1, -1, a2, TCG_TYPE_I64);
2932        } else {
2933            tcg_out_qemu_st(s, a0, a1, a2, args[3], args[4], TCG_TYPE_I64);
2934        }
2935        break;
2936    case INDEX_op_qemu_st_a32_i128:
2937    case INDEX_op_qemu_st_a64_i128:
2938        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
2939        tcg_out_qemu_st(s, a0, a1, a2, -1, args[3], TCG_TYPE_I128);
2940        break;
2941
2942    OP_32_64(mulu2):
2943        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_MUL, args[3]);
2944        break;
2945    OP_32_64(muls2):
2946        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_IMUL, args[3]);
2947        break;
2948    OP_32_64(add2):
2949        if (const_args[4]) {
2950            tgen_arithi(s, ARITH_ADD + rexw, a0, args[4], 1);
2951        } else {
2952            tgen_arithr(s, ARITH_ADD + rexw, a0, args[4]);
2953        }
2954        if (const_args[5]) {
2955            tgen_arithi(s, ARITH_ADC + rexw, a1, args[5], 1);
2956        } else {
2957            tgen_arithr(s, ARITH_ADC + rexw, a1, args[5]);
2958        }
2959        break;
2960    OP_32_64(sub2):
2961        if (const_args[4]) {
2962            tgen_arithi(s, ARITH_SUB + rexw, a0, args[4], 1);
2963        } else {
2964            tgen_arithr(s, ARITH_SUB + rexw, a0, args[4]);
2965        }
2966        if (const_args[5]) {
2967            tgen_arithi(s, ARITH_SBB + rexw, a1, args[5], 1);
2968        } else {
2969            tgen_arithr(s, ARITH_SBB + rexw, a1, args[5]);
2970        }
2971        break;
2972
2973#if TCG_TARGET_REG_BITS == 32
2974    case INDEX_op_brcond2_i32:
2975        tcg_out_brcond2(s, args, const_args, 0);
2976        break;
2977    case INDEX_op_setcond2_i32:
2978        tcg_out_setcond2(s, args, const_args);
2979        break;
2980#else /* TCG_TARGET_REG_BITS == 64 */
2981    case INDEX_op_ld32s_i64:
2982        tcg_out_modrm_offset(s, OPC_MOVSLQ, a0, a1, a2);
2983        break;
2984    case INDEX_op_ld_i64:
2985        tcg_out_ld(s, TCG_TYPE_I64, a0, a1, a2);
2986        break;
2987    case INDEX_op_st_i64:
2988        if (const_args[0]) {
2989            tcg_out_modrm_offset(s, OPC_MOVL_EvIz | P_REXW, 0, a1, a2);
2990            tcg_out32(s, a0);
2991        } else {
2992            tcg_out_st(s, TCG_TYPE_I64, a0, a1, a2);
2993        }
2994        break;
2995
2996    case INDEX_op_bswap64_i64:
2997        tcg_out_bswap64(s, a0);
2998        break;
2999    case INDEX_op_extrh_i64_i32:
3000        tcg_out_shifti(s, SHIFT_SHR + P_REXW, a0, 32);
3001        break;
3002#endif
3003
3004    OP_32_64(deposit):
3005        if (args[3] == 0 && args[4] == 8) {
3006            /* load bits 0..7 */
3007            if (const_a2) {
3008                tcg_out_opc(s, OPC_MOVB_Ib | P_REXB_RM | LOWREGMASK(a0),
3009                            0, a0, 0);
3010                tcg_out8(s, a2);
3011            } else {
3012                tcg_out_modrm(s, OPC_MOVB_EvGv | P_REXB_R | P_REXB_RM, a2, a0);
3013            }
3014        } else if (TCG_TARGET_REG_BITS == 32 && args[3] == 8 && args[4] == 8) {
3015            /* load bits 8..15 */
3016            if (const_a2) {
3017                tcg_out8(s, OPC_MOVB_Ib + a0 + 4);
3018                tcg_out8(s, a2);
3019            } else {
3020                tcg_out_modrm(s, OPC_MOVB_EvGv, a2, a0 + 4);
3021            }
3022        } else if (args[3] == 0 && args[4] == 16) {
3023            /* load bits 0..15 */
3024            if (const_a2) {
3025                tcg_out_opc(s, OPC_MOVL_Iv | P_DATA16 | LOWREGMASK(a0),
3026                            0, a0, 0);
3027                tcg_out16(s, a2);
3028            } else {
3029                tcg_out_modrm(s, OPC_MOVL_EvGv | P_DATA16, a2, a0);
3030            }
3031        } else {
3032            g_assert_not_reached();
3033        }
3034        break;
3035
3036    case INDEX_op_extract_i64:
3037        if (a2 + args[3] == 32) {
3038            if (a2 == 0) {
3039                tcg_out_ext32u(s, a0, a1);
3040                break;
3041            }
3042            /* This is a 32-bit zero-extending right shift.  */
3043            tcg_out_mov(s, TCG_TYPE_I32, a0, a1);
3044            tcg_out_shifti(s, SHIFT_SHR, a0, a2);
3045            break;
3046        }
3047        /* FALLTHRU */
3048    case INDEX_op_extract_i32:
3049        if (a2 == 0 && args[3] == 8) {
3050            tcg_out_ext8u(s, a0, a1);
3051        } else if (a2 == 0 && args[3] == 16) {
3052            tcg_out_ext16u(s, a0, a1);
3053        } else if (a2 == 8 && args[3] == 8) {
3054            /*
3055             * On the off-chance that we can use the high-byte registers.
3056             * Otherwise we emit the same ext16 + shift pattern that we
3057             * would have gotten from the normal tcg-op.c expansion.
3058             */
3059            if (a1 < 4 && a0 < 8) {
3060                tcg_out_modrm(s, OPC_MOVZBL, a0, a1 + 4);
3061            } else {
3062                tcg_out_ext16u(s, a0, a1);
3063                tcg_out_shifti(s, SHIFT_SHR, a0, 8);
3064            }
3065        } else {
3066            g_assert_not_reached();
3067        }
3068        break;
3069
3070    case INDEX_op_sextract_i64:
3071        if (a2 == 0 && args[3] == 8) {
3072            tcg_out_ext8s(s, TCG_TYPE_I64, a0, a1);
3073        } else if (a2 == 0 && args[3] == 16) {
3074            tcg_out_ext16s(s, TCG_TYPE_I64, a0, a1);
3075        } else if (a2 == 0 && args[3] == 32) {
3076            tcg_out_ext32s(s, a0, a1);
3077        } else {
3078            g_assert_not_reached();
3079        }
3080        break;
3081
3082    case INDEX_op_sextract_i32:
3083        if (a2 == 0 && args[3] == 8) {
3084            tcg_out_ext8s(s, TCG_TYPE_I32, a0, a1);
3085        } else if (a2 == 0 && args[3] == 16) {
3086            tcg_out_ext16s(s, TCG_TYPE_I32, a0, a1);
3087        } else if (a2 == 8 && args[3] == 8) {
3088            if (a1 < 4 && a0 < 8) {
3089                tcg_out_modrm(s, OPC_MOVSBL, a0, a1 + 4);
3090            } else {
3091                tcg_out_ext16s(s, TCG_TYPE_I32, a0, a1);
3092                tcg_out_shifti(s, SHIFT_SAR, a0, 8);
3093            }
3094        } else {
3095            g_assert_not_reached();
3096        }
3097        break;
3098
3099    OP_32_64(extract2):
3100        /* Note that SHRD outputs to the r/m operand.  */
3101        tcg_out_modrm(s, OPC_SHRD_Ib + rexw, a2, a0);
3102        tcg_out8(s, args[3]);
3103        break;
3104
3105    case INDEX_op_mb:
3106        tcg_out_mb(s, a0);
3107        break;
3108    case INDEX_op_mov_i32:  /* Always emitted via tcg_out_mov.  */
3109    case INDEX_op_mov_i64:
3110    case INDEX_op_call:     /* Always emitted via tcg_out_call.  */
3111    case INDEX_op_exit_tb:  /* Always emitted via tcg_out_exit_tb.  */
3112    case INDEX_op_goto_tb:  /* Always emitted via tcg_out_goto_tb.  */
3113    case INDEX_op_ext8s_i32:  /* Always emitted via tcg_reg_alloc_op.  */
3114    case INDEX_op_ext8s_i64:
3115    case INDEX_op_ext8u_i32:
3116    case INDEX_op_ext8u_i64:
3117    case INDEX_op_ext16s_i32:
3118    case INDEX_op_ext16s_i64:
3119    case INDEX_op_ext16u_i32:
3120    case INDEX_op_ext16u_i64:
3121    case INDEX_op_ext32s_i64:
3122    case INDEX_op_ext32u_i64:
3123    case INDEX_op_ext_i32_i64:
3124    case INDEX_op_extu_i32_i64:
3125    case INDEX_op_extrl_i64_i32:
3126    default:
3127        g_assert_not_reached();
3128    }
3129
3130#undef OP_32_64
3131}
3132
3133static int const umin_insn[4] = {
3134    OPC_PMINUB, OPC_PMINUW, OPC_PMINUD, OPC_VPMINUQ
3135};
3136
3137static int const umax_insn[4] = {
3138    OPC_PMAXUB, OPC_PMAXUW, OPC_PMAXUD, OPC_VPMAXUQ
3139};
3140
3141static bool tcg_out_cmp_vec_noinv(TCGContext *s, TCGType type, unsigned vece,
3142                                  TCGReg v0, TCGReg v1, TCGReg v2, TCGCond cond)
3143{
3144    static int const cmpeq_insn[4] = {
3145        OPC_PCMPEQB, OPC_PCMPEQW, OPC_PCMPEQD, OPC_PCMPEQQ
3146    };
3147    static int const cmpgt_insn[4] = {
3148        OPC_PCMPGTB, OPC_PCMPGTW, OPC_PCMPGTD, OPC_PCMPGTQ
3149    };
3150
3151    enum {
3152        NEED_INV  = 1,
3153        NEED_SWAP = 2,
3154        NEED_UMIN = 4,
3155        NEED_UMAX = 8,
3156        INVALID   = 16,
3157    };
3158    static const uint8_t cond_fixup[16] = {
3159        [0 ... 15] = INVALID,
3160        [TCG_COND_EQ] = 0,
3161        [TCG_COND_GT] = 0,
3162        [TCG_COND_NE] = NEED_INV,
3163        [TCG_COND_LE] = NEED_INV,
3164        [TCG_COND_LT] = NEED_SWAP,
3165        [TCG_COND_GE] = NEED_SWAP | NEED_INV,
3166        [TCG_COND_LEU] = NEED_UMIN,
3167        [TCG_COND_GTU] = NEED_UMIN | NEED_INV,
3168        [TCG_COND_GEU] = NEED_UMAX,
3169        [TCG_COND_LTU] = NEED_UMAX | NEED_INV,
3170    };
3171    int fixup = cond_fixup[cond];
3172
3173    assert(!(fixup & INVALID));
3174
3175    if (fixup & NEED_INV) {
3176        cond = tcg_invert_cond(cond);
3177    }
3178
3179    if (fixup & NEED_SWAP) {
3180        TCGReg swap = v1;
3181        v1 = v2;
3182        v2 = swap;
3183        cond = tcg_swap_cond(cond);
3184    }
3185
3186    if (fixup & (NEED_UMIN | NEED_UMAX)) {
3187        int op = (fixup & NEED_UMIN ? umin_insn[vece] : umax_insn[vece]);
3188
3189        /* avx2 does not have 64-bit min/max; adjusted during expand. */
3190        assert(vece <= MO_32);
3191
3192        tcg_out_vex_modrm_type(s, op, TCG_TMP_VEC, v1, v2, type);
3193        v2 = TCG_TMP_VEC;
3194        cond = TCG_COND_EQ;
3195    }
3196
3197    switch (cond) {
3198    case TCG_COND_EQ:
3199        tcg_out_vex_modrm_type(s, cmpeq_insn[vece], v0, v1, v2, type);
3200        break;
3201    case TCG_COND_GT:
3202        tcg_out_vex_modrm_type(s, cmpgt_insn[vece], v0, v1, v2, type);
3203        break;
3204    default:
3205        g_assert_not_reached();
3206    }
3207    return fixup & NEED_INV;
3208}
3209
3210static void tcg_out_cmp_vec_k1(TCGContext *s, TCGType type, unsigned vece,
3211                               TCGReg v1, TCGReg v2, TCGCond cond)
3212{
3213    static const int cmpm_insn[2][4] = {
3214        { OPC_VPCMPB, OPC_VPCMPW, OPC_VPCMPD, OPC_VPCMPQ },
3215        { OPC_VPCMPUB, OPC_VPCMPUW, OPC_VPCMPUD, OPC_VPCMPUQ }
3216    };
3217    static const int testm_insn[4] = {
3218        OPC_VPTESTMB, OPC_VPTESTMW, OPC_VPTESTMD, OPC_VPTESTMQ
3219    };
3220    static const int testnm_insn[4] = {
3221        OPC_VPTESTNMB, OPC_VPTESTNMW, OPC_VPTESTNMD, OPC_VPTESTNMQ
3222    };
3223
3224    static const int cond_ext[16] = {
3225        [TCG_COND_EQ] = 0,
3226        [TCG_COND_NE] = 4,
3227        [TCG_COND_LT] = 1,
3228        [TCG_COND_LTU] = 1,
3229        [TCG_COND_LE] = 2,
3230        [TCG_COND_LEU] = 2,
3231        [TCG_COND_NEVER] = 3,
3232        [TCG_COND_GE] = 5,
3233        [TCG_COND_GEU] = 5,
3234        [TCG_COND_GT] = 6,
3235        [TCG_COND_GTU] = 6,
3236        [TCG_COND_ALWAYS] = 7,
3237    };
3238
3239    switch (cond) {
3240    case TCG_COND_TSTNE:
3241        tcg_out_vex_modrm_type(s, testm_insn[vece], /* k1 */ 1, v1, v2, type);
3242        break;
3243    case TCG_COND_TSTEQ:
3244        tcg_out_vex_modrm_type(s, testnm_insn[vece], /* k1 */ 1, v1, v2, type);
3245        break;
3246    default:
3247        tcg_out_vex_modrm_type(s, cmpm_insn[is_unsigned_cond(cond)][vece],
3248                               /* k1 */ 1, v1, v2, type);
3249        tcg_out8(s, cond_ext[cond]);
3250        break;
3251    }
3252}
3253
3254static void tcg_out_k1_to_vec(TCGContext *s, TCGType type,
3255                              unsigned vece, TCGReg dest)
3256{
3257    static const int movm_insn[] = {
3258        OPC_VPMOVM2B, OPC_VPMOVM2W, OPC_VPMOVM2D, OPC_VPMOVM2Q
3259    };
3260    tcg_out_vex_modrm_type(s, movm_insn[vece], dest, 0, /* k1 */ 1, type);
3261}
3262
3263static void tcg_out_cmp_vec(TCGContext *s, TCGType type, unsigned vece,
3264                            TCGReg v0, TCGReg v1, TCGReg v2, TCGCond cond)
3265{
3266    /*
3267     * With avx512, we have a complete set of comparisons into mask.
3268     * Unless there's a single insn expansion for the comparision,
3269     * expand via a mask in k1.
3270     */
3271    if ((vece <= MO_16 ? have_avx512bw : have_avx512dq)
3272        && cond != TCG_COND_EQ
3273        && cond != TCG_COND_LT
3274        && cond != TCG_COND_GT) {
3275        tcg_out_cmp_vec_k1(s, type, vece, v1, v2, cond);
3276        tcg_out_k1_to_vec(s, type, vece, v0);
3277        return;
3278    }
3279
3280    if (tcg_out_cmp_vec_noinv(s, type, vece, v0, v1, v2, cond)) {
3281        tcg_out_dupi_vec(s, type, vece, TCG_TMP_VEC, -1);
3282        tcg_out_vex_modrm_type(s, OPC_PXOR, v0, v0, TCG_TMP_VEC, type);
3283    }
3284}
3285
3286static void tcg_out_cmpsel_vec_k1(TCGContext *s, TCGType type, unsigned vece,
3287                                  TCGReg v0, TCGReg c1, TCGReg c2,
3288                                  TCGReg v3, TCGReg v4, TCGCond cond)
3289{
3290    static const int vpblendm_insn[] = {
3291        OPC_VPBLENDMB, OPC_VPBLENDMW, OPC_VPBLENDMD, OPC_VPBLENDMQ
3292    };
3293    bool z = false;
3294
3295    /* Swap to place constant in V4 to take advantage of zero-masking. */
3296    if (!v3) {
3297        z = true;
3298        v3 = v4;
3299        cond = tcg_invert_cond(cond);
3300    }
3301
3302    tcg_out_cmp_vec_k1(s, type, vece, c1, c2, cond);
3303    tcg_out_evex_modrm_type(s, vpblendm_insn[vece], v0, v4, v3,
3304                            /* k1 */1, z, type);
3305}
3306
3307static void tcg_out_cmpsel_vec(TCGContext *s, TCGType type, unsigned vece,
3308                               TCGReg v0, TCGReg c1, TCGReg c2,
3309                               TCGReg v3, TCGReg v4, TCGCond cond)
3310{
3311    bool inv;
3312
3313    if (vece <= MO_16 ? have_avx512bw : have_avx512vl) {
3314        tcg_out_cmpsel_vec_k1(s, type, vece, v0, c1, c2, v3, v4, cond);
3315        return;
3316    }
3317
3318    inv = tcg_out_cmp_vec_noinv(s, type, vece, TCG_TMP_VEC, c1, c2, cond);
3319
3320    /*
3321     * Since XMM0 is 16, the only way we get 0 into V3
3322     * is via the constant zero constraint.
3323     */
3324    if (!v3) {
3325        if (inv) {
3326            tcg_out_vex_modrm_type(s, OPC_PAND, v0, TCG_TMP_VEC, v4, type);
3327        } else {
3328            tcg_out_vex_modrm_type(s, OPC_PANDN, v0, TCG_TMP_VEC, v4, type);
3329        }
3330    } else {
3331        if (inv) {
3332            TCGReg swap = v3;
3333            v3 = v4;
3334            v4 = swap;
3335        }
3336        tcg_out_vex_modrm_type(s, OPC_VPBLENDVB, v0, v4, v3, type);
3337        tcg_out8(s, (TCG_TMP_VEC - TCG_REG_XMM0) << 4);
3338    }
3339}
3340
3341static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
3342                           unsigned vecl, unsigned vece,
3343                           const TCGArg args[TCG_MAX_OP_ARGS],
3344                           const int const_args[TCG_MAX_OP_ARGS])
3345{
3346    static int const add_insn[4] = {
3347        OPC_PADDB, OPC_PADDW, OPC_PADDD, OPC_PADDQ
3348    };
3349    static int const ssadd_insn[4] = {
3350        OPC_PADDSB, OPC_PADDSW, OPC_UD2, OPC_UD2
3351    };
3352    static int const usadd_insn[4] = {
3353        OPC_PADDUB, OPC_PADDUW, OPC_UD2, OPC_UD2
3354    };
3355    static int const sub_insn[4] = {
3356        OPC_PSUBB, OPC_PSUBW, OPC_PSUBD, OPC_PSUBQ
3357    };
3358    static int const sssub_insn[4] = {
3359        OPC_PSUBSB, OPC_PSUBSW, OPC_UD2, OPC_UD2
3360    };
3361    static int const ussub_insn[4] = {
3362        OPC_PSUBUB, OPC_PSUBUW, OPC_UD2, OPC_UD2
3363    };
3364    static int const mul_insn[4] = {
3365        OPC_UD2, OPC_PMULLW, OPC_PMULLD, OPC_VPMULLQ
3366    };
3367    static int const shift_imm_insn[4] = {
3368        OPC_UD2, OPC_PSHIFTW_Ib, OPC_PSHIFTD_Ib, OPC_PSHIFTQ_Ib
3369    };
3370    static int const punpckl_insn[4] = {
3371        OPC_PUNPCKLBW, OPC_PUNPCKLWD, OPC_PUNPCKLDQ, OPC_PUNPCKLQDQ
3372    };
3373    static int const punpckh_insn[4] = {
3374        OPC_PUNPCKHBW, OPC_PUNPCKHWD, OPC_PUNPCKHDQ, OPC_PUNPCKHQDQ
3375    };
3376    static int const packss_insn[4] = {
3377        OPC_PACKSSWB, OPC_PACKSSDW, OPC_UD2, OPC_UD2
3378    };
3379    static int const packus_insn[4] = {
3380        OPC_PACKUSWB, OPC_PACKUSDW, OPC_UD2, OPC_UD2
3381    };
3382    static int const smin_insn[4] = {
3383        OPC_PMINSB, OPC_PMINSW, OPC_PMINSD, OPC_VPMINSQ
3384    };
3385    static int const smax_insn[4] = {
3386        OPC_PMAXSB, OPC_PMAXSW, OPC_PMAXSD, OPC_VPMAXSQ
3387    };
3388    static int const rotlv_insn[4] = {
3389        OPC_UD2, OPC_UD2, OPC_VPROLVD, OPC_VPROLVQ
3390    };
3391    static int const rotrv_insn[4] = {
3392        OPC_UD2, OPC_UD2, OPC_VPRORVD, OPC_VPRORVQ
3393    };
3394    static int const shlv_insn[4] = {
3395        OPC_UD2, OPC_VPSLLVW, OPC_VPSLLVD, OPC_VPSLLVQ
3396    };
3397    static int const shrv_insn[4] = {
3398        OPC_UD2, OPC_VPSRLVW, OPC_VPSRLVD, OPC_VPSRLVQ
3399    };
3400    static int const sarv_insn[4] = {
3401        OPC_UD2, OPC_VPSRAVW, OPC_VPSRAVD, OPC_VPSRAVQ
3402    };
3403    static int const shls_insn[4] = {
3404        OPC_UD2, OPC_PSLLW, OPC_PSLLD, OPC_PSLLQ
3405    };
3406    static int const shrs_insn[4] = {
3407        OPC_UD2, OPC_PSRLW, OPC_PSRLD, OPC_PSRLQ
3408    };
3409    static int const sars_insn[4] = {
3410        OPC_UD2, OPC_PSRAW, OPC_PSRAD, OPC_VPSRAQ
3411    };
3412    static int const vpshldi_insn[4] = {
3413        OPC_UD2, OPC_VPSHLDW, OPC_VPSHLDD, OPC_VPSHLDQ
3414    };
3415    static int const vpshldv_insn[4] = {
3416        OPC_UD2, OPC_VPSHLDVW, OPC_VPSHLDVD, OPC_VPSHLDVQ
3417    };
3418    static int const vpshrdv_insn[4] = {
3419        OPC_UD2, OPC_VPSHRDVW, OPC_VPSHRDVD, OPC_VPSHRDVQ
3420    };
3421    static int const abs_insn[4] = {
3422        OPC_PABSB, OPC_PABSW, OPC_PABSD, OPC_VPABSQ
3423    };
3424
3425    TCGType type = vecl + TCG_TYPE_V64;
3426    int insn, sub;
3427    TCGArg a0, a1, a2, a3;
3428
3429    a0 = args[0];
3430    a1 = args[1];
3431    a2 = args[2];
3432
3433    switch (opc) {
3434    case INDEX_op_add_vec:
3435        insn = add_insn[vece];
3436        goto gen_simd;
3437    case INDEX_op_ssadd_vec:
3438        insn = ssadd_insn[vece];
3439        goto gen_simd;
3440    case INDEX_op_usadd_vec:
3441        insn = usadd_insn[vece];
3442        goto gen_simd;
3443    case INDEX_op_sub_vec:
3444        insn = sub_insn[vece];
3445        goto gen_simd;
3446    case INDEX_op_sssub_vec:
3447        insn = sssub_insn[vece];
3448        goto gen_simd;
3449    case INDEX_op_ussub_vec:
3450        insn = ussub_insn[vece];
3451        goto gen_simd;
3452    case INDEX_op_mul_vec:
3453        insn = mul_insn[vece];
3454        goto gen_simd;
3455    case INDEX_op_and_vec:
3456        insn = OPC_PAND;
3457        goto gen_simd;
3458    case INDEX_op_or_vec:
3459        insn = OPC_POR;
3460        goto gen_simd;
3461    case INDEX_op_xor_vec:
3462        insn = OPC_PXOR;
3463        goto gen_simd;
3464    case INDEX_op_smin_vec:
3465        insn = smin_insn[vece];
3466        goto gen_simd;
3467    case INDEX_op_umin_vec:
3468        insn = umin_insn[vece];
3469        goto gen_simd;
3470    case INDEX_op_smax_vec:
3471        insn = smax_insn[vece];
3472        goto gen_simd;
3473    case INDEX_op_umax_vec:
3474        insn = umax_insn[vece];
3475        goto gen_simd;
3476    case INDEX_op_shlv_vec:
3477        insn = shlv_insn[vece];
3478        goto gen_simd;
3479    case INDEX_op_shrv_vec:
3480        insn = shrv_insn[vece];
3481        goto gen_simd;
3482    case INDEX_op_sarv_vec:
3483        insn = sarv_insn[vece];
3484        goto gen_simd;
3485    case INDEX_op_rotlv_vec:
3486        insn = rotlv_insn[vece];
3487        goto gen_simd;
3488    case INDEX_op_rotrv_vec:
3489        insn = rotrv_insn[vece];
3490        goto gen_simd;
3491    case INDEX_op_shls_vec:
3492        insn = shls_insn[vece];
3493        goto gen_simd;
3494    case INDEX_op_shrs_vec:
3495        insn = shrs_insn[vece];
3496        goto gen_simd;
3497    case INDEX_op_sars_vec:
3498        insn = sars_insn[vece];
3499        goto gen_simd;
3500    case INDEX_op_x86_punpckl_vec:
3501        insn = punpckl_insn[vece];
3502        goto gen_simd;
3503    case INDEX_op_x86_punpckh_vec:
3504        insn = punpckh_insn[vece];
3505        goto gen_simd;
3506    case INDEX_op_x86_packss_vec:
3507        insn = packss_insn[vece];
3508        goto gen_simd;
3509    case INDEX_op_x86_packus_vec:
3510        insn = packus_insn[vece];
3511        goto gen_simd;
3512    case INDEX_op_x86_vpshldv_vec:
3513        insn = vpshldv_insn[vece];
3514        a1 = a2;
3515        a2 = args[3];
3516        goto gen_simd;
3517    case INDEX_op_x86_vpshrdv_vec:
3518        insn = vpshrdv_insn[vece];
3519        a1 = a2;
3520        a2 = args[3];
3521        goto gen_simd;
3522#if TCG_TARGET_REG_BITS == 32
3523    case INDEX_op_dup2_vec:
3524        /* First merge the two 32-bit inputs to a single 64-bit element. */
3525        tcg_out_vex_modrm(s, OPC_PUNPCKLDQ, a0, a1, a2);
3526        /* Then replicate the 64-bit elements across the rest of the vector. */
3527        if (type != TCG_TYPE_V64) {
3528            tcg_out_dup_vec(s, type, MO_64, a0, a0);
3529        }
3530        break;
3531#endif
3532    case INDEX_op_abs_vec:
3533        insn = abs_insn[vece];
3534        a2 = a1;
3535        a1 = 0;
3536        goto gen_simd;
3537    gen_simd:
3538        tcg_debug_assert(insn != OPC_UD2);
3539        tcg_out_vex_modrm_type(s, insn, a0, a1, a2, type);
3540        break;
3541
3542    case INDEX_op_cmp_vec:
3543        tcg_out_cmp_vec(s, type, vece, a0, a1, a2, args[3]);
3544        break;
3545
3546    case INDEX_op_cmpsel_vec:
3547        tcg_out_cmpsel_vec(s, type, vece, a0, a1, a2,
3548                           args[3], args[4], args[5]);
3549        break;
3550
3551    case INDEX_op_andc_vec:
3552        insn = OPC_PANDN;
3553        tcg_out_vex_modrm_type(s, insn, a0, a2, a1, type);
3554        break;
3555
3556    case INDEX_op_shli_vec:
3557        insn = shift_imm_insn[vece];
3558        sub = 6;
3559        goto gen_shift;
3560    case INDEX_op_shri_vec:
3561        insn = shift_imm_insn[vece];
3562        sub = 2;
3563        goto gen_shift;
3564    case INDEX_op_sari_vec:
3565        if (vece == MO_64) {
3566            insn = OPC_PSHIFTD_Ib | P_VEXW | P_EVEX;
3567        } else {
3568            insn = shift_imm_insn[vece];
3569        }
3570        sub = 4;
3571        goto gen_shift;
3572    case INDEX_op_rotli_vec:
3573        insn = OPC_PSHIFTD_Ib | P_EVEX;  /* VPROL[DQ] */
3574        if (vece == MO_64) {
3575            insn |= P_VEXW;
3576        }
3577        sub = 1;
3578        goto gen_shift;
3579    gen_shift:
3580        tcg_debug_assert(vece != MO_8);
3581        tcg_out_vex_modrm_type(s, insn, sub, a0, a1, type);
3582        tcg_out8(s, a2);
3583        break;
3584
3585    case INDEX_op_ld_vec:
3586        tcg_out_ld(s, type, a0, a1, a2);
3587        break;
3588    case INDEX_op_st_vec:
3589        tcg_out_st(s, type, a0, a1, a2);
3590        break;
3591    case INDEX_op_dupm_vec:
3592        tcg_out_dupm_vec(s, type, vece, a0, a1, a2);
3593        break;
3594
3595    case INDEX_op_x86_shufps_vec:
3596        insn = OPC_SHUFPS;
3597        sub = args[3];
3598        goto gen_simd_imm8;
3599    case INDEX_op_x86_blend_vec:
3600        if (vece == MO_16) {
3601            insn = OPC_PBLENDW;
3602        } else if (vece == MO_32) {
3603            insn = (have_avx2 ? OPC_VPBLENDD : OPC_BLENDPS);
3604        } else {
3605            g_assert_not_reached();
3606        }
3607        sub = args[3];
3608        goto gen_simd_imm8;
3609    case INDEX_op_x86_vperm2i128_vec:
3610        insn = OPC_VPERM2I128;
3611        sub = args[3];
3612        goto gen_simd_imm8;
3613    case INDEX_op_x86_vpshldi_vec:
3614        insn = vpshldi_insn[vece];
3615        sub = args[3];
3616        goto gen_simd_imm8;
3617
3618    case INDEX_op_not_vec:
3619        insn = OPC_VPTERNLOGQ;
3620        a2 = a1;
3621        sub = 0x33; /* !B */
3622        goto gen_simd_imm8;
3623    case INDEX_op_nor_vec:
3624        insn = OPC_VPTERNLOGQ;
3625        sub = 0x11; /* norCB */
3626        goto gen_simd_imm8;
3627    case INDEX_op_nand_vec:
3628        insn = OPC_VPTERNLOGQ;
3629        sub = 0x77; /* nandCB */
3630        goto gen_simd_imm8;
3631    case INDEX_op_eqv_vec:
3632        insn = OPC_VPTERNLOGQ;
3633        sub = 0x99; /* xnorCB */
3634        goto gen_simd_imm8;
3635    case INDEX_op_orc_vec:
3636        insn = OPC_VPTERNLOGQ;
3637        sub = 0xdd; /* orB!C */
3638        goto gen_simd_imm8;
3639
3640    case INDEX_op_bitsel_vec:
3641        insn = OPC_VPTERNLOGQ;
3642        a3 = args[3];
3643        if (a0 == a1) {
3644            a1 = a2;
3645            a2 = a3;
3646            sub = 0xca; /* A?B:C */
3647        } else if (a0 == a2) {
3648            a2 = a3;
3649            sub = 0xe2; /* B?A:C */
3650        } else {
3651            tcg_out_mov(s, type, a0, a3);
3652            sub = 0xb8; /* B?C:A */
3653        }
3654        goto gen_simd_imm8;
3655
3656    gen_simd_imm8:
3657        tcg_debug_assert(insn != OPC_UD2);
3658        tcg_out_vex_modrm_type(s, insn, a0, a1, a2, type);
3659        tcg_out8(s, sub);
3660        break;
3661
3662    case INDEX_op_x86_psrldq_vec:
3663        tcg_out_vex_modrm(s, OPC_GRP14, 3, a0, a1);
3664        tcg_out8(s, a2);
3665        break;
3666
3667    case INDEX_op_mov_vec:  /* Always emitted via tcg_out_mov.  */
3668    case INDEX_op_dup_vec:  /* Always emitted via tcg_out_dup_vec.  */
3669    default:
3670        g_assert_not_reached();
3671    }
3672}
3673
3674static TCGConstraintSetIndex
3675tcg_target_op_def(TCGOpcode op, TCGType type, unsigned flags)
3676{
3677    switch (op) {
3678    case INDEX_op_goto_ptr:
3679        return C_O0_I1(r);
3680
3681    case INDEX_op_ld8u_i32:
3682    case INDEX_op_ld8u_i64:
3683    case INDEX_op_ld8s_i32:
3684    case INDEX_op_ld8s_i64:
3685    case INDEX_op_ld16u_i32:
3686    case INDEX_op_ld16u_i64:
3687    case INDEX_op_ld16s_i32:
3688    case INDEX_op_ld16s_i64:
3689    case INDEX_op_ld_i32:
3690    case INDEX_op_ld32u_i64:
3691    case INDEX_op_ld32s_i64:
3692    case INDEX_op_ld_i64:
3693        return C_O1_I1(r, r);
3694
3695    case INDEX_op_st8_i32:
3696    case INDEX_op_st8_i64:
3697        return C_O0_I2(qi, r);
3698
3699    case INDEX_op_st16_i32:
3700    case INDEX_op_st16_i64:
3701    case INDEX_op_st_i32:
3702    case INDEX_op_st32_i64:
3703        return C_O0_I2(ri, r);
3704
3705    case INDEX_op_st_i64:
3706        return C_O0_I2(re, r);
3707
3708    case INDEX_op_add_i32:
3709    case INDEX_op_add_i64:
3710        return C_O1_I2(r, r, re);
3711
3712    case INDEX_op_sub_i32:
3713    case INDEX_op_sub_i64:
3714    case INDEX_op_mul_i32:
3715    case INDEX_op_mul_i64:
3716    case INDEX_op_or_i32:
3717    case INDEX_op_or_i64:
3718    case INDEX_op_xor_i32:
3719    case INDEX_op_xor_i64:
3720        return C_O1_I2(r, 0, re);
3721
3722    case INDEX_op_and_i32:
3723    case INDEX_op_and_i64:
3724        return C_O1_I2(r, 0, reZ);
3725
3726    case INDEX_op_andc_i32:
3727    case INDEX_op_andc_i64:
3728        return C_O1_I2(r, r, rI);
3729
3730    case INDEX_op_shl_i32:
3731    case INDEX_op_shl_i64:
3732    case INDEX_op_shr_i32:
3733    case INDEX_op_shr_i64:
3734    case INDEX_op_sar_i32:
3735    case INDEX_op_sar_i64:
3736        return have_bmi2 ? C_O1_I2(r, r, ri) : C_O1_I2(r, 0, ci);
3737
3738    case INDEX_op_rotl_i32:
3739    case INDEX_op_rotl_i64:
3740    case INDEX_op_rotr_i32:
3741    case INDEX_op_rotr_i64:
3742        return C_O1_I2(r, 0, ci);
3743
3744    case INDEX_op_brcond_i32:
3745    case INDEX_op_brcond_i64:
3746        return C_O0_I2(r, reT);
3747
3748    case INDEX_op_bswap16_i32:
3749    case INDEX_op_bswap16_i64:
3750    case INDEX_op_bswap32_i32:
3751    case INDEX_op_bswap32_i64:
3752    case INDEX_op_bswap64_i64:
3753    case INDEX_op_neg_i32:
3754    case INDEX_op_neg_i64:
3755    case INDEX_op_not_i32:
3756    case INDEX_op_not_i64:
3757    case INDEX_op_extrh_i64_i32:
3758        return C_O1_I1(r, 0);
3759
3760    case INDEX_op_ext8s_i32:
3761    case INDEX_op_ext8s_i64:
3762    case INDEX_op_ext8u_i32:
3763    case INDEX_op_ext8u_i64:
3764        return C_O1_I1(r, q);
3765
3766    case INDEX_op_ext16s_i32:
3767    case INDEX_op_ext16s_i64:
3768    case INDEX_op_ext16u_i32:
3769    case INDEX_op_ext16u_i64:
3770    case INDEX_op_ext32s_i64:
3771    case INDEX_op_ext32u_i64:
3772    case INDEX_op_ext_i32_i64:
3773    case INDEX_op_extu_i32_i64:
3774    case INDEX_op_extrl_i64_i32:
3775    case INDEX_op_extract_i32:
3776    case INDEX_op_extract_i64:
3777    case INDEX_op_sextract_i32:
3778    case INDEX_op_sextract_i64:
3779    case INDEX_op_ctpop_i32:
3780    case INDEX_op_ctpop_i64:
3781        return C_O1_I1(r, r);
3782
3783    case INDEX_op_extract2_i32:
3784    case INDEX_op_extract2_i64:
3785        return C_O1_I2(r, 0, r);
3786
3787    case INDEX_op_deposit_i32:
3788    case INDEX_op_deposit_i64:
3789        return C_O1_I2(q, 0, qi);
3790
3791    case INDEX_op_setcond_i32:
3792    case INDEX_op_setcond_i64:
3793    case INDEX_op_negsetcond_i32:
3794    case INDEX_op_negsetcond_i64:
3795        return C_O1_I2(q, r, reT);
3796
3797    case INDEX_op_movcond_i32:
3798    case INDEX_op_movcond_i64:
3799        return C_O1_I4(r, r, reT, r, 0);
3800
3801    case INDEX_op_div2_i32:
3802    case INDEX_op_div2_i64:
3803    case INDEX_op_divu2_i32:
3804    case INDEX_op_divu2_i64:
3805        return C_O2_I3(a, d, 0, 1, r);
3806
3807    case INDEX_op_mulu2_i32:
3808    case INDEX_op_mulu2_i64:
3809    case INDEX_op_muls2_i32:
3810    case INDEX_op_muls2_i64:
3811        return C_O2_I2(a, d, a, r);
3812
3813    case INDEX_op_add2_i32:
3814    case INDEX_op_add2_i64:
3815    case INDEX_op_sub2_i32:
3816    case INDEX_op_sub2_i64:
3817        return C_N1_O1_I4(r, r, 0, 1, re, re);
3818
3819    case INDEX_op_ctz_i32:
3820    case INDEX_op_ctz_i64:
3821        return have_bmi1 ? C_N1_I2(r, r, rW) : C_N1_I2(r, r, r);
3822
3823    case INDEX_op_clz_i32:
3824    case INDEX_op_clz_i64:
3825        return have_lzcnt ? C_N1_I2(r, r, rW) : C_N1_I2(r, r, r);
3826
3827    case INDEX_op_qemu_ld_a32_i32:
3828        return C_O1_I1(r, L);
3829    case INDEX_op_qemu_ld_a64_i32:
3830        return TCG_TARGET_REG_BITS == 64 ? C_O1_I1(r, L) : C_O1_I2(r, L, L);
3831
3832    case INDEX_op_qemu_st_a32_i32:
3833        return C_O0_I2(L, L);
3834    case INDEX_op_qemu_st_a64_i32:
3835        return TCG_TARGET_REG_BITS == 64 ? C_O0_I2(L, L) : C_O0_I3(L, L, L);
3836    case INDEX_op_qemu_st8_a32_i32:
3837        return C_O0_I2(s, L);
3838    case INDEX_op_qemu_st8_a64_i32:
3839        return TCG_TARGET_REG_BITS == 64 ? C_O0_I2(s, L) : C_O0_I3(s, L, L);
3840
3841    case INDEX_op_qemu_ld_a32_i64:
3842        return TCG_TARGET_REG_BITS == 64 ? C_O1_I1(r, L) : C_O2_I1(r, r, L);
3843    case INDEX_op_qemu_ld_a64_i64:
3844        return TCG_TARGET_REG_BITS == 64 ? C_O1_I1(r, L) : C_O2_I2(r, r, L, L);
3845
3846    case INDEX_op_qemu_st_a32_i64:
3847        return TCG_TARGET_REG_BITS == 64 ? C_O0_I2(L, L) : C_O0_I3(L, L, L);
3848    case INDEX_op_qemu_st_a64_i64:
3849        return TCG_TARGET_REG_BITS == 64 ? C_O0_I2(L, L) : C_O0_I4(L, L, L, L);
3850
3851    case INDEX_op_qemu_ld_a32_i128:
3852    case INDEX_op_qemu_ld_a64_i128:
3853        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
3854        return C_O2_I1(r, r, L);
3855    case INDEX_op_qemu_st_a32_i128:
3856    case INDEX_op_qemu_st_a64_i128:
3857        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
3858        return C_O0_I3(L, L, L);
3859
3860    case INDEX_op_brcond2_i32:
3861        return C_O0_I4(r, r, ri, ri);
3862
3863    case INDEX_op_setcond2_i32:
3864        return C_O1_I4(r, r, r, ri, ri);
3865
3866    case INDEX_op_ld_vec:
3867    case INDEX_op_dupm_vec:
3868        return C_O1_I1(x, r);
3869
3870    case INDEX_op_st_vec:
3871        return C_O0_I2(x, r);
3872
3873    case INDEX_op_add_vec:
3874    case INDEX_op_sub_vec:
3875    case INDEX_op_mul_vec:
3876    case INDEX_op_and_vec:
3877    case INDEX_op_or_vec:
3878    case INDEX_op_xor_vec:
3879    case INDEX_op_andc_vec:
3880    case INDEX_op_orc_vec:
3881    case INDEX_op_nand_vec:
3882    case INDEX_op_nor_vec:
3883    case INDEX_op_eqv_vec:
3884    case INDEX_op_ssadd_vec:
3885    case INDEX_op_usadd_vec:
3886    case INDEX_op_sssub_vec:
3887    case INDEX_op_ussub_vec:
3888    case INDEX_op_smin_vec:
3889    case INDEX_op_umin_vec:
3890    case INDEX_op_smax_vec:
3891    case INDEX_op_umax_vec:
3892    case INDEX_op_shlv_vec:
3893    case INDEX_op_shrv_vec:
3894    case INDEX_op_sarv_vec:
3895    case INDEX_op_rotlv_vec:
3896    case INDEX_op_rotrv_vec:
3897    case INDEX_op_shls_vec:
3898    case INDEX_op_shrs_vec:
3899    case INDEX_op_sars_vec:
3900    case INDEX_op_cmp_vec:
3901    case INDEX_op_x86_shufps_vec:
3902    case INDEX_op_x86_blend_vec:
3903    case INDEX_op_x86_packss_vec:
3904    case INDEX_op_x86_packus_vec:
3905    case INDEX_op_x86_vperm2i128_vec:
3906    case INDEX_op_x86_punpckl_vec:
3907    case INDEX_op_x86_punpckh_vec:
3908    case INDEX_op_x86_vpshldi_vec:
3909#if TCG_TARGET_REG_BITS == 32
3910    case INDEX_op_dup2_vec:
3911#endif
3912        return C_O1_I2(x, x, x);
3913
3914    case INDEX_op_abs_vec:
3915    case INDEX_op_dup_vec:
3916    case INDEX_op_not_vec:
3917    case INDEX_op_shli_vec:
3918    case INDEX_op_shri_vec:
3919    case INDEX_op_sari_vec:
3920    case INDEX_op_rotli_vec:
3921    case INDEX_op_x86_psrldq_vec:
3922        return C_O1_I1(x, x);
3923
3924    case INDEX_op_x86_vpshldv_vec:
3925    case INDEX_op_x86_vpshrdv_vec:
3926        return C_O1_I3(x, 0, x, x);
3927
3928    case INDEX_op_bitsel_vec:
3929        return C_O1_I3(x, x, x, x);
3930    case INDEX_op_cmpsel_vec:
3931        return C_O1_I4(x, x, x, xO, x);
3932
3933    default:
3934        return C_NotImplemented;
3935    }
3936}
3937
3938int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
3939{
3940    switch (opc) {
3941    case INDEX_op_add_vec:
3942    case INDEX_op_sub_vec:
3943    case INDEX_op_and_vec:
3944    case INDEX_op_or_vec:
3945    case INDEX_op_xor_vec:
3946    case INDEX_op_andc_vec:
3947    case INDEX_op_orc_vec:
3948    case INDEX_op_nand_vec:
3949    case INDEX_op_nor_vec:
3950    case INDEX_op_eqv_vec:
3951    case INDEX_op_not_vec:
3952    case INDEX_op_bitsel_vec:
3953        return 1;
3954    case INDEX_op_cmp_vec:
3955    case INDEX_op_cmpsel_vec:
3956        return -1;
3957
3958    case INDEX_op_rotli_vec:
3959        return have_avx512vl && vece >= MO_32 ? 1 : -1;
3960
3961    case INDEX_op_shli_vec:
3962    case INDEX_op_shri_vec:
3963        /* We must expand the operation for MO_8.  */
3964        return vece == MO_8 ? -1 : 1;
3965
3966    case INDEX_op_sari_vec:
3967        switch (vece) {
3968        case MO_8:
3969            return -1;
3970        case MO_16:
3971        case MO_32:
3972            return 1;
3973        case MO_64:
3974            if (have_avx512vl) {
3975                return 1;
3976            }
3977            /*
3978             * We can emulate this for MO_64, but it does not pay off
3979             * unless we're producing at least 4 values.
3980             */
3981            return type >= TCG_TYPE_V256 ? -1 : 0;
3982        }
3983        return 0;
3984
3985    case INDEX_op_shls_vec:
3986    case INDEX_op_shrs_vec:
3987        return vece >= MO_16;
3988    case INDEX_op_sars_vec:
3989        switch (vece) {
3990        case MO_16:
3991        case MO_32:
3992            return 1;
3993        case MO_64:
3994            return have_avx512vl;
3995        }
3996        return 0;
3997    case INDEX_op_rotls_vec:
3998        return vece >= MO_16 ? -1 : 0;
3999
4000    case INDEX_op_shlv_vec:
4001    case INDEX_op_shrv_vec:
4002        switch (vece) {
4003        case MO_16:
4004            return have_avx512bw;
4005        case MO_32:
4006        case MO_64:
4007            return have_avx2;
4008        }
4009        return 0;
4010    case INDEX_op_sarv_vec:
4011        switch (vece) {
4012        case MO_16:
4013            return have_avx512bw;
4014        case MO_32:
4015            return have_avx2;
4016        case MO_64:
4017            return have_avx512vl;
4018        }
4019        return 0;
4020    case INDEX_op_rotlv_vec:
4021    case INDEX_op_rotrv_vec:
4022        switch (vece) {
4023        case MO_16:
4024            return have_avx512vbmi2 ? -1 : 0;
4025        case MO_32:
4026        case MO_64:
4027            return have_avx512vl ? 1 : have_avx2 ? -1 : 0;
4028        }
4029        return 0;
4030
4031    case INDEX_op_mul_vec:
4032        switch (vece) {
4033        case MO_8:
4034            return -1;
4035        case MO_64:
4036            return have_avx512dq;
4037        }
4038        return 1;
4039
4040    case INDEX_op_ssadd_vec:
4041    case INDEX_op_usadd_vec:
4042    case INDEX_op_sssub_vec:
4043    case INDEX_op_ussub_vec:
4044        return vece <= MO_16;
4045    case INDEX_op_smin_vec:
4046    case INDEX_op_smax_vec:
4047    case INDEX_op_umin_vec:
4048    case INDEX_op_umax_vec:
4049    case INDEX_op_abs_vec:
4050        return vece <= MO_32 || have_avx512vl;
4051
4052    default:
4053        return 0;
4054    }
4055}
4056
4057static void expand_vec_shi(TCGType type, unsigned vece, bool right,
4058                           TCGv_vec v0, TCGv_vec v1, TCGArg imm)
4059{
4060    uint8_t mask;
4061
4062    tcg_debug_assert(vece == MO_8);
4063    if (right) {
4064        mask = 0xff >> imm;
4065        tcg_gen_shri_vec(MO_16, v0, v1, imm);
4066    } else {
4067        mask = 0xff << imm;
4068        tcg_gen_shli_vec(MO_16, v0, v1, imm);
4069    }
4070    tcg_gen_and_vec(MO_8, v0, v0, tcg_constant_vec(type, MO_8, mask));
4071}
4072
4073static void expand_vec_sari(TCGType type, unsigned vece,
4074                            TCGv_vec v0, TCGv_vec v1, TCGArg imm)
4075{
4076    TCGv_vec t1, t2;
4077
4078    switch (vece) {
4079    case MO_8:
4080        /* Unpack to 16-bit, shift, and repack.  */
4081        t1 = tcg_temp_new_vec(type);
4082        t2 = tcg_temp_new_vec(type);
4083        vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
4084                  tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
4085        vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
4086                  tcgv_vec_arg(t2), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
4087        tcg_gen_sari_vec(MO_16, t1, t1, imm + 8);
4088        tcg_gen_sari_vec(MO_16, t2, t2, imm + 8);
4089        vec_gen_3(INDEX_op_x86_packss_vec, type, MO_8,
4090                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t2));
4091        tcg_temp_free_vec(t1);
4092        tcg_temp_free_vec(t2);
4093        break;
4094
4095    case MO_64:
4096        t1 = tcg_temp_new_vec(type);
4097        if (imm <= 32) {
4098            /*
4099             * We can emulate a small sign extend by performing an arithmetic
4100             * 32-bit shift and overwriting the high half of a 64-bit logical
4101             * shift.  Note that the ISA says shift of 32 is valid, but TCG
4102             * does not, so we have to bound the smaller shift -- we get the
4103             * same result in the high half either way.
4104             */
4105            tcg_gen_sari_vec(MO_32, t1, v1, MIN(imm, 31));
4106            tcg_gen_shri_vec(MO_64, v0, v1, imm);
4107            vec_gen_4(INDEX_op_x86_blend_vec, type, MO_32,
4108                      tcgv_vec_arg(v0), tcgv_vec_arg(v0),
4109                      tcgv_vec_arg(t1), 0xaa);
4110        } else {
4111            /* Otherwise we will need to use a compare vs 0 to produce
4112             * the sign-extend, shift and merge.
4113             */
4114            tcg_gen_cmp_vec(TCG_COND_GT, MO_64, t1,
4115                            tcg_constant_vec(type, MO_64, 0), v1);
4116            tcg_gen_shri_vec(MO_64, v0, v1, imm);
4117            tcg_gen_shli_vec(MO_64, t1, t1, 64 - imm);
4118            tcg_gen_or_vec(MO_64, v0, v0, t1);
4119        }
4120        tcg_temp_free_vec(t1);
4121        break;
4122
4123    default:
4124        g_assert_not_reached();
4125    }
4126}
4127
4128static void expand_vec_rotli(TCGType type, unsigned vece,
4129                             TCGv_vec v0, TCGv_vec v1, TCGArg imm)
4130{
4131    TCGv_vec t;
4132
4133    if (vece != MO_8 && have_avx512vbmi2) {
4134        vec_gen_4(INDEX_op_x86_vpshldi_vec, type, vece,
4135                  tcgv_vec_arg(v0), tcgv_vec_arg(v1), tcgv_vec_arg(v1), imm);
4136        return;
4137    }
4138
4139    t = tcg_temp_new_vec(type);
4140    tcg_gen_shli_vec(vece, t, v1, imm);
4141    tcg_gen_shri_vec(vece, v0, v1, (8 << vece) - imm);
4142    tcg_gen_or_vec(vece, v0, v0, t);
4143    tcg_temp_free_vec(t);
4144}
4145
4146static void expand_vec_rotv(TCGType type, unsigned vece, TCGv_vec v0,
4147                            TCGv_vec v1, TCGv_vec sh, bool right)
4148{
4149    TCGv_vec t;
4150
4151    if (have_avx512vbmi2) {
4152        vec_gen_4(right ? INDEX_op_x86_vpshrdv_vec : INDEX_op_x86_vpshldv_vec,
4153                  type, vece, tcgv_vec_arg(v0), tcgv_vec_arg(v1),
4154                  tcgv_vec_arg(v1), tcgv_vec_arg(sh));
4155        return;
4156    }
4157
4158    t = tcg_temp_new_vec(type);
4159    tcg_gen_dupi_vec(vece, t, 8 << vece);
4160    tcg_gen_sub_vec(vece, t, t, sh);
4161    if (right) {
4162        tcg_gen_shlv_vec(vece, t, v1, t);
4163        tcg_gen_shrv_vec(vece, v0, v1, sh);
4164    } else {
4165        tcg_gen_shrv_vec(vece, t, v1, t);
4166        tcg_gen_shlv_vec(vece, v0, v1, sh);
4167    }
4168    tcg_gen_or_vec(vece, v0, v0, t);
4169    tcg_temp_free_vec(t);
4170}
4171
4172static void expand_vec_rotls(TCGType type, unsigned vece,
4173                             TCGv_vec v0, TCGv_vec v1, TCGv_i32 lsh)
4174{
4175    TCGv_vec t = tcg_temp_new_vec(type);
4176
4177    tcg_debug_assert(vece != MO_8);
4178
4179    if (vece >= MO_32 ? have_avx512vl : have_avx512vbmi2) {
4180        tcg_gen_dup_i32_vec(vece, t, lsh);
4181        if (vece >= MO_32) {
4182            tcg_gen_rotlv_vec(vece, v0, v1, t);
4183        } else {
4184            expand_vec_rotv(type, vece, v0, v1, t, false);
4185        }
4186    } else {
4187        TCGv_i32 rsh = tcg_temp_new_i32();
4188
4189        tcg_gen_neg_i32(rsh, lsh);
4190        tcg_gen_andi_i32(rsh, rsh, (8 << vece) - 1);
4191        tcg_gen_shls_vec(vece, t, v1, lsh);
4192        tcg_gen_shrs_vec(vece, v0, v1, rsh);
4193        tcg_gen_or_vec(vece, v0, v0, t);
4194
4195        tcg_temp_free_i32(rsh);
4196    }
4197
4198    tcg_temp_free_vec(t);
4199}
4200
4201static void expand_vec_mul(TCGType type, unsigned vece,
4202                           TCGv_vec v0, TCGv_vec v1, TCGv_vec v2)
4203{
4204    TCGv_vec t1, t2, t3, t4, zero;
4205
4206    tcg_debug_assert(vece == MO_8);
4207
4208    /*
4209     * Unpack v1 bytes to words, 0 | x.
4210     * Unpack v2 bytes to words, y | 0.
4211     * This leaves the 8-bit result, x * y, with 8 bits of right padding.
4212     * Shift logical right by 8 bits to clear the high 8 bytes before
4213     * using an unsigned saturated pack.
4214     *
4215     * The difference between the V64, V128 and V256 cases is merely how
4216     * we distribute the expansion between temporaries.
4217     */
4218    switch (type) {
4219    case TCG_TYPE_V64:
4220        t1 = tcg_temp_new_vec(TCG_TYPE_V128);
4221        t2 = tcg_temp_new_vec(TCG_TYPE_V128);
4222        zero = tcg_constant_vec(TCG_TYPE_V128, MO_8, 0);
4223        vec_gen_3(INDEX_op_x86_punpckl_vec, TCG_TYPE_V128, MO_8,
4224                  tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
4225        vec_gen_3(INDEX_op_x86_punpckl_vec, TCG_TYPE_V128, MO_8,
4226                  tcgv_vec_arg(t2), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
4227        tcg_gen_mul_vec(MO_16, t1, t1, t2);
4228        tcg_gen_shri_vec(MO_16, t1, t1, 8);
4229        vec_gen_3(INDEX_op_x86_packus_vec, TCG_TYPE_V128, MO_8,
4230                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t1));
4231        tcg_temp_free_vec(t1);
4232        tcg_temp_free_vec(t2);
4233        break;
4234
4235    case TCG_TYPE_V128:
4236    case TCG_TYPE_V256:
4237        t1 = tcg_temp_new_vec(type);
4238        t2 = tcg_temp_new_vec(type);
4239        t3 = tcg_temp_new_vec(type);
4240        t4 = tcg_temp_new_vec(type);
4241        zero = tcg_constant_vec(TCG_TYPE_V128, MO_8, 0);
4242        vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
4243                  tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
4244        vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
4245                  tcgv_vec_arg(t2), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
4246        vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
4247                  tcgv_vec_arg(t3), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
4248        vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
4249                  tcgv_vec_arg(t4), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
4250        tcg_gen_mul_vec(MO_16, t1, t1, t2);
4251        tcg_gen_mul_vec(MO_16, t3, t3, t4);
4252        tcg_gen_shri_vec(MO_16, t1, t1, 8);
4253        tcg_gen_shri_vec(MO_16, t3, t3, 8);
4254        vec_gen_3(INDEX_op_x86_packus_vec, type, MO_8,
4255                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t3));
4256        tcg_temp_free_vec(t1);
4257        tcg_temp_free_vec(t2);
4258        tcg_temp_free_vec(t3);
4259        tcg_temp_free_vec(t4);
4260        break;
4261
4262    default:
4263        g_assert_not_reached();
4264    }
4265}
4266
4267static TCGCond expand_vec_cond(TCGType type, unsigned vece,
4268                               TCGArg *a1, TCGArg *a2, TCGCond cond)
4269{
4270    /*
4271     * Without AVX512, there are no 64-bit unsigned comparisons.
4272     * We must bias the inputs so that they become signed.
4273     * All other swapping and inversion are handled during code generation.
4274     */
4275    if (vece == MO_64 && !have_avx512dq && is_unsigned_cond(cond)) {
4276        TCGv_vec v1 = temp_tcgv_vec(arg_temp(*a1));
4277        TCGv_vec v2 = temp_tcgv_vec(arg_temp(*a2));
4278        TCGv_vec t1 = tcg_temp_new_vec(type);
4279        TCGv_vec t2 = tcg_temp_new_vec(type);
4280        TCGv_vec t3 = tcg_constant_vec(type, vece, 1ull << ((8 << vece) - 1));
4281
4282        tcg_gen_sub_vec(vece, t1, v1, t3);
4283        tcg_gen_sub_vec(vece, t2, v2, t3);
4284        *a1 = tcgv_vec_arg(t1);
4285        *a2 = tcgv_vec_arg(t2);
4286        cond = tcg_signed_cond(cond);
4287    }
4288    return cond;
4289}
4290
4291static void expand_vec_cmp(TCGType type, unsigned vece, TCGArg a0,
4292                           TCGArg a1, TCGArg a2, TCGCond cond)
4293{
4294    cond = expand_vec_cond(type, vece, &a1, &a2, cond);
4295    /* Expand directly; do not recurse.  */
4296    vec_gen_4(INDEX_op_cmp_vec, type, vece, a0, a1, a2, cond);
4297}
4298
4299static void expand_vec_cmpsel(TCGType type, unsigned vece, TCGArg a0,
4300                              TCGArg a1, TCGArg a2,
4301                              TCGArg a3, TCGArg a4, TCGCond cond)
4302{
4303    cond = expand_vec_cond(type, vece, &a1, &a2, cond);
4304    /* Expand directly; do not recurse.  */
4305    vec_gen_6(INDEX_op_cmpsel_vec, type, vece, a0, a1, a2, a3, a4, cond);
4306}
4307
4308void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
4309                       TCGArg a0, ...)
4310{
4311    va_list va;
4312    TCGArg a1, a2, a3, a4, a5;
4313    TCGv_vec v0, v1, v2;
4314
4315    va_start(va, a0);
4316    a1 = va_arg(va, TCGArg);
4317    a2 = va_arg(va, TCGArg);
4318    v0 = temp_tcgv_vec(arg_temp(a0));
4319    v1 = temp_tcgv_vec(arg_temp(a1));
4320
4321    switch (opc) {
4322    case INDEX_op_shli_vec:
4323        expand_vec_shi(type, vece, false, v0, v1, a2);
4324        break;
4325    case INDEX_op_shri_vec:
4326        expand_vec_shi(type, vece, true, v0, v1, a2);
4327        break;
4328    case INDEX_op_sari_vec:
4329        expand_vec_sari(type, vece, v0, v1, a2);
4330        break;
4331
4332    case INDEX_op_rotli_vec:
4333        expand_vec_rotli(type, vece, v0, v1, a2);
4334        break;
4335
4336    case INDEX_op_rotls_vec:
4337        expand_vec_rotls(type, vece, v0, v1, temp_tcgv_i32(arg_temp(a2)));
4338        break;
4339
4340    case INDEX_op_rotlv_vec:
4341        v2 = temp_tcgv_vec(arg_temp(a2));
4342        expand_vec_rotv(type, vece, v0, v1, v2, false);
4343        break;
4344    case INDEX_op_rotrv_vec:
4345        v2 = temp_tcgv_vec(arg_temp(a2));
4346        expand_vec_rotv(type, vece, v0, v1, v2, true);
4347        break;
4348
4349    case INDEX_op_mul_vec:
4350        v2 = temp_tcgv_vec(arg_temp(a2));
4351        expand_vec_mul(type, vece, v0, v1, v2);
4352        break;
4353
4354    case INDEX_op_cmp_vec:
4355        a3 = va_arg(va, TCGArg);
4356        expand_vec_cmp(type, vece, a0, a1, a2, a3);
4357        break;
4358
4359    case INDEX_op_cmpsel_vec:
4360        a3 = va_arg(va, TCGArg);
4361        a4 = va_arg(va, TCGArg);
4362        a5 = va_arg(va, TCGArg);
4363        expand_vec_cmpsel(type, vece, a0, a1, a2, a3, a4, a5);
4364        break;
4365
4366    default:
4367        break;
4368    }
4369
4370    va_end(va);
4371}
4372
4373static const int tcg_target_callee_save_regs[] = {
4374#if TCG_TARGET_REG_BITS == 64
4375    TCG_REG_RBP,
4376    TCG_REG_RBX,
4377#if defined(_WIN64)
4378    TCG_REG_RDI,
4379    TCG_REG_RSI,
4380#endif
4381    TCG_REG_R12,
4382    TCG_REG_R13,
4383    TCG_REG_R14, /* Currently used for the global env. */
4384    TCG_REG_R15,
4385#else
4386    TCG_REG_EBP, /* Currently used for the global env. */
4387    TCG_REG_EBX,
4388    TCG_REG_ESI,
4389    TCG_REG_EDI,
4390#endif
4391};
4392
4393/* Compute frame size via macros, to share between tcg_target_qemu_prologue
4394   and tcg_register_jit.  */
4395
4396#define PUSH_SIZE \
4397    ((1 + ARRAY_SIZE(tcg_target_callee_save_regs)) \
4398     * (TCG_TARGET_REG_BITS / 8))
4399
4400#define FRAME_SIZE \
4401    ((PUSH_SIZE \
4402      + TCG_STATIC_CALL_ARGS_SIZE \
4403      + CPU_TEMP_BUF_NLONGS * sizeof(long) \
4404      + TCG_TARGET_STACK_ALIGN - 1) \
4405     & ~(TCG_TARGET_STACK_ALIGN - 1))
4406
4407/* Generate global QEMU prologue and epilogue code */
4408static void tcg_target_qemu_prologue(TCGContext *s)
4409{
4410    int i, stack_addend;
4411
4412    /* TB prologue */
4413
4414    /* Reserve some stack space, also for TCG temps.  */
4415    stack_addend = FRAME_SIZE - PUSH_SIZE;
4416    tcg_set_frame(s, TCG_REG_CALL_STACK, TCG_STATIC_CALL_ARGS_SIZE,
4417                  CPU_TEMP_BUF_NLONGS * sizeof(long));
4418
4419    /* Save all callee saved registers.  */
4420    for (i = 0; i < ARRAY_SIZE(tcg_target_callee_save_regs); i++) {
4421        tcg_out_push(s, tcg_target_callee_save_regs[i]);
4422    }
4423
4424    if (!tcg_use_softmmu && guest_base) {
4425        int seg = setup_guest_base_seg();
4426        if (seg != 0) {
4427            x86_guest_base.seg = seg;
4428        } else if (guest_base == (int32_t)guest_base) {
4429            x86_guest_base.ofs = guest_base;
4430        } else {
4431            assert(TCG_TARGET_REG_BITS == 64);
4432            /* Choose R12 because, as a base, it requires a SIB byte. */
4433            x86_guest_base.index = TCG_REG_R12;
4434            tcg_out_movi(s, TCG_TYPE_PTR, x86_guest_base.index, guest_base);
4435            tcg_regset_set_reg(s->reserved_regs, x86_guest_base.index);
4436        }
4437    }
4438
4439    if (TCG_TARGET_REG_BITS == 32) {
4440        tcg_out_ld(s, TCG_TYPE_PTR, TCG_AREG0, TCG_REG_ESP,
4441                   (ARRAY_SIZE(tcg_target_callee_save_regs) + 1) * 4);
4442        tcg_out_addi(s, TCG_REG_ESP, -stack_addend);
4443        /* jmp *tb.  */
4444        tcg_out_modrm_offset(s, OPC_GRP5, EXT5_JMPN_Ev, TCG_REG_ESP,
4445                             (ARRAY_SIZE(tcg_target_callee_save_regs) + 2) * 4
4446                             + stack_addend);
4447    } else {
4448        tcg_out_mov(s, TCG_TYPE_PTR, TCG_AREG0, tcg_target_call_iarg_regs[0]);
4449        tcg_out_addi(s, TCG_REG_ESP, -stack_addend);
4450        /* jmp *tb.  */
4451        tcg_out_modrm(s, OPC_GRP5, EXT5_JMPN_Ev, tcg_target_call_iarg_regs[1]);
4452    }
4453
4454    /*
4455     * Return path for goto_ptr. Set return value to 0, a-la exit_tb,
4456     * and fall through to the rest of the epilogue.
4457     */
4458    tcg_code_gen_epilogue = tcg_splitwx_to_rx(s->code_ptr);
4459    tcg_out_movi(s, TCG_TYPE_REG, TCG_REG_EAX, 0);
4460
4461    /* TB epilogue */
4462    tb_ret_addr = tcg_splitwx_to_rx(s->code_ptr);
4463
4464    tcg_out_addi(s, TCG_REG_CALL_STACK, stack_addend);
4465
4466    if (have_avx2) {
4467        tcg_out_vex_opc(s, OPC_VZEROUPPER, 0, 0, 0, 0);
4468    }
4469    for (i = ARRAY_SIZE(tcg_target_callee_save_regs) - 1; i >= 0; i--) {
4470        tcg_out_pop(s, tcg_target_callee_save_regs[i]);
4471    }
4472    tcg_out_opc(s, OPC_RET, 0, 0, 0);
4473}
4474
4475static void tcg_out_tb_start(TCGContext *s)
4476{
4477    /* nothing to do */
4478}
4479
4480static void tcg_out_nop_fill(tcg_insn_unit *p, int count)
4481{
4482    memset(p, 0x90, count);
4483}
4484
4485static void tcg_target_init(TCGContext *s)
4486{
4487    tcg_target_available_regs[TCG_TYPE_I32] = ALL_GENERAL_REGS;
4488    if (TCG_TARGET_REG_BITS == 64) {
4489        tcg_target_available_regs[TCG_TYPE_I64] = ALL_GENERAL_REGS;
4490    }
4491    if (have_avx1) {
4492        tcg_target_available_regs[TCG_TYPE_V64] = ALL_VECTOR_REGS;
4493        tcg_target_available_regs[TCG_TYPE_V128] = ALL_VECTOR_REGS;
4494    }
4495    if (have_avx2) {
4496        tcg_target_available_regs[TCG_TYPE_V256] = ALL_VECTOR_REGS;
4497    }
4498
4499    tcg_target_call_clobber_regs = ALL_VECTOR_REGS;
4500    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_EAX);
4501    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_EDX);
4502    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_ECX);
4503    if (TCG_TARGET_REG_BITS == 64) {
4504#if !defined(_WIN64)
4505        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_RDI);
4506        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_RSI);
4507#endif
4508        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R8);
4509        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R9);
4510        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R10);
4511        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R11);
4512    }
4513
4514    s->reserved_regs = 0;
4515    tcg_regset_set_reg(s->reserved_regs, TCG_REG_CALL_STACK);
4516    tcg_regset_set_reg(s->reserved_regs, TCG_TMP_VEC);
4517#ifdef _WIN64
4518    /* These are call saved, and we don't save them, so don't use them. */
4519    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM6);
4520    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM7);
4521    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM8);
4522    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM9);
4523    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM10);
4524    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM11);
4525    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM12);
4526    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM13);
4527    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM14);
4528    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM15);
4529#endif
4530}
4531
4532typedef struct {
4533    DebugFrameHeader h;
4534    uint8_t fde_def_cfa[4];
4535    uint8_t fde_reg_ofs[14];
4536} DebugFrame;
4537
4538/* We're expecting a 2 byte uleb128 encoded value.  */
4539QEMU_BUILD_BUG_ON(FRAME_SIZE >= (1 << 14));
4540
4541#if !defined(__ELF__)
4542    /* Host machine without ELF. */
4543#elif TCG_TARGET_REG_BITS == 64
4544#define ELF_HOST_MACHINE EM_X86_64
4545static const DebugFrame debug_frame = {
4546    .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
4547    .h.cie.id = -1,
4548    .h.cie.version = 1,
4549    .h.cie.code_align = 1,
4550    .h.cie.data_align = 0x78,             /* sleb128 -8 */
4551    .h.cie.return_column = 16,
4552
4553    /* Total FDE size does not include the "len" member.  */
4554    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
4555
4556    .fde_def_cfa = {
4557        12, 7,                          /* DW_CFA_def_cfa %rsp, ... */
4558        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
4559        (FRAME_SIZE >> 7)
4560    },
4561    .fde_reg_ofs = {
4562        0x90, 1,                        /* DW_CFA_offset, %rip, -8 */
4563        /* The following ordering must match tcg_target_callee_save_regs.  */
4564        0x86, 2,                        /* DW_CFA_offset, %rbp, -16 */
4565        0x83, 3,                        /* DW_CFA_offset, %rbx, -24 */
4566        0x8c, 4,                        /* DW_CFA_offset, %r12, -32 */
4567        0x8d, 5,                        /* DW_CFA_offset, %r13, -40 */
4568        0x8e, 6,                        /* DW_CFA_offset, %r14, -48 */
4569        0x8f, 7,                        /* DW_CFA_offset, %r15, -56 */
4570    }
4571};
4572#else
4573#define ELF_HOST_MACHINE EM_386
4574static const DebugFrame debug_frame = {
4575    .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
4576    .h.cie.id = -1,
4577    .h.cie.version = 1,
4578    .h.cie.code_align = 1,
4579    .h.cie.data_align = 0x7c,             /* sleb128 -4 */
4580    .h.cie.return_column = 8,
4581
4582    /* Total FDE size does not include the "len" member.  */
4583    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
4584
4585    .fde_def_cfa = {
4586        12, 4,                          /* DW_CFA_def_cfa %esp, ... */
4587        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
4588        (FRAME_SIZE >> 7)
4589    },
4590    .fde_reg_ofs = {
4591        0x88, 1,                        /* DW_CFA_offset, %eip, -4 */
4592        /* The following ordering must match tcg_target_callee_save_regs.  */
4593        0x85, 2,                        /* DW_CFA_offset, %ebp, -8 */
4594        0x83, 3,                        /* DW_CFA_offset, %ebx, -12 */
4595        0x86, 4,                        /* DW_CFA_offset, %esi, -16 */
4596        0x87, 5,                        /* DW_CFA_offset, %edi, -20 */
4597    }
4598};
4599#endif
4600
4601#if defined(ELF_HOST_MACHINE)
4602void tcg_register_jit(const void *buf, size_t buf_size)
4603{
4604    tcg_register_jit_int(buf, buf_size, &debug_frame, sizeof(debug_frame));
4605}
4606#endif
4607