xref: /openbmc/qemu/tcg/i386/tcg-target.c.inc (revision 86d063fa)
1/*
2 * Tiny Code Generator for QEMU
3 *
4 * Copyright (c) 2008 Fabrice Bellard
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to deal
8 * in the Software without restriction, including without limitation the rights
9 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
10 * copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
22 * THE SOFTWARE.
23 */
24
25#include "../tcg-ldst.c.inc"
26#include "../tcg-pool.c.inc"
27
28#ifdef CONFIG_DEBUG_TCG
29static const char * const tcg_target_reg_names[TCG_TARGET_NB_REGS] = {
30#if TCG_TARGET_REG_BITS == 64
31    "%rax", "%rcx", "%rdx", "%rbx", "%rsp", "%rbp", "%rsi", "%rdi",
32#else
33    "%eax", "%ecx", "%edx", "%ebx", "%esp", "%ebp", "%esi", "%edi",
34#endif
35    "%r8",  "%r9",  "%r10", "%r11", "%r12", "%r13", "%r14", "%r15",
36    "%xmm0", "%xmm1", "%xmm2", "%xmm3", "%xmm4", "%xmm5", "%xmm6", "%xmm7",
37#if TCG_TARGET_REG_BITS == 64
38    "%xmm8", "%xmm9", "%xmm10", "%xmm11",
39    "%xmm12", "%xmm13", "%xmm14", "%xmm15",
40#endif
41};
42#endif
43
44static const int tcg_target_reg_alloc_order[] = {
45#if TCG_TARGET_REG_BITS == 64
46    TCG_REG_RBP,
47    TCG_REG_RBX,
48    TCG_REG_R12,
49    TCG_REG_R13,
50    TCG_REG_R14,
51    TCG_REG_R15,
52    TCG_REG_R10,
53    TCG_REG_R11,
54    TCG_REG_R9,
55    TCG_REG_R8,
56    TCG_REG_RCX,
57    TCG_REG_RDX,
58    TCG_REG_RSI,
59    TCG_REG_RDI,
60    TCG_REG_RAX,
61#else
62    TCG_REG_EBX,
63    TCG_REG_ESI,
64    TCG_REG_EDI,
65    TCG_REG_EBP,
66    TCG_REG_ECX,
67    TCG_REG_EDX,
68    TCG_REG_EAX,
69#endif
70    TCG_REG_XMM0,
71    TCG_REG_XMM1,
72    TCG_REG_XMM2,
73    TCG_REG_XMM3,
74    TCG_REG_XMM4,
75    TCG_REG_XMM5,
76#ifndef _WIN64
77    /* The Win64 ABI has xmm6-xmm15 as caller-saves, and we do not save
78       any of them.  Therefore only allow xmm0-xmm5 to be allocated.  */
79    TCG_REG_XMM6,
80    TCG_REG_XMM7,
81#if TCG_TARGET_REG_BITS == 64
82    TCG_REG_XMM8,
83    TCG_REG_XMM9,
84    TCG_REG_XMM10,
85    TCG_REG_XMM11,
86    TCG_REG_XMM12,
87    TCG_REG_XMM13,
88    TCG_REG_XMM14,
89    TCG_REG_XMM15,
90#endif
91#endif
92};
93
94static const int tcg_target_call_iarg_regs[] = {
95#if TCG_TARGET_REG_BITS == 64
96#if defined(_WIN64)
97    TCG_REG_RCX,
98    TCG_REG_RDX,
99#else
100    TCG_REG_RDI,
101    TCG_REG_RSI,
102    TCG_REG_RDX,
103    TCG_REG_RCX,
104#endif
105    TCG_REG_R8,
106    TCG_REG_R9,
107#else
108    /* 32 bit mode uses stack based calling convention (GCC default). */
109#endif
110};
111
112static TCGReg tcg_target_call_oarg_reg(TCGCallReturnKind kind, int slot)
113{
114    switch (kind) {
115    case TCG_CALL_RET_NORMAL:
116        tcg_debug_assert(slot >= 0 && slot <= 1);
117        return slot ? TCG_REG_EDX : TCG_REG_EAX;
118#ifdef _WIN64
119    case TCG_CALL_RET_BY_VEC:
120        tcg_debug_assert(slot == 0);
121        return TCG_REG_XMM0;
122#endif
123    default:
124        g_assert_not_reached();
125    }
126}
127
128/* Constants we accept.  */
129#define TCG_CT_CONST_S32 0x100
130#define TCG_CT_CONST_U32 0x200
131#define TCG_CT_CONST_I32 0x400
132#define TCG_CT_CONST_WSZ 0x800
133
134/* Registers used with L constraint, which are the first argument
135   registers on x86_64, and two random call clobbered registers on
136   i386. */
137#if TCG_TARGET_REG_BITS == 64
138# define TCG_REG_L0 tcg_target_call_iarg_regs[0]
139# define TCG_REG_L1 tcg_target_call_iarg_regs[1]
140#else
141# define TCG_REG_L0 TCG_REG_EAX
142# define TCG_REG_L1 TCG_REG_EDX
143#endif
144
145#define ALL_BYTEH_REGS         0x0000000fu
146#if TCG_TARGET_REG_BITS == 64
147# define ALL_GENERAL_REGS      0x0000ffffu
148# define ALL_VECTOR_REGS       0xffff0000u
149# define ALL_BYTEL_REGS        ALL_GENERAL_REGS
150#else
151# define ALL_GENERAL_REGS      0x000000ffu
152# define ALL_VECTOR_REGS       0x00ff0000u
153# define ALL_BYTEL_REGS        ALL_BYTEH_REGS
154#endif
155#ifdef CONFIG_SOFTMMU
156# define SOFTMMU_RESERVE_REGS  ((1 << TCG_REG_L0) | (1 << TCG_REG_L1))
157#else
158# define SOFTMMU_RESERVE_REGS  0
159#endif
160
161/* The host compiler should supply <cpuid.h> to enable runtime features
162   detection, as we're not going to go so far as our own inline assembly.
163   If not available, default values will be assumed.  */
164#if defined(CONFIG_CPUID_H)
165#include "qemu/cpuid.h"
166#endif
167
168/* For 64-bit, we always know that CMOV is available.  */
169#if TCG_TARGET_REG_BITS == 64
170# define have_cmov 1
171#elif defined(CONFIG_CPUID_H)
172static bool have_cmov;
173#else
174# define have_cmov 0
175#endif
176
177/* We need these symbols in tcg-target.h, and we can't properly conditionalize
178   it there.  Therefore we always define the variable.  */
179bool have_bmi1;
180bool have_popcnt;
181bool have_avx1;
182bool have_avx2;
183bool have_avx512bw;
184bool have_avx512dq;
185bool have_avx512vbmi2;
186bool have_avx512vl;
187bool have_movbe;
188
189#ifdef CONFIG_CPUID_H
190static bool have_bmi2;
191static bool have_lzcnt;
192#else
193# define have_bmi2 0
194# define have_lzcnt 0
195#endif
196
197static const tcg_insn_unit *tb_ret_addr;
198
199static bool patch_reloc(tcg_insn_unit *code_ptr, int type,
200                        intptr_t value, intptr_t addend)
201{
202    value += addend;
203    switch(type) {
204    case R_386_PC32:
205        value -= (uintptr_t)tcg_splitwx_to_rx(code_ptr);
206        if (value != (int32_t)value) {
207            return false;
208        }
209        /* FALLTHRU */
210    case R_386_32:
211        tcg_patch32(code_ptr, value);
212        break;
213    case R_386_PC8:
214        value -= (uintptr_t)tcg_splitwx_to_rx(code_ptr);
215        if (value != (int8_t)value) {
216            return false;
217        }
218        tcg_patch8(code_ptr, value);
219        break;
220    default:
221        tcg_abort();
222    }
223    return true;
224}
225
226/* test if a constant matches the constraint */
227static bool tcg_target_const_match(int64_t val, TCGType type, int ct)
228{
229    if (ct & TCG_CT_CONST) {
230        return 1;
231    }
232    if (type == TCG_TYPE_I32) {
233        if (ct & (TCG_CT_CONST_S32 | TCG_CT_CONST_U32 | TCG_CT_CONST_I32)) {
234            return 1;
235        }
236    } else {
237        if ((ct & TCG_CT_CONST_S32) && val == (int32_t)val) {
238            return 1;
239        }
240        if ((ct & TCG_CT_CONST_U32) && val == (uint32_t)val) {
241            return 1;
242        }
243        if ((ct & TCG_CT_CONST_I32) && ~val == (int32_t)~val) {
244            return 1;
245        }
246    }
247    if ((ct & TCG_CT_CONST_WSZ) && val == (type == TCG_TYPE_I32 ? 32 : 64)) {
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_EvIz	(0x81)
277#define OPC_ARITH_EvIb	(0x83)
278#define OPC_ARITH_GvEv	(0x03)		/* ... plus (ARITH_FOO << 3) */
279#define OPC_ANDN        (0xf2 | P_EXT38)
280#define OPC_ADD_GvEv	(OPC_ARITH_GvEv | (ARITH_ADD << 3))
281#define OPC_AND_GvEv    (OPC_ARITH_GvEv | (ARITH_AND << 3))
282#define OPC_BLENDPS     (0x0c | P_EXT3A | P_DATA16)
283#define OPC_BSF         (0xbc | P_EXT)
284#define OPC_BSR         (0xbd | P_EXT)
285#define OPC_BSWAP	(0xc8 | P_EXT)
286#define OPC_CALL_Jz	(0xe8)
287#define OPC_CMOVCC      (0x40 | P_EXT)  /* ... plus condition code */
288#define OPC_CMP_GvEv	(OPC_ARITH_GvEv | (ARITH_CMP << 3))
289#define OPC_DEC_r32	(0x48)
290#define OPC_IMUL_GvEv	(0xaf | P_EXT)
291#define OPC_IMUL_GvEvIb	(0x6b)
292#define OPC_IMUL_GvEvIz	(0x69)
293#define OPC_INC_r32	(0x40)
294#define OPC_JCC_long	(0x80 | P_EXT)	/* ... plus condition code */
295#define OPC_JCC_short	(0x70)		/* ... plus condition code */
296#define OPC_JMP_long	(0xe9)
297#define OPC_JMP_short	(0xeb)
298#define OPC_LEA         (0x8d)
299#define OPC_LZCNT       (0xbd | P_EXT | P_SIMDF3)
300#define OPC_MOVB_EvGv	(0x88)		/* stores, more or less */
301#define OPC_MOVL_EvGv	(0x89)		/* stores, more or less */
302#define OPC_MOVL_GvEv	(0x8b)		/* loads, more or less */
303#define OPC_MOVB_EvIz   (0xc6)
304#define OPC_MOVL_EvIz	(0xc7)
305#define OPC_MOVL_Iv     (0xb8)
306#define OPC_MOVBE_GyMy  (0xf0 | P_EXT38)
307#define OPC_MOVBE_MyGy  (0xf1 | P_EXT38)
308#define OPC_MOVD_VyEy   (0x6e | P_EXT | P_DATA16)
309#define OPC_MOVD_EyVy   (0x7e | P_EXT | P_DATA16)
310#define OPC_MOVDDUP     (0x12 | P_EXT | P_SIMDF2)
311#define OPC_MOVDQA_VxWx (0x6f | P_EXT | P_DATA16)
312#define OPC_MOVDQA_WxVx (0x7f | P_EXT | P_DATA16)
313#define OPC_MOVDQU_VxWx (0x6f | P_EXT | P_SIMDF3)
314#define OPC_MOVDQU_WxVx (0x7f | P_EXT | P_SIMDF3)
315#define OPC_MOVQ_VqWq   (0x7e | P_EXT | P_SIMDF3)
316#define OPC_MOVQ_WqVq   (0xd6 | P_EXT | P_DATA16)
317#define OPC_MOVSBL	(0xbe | P_EXT)
318#define OPC_MOVSWL	(0xbf | P_EXT)
319#define OPC_MOVSLQ	(0x63 | P_REXW)
320#define OPC_MOVZBL	(0xb6 | P_EXT)
321#define OPC_MOVZWL	(0xb7 | P_EXT)
322#define OPC_PABSB       (0x1c | P_EXT38 | P_DATA16)
323#define OPC_PABSW       (0x1d | P_EXT38 | P_DATA16)
324#define OPC_PABSD       (0x1e | P_EXT38 | P_DATA16)
325#define OPC_VPABSQ      (0x1f | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
326#define OPC_PACKSSDW    (0x6b | P_EXT | P_DATA16)
327#define OPC_PACKSSWB    (0x63 | P_EXT | P_DATA16)
328#define OPC_PACKUSDW    (0x2b | P_EXT38 | P_DATA16)
329#define OPC_PACKUSWB    (0x67 | P_EXT | P_DATA16)
330#define OPC_PADDB       (0xfc | P_EXT | P_DATA16)
331#define OPC_PADDW       (0xfd | P_EXT | P_DATA16)
332#define OPC_PADDD       (0xfe | P_EXT | P_DATA16)
333#define OPC_PADDQ       (0xd4 | P_EXT | P_DATA16)
334#define OPC_PADDSB      (0xec | P_EXT | P_DATA16)
335#define OPC_PADDSW      (0xed | P_EXT | P_DATA16)
336#define OPC_PADDUB      (0xdc | P_EXT | P_DATA16)
337#define OPC_PADDUW      (0xdd | P_EXT | P_DATA16)
338#define OPC_PAND        (0xdb | P_EXT | P_DATA16)
339#define OPC_PANDN       (0xdf | P_EXT | P_DATA16)
340#define OPC_PBLENDW     (0x0e | P_EXT3A | P_DATA16)
341#define OPC_PCMPEQB     (0x74 | P_EXT | P_DATA16)
342#define OPC_PCMPEQW     (0x75 | P_EXT | P_DATA16)
343#define OPC_PCMPEQD     (0x76 | P_EXT | P_DATA16)
344#define OPC_PCMPEQQ     (0x29 | P_EXT38 | P_DATA16)
345#define OPC_PCMPGTB     (0x64 | P_EXT | P_DATA16)
346#define OPC_PCMPGTW     (0x65 | P_EXT | P_DATA16)
347#define OPC_PCMPGTD     (0x66 | P_EXT | P_DATA16)
348#define OPC_PCMPGTQ     (0x37 | P_EXT38 | P_DATA16)
349#define OPC_PMAXSB      (0x3c | P_EXT38 | P_DATA16)
350#define OPC_PMAXSW      (0xee | P_EXT | P_DATA16)
351#define OPC_PMAXSD      (0x3d | P_EXT38 | P_DATA16)
352#define OPC_VPMAXSQ     (0x3d | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
353#define OPC_PMAXUB      (0xde | P_EXT | P_DATA16)
354#define OPC_PMAXUW      (0x3e | P_EXT38 | P_DATA16)
355#define OPC_PMAXUD      (0x3f | P_EXT38 | P_DATA16)
356#define OPC_VPMAXUQ     (0x3f | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
357#define OPC_PMINSB      (0x38 | P_EXT38 | P_DATA16)
358#define OPC_PMINSW      (0xea | P_EXT | P_DATA16)
359#define OPC_PMINSD      (0x39 | P_EXT38 | P_DATA16)
360#define OPC_VPMINSQ     (0x39 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
361#define OPC_PMINUB      (0xda | P_EXT | P_DATA16)
362#define OPC_PMINUW      (0x3a | P_EXT38 | P_DATA16)
363#define OPC_PMINUD      (0x3b | P_EXT38 | P_DATA16)
364#define OPC_VPMINUQ     (0x3b | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
365#define OPC_PMOVSXBW    (0x20 | P_EXT38 | P_DATA16)
366#define OPC_PMOVSXWD    (0x23 | P_EXT38 | P_DATA16)
367#define OPC_PMOVSXDQ    (0x25 | P_EXT38 | P_DATA16)
368#define OPC_PMOVZXBW    (0x30 | P_EXT38 | P_DATA16)
369#define OPC_PMOVZXWD    (0x33 | P_EXT38 | P_DATA16)
370#define OPC_PMOVZXDQ    (0x35 | P_EXT38 | P_DATA16)
371#define OPC_PMULLW      (0xd5 | P_EXT | P_DATA16)
372#define OPC_PMULLD      (0x40 | P_EXT38 | P_DATA16)
373#define OPC_VPMULLQ     (0x40 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
374#define OPC_POR         (0xeb | P_EXT | P_DATA16)
375#define OPC_PSHUFB      (0x00 | P_EXT38 | P_DATA16)
376#define OPC_PSHUFD      (0x70 | P_EXT | P_DATA16)
377#define OPC_PSHUFLW     (0x70 | P_EXT | P_SIMDF2)
378#define OPC_PSHUFHW     (0x70 | P_EXT | P_SIMDF3)
379#define OPC_PSHIFTW_Ib  (0x71 | P_EXT | P_DATA16) /* /2 /6 /4 */
380#define OPC_PSHIFTD_Ib  (0x72 | P_EXT | P_DATA16) /* /1 /2 /6 /4 */
381#define OPC_PSHIFTQ_Ib  (0x73 | P_EXT | P_DATA16) /* /2 /6 /4 */
382#define OPC_PSLLW       (0xf1 | P_EXT | P_DATA16)
383#define OPC_PSLLD       (0xf2 | P_EXT | P_DATA16)
384#define OPC_PSLLQ       (0xf3 | P_EXT | P_DATA16)
385#define OPC_PSRAW       (0xe1 | P_EXT | P_DATA16)
386#define OPC_PSRAD       (0xe2 | P_EXT | P_DATA16)
387#define OPC_VPSRAQ      (0xe2 | P_EXT | P_DATA16 | P_VEXW | P_EVEX)
388#define OPC_PSRLW       (0xd1 | P_EXT | P_DATA16)
389#define OPC_PSRLD       (0xd2 | P_EXT | P_DATA16)
390#define OPC_PSRLQ       (0xd3 | P_EXT | P_DATA16)
391#define OPC_PSUBB       (0xf8 | P_EXT | P_DATA16)
392#define OPC_PSUBW       (0xf9 | P_EXT | P_DATA16)
393#define OPC_PSUBD       (0xfa | P_EXT | P_DATA16)
394#define OPC_PSUBQ       (0xfb | P_EXT | P_DATA16)
395#define OPC_PSUBSB      (0xe8 | P_EXT | P_DATA16)
396#define OPC_PSUBSW      (0xe9 | P_EXT | P_DATA16)
397#define OPC_PSUBUB      (0xd8 | P_EXT | P_DATA16)
398#define OPC_PSUBUW      (0xd9 | P_EXT | P_DATA16)
399#define OPC_PUNPCKLBW   (0x60 | P_EXT | P_DATA16)
400#define OPC_PUNPCKLWD   (0x61 | P_EXT | P_DATA16)
401#define OPC_PUNPCKLDQ   (0x62 | P_EXT | P_DATA16)
402#define OPC_PUNPCKLQDQ  (0x6c | P_EXT | P_DATA16)
403#define OPC_PUNPCKHBW   (0x68 | P_EXT | P_DATA16)
404#define OPC_PUNPCKHWD   (0x69 | P_EXT | P_DATA16)
405#define OPC_PUNPCKHDQ   (0x6a | P_EXT | P_DATA16)
406#define OPC_PUNPCKHQDQ  (0x6d | P_EXT | P_DATA16)
407#define OPC_PXOR        (0xef | P_EXT | P_DATA16)
408#define OPC_POP_r32	(0x58)
409#define OPC_POPCNT      (0xb8 | P_EXT | P_SIMDF3)
410#define OPC_PUSH_r32	(0x50)
411#define OPC_PUSH_Iv	(0x68)
412#define OPC_PUSH_Ib	(0x6a)
413#define OPC_RET		(0xc3)
414#define OPC_SETCC	(0x90 | P_EXT | P_REXB_RM) /* ... plus cc */
415#define OPC_SHIFT_1	(0xd1)
416#define OPC_SHIFT_Ib	(0xc1)
417#define OPC_SHIFT_cl	(0xd3)
418#define OPC_SARX        (0xf7 | P_EXT38 | P_SIMDF3)
419#define OPC_SHUFPS      (0xc6 | P_EXT)
420#define OPC_SHLX        (0xf7 | P_EXT38 | P_DATA16)
421#define OPC_SHRX        (0xf7 | P_EXT38 | P_SIMDF2)
422#define OPC_SHRD_Ib     (0xac | P_EXT)
423#define OPC_TESTL	(0x85)
424#define OPC_TZCNT       (0xbc | P_EXT | P_SIMDF3)
425#define OPC_UD2         (0x0b | P_EXT)
426#define OPC_VPBLENDD    (0x02 | P_EXT3A | P_DATA16)
427#define OPC_VPBLENDVB   (0x4c | P_EXT3A | P_DATA16)
428#define OPC_VPINSRB     (0x20 | P_EXT3A | P_DATA16)
429#define OPC_VPINSRW     (0xc4 | P_EXT | P_DATA16)
430#define OPC_VBROADCASTSS (0x18 | P_EXT38 | P_DATA16)
431#define OPC_VBROADCASTSD (0x19 | P_EXT38 | P_DATA16)
432#define OPC_VPBROADCASTB (0x78 | P_EXT38 | P_DATA16)
433#define OPC_VPBROADCASTW (0x79 | P_EXT38 | P_DATA16)
434#define OPC_VPBROADCASTD (0x58 | P_EXT38 | P_DATA16)
435#define OPC_VPBROADCASTQ (0x59 | P_EXT38 | P_DATA16)
436#define OPC_VPERMQ      (0x00 | P_EXT3A | P_DATA16 | P_VEXW)
437#define OPC_VPERM2I128  (0x46 | P_EXT3A | P_DATA16 | P_VEXL)
438#define OPC_VPROLVD     (0x15 | P_EXT38 | P_DATA16 | P_EVEX)
439#define OPC_VPROLVQ     (0x15 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
440#define OPC_VPRORVD     (0x14 | P_EXT38 | P_DATA16 | P_EVEX)
441#define OPC_VPRORVQ     (0x14 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
442#define OPC_VPSHLDW     (0x70 | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
443#define OPC_VPSHLDD     (0x71 | P_EXT3A | P_DATA16 | P_EVEX)
444#define OPC_VPSHLDQ     (0x71 | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
445#define OPC_VPSHLDVW    (0x70 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
446#define OPC_VPSHLDVD    (0x71 | P_EXT38 | P_DATA16 | P_EVEX)
447#define OPC_VPSHLDVQ    (0x71 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
448#define OPC_VPSHRDVW    (0x72 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
449#define OPC_VPSHRDVD    (0x73 | P_EXT38 | P_DATA16 | P_EVEX)
450#define OPC_VPSHRDVQ    (0x73 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
451#define OPC_VPSLLVW     (0x12 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
452#define OPC_VPSLLVD     (0x47 | P_EXT38 | P_DATA16)
453#define OPC_VPSLLVQ     (0x47 | P_EXT38 | P_DATA16 | P_VEXW)
454#define OPC_VPSRAVW     (0x11 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
455#define OPC_VPSRAVD     (0x46 | P_EXT38 | P_DATA16)
456#define OPC_VPSRAVQ     (0x46 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
457#define OPC_VPSRLVW     (0x10 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
458#define OPC_VPSRLVD     (0x45 | P_EXT38 | P_DATA16)
459#define OPC_VPSRLVQ     (0x45 | P_EXT38 | P_DATA16 | P_VEXW)
460#define OPC_VPTERNLOGQ  (0x25 | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
461#define OPC_VZEROUPPER  (0x77 | P_EXT)
462#define OPC_XCHG_ax_r32	(0x90)
463
464#define OPC_GRP3_Eb     (0xf6)
465#define OPC_GRP3_Ev     (0xf7)
466#define OPC_GRP5        (0xff)
467#define OPC_GRP14       (0x73 | P_EXT | P_DATA16)
468
469/* Group 1 opcode extensions for 0x80-0x83.
470   These are also used as modifiers for OPC_ARITH.  */
471#define ARITH_ADD 0
472#define ARITH_OR  1
473#define ARITH_ADC 2
474#define ARITH_SBB 3
475#define ARITH_AND 4
476#define ARITH_SUB 5
477#define ARITH_XOR 6
478#define ARITH_CMP 7
479
480/* Group 2 opcode extensions for 0xc0, 0xc1, 0xd0-0xd3.  */
481#define SHIFT_ROL 0
482#define SHIFT_ROR 1
483#define SHIFT_SHL 4
484#define SHIFT_SHR 5
485#define SHIFT_SAR 7
486
487/* Group 3 opcode extensions for 0xf6, 0xf7.  To be used with OPC_GRP3.  */
488#define EXT3_TESTi 0
489#define EXT3_NOT   2
490#define EXT3_NEG   3
491#define EXT3_MUL   4
492#define EXT3_IMUL  5
493#define EXT3_DIV   6
494#define EXT3_IDIV  7
495
496/* Group 5 opcode extensions for 0xff.  To be used with OPC_GRP5.  */
497#define EXT5_INC_Ev	0
498#define EXT5_DEC_Ev	1
499#define EXT5_CALLN_Ev	2
500#define EXT5_JMPN_Ev	4
501
502/* Condition codes to be added to OPC_JCC_{long,short}.  */
503#define JCC_JMP (-1)
504#define JCC_JO  0x0
505#define JCC_JNO 0x1
506#define JCC_JB  0x2
507#define JCC_JAE 0x3
508#define JCC_JE  0x4
509#define JCC_JNE 0x5
510#define JCC_JBE 0x6
511#define JCC_JA  0x7
512#define JCC_JS  0x8
513#define JCC_JNS 0x9
514#define JCC_JP  0xa
515#define JCC_JNP 0xb
516#define JCC_JL  0xc
517#define JCC_JGE 0xd
518#define JCC_JLE 0xe
519#define JCC_JG  0xf
520
521static const uint8_t tcg_cond_to_jcc[] = {
522    [TCG_COND_EQ] = JCC_JE,
523    [TCG_COND_NE] = JCC_JNE,
524    [TCG_COND_LT] = JCC_JL,
525    [TCG_COND_GE] = JCC_JGE,
526    [TCG_COND_LE] = JCC_JLE,
527    [TCG_COND_GT] = JCC_JG,
528    [TCG_COND_LTU] = JCC_JB,
529    [TCG_COND_GEU] = JCC_JAE,
530    [TCG_COND_LEU] = JCC_JBE,
531    [TCG_COND_GTU] = JCC_JA,
532};
533
534#if TCG_TARGET_REG_BITS == 64
535static void tcg_out_opc(TCGContext *s, int opc, int r, int rm, int x)
536{
537    int rex;
538
539    if (opc & P_GS) {
540        tcg_out8(s, 0x65);
541    }
542    if (opc & P_DATA16) {
543        /* We should never be asking for both 16 and 64-bit operation.  */
544        tcg_debug_assert((opc & P_REXW) == 0);
545        tcg_out8(s, 0x66);
546    }
547    if (opc & P_SIMDF3) {
548        tcg_out8(s, 0xf3);
549    } else if (opc & P_SIMDF2) {
550        tcg_out8(s, 0xf2);
551    }
552
553    rex = 0;
554    rex |= (opc & P_REXW) ? 0x8 : 0x0;  /* REX.W */
555    rex |= (r & 8) >> 1;                /* REX.R */
556    rex |= (x & 8) >> 2;                /* REX.X */
557    rex |= (rm & 8) >> 3;               /* REX.B */
558
559    /* P_REXB_{R,RM} indicates that the given register is the low byte.
560       For %[abcd]l we need no REX prefix, but for %{si,di,bp,sp}l we do,
561       as otherwise the encoding indicates %[abcd]h.  Note that the values
562       that are ORed in merely indicate that the REX byte must be present;
563       those bits get discarded in output.  */
564    rex |= opc & (r >= 4 ? P_REXB_R : 0);
565    rex |= opc & (rm >= 4 ? P_REXB_RM : 0);
566
567    if (rex) {
568        tcg_out8(s, (uint8_t)(rex | 0x40));
569    }
570
571    if (opc & (P_EXT | P_EXT38 | P_EXT3A)) {
572        tcg_out8(s, 0x0f);
573        if (opc & P_EXT38) {
574            tcg_out8(s, 0x38);
575        } else if (opc & P_EXT3A) {
576            tcg_out8(s, 0x3a);
577        }
578    }
579
580    tcg_out8(s, opc);
581}
582#else
583static void tcg_out_opc(TCGContext *s, int opc)
584{
585    if (opc & P_DATA16) {
586        tcg_out8(s, 0x66);
587    }
588    if (opc & P_SIMDF3) {
589        tcg_out8(s, 0xf3);
590    } else if (opc & P_SIMDF2) {
591        tcg_out8(s, 0xf2);
592    }
593    if (opc & (P_EXT | P_EXT38 | P_EXT3A)) {
594        tcg_out8(s, 0x0f);
595        if (opc & P_EXT38) {
596            tcg_out8(s, 0x38);
597        } else if (opc & P_EXT3A) {
598            tcg_out8(s, 0x3a);
599        }
600    }
601    tcg_out8(s, opc);
602}
603/* Discard the register arguments to tcg_out_opc early, so as not to penalize
604   the 32-bit compilation paths.  This method works with all versions of gcc,
605   whereas relying on optimization may not be able to exclude them.  */
606#define tcg_out_opc(s, opc, r, rm, x)  (tcg_out_opc)(s, opc)
607#endif
608
609static void tcg_out_modrm(TCGContext *s, int opc, int r, int rm)
610{
611    tcg_out_opc(s, opc, r, rm, 0);
612    tcg_out8(s, 0xc0 | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
613}
614
615static void tcg_out_vex_opc(TCGContext *s, int opc, int r, int v,
616                            int rm, int index)
617{
618    int tmp;
619
620    /* Use the two byte form if possible, which cannot encode
621       VEX.W, VEX.B, VEX.X, or an m-mmmm field other than P_EXT.  */
622    if ((opc & (P_EXT | P_EXT38 | P_EXT3A | P_VEXW)) == P_EXT
623        && ((rm | index) & 8) == 0) {
624        /* Two byte VEX prefix.  */
625        tcg_out8(s, 0xc5);
626
627        tmp = (r & 8 ? 0 : 0x80);              /* VEX.R */
628    } else {
629        /* Three byte VEX prefix.  */
630        tcg_out8(s, 0xc4);
631
632        /* VEX.m-mmmm */
633        if (opc & P_EXT3A) {
634            tmp = 3;
635        } else if (opc & P_EXT38) {
636            tmp = 2;
637        } else if (opc & P_EXT) {
638            tmp = 1;
639        } else {
640            g_assert_not_reached();
641        }
642        tmp |= (r & 8 ? 0 : 0x80);             /* VEX.R */
643        tmp |= (index & 8 ? 0 : 0x40);         /* VEX.X */
644        tmp |= (rm & 8 ? 0 : 0x20);            /* VEX.B */
645        tcg_out8(s, tmp);
646
647        tmp = (opc & P_VEXW ? 0x80 : 0);       /* VEX.W */
648    }
649
650    tmp |= (opc & P_VEXL ? 0x04 : 0);      /* VEX.L */
651    /* VEX.pp */
652    if (opc & P_DATA16) {
653        tmp |= 1;                          /* 0x66 */
654    } else if (opc & P_SIMDF3) {
655        tmp |= 2;                          /* 0xf3 */
656    } else if (opc & P_SIMDF2) {
657        tmp |= 3;                          /* 0xf2 */
658    }
659    tmp |= (~v & 15) << 3;                 /* VEX.vvvv */
660    tcg_out8(s, tmp);
661    tcg_out8(s, opc);
662}
663
664static void tcg_out_evex_opc(TCGContext *s, int opc, int r, int v,
665                             int rm, int index)
666{
667    /* The entire 4-byte evex prefix; with R' and V' set. */
668    uint32_t p = 0x08041062;
669    int mm, pp;
670
671    tcg_debug_assert(have_avx512vl);
672
673    /* EVEX.mm */
674    if (opc & P_EXT3A) {
675        mm = 3;
676    } else if (opc & P_EXT38) {
677        mm = 2;
678    } else if (opc & P_EXT) {
679        mm = 1;
680    } else {
681        g_assert_not_reached();
682    }
683
684    /* EVEX.pp */
685    if (opc & P_DATA16) {
686        pp = 1;                          /* 0x66 */
687    } else if (opc & P_SIMDF3) {
688        pp = 2;                          /* 0xf3 */
689    } else if (opc & P_SIMDF2) {
690        pp = 3;                          /* 0xf2 */
691    } else {
692        pp = 0;
693    }
694
695    p = deposit32(p, 8, 2, mm);
696    p = deposit32(p, 13, 1, (rm & 8) == 0);             /* EVEX.RXB.B */
697    p = deposit32(p, 14, 1, (index & 8) == 0);          /* EVEX.RXB.X */
698    p = deposit32(p, 15, 1, (r & 8) == 0);              /* EVEX.RXB.R */
699    p = deposit32(p, 16, 2, pp);
700    p = deposit32(p, 19, 4, ~v);
701    p = deposit32(p, 23, 1, (opc & P_VEXW) != 0);
702    p = deposit32(p, 29, 2, (opc & P_VEXL) != 0);
703
704    tcg_out32(s, p);
705    tcg_out8(s, opc);
706}
707
708static void tcg_out_vex_modrm(TCGContext *s, int opc, int r, int v, int rm)
709{
710    if (opc & P_EVEX) {
711        tcg_out_evex_opc(s, opc, r, v, rm, 0);
712    } else {
713        tcg_out_vex_opc(s, opc, r, v, rm, 0);
714    }
715    tcg_out8(s, 0xc0 | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
716}
717
718/* Output an opcode with a full "rm + (index<<shift) + offset" address mode.
719   We handle either RM and INDEX missing with a negative value.  In 64-bit
720   mode for absolute addresses, ~RM is the size of the immediate operand
721   that will follow the instruction.  */
722
723static void tcg_out_sib_offset(TCGContext *s, int r, int rm, int index,
724                               int shift, intptr_t offset)
725{
726    int mod, len;
727
728    if (index < 0 && rm < 0) {
729        if (TCG_TARGET_REG_BITS == 64) {
730            /* Try for a rip-relative addressing mode.  This has replaced
731               the 32-bit-mode absolute addressing encoding.  */
732            intptr_t pc = (intptr_t)s->code_ptr + 5 + ~rm;
733            intptr_t disp = offset - pc;
734            if (disp == (int32_t)disp) {
735                tcg_out8(s, (LOWREGMASK(r) << 3) | 5);
736                tcg_out32(s, disp);
737                return;
738            }
739
740            /* Try for an absolute address encoding.  This requires the
741               use of the MODRM+SIB encoding and is therefore larger than
742               rip-relative addressing.  */
743            if (offset == (int32_t)offset) {
744                tcg_out8(s, (LOWREGMASK(r) << 3) | 4);
745                tcg_out8(s, (4 << 3) | 5);
746                tcg_out32(s, offset);
747                return;
748            }
749
750            /* ??? The memory isn't directly addressable.  */
751            g_assert_not_reached();
752        } else {
753            /* Absolute address.  */
754            tcg_out8(s, (r << 3) | 5);
755            tcg_out32(s, offset);
756            return;
757        }
758    }
759
760    /* Find the length of the immediate addend.  Note that the encoding
761       that would be used for (%ebp) indicates absolute addressing.  */
762    if (rm < 0) {
763        mod = 0, len = 4, rm = 5;
764    } else if (offset == 0 && LOWREGMASK(rm) != TCG_REG_EBP) {
765        mod = 0, len = 0;
766    } else if (offset == (int8_t)offset) {
767        mod = 0x40, len = 1;
768    } else {
769        mod = 0x80, len = 4;
770    }
771
772    /* Use a single byte MODRM format if possible.  Note that the encoding
773       that would be used for %esp is the escape to the two byte form.  */
774    if (index < 0 && LOWREGMASK(rm) != TCG_REG_ESP) {
775        /* Single byte MODRM format.  */
776        tcg_out8(s, mod | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
777    } else {
778        /* Two byte MODRM+SIB format.  */
779
780        /* Note that the encoding that would place %esp into the index
781           field indicates no index register.  In 64-bit mode, the REX.X
782           bit counts, so %r12 can be used as the index.  */
783        if (index < 0) {
784            index = 4;
785        } else {
786            tcg_debug_assert(index != TCG_REG_ESP);
787        }
788
789        tcg_out8(s, mod | (LOWREGMASK(r) << 3) | 4);
790        tcg_out8(s, (shift << 6) | (LOWREGMASK(index) << 3) | LOWREGMASK(rm));
791    }
792
793    if (len == 1) {
794        tcg_out8(s, offset);
795    } else if (len == 4) {
796        tcg_out32(s, offset);
797    }
798}
799
800static void tcg_out_modrm_sib_offset(TCGContext *s, int opc, int r, int rm,
801                                     int index, int shift, intptr_t offset)
802{
803    tcg_out_opc(s, opc, r, rm < 0 ? 0 : rm, index < 0 ? 0 : index);
804    tcg_out_sib_offset(s, r, rm, index, shift, offset);
805}
806
807static void tcg_out_vex_modrm_sib_offset(TCGContext *s, int opc, int r, int v,
808                                         int rm, int index, int shift,
809                                         intptr_t offset)
810{
811    tcg_out_vex_opc(s, opc, r, v, rm < 0 ? 0 : rm, index < 0 ? 0 : index);
812    tcg_out_sib_offset(s, r, rm, index, shift, offset);
813}
814
815/* A simplification of the above with no index or shift.  */
816static inline void tcg_out_modrm_offset(TCGContext *s, int opc, int r,
817                                        int rm, intptr_t offset)
818{
819    tcg_out_modrm_sib_offset(s, opc, r, rm, -1, 0, offset);
820}
821
822static inline void tcg_out_vex_modrm_offset(TCGContext *s, int opc, int r,
823                                            int v, int rm, intptr_t offset)
824{
825    tcg_out_vex_modrm_sib_offset(s, opc, r, v, rm, -1, 0, offset);
826}
827
828/* Output an opcode with an expected reference to the constant pool.  */
829static inline void tcg_out_modrm_pool(TCGContext *s, int opc, int r)
830{
831    tcg_out_opc(s, opc, r, 0, 0);
832    /* Absolute for 32-bit, pc-relative for 64-bit.  */
833    tcg_out8(s, LOWREGMASK(r) << 3 | 5);
834    tcg_out32(s, 0);
835}
836
837/* Output an opcode with an expected reference to the constant pool.  */
838static inline void tcg_out_vex_modrm_pool(TCGContext *s, int opc, int r)
839{
840    tcg_out_vex_opc(s, opc, r, 0, 0, 0);
841    /* Absolute for 32-bit, pc-relative for 64-bit.  */
842    tcg_out8(s, LOWREGMASK(r) << 3 | 5);
843    tcg_out32(s, 0);
844}
845
846/* Generate dest op= src.  Uses the same ARITH_* codes as tgen_arithi.  */
847static inline void tgen_arithr(TCGContext *s, int subop, int dest, int src)
848{
849    /* Propagate an opcode prefix, such as P_REXW.  */
850    int ext = subop & ~0x7;
851    subop &= 0x7;
852
853    tcg_out_modrm(s, OPC_ARITH_GvEv + (subop << 3) + ext, dest, src);
854}
855
856static bool tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
857{
858    int rexw = 0;
859
860    if (arg == ret) {
861        return true;
862    }
863    switch (type) {
864    case TCG_TYPE_I64:
865        rexw = P_REXW;
866        /* fallthru */
867    case TCG_TYPE_I32:
868        if (ret < 16) {
869            if (arg < 16) {
870                tcg_out_modrm(s, OPC_MOVL_GvEv + rexw, ret, arg);
871            } else {
872                tcg_out_vex_modrm(s, OPC_MOVD_EyVy + rexw, arg, 0, ret);
873            }
874        } else {
875            if (arg < 16) {
876                tcg_out_vex_modrm(s, OPC_MOVD_VyEy + rexw, ret, 0, arg);
877            } else {
878                tcg_out_vex_modrm(s, OPC_MOVQ_VqWq, ret, 0, arg);
879            }
880        }
881        break;
882
883    case TCG_TYPE_V64:
884        tcg_debug_assert(ret >= 16 && arg >= 16);
885        tcg_out_vex_modrm(s, OPC_MOVQ_VqWq, ret, 0, arg);
886        break;
887    case TCG_TYPE_V128:
888        tcg_debug_assert(ret >= 16 && arg >= 16);
889        tcg_out_vex_modrm(s, OPC_MOVDQA_VxWx, ret, 0, arg);
890        break;
891    case TCG_TYPE_V256:
892        tcg_debug_assert(ret >= 16 && arg >= 16);
893        tcg_out_vex_modrm(s, OPC_MOVDQA_VxWx | P_VEXL, ret, 0, arg);
894        break;
895
896    default:
897        g_assert_not_reached();
898    }
899    return true;
900}
901
902static const int avx2_dup_insn[4] = {
903    OPC_VPBROADCASTB, OPC_VPBROADCASTW,
904    OPC_VPBROADCASTD, OPC_VPBROADCASTQ,
905};
906
907static bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece,
908                            TCGReg r, TCGReg a)
909{
910    if (have_avx2) {
911        int vex_l = (type == TCG_TYPE_V256 ? P_VEXL : 0);
912        tcg_out_vex_modrm(s, avx2_dup_insn[vece] + vex_l, r, 0, a);
913    } else {
914        switch (vece) {
915        case MO_8:
916            /* ??? With zero in a register, use PSHUFB.  */
917            tcg_out_vex_modrm(s, OPC_PUNPCKLBW, r, a, a);
918            a = r;
919            /* FALLTHRU */
920        case MO_16:
921            tcg_out_vex_modrm(s, OPC_PUNPCKLWD, r, a, a);
922            a = r;
923            /* FALLTHRU */
924        case MO_32:
925            tcg_out_vex_modrm(s, OPC_PSHUFD, r, 0, a);
926            /* imm8 operand: all output lanes selected from input lane 0.  */
927            tcg_out8(s, 0);
928            break;
929        case MO_64:
930            tcg_out_vex_modrm(s, OPC_PUNPCKLQDQ, r, a, a);
931            break;
932        default:
933            g_assert_not_reached();
934        }
935    }
936    return true;
937}
938
939static bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece,
940                             TCGReg r, TCGReg base, intptr_t offset)
941{
942    if (have_avx2) {
943        int vex_l = (type == TCG_TYPE_V256 ? P_VEXL : 0);
944        tcg_out_vex_modrm_offset(s, avx2_dup_insn[vece] + vex_l,
945                                 r, 0, base, offset);
946    } else {
947        switch (vece) {
948        case MO_64:
949            tcg_out_vex_modrm_offset(s, OPC_MOVDDUP, r, 0, base, offset);
950            break;
951        case MO_32:
952            tcg_out_vex_modrm_offset(s, OPC_VBROADCASTSS, r, 0, base, offset);
953            break;
954        case MO_16:
955            tcg_out_vex_modrm_offset(s, OPC_VPINSRW, r, r, base, offset);
956            tcg_out8(s, 0); /* imm8 */
957            tcg_out_dup_vec(s, type, vece, r, r);
958            break;
959        case MO_8:
960            tcg_out_vex_modrm_offset(s, OPC_VPINSRB, r, r, base, offset);
961            tcg_out8(s, 0); /* imm8 */
962            tcg_out_dup_vec(s, type, vece, r, r);
963            break;
964        default:
965            g_assert_not_reached();
966        }
967    }
968    return true;
969}
970
971static void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece,
972                             TCGReg ret, int64_t arg)
973{
974    int vex_l = (type == TCG_TYPE_V256 ? P_VEXL : 0);
975
976    if (arg == 0) {
977        tcg_out_vex_modrm(s, OPC_PXOR, ret, ret, ret);
978        return;
979    }
980    if (arg == -1) {
981        tcg_out_vex_modrm(s, OPC_PCMPEQB + vex_l, ret, ret, ret);
982        return;
983    }
984
985    if (TCG_TARGET_REG_BITS == 32 && vece < MO_64) {
986        if (have_avx2) {
987            tcg_out_vex_modrm_pool(s, OPC_VPBROADCASTD + vex_l, ret);
988        } else {
989            tcg_out_vex_modrm_pool(s, OPC_VBROADCASTSS, ret);
990        }
991        new_pool_label(s, arg, R_386_32, s->code_ptr - 4, 0);
992    } else {
993        if (type == TCG_TYPE_V64) {
994            tcg_out_vex_modrm_pool(s, OPC_MOVQ_VqWq, ret);
995        } else if (have_avx2) {
996            tcg_out_vex_modrm_pool(s, OPC_VPBROADCASTQ + vex_l, ret);
997        } else {
998            tcg_out_vex_modrm_pool(s, OPC_MOVDDUP, ret);
999        }
1000        if (TCG_TARGET_REG_BITS == 64) {
1001            new_pool_label(s, arg, R_386_PC32, s->code_ptr - 4, -4);
1002        } else {
1003            new_pool_l2(s, R_386_32, s->code_ptr - 4, 0, arg, arg >> 32);
1004        }
1005    }
1006}
1007
1008static void tcg_out_movi_vec(TCGContext *s, TCGType type,
1009                             TCGReg ret, tcg_target_long arg)
1010{
1011    if (arg == 0) {
1012        tcg_out_vex_modrm(s, OPC_PXOR, ret, ret, ret);
1013        return;
1014    }
1015    if (arg == -1) {
1016        tcg_out_vex_modrm(s, OPC_PCMPEQB, ret, ret, ret);
1017        return;
1018    }
1019
1020    int rexw = (type == TCG_TYPE_I32 ? 0 : P_REXW);
1021    tcg_out_vex_modrm_pool(s, OPC_MOVD_VyEy + rexw, ret);
1022    if (TCG_TARGET_REG_BITS == 64) {
1023        new_pool_label(s, arg, R_386_PC32, s->code_ptr - 4, -4);
1024    } else {
1025        new_pool_label(s, arg, R_386_32, s->code_ptr - 4, 0);
1026    }
1027}
1028
1029static void tcg_out_movi_int(TCGContext *s, TCGType type,
1030                             TCGReg ret, tcg_target_long arg)
1031{
1032    tcg_target_long diff;
1033
1034    if (arg == 0) {
1035        tgen_arithr(s, ARITH_XOR, ret, ret);
1036        return;
1037    }
1038    if (arg == (uint32_t)arg || type == TCG_TYPE_I32) {
1039        tcg_out_opc(s, OPC_MOVL_Iv + LOWREGMASK(ret), 0, ret, 0);
1040        tcg_out32(s, arg);
1041        return;
1042    }
1043    if (arg == (int32_t)arg) {
1044        tcg_out_modrm(s, OPC_MOVL_EvIz + P_REXW, 0, ret);
1045        tcg_out32(s, arg);
1046        return;
1047    }
1048
1049    /* Try a 7 byte pc-relative lea before the 10 byte movq.  */
1050    diff = tcg_pcrel_diff(s, (const void *)arg) - 7;
1051    if (diff == (int32_t)diff) {
1052        tcg_out_opc(s, OPC_LEA | P_REXW, ret, 0, 0);
1053        tcg_out8(s, (LOWREGMASK(ret) << 3) | 5);
1054        tcg_out32(s, diff);
1055        return;
1056    }
1057
1058    tcg_out_opc(s, OPC_MOVL_Iv + P_REXW + LOWREGMASK(ret), 0, ret, 0);
1059    tcg_out64(s, arg);
1060}
1061
1062static void tcg_out_movi(TCGContext *s, TCGType type,
1063                         TCGReg ret, tcg_target_long arg)
1064{
1065    switch (type) {
1066    case TCG_TYPE_I32:
1067#if TCG_TARGET_REG_BITS == 64
1068    case TCG_TYPE_I64:
1069#endif
1070        if (ret < 16) {
1071            tcg_out_movi_int(s, type, ret, arg);
1072        } else {
1073            tcg_out_movi_vec(s, type, ret, arg);
1074        }
1075        break;
1076    default:
1077        g_assert_not_reached();
1078    }
1079}
1080
1081static void tcg_out_addi_ptr(TCGContext *s, TCGReg rd, TCGReg rs,
1082                             tcg_target_long imm)
1083{
1084    /* This function is only used for passing structs by reference. */
1085    tcg_debug_assert(imm == (int32_t)imm);
1086    tcg_out_modrm_offset(s, OPC_LEA, rd, rs, imm);
1087}
1088
1089static inline void tcg_out_pushi(TCGContext *s, tcg_target_long val)
1090{
1091    if (val == (int8_t)val) {
1092        tcg_out_opc(s, OPC_PUSH_Ib, 0, 0, 0);
1093        tcg_out8(s, val);
1094    } else if (val == (int32_t)val) {
1095        tcg_out_opc(s, OPC_PUSH_Iv, 0, 0, 0);
1096        tcg_out32(s, val);
1097    } else {
1098        tcg_abort();
1099    }
1100}
1101
1102static inline void tcg_out_mb(TCGContext *s, TCGArg a0)
1103{
1104    /* Given the strength of x86 memory ordering, we only need care for
1105       store-load ordering.  Experimentally, "lock orl $0,0(%esp)" is
1106       faster than "mfence", so don't bother with the sse insn.  */
1107    if (a0 & TCG_MO_ST_LD) {
1108        tcg_out8(s, 0xf0);
1109        tcg_out_modrm_offset(s, OPC_ARITH_EvIb, ARITH_OR, TCG_REG_ESP, 0);
1110        tcg_out8(s, 0);
1111    }
1112}
1113
1114static inline void tcg_out_push(TCGContext *s, int reg)
1115{
1116    tcg_out_opc(s, OPC_PUSH_r32 + LOWREGMASK(reg), 0, reg, 0);
1117}
1118
1119static inline void tcg_out_pop(TCGContext *s, int reg)
1120{
1121    tcg_out_opc(s, OPC_POP_r32 + LOWREGMASK(reg), 0, reg, 0);
1122}
1123
1124static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg ret,
1125                       TCGReg arg1, intptr_t arg2)
1126{
1127    switch (type) {
1128    case TCG_TYPE_I32:
1129        if (ret < 16) {
1130            tcg_out_modrm_offset(s, OPC_MOVL_GvEv, ret, arg1, arg2);
1131        } else {
1132            tcg_out_vex_modrm_offset(s, OPC_MOVD_VyEy, ret, 0, arg1, arg2);
1133        }
1134        break;
1135    case TCG_TYPE_I64:
1136        if (ret < 16) {
1137            tcg_out_modrm_offset(s, OPC_MOVL_GvEv | P_REXW, ret, arg1, arg2);
1138            break;
1139        }
1140        /* FALLTHRU */
1141    case TCG_TYPE_V64:
1142        /* There is no instruction that can validate 8-byte alignment.  */
1143        tcg_debug_assert(ret >= 16);
1144        tcg_out_vex_modrm_offset(s, OPC_MOVQ_VqWq, ret, 0, arg1, arg2);
1145        break;
1146    case TCG_TYPE_V128:
1147        /*
1148         * The gvec infrastructure is asserts that v128 vector loads
1149         * and stores use a 16-byte aligned offset.  Validate that the
1150         * final pointer is aligned by using an insn that will SIGSEGV.
1151         */
1152        tcg_debug_assert(ret >= 16);
1153        tcg_out_vex_modrm_offset(s, OPC_MOVDQA_VxWx, ret, 0, arg1, arg2);
1154        break;
1155    case TCG_TYPE_V256:
1156        /*
1157         * The gvec infrastructure only requires 16-byte alignment,
1158         * so here we must use an unaligned load.
1159         */
1160        tcg_debug_assert(ret >= 16);
1161        tcg_out_vex_modrm_offset(s, OPC_MOVDQU_VxWx | P_VEXL,
1162                                 ret, 0, arg1, arg2);
1163        break;
1164    default:
1165        g_assert_not_reached();
1166    }
1167}
1168
1169static void tcg_out_st(TCGContext *s, TCGType type, TCGReg arg,
1170                       TCGReg arg1, intptr_t arg2)
1171{
1172    switch (type) {
1173    case TCG_TYPE_I32:
1174        if (arg < 16) {
1175            tcg_out_modrm_offset(s, OPC_MOVL_EvGv, arg, arg1, arg2);
1176        } else {
1177            tcg_out_vex_modrm_offset(s, OPC_MOVD_EyVy, arg, 0, arg1, arg2);
1178        }
1179        break;
1180    case TCG_TYPE_I64:
1181        if (arg < 16) {
1182            tcg_out_modrm_offset(s, OPC_MOVL_EvGv | P_REXW, arg, arg1, arg2);
1183            break;
1184        }
1185        /* FALLTHRU */
1186    case TCG_TYPE_V64:
1187        /* There is no instruction that can validate 8-byte alignment.  */
1188        tcg_debug_assert(arg >= 16);
1189        tcg_out_vex_modrm_offset(s, OPC_MOVQ_WqVq, arg, 0, arg1, arg2);
1190        break;
1191    case TCG_TYPE_V128:
1192        /*
1193         * The gvec infrastructure is asserts that v128 vector loads
1194         * and stores use a 16-byte aligned offset.  Validate that the
1195         * final pointer is aligned by using an insn that will SIGSEGV.
1196         *
1197         * This specific instance is also used by TCG_CALL_RET_BY_VEC,
1198         * for _WIN64, which must have SSE2 but may not have AVX.
1199         */
1200        tcg_debug_assert(arg >= 16);
1201        if (have_avx1) {
1202            tcg_out_vex_modrm_offset(s, OPC_MOVDQA_WxVx, arg, 0, arg1, arg2);
1203        } else {
1204            tcg_out_modrm_offset(s, OPC_MOVDQA_WxVx, arg, arg1, arg2);
1205        }
1206        break;
1207    case TCG_TYPE_V256:
1208        /*
1209         * The gvec infrastructure only requires 16-byte alignment,
1210         * so here we must use an unaligned store.
1211         */
1212        tcg_debug_assert(arg >= 16);
1213        tcg_out_vex_modrm_offset(s, OPC_MOVDQU_WxVx | P_VEXL,
1214                                 arg, 0, arg1, arg2);
1215        break;
1216    default:
1217        g_assert_not_reached();
1218    }
1219}
1220
1221static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val,
1222                        TCGReg base, intptr_t ofs)
1223{
1224    int rexw = 0;
1225    if (TCG_TARGET_REG_BITS == 64 && type == TCG_TYPE_I64) {
1226        if (val != (int32_t)val) {
1227            return false;
1228        }
1229        rexw = P_REXW;
1230    } else if (type != TCG_TYPE_I32) {
1231        return false;
1232    }
1233    tcg_out_modrm_offset(s, OPC_MOVL_EvIz | rexw, 0, base, ofs);
1234    tcg_out32(s, val);
1235    return true;
1236}
1237
1238static void tcg_out_shifti(TCGContext *s, int subopc, int reg, int count)
1239{
1240    /* Propagate an opcode prefix, such as P_DATA16.  */
1241    int ext = subopc & ~0x7;
1242    subopc &= 0x7;
1243
1244    if (count == 1) {
1245        tcg_out_modrm(s, OPC_SHIFT_1 + ext, subopc, reg);
1246    } else {
1247        tcg_out_modrm(s, OPC_SHIFT_Ib + ext, subopc, reg);
1248        tcg_out8(s, count);
1249    }
1250}
1251
1252static inline void tcg_out_bswap32(TCGContext *s, int reg)
1253{
1254    tcg_out_opc(s, OPC_BSWAP + LOWREGMASK(reg), 0, reg, 0);
1255}
1256
1257static inline void tcg_out_rolw_8(TCGContext *s, int reg)
1258{
1259    tcg_out_shifti(s, SHIFT_ROL + P_DATA16, reg, 8);
1260}
1261
1262static inline void tcg_out_ext8u(TCGContext *s, int dest, int src)
1263{
1264    /* movzbl */
1265    tcg_debug_assert(src < 4 || TCG_TARGET_REG_BITS == 64);
1266    tcg_out_modrm(s, OPC_MOVZBL + P_REXB_RM, dest, src);
1267}
1268
1269static void tcg_out_ext8s(TCGContext *s, int dest, int src, int rexw)
1270{
1271    /* movsbl */
1272    tcg_debug_assert(src < 4 || TCG_TARGET_REG_BITS == 64);
1273    tcg_out_modrm(s, OPC_MOVSBL + P_REXB_RM + rexw, dest, src);
1274}
1275
1276static inline void tcg_out_ext16u(TCGContext *s, int dest, int src)
1277{
1278    /* movzwl */
1279    tcg_out_modrm(s, OPC_MOVZWL, dest, src);
1280}
1281
1282static inline void tcg_out_ext16s(TCGContext *s, int dest, int src, int rexw)
1283{
1284    /* movsw[lq] */
1285    tcg_out_modrm(s, OPC_MOVSWL + rexw, dest, src);
1286}
1287
1288static inline void tcg_out_ext32u(TCGContext *s, int dest, int src)
1289{
1290    /* 32-bit mov zero extends.  */
1291    tcg_out_modrm(s, OPC_MOVL_GvEv, dest, src);
1292}
1293
1294static inline void tcg_out_ext32s(TCGContext *s, int dest, int src)
1295{
1296    tcg_out_modrm(s, OPC_MOVSLQ, dest, src);
1297}
1298
1299static inline void tcg_out_bswap64(TCGContext *s, int reg)
1300{
1301    tcg_out_opc(s, OPC_BSWAP + P_REXW + LOWREGMASK(reg), 0, reg, 0);
1302}
1303
1304static void tgen_arithi(TCGContext *s, int c, int r0,
1305                        tcg_target_long val, int cf)
1306{
1307    int rexw = 0;
1308
1309    if (TCG_TARGET_REG_BITS == 64) {
1310        rexw = c & -8;
1311        c &= 7;
1312    }
1313
1314    /* ??? While INC is 2 bytes shorter than ADDL $1, they also induce
1315       partial flags update stalls on Pentium4 and are not recommended
1316       by current Intel optimization manuals.  */
1317    if (!cf && (c == ARITH_ADD || c == ARITH_SUB) && (val == 1 || val == -1)) {
1318        int is_inc = (c == ARITH_ADD) ^ (val < 0);
1319        if (TCG_TARGET_REG_BITS == 64) {
1320            /* The single-byte increment encodings are re-tasked as the
1321               REX prefixes.  Use the MODRM encoding.  */
1322            tcg_out_modrm(s, OPC_GRP5 + rexw,
1323                          (is_inc ? EXT5_INC_Ev : EXT5_DEC_Ev), r0);
1324        } else {
1325            tcg_out8(s, (is_inc ? OPC_INC_r32 : OPC_DEC_r32) + r0);
1326        }
1327        return;
1328    }
1329
1330    if (c == ARITH_AND) {
1331        if (TCG_TARGET_REG_BITS == 64) {
1332            if (val == 0xffffffffu) {
1333                tcg_out_ext32u(s, r0, r0);
1334                return;
1335            }
1336            if (val == (uint32_t)val) {
1337                /* AND with no high bits set can use a 32-bit operation.  */
1338                rexw = 0;
1339            }
1340        }
1341        if (val == 0xffu && (r0 < 4 || TCG_TARGET_REG_BITS == 64)) {
1342            tcg_out_ext8u(s, r0, r0);
1343            return;
1344        }
1345        if (val == 0xffffu) {
1346            tcg_out_ext16u(s, r0, r0);
1347            return;
1348        }
1349    }
1350
1351    if (val == (int8_t)val) {
1352        tcg_out_modrm(s, OPC_ARITH_EvIb + rexw, c, r0);
1353        tcg_out8(s, val);
1354        return;
1355    }
1356    if (rexw == 0 || val == (int32_t)val) {
1357        tcg_out_modrm(s, OPC_ARITH_EvIz + rexw, c, r0);
1358        tcg_out32(s, val);
1359        return;
1360    }
1361
1362    tcg_abort();
1363}
1364
1365static void tcg_out_addi(TCGContext *s, int reg, tcg_target_long val)
1366{
1367    if (val != 0) {
1368        tgen_arithi(s, ARITH_ADD + P_REXW, reg, val, 0);
1369    }
1370}
1371
1372/* Use SMALL != 0 to force a short forward branch.  */
1373static void tcg_out_jxx(TCGContext *s, int opc, TCGLabel *l, int small)
1374{
1375    int32_t val, val1;
1376
1377    if (l->has_value) {
1378        val = tcg_pcrel_diff(s, l->u.value_ptr);
1379        val1 = val - 2;
1380        if ((int8_t)val1 == val1) {
1381            if (opc == -1) {
1382                tcg_out8(s, OPC_JMP_short);
1383            } else {
1384                tcg_out8(s, OPC_JCC_short + opc);
1385            }
1386            tcg_out8(s, val1);
1387        } else {
1388            if (small) {
1389                tcg_abort();
1390            }
1391            if (opc == -1) {
1392                tcg_out8(s, OPC_JMP_long);
1393                tcg_out32(s, val - 5);
1394            } else {
1395                tcg_out_opc(s, OPC_JCC_long + opc, 0, 0, 0);
1396                tcg_out32(s, val - 6);
1397            }
1398        }
1399    } else if (small) {
1400        if (opc == -1) {
1401            tcg_out8(s, OPC_JMP_short);
1402        } else {
1403            tcg_out8(s, OPC_JCC_short + opc);
1404        }
1405        tcg_out_reloc(s, s->code_ptr, R_386_PC8, l, -1);
1406        s->code_ptr += 1;
1407    } else {
1408        if (opc == -1) {
1409            tcg_out8(s, OPC_JMP_long);
1410        } else {
1411            tcg_out_opc(s, OPC_JCC_long + opc, 0, 0, 0);
1412        }
1413        tcg_out_reloc(s, s->code_ptr, R_386_PC32, l, -4);
1414        s->code_ptr += 4;
1415    }
1416}
1417
1418static void tcg_out_cmp(TCGContext *s, TCGArg arg1, TCGArg arg2,
1419                        int const_arg2, int rexw)
1420{
1421    if (const_arg2) {
1422        if (arg2 == 0) {
1423            /* test r, r */
1424            tcg_out_modrm(s, OPC_TESTL + rexw, arg1, arg1);
1425        } else {
1426            tgen_arithi(s, ARITH_CMP + rexw, arg1, arg2, 0);
1427        }
1428    } else {
1429        tgen_arithr(s, ARITH_CMP + rexw, arg1, arg2);
1430    }
1431}
1432
1433static void tcg_out_brcond32(TCGContext *s, TCGCond cond,
1434                             TCGArg arg1, TCGArg arg2, int const_arg2,
1435                             TCGLabel *label, int small)
1436{
1437    tcg_out_cmp(s, arg1, arg2, const_arg2, 0);
1438    tcg_out_jxx(s, tcg_cond_to_jcc[cond], label, small);
1439}
1440
1441#if TCG_TARGET_REG_BITS == 64
1442static void tcg_out_brcond64(TCGContext *s, TCGCond cond,
1443                             TCGArg arg1, TCGArg arg2, int const_arg2,
1444                             TCGLabel *label, int small)
1445{
1446    tcg_out_cmp(s, arg1, arg2, const_arg2, P_REXW);
1447    tcg_out_jxx(s, tcg_cond_to_jcc[cond], label, small);
1448}
1449#else
1450/* XXX: we implement it at the target level to avoid having to
1451   handle cross basic blocks temporaries */
1452static void tcg_out_brcond2(TCGContext *s, const TCGArg *args,
1453                            const int *const_args, int small)
1454{
1455    TCGLabel *label_next = gen_new_label();
1456    TCGLabel *label_this = arg_label(args[5]);
1457
1458    switch(args[4]) {
1459    case TCG_COND_EQ:
1460        tcg_out_brcond32(s, TCG_COND_NE, args[0], args[2], const_args[2],
1461                         label_next, 1);
1462        tcg_out_brcond32(s, TCG_COND_EQ, args[1], args[3], const_args[3],
1463                         label_this, small);
1464        break;
1465    case TCG_COND_NE:
1466        tcg_out_brcond32(s, TCG_COND_NE, args[0], args[2], const_args[2],
1467                         label_this, small);
1468        tcg_out_brcond32(s, TCG_COND_NE, args[1], args[3], const_args[3],
1469                         label_this, small);
1470        break;
1471    case TCG_COND_LT:
1472        tcg_out_brcond32(s, TCG_COND_LT, args[1], args[3], const_args[3],
1473                         label_this, small);
1474        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1475        tcg_out_brcond32(s, TCG_COND_LTU, args[0], args[2], const_args[2],
1476                         label_this, small);
1477        break;
1478    case TCG_COND_LE:
1479        tcg_out_brcond32(s, TCG_COND_LT, args[1], args[3], const_args[3],
1480                         label_this, small);
1481        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1482        tcg_out_brcond32(s, TCG_COND_LEU, args[0], args[2], const_args[2],
1483                         label_this, small);
1484        break;
1485    case TCG_COND_GT:
1486        tcg_out_brcond32(s, TCG_COND_GT, args[1], args[3], const_args[3],
1487                         label_this, small);
1488        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1489        tcg_out_brcond32(s, TCG_COND_GTU, args[0], args[2], const_args[2],
1490                         label_this, small);
1491        break;
1492    case TCG_COND_GE:
1493        tcg_out_brcond32(s, TCG_COND_GT, args[1], args[3], const_args[3],
1494                         label_this, small);
1495        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1496        tcg_out_brcond32(s, TCG_COND_GEU, args[0], args[2], const_args[2],
1497                         label_this, small);
1498        break;
1499    case TCG_COND_LTU:
1500        tcg_out_brcond32(s, TCG_COND_LTU, args[1], args[3], const_args[3],
1501                         label_this, small);
1502        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1503        tcg_out_brcond32(s, TCG_COND_LTU, args[0], args[2], const_args[2],
1504                         label_this, small);
1505        break;
1506    case TCG_COND_LEU:
1507        tcg_out_brcond32(s, TCG_COND_LTU, args[1], args[3], const_args[3],
1508                         label_this, small);
1509        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1510        tcg_out_brcond32(s, TCG_COND_LEU, args[0], args[2], const_args[2],
1511                         label_this, small);
1512        break;
1513    case TCG_COND_GTU:
1514        tcg_out_brcond32(s, TCG_COND_GTU, args[1], args[3], const_args[3],
1515                         label_this, small);
1516        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1517        tcg_out_brcond32(s, TCG_COND_GTU, args[0], args[2], const_args[2],
1518                         label_this, small);
1519        break;
1520    case TCG_COND_GEU:
1521        tcg_out_brcond32(s, TCG_COND_GTU, args[1], args[3], const_args[3],
1522                         label_this, small);
1523        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1524        tcg_out_brcond32(s, TCG_COND_GEU, args[0], args[2], const_args[2],
1525                         label_this, small);
1526        break;
1527    default:
1528        tcg_abort();
1529    }
1530    tcg_out_label(s, label_next);
1531}
1532#endif
1533
1534static void tcg_out_setcond32(TCGContext *s, TCGCond cond, TCGArg dest,
1535                              TCGArg arg1, TCGArg arg2, int const_arg2)
1536{
1537    tcg_out_cmp(s, arg1, arg2, const_arg2, 0);
1538    tcg_out_modrm(s, OPC_SETCC | tcg_cond_to_jcc[cond], 0, dest);
1539    tcg_out_ext8u(s, dest, dest);
1540}
1541
1542#if TCG_TARGET_REG_BITS == 64
1543static void tcg_out_setcond64(TCGContext *s, TCGCond cond, TCGArg dest,
1544                              TCGArg arg1, TCGArg arg2, int const_arg2)
1545{
1546    tcg_out_cmp(s, arg1, arg2, const_arg2, P_REXW);
1547    tcg_out_modrm(s, OPC_SETCC | tcg_cond_to_jcc[cond], 0, dest);
1548    tcg_out_ext8u(s, dest, dest);
1549}
1550#else
1551static void tcg_out_setcond2(TCGContext *s, const TCGArg *args,
1552                             const int *const_args)
1553{
1554    TCGArg new_args[6];
1555    TCGLabel *label_true, *label_over;
1556
1557    memcpy(new_args, args+1, 5*sizeof(TCGArg));
1558
1559    if (args[0] == args[1] || args[0] == args[2]
1560        || (!const_args[3] && args[0] == args[3])
1561        || (!const_args[4] && args[0] == args[4])) {
1562        /* When the destination overlaps with one of the argument
1563           registers, don't do anything tricky.  */
1564        label_true = gen_new_label();
1565        label_over = gen_new_label();
1566
1567        new_args[5] = label_arg(label_true);
1568        tcg_out_brcond2(s, new_args, const_args+1, 1);
1569
1570        tcg_out_movi(s, TCG_TYPE_I32, args[0], 0);
1571        tcg_out_jxx(s, JCC_JMP, label_over, 1);
1572        tcg_out_label(s, label_true);
1573
1574        tcg_out_movi(s, TCG_TYPE_I32, args[0], 1);
1575        tcg_out_label(s, label_over);
1576    } else {
1577        /* When the destination does not overlap one of the arguments,
1578           clear the destination first, jump if cond false, and emit an
1579           increment in the true case.  This results in smaller code.  */
1580
1581        tcg_out_movi(s, TCG_TYPE_I32, args[0], 0);
1582
1583        label_over = gen_new_label();
1584        new_args[4] = tcg_invert_cond(new_args[4]);
1585        new_args[5] = label_arg(label_over);
1586        tcg_out_brcond2(s, new_args, const_args+1, 1);
1587
1588        tgen_arithi(s, ARITH_ADD, args[0], 1, 0);
1589        tcg_out_label(s, label_over);
1590    }
1591}
1592#endif
1593
1594static void tcg_out_cmov(TCGContext *s, TCGCond cond, int rexw,
1595                         TCGReg dest, TCGReg v1)
1596{
1597    if (have_cmov) {
1598        tcg_out_modrm(s, OPC_CMOVCC | tcg_cond_to_jcc[cond] | rexw, dest, v1);
1599    } else {
1600        TCGLabel *over = gen_new_label();
1601        tcg_out_jxx(s, tcg_cond_to_jcc[tcg_invert_cond(cond)], over, 1);
1602        tcg_out_mov(s, TCG_TYPE_I32, dest, v1);
1603        tcg_out_label(s, over);
1604    }
1605}
1606
1607static void tcg_out_movcond32(TCGContext *s, TCGCond cond, TCGReg dest,
1608                              TCGReg c1, TCGArg c2, int const_c2,
1609                              TCGReg v1)
1610{
1611    tcg_out_cmp(s, c1, c2, const_c2, 0);
1612    tcg_out_cmov(s, cond, 0, dest, v1);
1613}
1614
1615#if TCG_TARGET_REG_BITS == 64
1616static void tcg_out_movcond64(TCGContext *s, TCGCond cond, TCGReg dest,
1617                              TCGReg c1, TCGArg c2, int const_c2,
1618                              TCGReg v1)
1619{
1620    tcg_out_cmp(s, c1, c2, const_c2, P_REXW);
1621    tcg_out_cmov(s, cond, P_REXW, dest, v1);
1622}
1623#endif
1624
1625static void tcg_out_ctz(TCGContext *s, int rexw, TCGReg dest, TCGReg arg1,
1626                        TCGArg arg2, bool const_a2)
1627{
1628    if (have_bmi1) {
1629        tcg_out_modrm(s, OPC_TZCNT + rexw, dest, arg1);
1630        if (const_a2) {
1631            tcg_debug_assert(arg2 == (rexw ? 64 : 32));
1632        } else {
1633            tcg_debug_assert(dest != arg2);
1634            tcg_out_cmov(s, TCG_COND_LTU, rexw, dest, arg2);
1635        }
1636    } else {
1637        tcg_debug_assert(dest != arg2);
1638        tcg_out_modrm(s, OPC_BSF + rexw, dest, arg1);
1639        tcg_out_cmov(s, TCG_COND_EQ, rexw, dest, arg2);
1640    }
1641}
1642
1643static void tcg_out_clz(TCGContext *s, int rexw, TCGReg dest, TCGReg arg1,
1644                        TCGArg arg2, bool const_a2)
1645{
1646    if (have_lzcnt) {
1647        tcg_out_modrm(s, OPC_LZCNT + rexw, dest, arg1);
1648        if (const_a2) {
1649            tcg_debug_assert(arg2 == (rexw ? 64 : 32));
1650        } else {
1651            tcg_debug_assert(dest != arg2);
1652            tcg_out_cmov(s, TCG_COND_LTU, rexw, dest, arg2);
1653        }
1654    } else {
1655        tcg_debug_assert(!const_a2);
1656        tcg_debug_assert(dest != arg1);
1657        tcg_debug_assert(dest != arg2);
1658
1659        /* Recall that the output of BSR is the index not the count.  */
1660        tcg_out_modrm(s, OPC_BSR + rexw, dest, arg1);
1661        tgen_arithi(s, ARITH_XOR + rexw, dest, rexw ? 63 : 31, 0);
1662
1663        /* Since we have destroyed the flags from BSR, we have to re-test.  */
1664        tcg_out_cmp(s, arg1, 0, 1, rexw);
1665        tcg_out_cmov(s, TCG_COND_EQ, rexw, dest, arg2);
1666    }
1667}
1668
1669static void tcg_out_branch(TCGContext *s, int call, const tcg_insn_unit *dest)
1670{
1671    intptr_t disp = tcg_pcrel_diff(s, dest) - 5;
1672
1673    if (disp == (int32_t)disp) {
1674        tcg_out_opc(s, call ? OPC_CALL_Jz : OPC_JMP_long, 0, 0, 0);
1675        tcg_out32(s, disp);
1676    } else {
1677        /* rip-relative addressing into the constant pool.
1678           This is 6 + 8 = 14 bytes, as compared to using an
1679           immediate load 10 + 6 = 16 bytes, plus we may
1680           be able to re-use the pool constant for more calls.  */
1681        tcg_out_opc(s, OPC_GRP5, 0, 0, 0);
1682        tcg_out8(s, (call ? EXT5_CALLN_Ev : EXT5_JMPN_Ev) << 3 | 5);
1683        new_pool_label(s, (uintptr_t)dest, R_386_PC32, s->code_ptr, -4);
1684        tcg_out32(s, 0);
1685    }
1686}
1687
1688static void tcg_out_call(TCGContext *s, const tcg_insn_unit *dest,
1689                         const TCGHelperInfo *info)
1690{
1691    tcg_out_branch(s, 1, dest);
1692
1693#ifndef _WIN32
1694    if (TCG_TARGET_REG_BITS == 32 && info->out_kind == TCG_CALL_RET_BY_REF) {
1695        /*
1696         * The sysv i386 abi for struct return places a reference as the
1697         * first argument of the stack, and pops that argument with the
1698         * return statement.  Since we want to retain the aligned stack
1699         * pointer for the callee, we do not want to actually push that
1700         * argument before the call but rely on the normal store to the
1701         * stack slot.  But we do need to compensate for the pop in order
1702         * to reset our correct stack pointer value.
1703         * Pushing a garbage value back onto the stack is quickest.
1704         */
1705        tcg_out_push(s, TCG_REG_EAX);
1706    }
1707#endif
1708}
1709
1710static void tcg_out_jmp(TCGContext *s, const tcg_insn_unit *dest)
1711{
1712    tcg_out_branch(s, 0, dest);
1713}
1714
1715static void tcg_out_nopn(TCGContext *s, int n)
1716{
1717    int i;
1718    /* Emit 1 or 2 operand size prefixes for the standard one byte nop,
1719     * "xchg %eax,%eax", forming "xchg %ax,%ax". All cores accept the
1720     * duplicate prefix, and all of the interesting recent cores can
1721     * decode and discard the duplicates in a single cycle.
1722     */
1723    tcg_debug_assert(n >= 1);
1724    for (i = 1; i < n; ++i) {
1725        tcg_out8(s, 0x66);
1726    }
1727    tcg_out8(s, 0x90);
1728}
1729
1730#if defined(CONFIG_SOFTMMU)
1731/* helper signature: helper_ret_ld_mmu(CPUState *env, target_ulong addr,
1732 *                                     int mmu_idx, uintptr_t ra)
1733 */
1734static void * const qemu_ld_helpers[(MO_SIZE | MO_BSWAP) + 1] = {
1735    [MO_UB]   = helper_ret_ldub_mmu,
1736    [MO_LEUW] = helper_le_lduw_mmu,
1737    [MO_LEUL] = helper_le_ldul_mmu,
1738    [MO_LEUQ] = helper_le_ldq_mmu,
1739    [MO_BEUW] = helper_be_lduw_mmu,
1740    [MO_BEUL] = helper_be_ldul_mmu,
1741    [MO_BEUQ] = helper_be_ldq_mmu,
1742};
1743
1744/* helper signature: helper_ret_st_mmu(CPUState *env, target_ulong addr,
1745 *                                     uintxx_t val, int mmu_idx, uintptr_t ra)
1746 */
1747static void * const qemu_st_helpers[(MO_SIZE | MO_BSWAP) + 1] = {
1748    [MO_UB]   = helper_ret_stb_mmu,
1749    [MO_LEUW] = helper_le_stw_mmu,
1750    [MO_LEUL] = helper_le_stl_mmu,
1751    [MO_LEUQ] = helper_le_stq_mmu,
1752    [MO_BEUW] = helper_be_stw_mmu,
1753    [MO_BEUL] = helper_be_stl_mmu,
1754    [MO_BEUQ] = helper_be_stq_mmu,
1755};
1756
1757/* Perform the TLB load and compare.
1758
1759   Inputs:
1760   ADDRLO and ADDRHI contain the low and high part of the address.
1761
1762   MEM_INDEX and S_BITS are the memory context and log2 size of the load.
1763
1764   WHICH is the offset into the CPUTLBEntry structure of the slot to read.
1765   This should be offsetof addr_read or addr_write.
1766
1767   Outputs:
1768   LABEL_PTRS is filled with 1 (32-bit addresses) or 2 (64-bit addresses)
1769   positions of the displacements of forward jumps to the TLB miss case.
1770
1771   Second argument register is loaded with the low part of the address.
1772   In the TLB hit case, it has been adjusted as indicated by the TLB
1773   and so is a host address.  In the TLB miss case, it continues to
1774   hold a guest address.
1775
1776   First argument register is clobbered.  */
1777
1778static inline void tcg_out_tlb_load(TCGContext *s, TCGReg addrlo, TCGReg addrhi,
1779                                    int mem_index, MemOp opc,
1780                                    tcg_insn_unit **label_ptr, int which)
1781{
1782    const TCGReg r0 = TCG_REG_L0;
1783    const TCGReg r1 = TCG_REG_L1;
1784    TCGType ttype = TCG_TYPE_I32;
1785    TCGType tlbtype = TCG_TYPE_I32;
1786    int trexw = 0, hrexw = 0, tlbrexw = 0;
1787    unsigned a_bits = get_alignment_bits(opc);
1788    unsigned s_bits = opc & MO_SIZE;
1789    unsigned a_mask = (1 << a_bits) - 1;
1790    unsigned s_mask = (1 << s_bits) - 1;
1791    target_ulong tlb_mask;
1792
1793    if (TCG_TARGET_REG_BITS == 64) {
1794        if (TARGET_LONG_BITS == 64) {
1795            ttype = TCG_TYPE_I64;
1796            trexw = P_REXW;
1797        }
1798        if (TCG_TYPE_PTR == TCG_TYPE_I64) {
1799            hrexw = P_REXW;
1800            if (TARGET_PAGE_BITS + CPU_TLB_DYN_MAX_BITS > 32) {
1801                tlbtype = TCG_TYPE_I64;
1802                tlbrexw = P_REXW;
1803            }
1804        }
1805    }
1806
1807    tcg_out_mov(s, tlbtype, r0, addrlo);
1808    tcg_out_shifti(s, SHIFT_SHR + tlbrexw, r0,
1809                   TARGET_PAGE_BITS - CPU_TLB_ENTRY_BITS);
1810
1811    tcg_out_modrm_offset(s, OPC_AND_GvEv + trexw, r0, TCG_AREG0,
1812                         TLB_MASK_TABLE_OFS(mem_index) +
1813                         offsetof(CPUTLBDescFast, mask));
1814
1815    tcg_out_modrm_offset(s, OPC_ADD_GvEv + hrexw, r0, TCG_AREG0,
1816                         TLB_MASK_TABLE_OFS(mem_index) +
1817                         offsetof(CPUTLBDescFast, table));
1818
1819    /* If the required alignment is at least as large as the access, simply
1820       copy the address and mask.  For lesser alignments, check that we don't
1821       cross pages for the complete access.  */
1822    if (a_bits >= s_bits) {
1823        tcg_out_mov(s, ttype, r1, addrlo);
1824    } else {
1825        tcg_out_modrm_offset(s, OPC_LEA + trexw, r1, addrlo, s_mask - a_mask);
1826    }
1827    tlb_mask = (target_ulong)TARGET_PAGE_MASK | a_mask;
1828    tgen_arithi(s, ARITH_AND + trexw, r1, tlb_mask, 0);
1829
1830    /* cmp 0(r0), r1 */
1831    tcg_out_modrm_offset(s, OPC_CMP_GvEv + trexw, r1, r0, which);
1832
1833    /* Prepare for both the fast path add of the tlb addend, and the slow
1834       path function argument setup.  */
1835    tcg_out_mov(s, ttype, r1, addrlo);
1836
1837    /* jne slow_path */
1838    tcg_out_opc(s, OPC_JCC_long + JCC_JNE, 0, 0, 0);
1839    label_ptr[0] = s->code_ptr;
1840    s->code_ptr += 4;
1841
1842    if (TARGET_LONG_BITS > TCG_TARGET_REG_BITS) {
1843        /* cmp 4(r0), addrhi */
1844        tcg_out_modrm_offset(s, OPC_CMP_GvEv, addrhi, r0, which + 4);
1845
1846        /* jne slow_path */
1847        tcg_out_opc(s, OPC_JCC_long + JCC_JNE, 0, 0, 0);
1848        label_ptr[1] = s->code_ptr;
1849        s->code_ptr += 4;
1850    }
1851
1852    /* TLB Hit.  */
1853
1854    /* add addend(r0), r1 */
1855    tcg_out_modrm_offset(s, OPC_ADD_GvEv + hrexw, r1, r0,
1856                         offsetof(CPUTLBEntry, addend));
1857}
1858
1859/*
1860 * Record the context of a call to the out of line helper code for the slow path
1861 * for a load or store, so that we can later generate the correct helper code
1862 */
1863static void add_qemu_ldst_label(TCGContext *s, bool is_ld, bool is_64,
1864                                MemOpIdx oi,
1865                                TCGReg datalo, TCGReg datahi,
1866                                TCGReg addrlo, TCGReg addrhi,
1867                                tcg_insn_unit *raddr,
1868                                tcg_insn_unit **label_ptr)
1869{
1870    TCGLabelQemuLdst *label = new_ldst_label(s);
1871
1872    label->is_ld = is_ld;
1873    label->oi = oi;
1874    label->type = is_64 ? TCG_TYPE_I64 : TCG_TYPE_I32;
1875    label->datalo_reg = datalo;
1876    label->datahi_reg = datahi;
1877    label->addrlo_reg = addrlo;
1878    label->addrhi_reg = addrhi;
1879    label->raddr = tcg_splitwx_to_rx(raddr);
1880    label->label_ptr[0] = label_ptr[0];
1881    if (TARGET_LONG_BITS > TCG_TARGET_REG_BITS) {
1882        label->label_ptr[1] = label_ptr[1];
1883    }
1884}
1885
1886/*
1887 * Generate code for the slow path for a load at the end of block
1888 */
1889static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
1890{
1891    MemOpIdx oi = l->oi;
1892    MemOp opc = get_memop(oi);
1893    TCGReg data_reg;
1894    tcg_insn_unit **label_ptr = &l->label_ptr[0];
1895    int rexw = (l->type == TCG_TYPE_I64 ? P_REXW : 0);
1896
1897    /* resolve label address */
1898    tcg_patch32(label_ptr[0], s->code_ptr - label_ptr[0] - 4);
1899    if (TARGET_LONG_BITS > TCG_TARGET_REG_BITS) {
1900        tcg_patch32(label_ptr[1], s->code_ptr - label_ptr[1] - 4);
1901    }
1902
1903    if (TCG_TARGET_REG_BITS == 32) {
1904        int ofs = 0;
1905
1906        tcg_out_st(s, TCG_TYPE_PTR, TCG_AREG0, TCG_REG_ESP, ofs);
1907        ofs += 4;
1908
1909        tcg_out_st(s, TCG_TYPE_I32, l->addrlo_reg, TCG_REG_ESP, ofs);
1910        ofs += 4;
1911
1912        if (TARGET_LONG_BITS == 64) {
1913            tcg_out_st(s, TCG_TYPE_I32, l->addrhi_reg, TCG_REG_ESP, ofs);
1914            ofs += 4;
1915        }
1916
1917        tcg_out_sti(s, TCG_TYPE_I32, oi, TCG_REG_ESP, ofs);
1918        ofs += 4;
1919
1920        tcg_out_sti(s, TCG_TYPE_PTR, (uintptr_t)l->raddr, TCG_REG_ESP, ofs);
1921    } else {
1922        tcg_out_mov(s, TCG_TYPE_PTR, tcg_target_call_iarg_regs[0], TCG_AREG0);
1923        /* The second argument is already loaded with addrlo.  */
1924        tcg_out_movi(s, TCG_TYPE_I32, tcg_target_call_iarg_regs[2], oi);
1925        tcg_out_movi(s, TCG_TYPE_PTR, tcg_target_call_iarg_regs[3],
1926                     (uintptr_t)l->raddr);
1927    }
1928
1929    tcg_out_branch(s, 1, qemu_ld_helpers[opc & (MO_BSWAP | MO_SIZE)]);
1930
1931    data_reg = l->datalo_reg;
1932    switch (opc & MO_SSIZE) {
1933    case MO_SB:
1934        tcg_out_ext8s(s, data_reg, TCG_REG_EAX, rexw);
1935        break;
1936    case MO_SW:
1937        tcg_out_ext16s(s, data_reg, TCG_REG_EAX, rexw);
1938        break;
1939#if TCG_TARGET_REG_BITS == 64
1940    case MO_SL:
1941        tcg_out_ext32s(s, data_reg, TCG_REG_EAX);
1942        break;
1943#endif
1944    case MO_UB:
1945    case MO_UW:
1946        /* Note that the helpers have zero-extended to tcg_target_long.  */
1947    case MO_UL:
1948        tcg_out_mov(s, TCG_TYPE_I32, data_reg, TCG_REG_EAX);
1949        break;
1950    case MO_UQ:
1951        if (TCG_TARGET_REG_BITS == 64) {
1952            tcg_out_mov(s, TCG_TYPE_I64, data_reg, TCG_REG_RAX);
1953        } else if (data_reg == TCG_REG_EDX) {
1954            /* xchg %edx, %eax */
1955            tcg_out_opc(s, OPC_XCHG_ax_r32 + TCG_REG_EDX, 0, 0, 0);
1956            tcg_out_mov(s, TCG_TYPE_I32, l->datahi_reg, TCG_REG_EAX);
1957        } else {
1958            tcg_out_mov(s, TCG_TYPE_I32, data_reg, TCG_REG_EAX);
1959            tcg_out_mov(s, TCG_TYPE_I32, l->datahi_reg, TCG_REG_EDX);
1960        }
1961        break;
1962    default:
1963        tcg_abort();
1964    }
1965
1966    /* Jump to the code corresponding to next IR of qemu_st */
1967    tcg_out_jmp(s, l->raddr);
1968    return true;
1969}
1970
1971/*
1972 * Generate code for the slow path for a store at the end of block
1973 */
1974static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
1975{
1976    MemOpIdx oi = l->oi;
1977    MemOp opc = get_memop(oi);
1978    MemOp s_bits = opc & MO_SIZE;
1979    tcg_insn_unit **label_ptr = &l->label_ptr[0];
1980    TCGReg retaddr;
1981
1982    /* resolve label address */
1983    tcg_patch32(label_ptr[0], s->code_ptr - label_ptr[0] - 4);
1984    if (TARGET_LONG_BITS > TCG_TARGET_REG_BITS) {
1985        tcg_patch32(label_ptr[1], s->code_ptr - label_ptr[1] - 4);
1986    }
1987
1988    if (TCG_TARGET_REG_BITS == 32) {
1989        int ofs = 0;
1990
1991        tcg_out_st(s, TCG_TYPE_PTR, TCG_AREG0, TCG_REG_ESP, ofs);
1992        ofs += 4;
1993
1994        tcg_out_st(s, TCG_TYPE_I32, l->addrlo_reg, TCG_REG_ESP, ofs);
1995        ofs += 4;
1996
1997        if (TARGET_LONG_BITS == 64) {
1998            tcg_out_st(s, TCG_TYPE_I32, l->addrhi_reg, TCG_REG_ESP, ofs);
1999            ofs += 4;
2000        }
2001
2002        tcg_out_st(s, TCG_TYPE_I32, l->datalo_reg, TCG_REG_ESP, ofs);
2003        ofs += 4;
2004
2005        if (s_bits == MO_64) {
2006            tcg_out_st(s, TCG_TYPE_I32, l->datahi_reg, TCG_REG_ESP, ofs);
2007            ofs += 4;
2008        }
2009
2010        tcg_out_sti(s, TCG_TYPE_I32, oi, TCG_REG_ESP, ofs);
2011        ofs += 4;
2012
2013        retaddr = TCG_REG_EAX;
2014        tcg_out_movi(s, TCG_TYPE_PTR, retaddr, (uintptr_t)l->raddr);
2015        tcg_out_st(s, TCG_TYPE_PTR, retaddr, TCG_REG_ESP, ofs);
2016    } else {
2017        tcg_out_mov(s, TCG_TYPE_PTR, tcg_target_call_iarg_regs[0], TCG_AREG0);
2018        /* The second argument is already loaded with addrlo.  */
2019        tcg_out_mov(s, (s_bits == MO_64 ? TCG_TYPE_I64 : TCG_TYPE_I32),
2020                    tcg_target_call_iarg_regs[2], l->datalo_reg);
2021        tcg_out_movi(s, TCG_TYPE_I32, tcg_target_call_iarg_regs[3], oi);
2022
2023        if (ARRAY_SIZE(tcg_target_call_iarg_regs) > 4) {
2024            retaddr = tcg_target_call_iarg_regs[4];
2025            tcg_out_movi(s, TCG_TYPE_PTR, retaddr, (uintptr_t)l->raddr);
2026        } else {
2027            retaddr = TCG_REG_RAX;
2028            tcg_out_movi(s, TCG_TYPE_PTR, retaddr, (uintptr_t)l->raddr);
2029            tcg_out_st(s, TCG_TYPE_PTR, retaddr, TCG_REG_ESP,
2030                       TCG_TARGET_CALL_STACK_OFFSET);
2031        }
2032    }
2033
2034    /* "Tail call" to the helper, with the return address back inline.  */
2035    tcg_out_push(s, retaddr);
2036    tcg_out_jmp(s, qemu_st_helpers[opc & (MO_BSWAP | MO_SIZE)]);
2037    return true;
2038}
2039#else
2040
2041static void tcg_out_test_alignment(TCGContext *s, bool is_ld, TCGReg addrlo,
2042                                   TCGReg addrhi, unsigned a_bits)
2043{
2044    unsigned a_mask = (1 << a_bits) - 1;
2045    TCGLabelQemuLdst *label;
2046
2047    /*
2048     * We are expecting a_bits to max out at 7, so we can usually use testb.
2049     * For i686, we have to use testl for %esi/%edi.
2050     */
2051    if (a_mask <= 0xff && (TCG_TARGET_REG_BITS == 64 || addrlo < 4)) {
2052        tcg_out_modrm(s, OPC_GRP3_Eb | P_REXB_RM, EXT3_TESTi, addrlo);
2053        tcg_out8(s, a_mask);
2054    } else {
2055        tcg_out_modrm(s, OPC_GRP3_Ev, EXT3_TESTi, addrlo);
2056        tcg_out32(s, a_mask);
2057    }
2058
2059    /* jne slow_path */
2060    tcg_out_opc(s, OPC_JCC_long + JCC_JNE, 0, 0, 0);
2061
2062    label = new_ldst_label(s);
2063    label->is_ld = is_ld;
2064    label->addrlo_reg = addrlo;
2065    label->addrhi_reg = addrhi;
2066    label->raddr = tcg_splitwx_to_rx(s->code_ptr + 4);
2067    label->label_ptr[0] = s->code_ptr;
2068
2069    s->code_ptr += 4;
2070}
2071
2072static bool tcg_out_fail_alignment(TCGContext *s, TCGLabelQemuLdst *l)
2073{
2074    /* resolve label address */
2075    tcg_patch32(l->label_ptr[0], s->code_ptr - l->label_ptr[0] - 4);
2076
2077    if (TCG_TARGET_REG_BITS == 32) {
2078        int ofs = 0;
2079
2080        tcg_out_st(s, TCG_TYPE_PTR, TCG_AREG0, TCG_REG_ESP, ofs);
2081        ofs += 4;
2082
2083        tcg_out_st(s, TCG_TYPE_I32, l->addrlo_reg, TCG_REG_ESP, ofs);
2084        ofs += 4;
2085        if (TARGET_LONG_BITS == 64) {
2086            tcg_out_st(s, TCG_TYPE_I32, l->addrhi_reg, TCG_REG_ESP, ofs);
2087            ofs += 4;
2088        }
2089
2090        tcg_out_pushi(s, (uintptr_t)l->raddr);
2091    } else {
2092        tcg_out_mov(s, TCG_TYPE_TL, tcg_target_call_iarg_regs[1],
2093                    l->addrlo_reg);
2094        tcg_out_mov(s, TCG_TYPE_PTR, tcg_target_call_iarg_regs[0], TCG_AREG0);
2095
2096        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_RAX, (uintptr_t)l->raddr);
2097        tcg_out_push(s, TCG_REG_RAX);
2098    }
2099
2100    /* "Tail call" to the helper, with the return address back inline. */
2101    tcg_out_jmp(s, (const void *)(l->is_ld ? helper_unaligned_ld
2102                                  : helper_unaligned_st));
2103    return true;
2104}
2105
2106static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
2107{
2108    return tcg_out_fail_alignment(s, l);
2109}
2110
2111static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
2112{
2113    return tcg_out_fail_alignment(s, l);
2114}
2115
2116#if TCG_TARGET_REG_BITS == 32
2117# define x86_guest_base_seg     0
2118# define x86_guest_base_index   -1
2119# define x86_guest_base_offset  guest_base
2120#else
2121static int x86_guest_base_seg;
2122static int x86_guest_base_index = -1;
2123static int32_t x86_guest_base_offset;
2124# if defined(__x86_64__) && defined(__linux__)
2125#  include <asm/prctl.h>
2126#  include <sys/prctl.h>
2127int arch_prctl(int code, unsigned long addr);
2128static inline int setup_guest_base_seg(void)
2129{
2130    if (arch_prctl(ARCH_SET_GS, guest_base) == 0) {
2131        return P_GS;
2132    }
2133    return 0;
2134}
2135# elif defined (__FreeBSD__) || defined (__FreeBSD_kernel__)
2136#  include <machine/sysarch.h>
2137static inline int setup_guest_base_seg(void)
2138{
2139    if (sysarch(AMD64_SET_GSBASE, &guest_base) == 0) {
2140        return P_GS;
2141    }
2142    return 0;
2143}
2144# else
2145static inline int setup_guest_base_seg(void)
2146{
2147    return 0;
2148}
2149# endif
2150#endif
2151#endif /* SOFTMMU */
2152
2153static void tcg_out_qemu_ld_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
2154                                   TCGReg base, int index, intptr_t ofs,
2155                                   int seg, bool is64, MemOp memop)
2156{
2157    bool use_movbe = false;
2158    int rexw = is64 * P_REXW;
2159    int movop = OPC_MOVL_GvEv;
2160
2161    /* Do big-endian loads with movbe.  */
2162    if (memop & MO_BSWAP) {
2163        tcg_debug_assert(have_movbe);
2164        use_movbe = true;
2165        movop = OPC_MOVBE_GyMy;
2166    }
2167
2168    switch (memop & MO_SSIZE) {
2169    case MO_UB:
2170        tcg_out_modrm_sib_offset(s, OPC_MOVZBL + seg, datalo,
2171                                 base, index, 0, ofs);
2172        break;
2173    case MO_SB:
2174        tcg_out_modrm_sib_offset(s, OPC_MOVSBL + rexw + seg, datalo,
2175                                 base, index, 0, ofs);
2176        break;
2177    case MO_UW:
2178        if (use_movbe) {
2179            /* There is no extending movbe; only low 16-bits are modified.  */
2180            if (datalo != base && datalo != index) {
2181                /* XOR breaks dependency chains.  */
2182                tgen_arithr(s, ARITH_XOR, datalo, datalo);
2183                tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + P_DATA16 + seg,
2184                                         datalo, base, index, 0, ofs);
2185            } else {
2186                tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + P_DATA16 + seg,
2187                                         datalo, base, index, 0, ofs);
2188                tcg_out_ext16u(s, datalo, datalo);
2189            }
2190        } else {
2191            tcg_out_modrm_sib_offset(s, OPC_MOVZWL + seg, datalo,
2192                                     base, index, 0, ofs);
2193        }
2194        break;
2195    case MO_SW:
2196        if (use_movbe) {
2197            tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + P_DATA16 + seg,
2198                                     datalo, base, index, 0, ofs);
2199            tcg_out_ext16s(s, datalo, datalo, rexw);
2200        } else {
2201            tcg_out_modrm_sib_offset(s, OPC_MOVSWL + rexw + seg,
2202                                     datalo, base, index, 0, ofs);
2203        }
2204        break;
2205    case MO_UL:
2206        tcg_out_modrm_sib_offset(s, movop + seg, datalo, base, index, 0, ofs);
2207        break;
2208#if TCG_TARGET_REG_BITS == 64
2209    case MO_SL:
2210        if (use_movbe) {
2211            tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + seg, datalo,
2212                                     base, index, 0, ofs);
2213            tcg_out_ext32s(s, datalo, datalo);
2214        } else {
2215            tcg_out_modrm_sib_offset(s, OPC_MOVSLQ + seg, datalo,
2216                                     base, index, 0, ofs);
2217        }
2218        break;
2219#endif
2220    case MO_UQ:
2221        if (TCG_TARGET_REG_BITS == 64) {
2222            tcg_out_modrm_sib_offset(s, movop + P_REXW + seg, datalo,
2223                                     base, index, 0, ofs);
2224        } else {
2225            if (use_movbe) {
2226                TCGReg t = datalo;
2227                datalo = datahi;
2228                datahi = t;
2229            }
2230            if (base != datalo) {
2231                tcg_out_modrm_sib_offset(s, movop + seg, datalo,
2232                                         base, index, 0, ofs);
2233                tcg_out_modrm_sib_offset(s, movop + seg, datahi,
2234                                         base, index, 0, ofs + 4);
2235            } else {
2236                tcg_out_modrm_sib_offset(s, movop + seg, datahi,
2237                                         base, index, 0, ofs + 4);
2238                tcg_out_modrm_sib_offset(s, movop + seg, datalo,
2239                                         base, index, 0, ofs);
2240            }
2241        }
2242        break;
2243    default:
2244        g_assert_not_reached();
2245    }
2246}
2247
2248/* XXX: qemu_ld and qemu_st could be modified to clobber only EDX and
2249   EAX. It will be useful once fixed registers globals are less
2250   common. */
2251static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args, bool is64)
2252{
2253    TCGReg datalo, datahi, addrlo;
2254    TCGReg addrhi __attribute__((unused));
2255    MemOpIdx oi;
2256    MemOp opc;
2257#if defined(CONFIG_SOFTMMU)
2258    int mem_index;
2259    tcg_insn_unit *label_ptr[2];
2260#else
2261    unsigned a_bits;
2262#endif
2263
2264    datalo = *args++;
2265    datahi = (TCG_TARGET_REG_BITS == 32 && is64 ? *args++ : 0);
2266    addrlo = *args++;
2267    addrhi = (TARGET_LONG_BITS > TCG_TARGET_REG_BITS ? *args++ : 0);
2268    oi = *args++;
2269    opc = get_memop(oi);
2270
2271#if defined(CONFIG_SOFTMMU)
2272    mem_index = get_mmuidx(oi);
2273
2274    tcg_out_tlb_load(s, addrlo, addrhi, mem_index, opc,
2275                     label_ptr, offsetof(CPUTLBEntry, addr_read));
2276
2277    /* TLB Hit.  */
2278    tcg_out_qemu_ld_direct(s, datalo, datahi, TCG_REG_L1, -1, 0, 0, is64, opc);
2279
2280    /* Record the current context of a load into ldst label */
2281    add_qemu_ldst_label(s, true, is64, oi, datalo, datahi, addrlo, addrhi,
2282                        s->code_ptr, label_ptr);
2283#else
2284    a_bits = get_alignment_bits(opc);
2285    if (a_bits) {
2286        tcg_out_test_alignment(s, true, addrlo, addrhi, a_bits);
2287    }
2288
2289    tcg_out_qemu_ld_direct(s, datalo, datahi, addrlo, x86_guest_base_index,
2290                           x86_guest_base_offset, x86_guest_base_seg,
2291                           is64, opc);
2292#endif
2293}
2294
2295static void tcg_out_qemu_st_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
2296                                   TCGReg base, int index, intptr_t ofs,
2297                                   int seg, MemOp memop)
2298{
2299    bool use_movbe = false;
2300    int movop = OPC_MOVL_EvGv;
2301
2302    /*
2303     * Do big-endian stores with movbe or softmmu.
2304     * User-only without movbe will have its swapping done generically.
2305     */
2306    if (memop & MO_BSWAP) {
2307        tcg_debug_assert(have_movbe);
2308        use_movbe = true;
2309        movop = OPC_MOVBE_MyGy;
2310    }
2311
2312    switch (memop & MO_SIZE) {
2313    case MO_8:
2314        /* This is handled with constraints on INDEX_op_qemu_st8_i32. */
2315        tcg_debug_assert(TCG_TARGET_REG_BITS == 64 || datalo < 4);
2316        tcg_out_modrm_sib_offset(s, OPC_MOVB_EvGv + P_REXB_R + seg,
2317                                 datalo, base, index, 0, ofs);
2318        break;
2319    case MO_16:
2320        tcg_out_modrm_sib_offset(s, movop + P_DATA16 + seg, datalo,
2321                                 base, index, 0, ofs);
2322        break;
2323    case MO_32:
2324        tcg_out_modrm_sib_offset(s, movop + seg, datalo, base, index, 0, ofs);
2325        break;
2326    case MO_64:
2327        if (TCG_TARGET_REG_BITS == 64) {
2328            tcg_out_modrm_sib_offset(s, movop + P_REXW + seg, datalo,
2329                                     base, index, 0, ofs);
2330        } else {
2331            if (use_movbe) {
2332                TCGReg t = datalo;
2333                datalo = datahi;
2334                datahi = t;
2335            }
2336            tcg_out_modrm_sib_offset(s, movop + seg, datalo,
2337                                     base, index, 0, ofs);
2338            tcg_out_modrm_sib_offset(s, movop + seg, datahi,
2339                                     base, index, 0, ofs + 4);
2340        }
2341        break;
2342    default:
2343        g_assert_not_reached();
2344    }
2345}
2346
2347static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args, bool is64)
2348{
2349    TCGReg datalo, datahi, addrlo;
2350    TCGReg addrhi __attribute__((unused));
2351    MemOpIdx oi;
2352    MemOp opc;
2353#if defined(CONFIG_SOFTMMU)
2354    int mem_index;
2355    tcg_insn_unit *label_ptr[2];
2356#else
2357    unsigned a_bits;
2358#endif
2359
2360    datalo = *args++;
2361    datahi = (TCG_TARGET_REG_BITS == 32 && is64 ? *args++ : 0);
2362    addrlo = *args++;
2363    addrhi = (TARGET_LONG_BITS > TCG_TARGET_REG_BITS ? *args++ : 0);
2364    oi = *args++;
2365    opc = get_memop(oi);
2366
2367#if defined(CONFIG_SOFTMMU)
2368    mem_index = get_mmuidx(oi);
2369
2370    tcg_out_tlb_load(s, addrlo, addrhi, mem_index, opc,
2371                     label_ptr, offsetof(CPUTLBEntry, addr_write));
2372
2373    /* TLB Hit.  */
2374    tcg_out_qemu_st_direct(s, datalo, datahi, TCG_REG_L1, -1, 0, 0, opc);
2375
2376    /* Record the current context of a store into ldst label */
2377    add_qemu_ldst_label(s, false, is64, oi, datalo, datahi, addrlo, addrhi,
2378                        s->code_ptr, label_ptr);
2379#else
2380    a_bits = get_alignment_bits(opc);
2381    if (a_bits) {
2382        tcg_out_test_alignment(s, false, addrlo, addrhi, a_bits);
2383    }
2384
2385    tcg_out_qemu_st_direct(s, datalo, datahi, addrlo, x86_guest_base_index,
2386                           x86_guest_base_offset, x86_guest_base_seg, opc);
2387#endif
2388}
2389
2390static void tcg_out_exit_tb(TCGContext *s, uintptr_t a0)
2391{
2392    /* Reuse the zeroing that exists for goto_ptr.  */
2393    if (a0 == 0) {
2394        tcg_out_jmp(s, tcg_code_gen_epilogue);
2395    } else {
2396        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_EAX, a0);
2397        tcg_out_jmp(s, tb_ret_addr);
2398    }
2399}
2400
2401static void tcg_out_goto_tb(TCGContext *s, int which)
2402{
2403    /*
2404     * Jump displacement must be aligned for atomic patching;
2405     * see if we need to add extra nops before jump
2406     */
2407    int gap = QEMU_ALIGN_PTR_UP(s->code_ptr + 1, 4) - s->code_ptr;
2408    if (gap != 1) {
2409        tcg_out_nopn(s, gap - 1);
2410    }
2411    tcg_out8(s, OPC_JMP_long); /* jmp im */
2412    set_jmp_insn_offset(s, which);
2413    tcg_out32(s, 0);
2414    set_jmp_reset_offset(s, which);
2415}
2416
2417void tb_target_set_jmp_target(const TranslationBlock *tb, int n,
2418                              uintptr_t jmp_rx, uintptr_t jmp_rw)
2419{
2420    /* patch the branch destination */
2421    uintptr_t addr = tb->jmp_target_addr[n];
2422    qatomic_set((int32_t *)jmp_rw, addr - (jmp_rx + 4));
2423    /* no need to flush icache explicitly */
2424}
2425
2426static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
2427                              const TCGArg args[TCG_MAX_OP_ARGS],
2428                              const int const_args[TCG_MAX_OP_ARGS])
2429{
2430    TCGArg a0, a1, a2;
2431    int c, const_a2, vexop, rexw = 0;
2432
2433#if TCG_TARGET_REG_BITS == 64
2434# define OP_32_64(x) \
2435        case glue(glue(INDEX_op_, x), _i64): \
2436            rexw = P_REXW; /* FALLTHRU */    \
2437        case glue(glue(INDEX_op_, x), _i32)
2438#else
2439# define OP_32_64(x) \
2440        case glue(glue(INDEX_op_, x), _i32)
2441#endif
2442
2443    /* Hoist the loads of the most common arguments.  */
2444    a0 = args[0];
2445    a1 = args[1];
2446    a2 = args[2];
2447    const_a2 = const_args[2];
2448
2449    switch (opc) {
2450    case INDEX_op_goto_ptr:
2451        /* jmp to the given host address (could be epilogue) */
2452        tcg_out_modrm(s, OPC_GRP5, EXT5_JMPN_Ev, a0);
2453        break;
2454    case INDEX_op_br:
2455        tcg_out_jxx(s, JCC_JMP, arg_label(a0), 0);
2456        break;
2457    OP_32_64(ld8u):
2458        /* Note that we can ignore REXW for the zero-extend to 64-bit.  */
2459        tcg_out_modrm_offset(s, OPC_MOVZBL, a0, a1, a2);
2460        break;
2461    OP_32_64(ld8s):
2462        tcg_out_modrm_offset(s, OPC_MOVSBL + rexw, a0, a1, a2);
2463        break;
2464    OP_32_64(ld16u):
2465        /* Note that we can ignore REXW for the zero-extend to 64-bit.  */
2466        tcg_out_modrm_offset(s, OPC_MOVZWL, a0, a1, a2);
2467        break;
2468    OP_32_64(ld16s):
2469        tcg_out_modrm_offset(s, OPC_MOVSWL + rexw, a0, a1, a2);
2470        break;
2471#if TCG_TARGET_REG_BITS == 64
2472    case INDEX_op_ld32u_i64:
2473#endif
2474    case INDEX_op_ld_i32:
2475        tcg_out_ld(s, TCG_TYPE_I32, a0, a1, a2);
2476        break;
2477
2478    OP_32_64(st8):
2479        if (const_args[0]) {
2480            tcg_out_modrm_offset(s, OPC_MOVB_EvIz, 0, a1, a2);
2481            tcg_out8(s, a0);
2482        } else {
2483            tcg_out_modrm_offset(s, OPC_MOVB_EvGv | P_REXB_R, a0, a1, a2);
2484        }
2485        break;
2486    OP_32_64(st16):
2487        if (const_args[0]) {
2488            tcg_out_modrm_offset(s, OPC_MOVL_EvIz | P_DATA16, 0, a1, a2);
2489            tcg_out16(s, a0);
2490        } else {
2491            tcg_out_modrm_offset(s, OPC_MOVL_EvGv | P_DATA16, a0, a1, a2);
2492        }
2493        break;
2494#if TCG_TARGET_REG_BITS == 64
2495    case INDEX_op_st32_i64:
2496#endif
2497    case INDEX_op_st_i32:
2498        if (const_args[0]) {
2499            tcg_out_modrm_offset(s, OPC_MOVL_EvIz, 0, a1, a2);
2500            tcg_out32(s, a0);
2501        } else {
2502            tcg_out_st(s, TCG_TYPE_I32, a0, a1, a2);
2503        }
2504        break;
2505
2506    OP_32_64(add):
2507        /* For 3-operand addition, use LEA.  */
2508        if (a0 != a1) {
2509            TCGArg c3 = 0;
2510            if (const_a2) {
2511                c3 = a2, a2 = -1;
2512            } else if (a0 == a2) {
2513                /* Watch out for dest = src + dest, since we've removed
2514                   the matching constraint on the add.  */
2515                tgen_arithr(s, ARITH_ADD + rexw, a0, a1);
2516                break;
2517            }
2518
2519            tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, a1, a2, 0, c3);
2520            break;
2521        }
2522        c = ARITH_ADD;
2523        goto gen_arith;
2524    OP_32_64(sub):
2525        c = ARITH_SUB;
2526        goto gen_arith;
2527    OP_32_64(and):
2528        c = ARITH_AND;
2529        goto gen_arith;
2530    OP_32_64(or):
2531        c = ARITH_OR;
2532        goto gen_arith;
2533    OP_32_64(xor):
2534        c = ARITH_XOR;
2535        goto gen_arith;
2536    gen_arith:
2537        if (const_a2) {
2538            tgen_arithi(s, c + rexw, a0, a2, 0);
2539        } else {
2540            tgen_arithr(s, c + rexw, a0, a2);
2541        }
2542        break;
2543
2544    OP_32_64(andc):
2545        if (const_a2) {
2546            tcg_out_mov(s, rexw ? TCG_TYPE_I64 : TCG_TYPE_I32, a0, a1);
2547            tgen_arithi(s, ARITH_AND + rexw, a0, ~a2, 0);
2548        } else {
2549            tcg_out_vex_modrm(s, OPC_ANDN + rexw, a0, a2, a1);
2550        }
2551        break;
2552
2553    OP_32_64(mul):
2554        if (const_a2) {
2555            int32_t val;
2556            val = a2;
2557            if (val == (int8_t)val) {
2558                tcg_out_modrm(s, OPC_IMUL_GvEvIb + rexw, a0, a0);
2559                tcg_out8(s, val);
2560            } else {
2561                tcg_out_modrm(s, OPC_IMUL_GvEvIz + rexw, a0, a0);
2562                tcg_out32(s, val);
2563            }
2564        } else {
2565            tcg_out_modrm(s, OPC_IMUL_GvEv + rexw, a0, a2);
2566        }
2567        break;
2568
2569    OP_32_64(div2):
2570        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_IDIV, args[4]);
2571        break;
2572    OP_32_64(divu2):
2573        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_DIV, args[4]);
2574        break;
2575
2576    OP_32_64(shl):
2577        /* For small constant 3-operand shift, use LEA.  */
2578        if (const_a2 && a0 != a1 && (a2 - 1) < 3) {
2579            if (a2 - 1 == 0) {
2580                /* shl $1,a1,a0 -> lea (a1,a1),a0 */
2581                tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, a1, a1, 0, 0);
2582            } else {
2583                /* shl $n,a1,a0 -> lea 0(,a1,n),a0 */
2584                tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, -1, a1, a2, 0);
2585            }
2586            break;
2587        }
2588        c = SHIFT_SHL;
2589        vexop = OPC_SHLX;
2590        goto gen_shift_maybe_vex;
2591    OP_32_64(shr):
2592        c = SHIFT_SHR;
2593        vexop = OPC_SHRX;
2594        goto gen_shift_maybe_vex;
2595    OP_32_64(sar):
2596        c = SHIFT_SAR;
2597        vexop = OPC_SARX;
2598        goto gen_shift_maybe_vex;
2599    OP_32_64(rotl):
2600        c = SHIFT_ROL;
2601        goto gen_shift;
2602    OP_32_64(rotr):
2603        c = SHIFT_ROR;
2604        goto gen_shift;
2605    gen_shift_maybe_vex:
2606        if (have_bmi2) {
2607            if (!const_a2) {
2608                tcg_out_vex_modrm(s, vexop + rexw, a0, a2, a1);
2609                break;
2610            }
2611            tcg_out_mov(s, rexw ? TCG_TYPE_I64 : TCG_TYPE_I32, a0, a1);
2612        }
2613        /* FALLTHRU */
2614    gen_shift:
2615        if (const_a2) {
2616            tcg_out_shifti(s, c + rexw, a0, a2);
2617        } else {
2618            tcg_out_modrm(s, OPC_SHIFT_cl + rexw, c, a0);
2619        }
2620        break;
2621
2622    OP_32_64(ctz):
2623        tcg_out_ctz(s, rexw, args[0], args[1], args[2], const_args[2]);
2624        break;
2625    OP_32_64(clz):
2626        tcg_out_clz(s, rexw, args[0], args[1], args[2], const_args[2]);
2627        break;
2628    OP_32_64(ctpop):
2629        tcg_out_modrm(s, OPC_POPCNT + rexw, a0, a1);
2630        break;
2631
2632    case INDEX_op_brcond_i32:
2633        tcg_out_brcond32(s, a2, a0, a1, const_args[1], arg_label(args[3]), 0);
2634        break;
2635    case INDEX_op_setcond_i32:
2636        tcg_out_setcond32(s, args[3], a0, a1, a2, const_a2);
2637        break;
2638    case INDEX_op_movcond_i32:
2639        tcg_out_movcond32(s, args[5], a0, a1, a2, const_a2, args[3]);
2640        break;
2641
2642    OP_32_64(bswap16):
2643        if (a2 & TCG_BSWAP_OS) {
2644            /* Output must be sign-extended. */
2645            if (rexw) {
2646                tcg_out_bswap64(s, a0);
2647                tcg_out_shifti(s, SHIFT_SAR + rexw, a0, 48);
2648            } else {
2649                tcg_out_bswap32(s, a0);
2650                tcg_out_shifti(s, SHIFT_SAR, a0, 16);
2651            }
2652        } else if ((a2 & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
2653            /* Output must be zero-extended, but input isn't. */
2654            tcg_out_bswap32(s, a0);
2655            tcg_out_shifti(s, SHIFT_SHR, a0, 16);
2656        } else {
2657            tcg_out_rolw_8(s, a0);
2658        }
2659        break;
2660    OP_32_64(bswap32):
2661        tcg_out_bswap32(s, a0);
2662        if (rexw && (a2 & TCG_BSWAP_OS)) {
2663            tcg_out_ext32s(s, a0, a0);
2664        }
2665        break;
2666
2667    OP_32_64(neg):
2668        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NEG, a0);
2669        break;
2670    OP_32_64(not):
2671        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NOT, a0);
2672        break;
2673
2674    OP_32_64(ext8s):
2675        tcg_out_ext8s(s, a0, a1, rexw);
2676        break;
2677    OP_32_64(ext16s):
2678        tcg_out_ext16s(s, a0, a1, rexw);
2679        break;
2680    OP_32_64(ext8u):
2681        tcg_out_ext8u(s, a0, a1);
2682        break;
2683    OP_32_64(ext16u):
2684        tcg_out_ext16u(s, a0, a1);
2685        break;
2686
2687    case INDEX_op_qemu_ld_i32:
2688        tcg_out_qemu_ld(s, args, 0);
2689        break;
2690    case INDEX_op_qemu_ld_i64:
2691        tcg_out_qemu_ld(s, args, 1);
2692        break;
2693    case INDEX_op_qemu_st_i32:
2694    case INDEX_op_qemu_st8_i32:
2695        tcg_out_qemu_st(s, args, 0);
2696        break;
2697    case INDEX_op_qemu_st_i64:
2698        tcg_out_qemu_st(s, args, 1);
2699        break;
2700
2701    OP_32_64(mulu2):
2702        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_MUL, args[3]);
2703        break;
2704    OP_32_64(muls2):
2705        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_IMUL, args[3]);
2706        break;
2707    OP_32_64(add2):
2708        if (const_args[4]) {
2709            tgen_arithi(s, ARITH_ADD + rexw, a0, args[4], 1);
2710        } else {
2711            tgen_arithr(s, ARITH_ADD + rexw, a0, args[4]);
2712        }
2713        if (const_args[5]) {
2714            tgen_arithi(s, ARITH_ADC + rexw, a1, args[5], 1);
2715        } else {
2716            tgen_arithr(s, ARITH_ADC + rexw, a1, args[5]);
2717        }
2718        break;
2719    OP_32_64(sub2):
2720        if (const_args[4]) {
2721            tgen_arithi(s, ARITH_SUB + rexw, a0, args[4], 1);
2722        } else {
2723            tgen_arithr(s, ARITH_SUB + rexw, a0, args[4]);
2724        }
2725        if (const_args[5]) {
2726            tgen_arithi(s, ARITH_SBB + rexw, a1, args[5], 1);
2727        } else {
2728            tgen_arithr(s, ARITH_SBB + rexw, a1, args[5]);
2729        }
2730        break;
2731
2732#if TCG_TARGET_REG_BITS == 32
2733    case INDEX_op_brcond2_i32:
2734        tcg_out_brcond2(s, args, const_args, 0);
2735        break;
2736    case INDEX_op_setcond2_i32:
2737        tcg_out_setcond2(s, args, const_args);
2738        break;
2739#else /* TCG_TARGET_REG_BITS == 64 */
2740    case INDEX_op_ld32s_i64:
2741        tcg_out_modrm_offset(s, OPC_MOVSLQ, a0, a1, a2);
2742        break;
2743    case INDEX_op_ld_i64:
2744        tcg_out_ld(s, TCG_TYPE_I64, a0, a1, a2);
2745        break;
2746    case INDEX_op_st_i64:
2747        if (const_args[0]) {
2748            tcg_out_modrm_offset(s, OPC_MOVL_EvIz | P_REXW, 0, a1, a2);
2749            tcg_out32(s, a0);
2750        } else {
2751            tcg_out_st(s, TCG_TYPE_I64, a0, a1, a2);
2752        }
2753        break;
2754
2755    case INDEX_op_brcond_i64:
2756        tcg_out_brcond64(s, a2, a0, a1, const_args[1], arg_label(args[3]), 0);
2757        break;
2758    case INDEX_op_setcond_i64:
2759        tcg_out_setcond64(s, args[3], a0, a1, a2, const_a2);
2760        break;
2761    case INDEX_op_movcond_i64:
2762        tcg_out_movcond64(s, args[5], a0, a1, a2, const_a2, args[3]);
2763        break;
2764
2765    case INDEX_op_bswap64_i64:
2766        tcg_out_bswap64(s, a0);
2767        break;
2768    case INDEX_op_extu_i32_i64:
2769    case INDEX_op_ext32u_i64:
2770    case INDEX_op_extrl_i64_i32:
2771        tcg_out_ext32u(s, a0, a1);
2772        break;
2773    case INDEX_op_ext_i32_i64:
2774    case INDEX_op_ext32s_i64:
2775        tcg_out_ext32s(s, a0, a1);
2776        break;
2777    case INDEX_op_extrh_i64_i32:
2778        tcg_out_shifti(s, SHIFT_SHR + P_REXW, a0, 32);
2779        break;
2780#endif
2781
2782    OP_32_64(deposit):
2783        if (args[3] == 0 && args[4] == 8) {
2784            /* load bits 0..7 */
2785            tcg_out_modrm(s, OPC_MOVB_EvGv | P_REXB_R | P_REXB_RM, a2, a0);
2786        } else if (args[3] == 8 && args[4] == 8) {
2787            /* load bits 8..15 */
2788            tcg_out_modrm(s, OPC_MOVB_EvGv, a2, a0 + 4);
2789        } else if (args[3] == 0 && args[4] == 16) {
2790            /* load bits 0..15 */
2791            tcg_out_modrm(s, OPC_MOVL_EvGv | P_DATA16, a2, a0);
2792        } else {
2793            tcg_abort();
2794        }
2795        break;
2796
2797    case INDEX_op_extract_i64:
2798        if (a2 + args[3] == 32) {
2799            /* This is a 32-bit zero-extending right shift.  */
2800            tcg_out_mov(s, TCG_TYPE_I32, a0, a1);
2801            tcg_out_shifti(s, SHIFT_SHR, a0, a2);
2802            break;
2803        }
2804        /* FALLTHRU */
2805    case INDEX_op_extract_i32:
2806        /* On the off-chance that we can use the high-byte registers.
2807           Otherwise we emit the same ext16 + shift pattern that we
2808           would have gotten from the normal tcg-op.c expansion.  */
2809        tcg_debug_assert(a2 == 8 && args[3] == 8);
2810        if (a1 < 4 && a0 < 8) {
2811            tcg_out_modrm(s, OPC_MOVZBL, a0, a1 + 4);
2812        } else {
2813            tcg_out_ext16u(s, a0, a1);
2814            tcg_out_shifti(s, SHIFT_SHR, a0, 8);
2815        }
2816        break;
2817
2818    case INDEX_op_sextract_i32:
2819        /* We don't implement sextract_i64, as we cannot sign-extend to
2820           64-bits without using the REX prefix that explicitly excludes
2821           access to the high-byte registers.  */
2822        tcg_debug_assert(a2 == 8 && args[3] == 8);
2823        if (a1 < 4 && a0 < 8) {
2824            tcg_out_modrm(s, OPC_MOVSBL, a0, a1 + 4);
2825        } else {
2826            tcg_out_ext16s(s, a0, a1, 0);
2827            tcg_out_shifti(s, SHIFT_SAR, a0, 8);
2828        }
2829        break;
2830
2831    OP_32_64(extract2):
2832        /* Note that SHRD outputs to the r/m operand.  */
2833        tcg_out_modrm(s, OPC_SHRD_Ib + rexw, a2, a0);
2834        tcg_out8(s, args[3]);
2835        break;
2836
2837    case INDEX_op_mb:
2838        tcg_out_mb(s, a0);
2839        break;
2840    case INDEX_op_mov_i32:  /* Always emitted via tcg_out_mov.  */
2841    case INDEX_op_mov_i64:
2842    case INDEX_op_call:     /* Always emitted via tcg_out_call.  */
2843    case INDEX_op_exit_tb:  /* Always emitted via tcg_out_exit_tb.  */
2844    case INDEX_op_goto_tb:  /* Always emitted via tcg_out_goto_tb.  */
2845    default:
2846        tcg_abort();
2847    }
2848
2849#undef OP_32_64
2850}
2851
2852static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
2853                           unsigned vecl, unsigned vece,
2854                           const TCGArg args[TCG_MAX_OP_ARGS],
2855                           const int const_args[TCG_MAX_OP_ARGS])
2856{
2857    static int const add_insn[4] = {
2858        OPC_PADDB, OPC_PADDW, OPC_PADDD, OPC_PADDQ
2859    };
2860    static int const ssadd_insn[4] = {
2861        OPC_PADDSB, OPC_PADDSW, OPC_UD2, OPC_UD2
2862    };
2863    static int const usadd_insn[4] = {
2864        OPC_PADDUB, OPC_PADDUW, OPC_UD2, OPC_UD2
2865    };
2866    static int const sub_insn[4] = {
2867        OPC_PSUBB, OPC_PSUBW, OPC_PSUBD, OPC_PSUBQ
2868    };
2869    static int const sssub_insn[4] = {
2870        OPC_PSUBSB, OPC_PSUBSW, OPC_UD2, OPC_UD2
2871    };
2872    static int const ussub_insn[4] = {
2873        OPC_PSUBUB, OPC_PSUBUW, OPC_UD2, OPC_UD2
2874    };
2875    static int const mul_insn[4] = {
2876        OPC_UD2, OPC_PMULLW, OPC_PMULLD, OPC_VPMULLQ
2877    };
2878    static int const shift_imm_insn[4] = {
2879        OPC_UD2, OPC_PSHIFTW_Ib, OPC_PSHIFTD_Ib, OPC_PSHIFTQ_Ib
2880    };
2881    static int const cmpeq_insn[4] = {
2882        OPC_PCMPEQB, OPC_PCMPEQW, OPC_PCMPEQD, OPC_PCMPEQQ
2883    };
2884    static int const cmpgt_insn[4] = {
2885        OPC_PCMPGTB, OPC_PCMPGTW, OPC_PCMPGTD, OPC_PCMPGTQ
2886    };
2887    static int const punpckl_insn[4] = {
2888        OPC_PUNPCKLBW, OPC_PUNPCKLWD, OPC_PUNPCKLDQ, OPC_PUNPCKLQDQ
2889    };
2890    static int const punpckh_insn[4] = {
2891        OPC_PUNPCKHBW, OPC_PUNPCKHWD, OPC_PUNPCKHDQ, OPC_PUNPCKHQDQ
2892    };
2893    static int const packss_insn[4] = {
2894        OPC_PACKSSWB, OPC_PACKSSDW, OPC_UD2, OPC_UD2
2895    };
2896    static int const packus_insn[4] = {
2897        OPC_PACKUSWB, OPC_PACKUSDW, OPC_UD2, OPC_UD2
2898    };
2899    static int const smin_insn[4] = {
2900        OPC_PMINSB, OPC_PMINSW, OPC_PMINSD, OPC_VPMINSQ
2901    };
2902    static int const smax_insn[4] = {
2903        OPC_PMAXSB, OPC_PMAXSW, OPC_PMAXSD, OPC_VPMAXSQ
2904    };
2905    static int const umin_insn[4] = {
2906        OPC_PMINUB, OPC_PMINUW, OPC_PMINUD, OPC_VPMINUQ
2907    };
2908    static int const umax_insn[4] = {
2909        OPC_PMAXUB, OPC_PMAXUW, OPC_PMAXUD, OPC_VPMAXUQ
2910    };
2911    static int const rotlv_insn[4] = {
2912        OPC_UD2, OPC_UD2, OPC_VPROLVD, OPC_VPROLVQ
2913    };
2914    static int const rotrv_insn[4] = {
2915        OPC_UD2, OPC_UD2, OPC_VPRORVD, OPC_VPRORVQ
2916    };
2917    static int const shlv_insn[4] = {
2918        OPC_UD2, OPC_VPSLLVW, OPC_VPSLLVD, OPC_VPSLLVQ
2919    };
2920    static int const shrv_insn[4] = {
2921        OPC_UD2, OPC_VPSRLVW, OPC_VPSRLVD, OPC_VPSRLVQ
2922    };
2923    static int const sarv_insn[4] = {
2924        OPC_UD2, OPC_VPSRAVW, OPC_VPSRAVD, OPC_VPSRAVQ
2925    };
2926    static int const shls_insn[4] = {
2927        OPC_UD2, OPC_PSLLW, OPC_PSLLD, OPC_PSLLQ
2928    };
2929    static int const shrs_insn[4] = {
2930        OPC_UD2, OPC_PSRLW, OPC_PSRLD, OPC_PSRLQ
2931    };
2932    static int const sars_insn[4] = {
2933        OPC_UD2, OPC_PSRAW, OPC_PSRAD, OPC_VPSRAQ
2934    };
2935    static int const vpshldi_insn[4] = {
2936        OPC_UD2, OPC_VPSHLDW, OPC_VPSHLDD, OPC_VPSHLDQ
2937    };
2938    static int const vpshldv_insn[4] = {
2939        OPC_UD2, OPC_VPSHLDVW, OPC_VPSHLDVD, OPC_VPSHLDVQ
2940    };
2941    static int const vpshrdv_insn[4] = {
2942        OPC_UD2, OPC_VPSHRDVW, OPC_VPSHRDVD, OPC_VPSHRDVQ
2943    };
2944    static int const abs_insn[4] = {
2945        OPC_PABSB, OPC_PABSW, OPC_PABSD, OPC_VPABSQ
2946    };
2947
2948    TCGType type = vecl + TCG_TYPE_V64;
2949    int insn, sub;
2950    TCGArg a0, a1, a2, a3;
2951
2952    a0 = args[0];
2953    a1 = args[1];
2954    a2 = args[2];
2955
2956    switch (opc) {
2957    case INDEX_op_add_vec:
2958        insn = add_insn[vece];
2959        goto gen_simd;
2960    case INDEX_op_ssadd_vec:
2961        insn = ssadd_insn[vece];
2962        goto gen_simd;
2963    case INDEX_op_usadd_vec:
2964        insn = usadd_insn[vece];
2965        goto gen_simd;
2966    case INDEX_op_sub_vec:
2967        insn = sub_insn[vece];
2968        goto gen_simd;
2969    case INDEX_op_sssub_vec:
2970        insn = sssub_insn[vece];
2971        goto gen_simd;
2972    case INDEX_op_ussub_vec:
2973        insn = ussub_insn[vece];
2974        goto gen_simd;
2975    case INDEX_op_mul_vec:
2976        insn = mul_insn[vece];
2977        goto gen_simd;
2978    case INDEX_op_and_vec:
2979        insn = OPC_PAND;
2980        goto gen_simd;
2981    case INDEX_op_or_vec:
2982        insn = OPC_POR;
2983        goto gen_simd;
2984    case INDEX_op_xor_vec:
2985        insn = OPC_PXOR;
2986        goto gen_simd;
2987    case INDEX_op_smin_vec:
2988        insn = smin_insn[vece];
2989        goto gen_simd;
2990    case INDEX_op_umin_vec:
2991        insn = umin_insn[vece];
2992        goto gen_simd;
2993    case INDEX_op_smax_vec:
2994        insn = smax_insn[vece];
2995        goto gen_simd;
2996    case INDEX_op_umax_vec:
2997        insn = umax_insn[vece];
2998        goto gen_simd;
2999    case INDEX_op_shlv_vec:
3000        insn = shlv_insn[vece];
3001        goto gen_simd;
3002    case INDEX_op_shrv_vec:
3003        insn = shrv_insn[vece];
3004        goto gen_simd;
3005    case INDEX_op_sarv_vec:
3006        insn = sarv_insn[vece];
3007        goto gen_simd;
3008    case INDEX_op_rotlv_vec:
3009        insn = rotlv_insn[vece];
3010        goto gen_simd;
3011    case INDEX_op_rotrv_vec:
3012        insn = rotrv_insn[vece];
3013        goto gen_simd;
3014    case INDEX_op_shls_vec:
3015        insn = shls_insn[vece];
3016        goto gen_simd;
3017    case INDEX_op_shrs_vec:
3018        insn = shrs_insn[vece];
3019        goto gen_simd;
3020    case INDEX_op_sars_vec:
3021        insn = sars_insn[vece];
3022        goto gen_simd;
3023    case INDEX_op_x86_punpckl_vec:
3024        insn = punpckl_insn[vece];
3025        goto gen_simd;
3026    case INDEX_op_x86_punpckh_vec:
3027        insn = punpckh_insn[vece];
3028        goto gen_simd;
3029    case INDEX_op_x86_packss_vec:
3030        insn = packss_insn[vece];
3031        goto gen_simd;
3032    case INDEX_op_x86_packus_vec:
3033        insn = packus_insn[vece];
3034        goto gen_simd;
3035    case INDEX_op_x86_vpshldv_vec:
3036        insn = vpshldv_insn[vece];
3037        a1 = a2;
3038        a2 = args[3];
3039        goto gen_simd;
3040    case INDEX_op_x86_vpshrdv_vec:
3041        insn = vpshrdv_insn[vece];
3042        a1 = a2;
3043        a2 = args[3];
3044        goto gen_simd;
3045#if TCG_TARGET_REG_BITS == 32
3046    case INDEX_op_dup2_vec:
3047        /* First merge the two 32-bit inputs to a single 64-bit element. */
3048        tcg_out_vex_modrm(s, OPC_PUNPCKLDQ, a0, a1, a2);
3049        /* Then replicate the 64-bit elements across the rest of the vector. */
3050        if (type != TCG_TYPE_V64) {
3051            tcg_out_dup_vec(s, type, MO_64, a0, a0);
3052        }
3053        break;
3054#endif
3055    case INDEX_op_abs_vec:
3056        insn = abs_insn[vece];
3057        a2 = a1;
3058        a1 = 0;
3059        goto gen_simd;
3060    gen_simd:
3061        tcg_debug_assert(insn != OPC_UD2);
3062        if (type == TCG_TYPE_V256) {
3063            insn |= P_VEXL;
3064        }
3065        tcg_out_vex_modrm(s, insn, a0, a1, a2);
3066        break;
3067
3068    case INDEX_op_cmp_vec:
3069        sub = args[3];
3070        if (sub == TCG_COND_EQ) {
3071            insn = cmpeq_insn[vece];
3072        } else if (sub == TCG_COND_GT) {
3073            insn = cmpgt_insn[vece];
3074        } else {
3075            g_assert_not_reached();
3076        }
3077        goto gen_simd;
3078
3079    case INDEX_op_andc_vec:
3080        insn = OPC_PANDN;
3081        if (type == TCG_TYPE_V256) {
3082            insn |= P_VEXL;
3083        }
3084        tcg_out_vex_modrm(s, insn, a0, a2, a1);
3085        break;
3086
3087    case INDEX_op_shli_vec:
3088        insn = shift_imm_insn[vece];
3089        sub = 6;
3090        goto gen_shift;
3091    case INDEX_op_shri_vec:
3092        insn = shift_imm_insn[vece];
3093        sub = 2;
3094        goto gen_shift;
3095    case INDEX_op_sari_vec:
3096        if (vece == MO_64) {
3097            insn = OPC_PSHIFTD_Ib | P_VEXW | P_EVEX;
3098        } else {
3099            insn = shift_imm_insn[vece];
3100        }
3101        sub = 4;
3102        goto gen_shift;
3103    case INDEX_op_rotli_vec:
3104        insn = OPC_PSHIFTD_Ib | P_EVEX;  /* VPROL[DQ] */
3105        if (vece == MO_64) {
3106            insn |= P_VEXW;
3107        }
3108        sub = 1;
3109        goto gen_shift;
3110    gen_shift:
3111        tcg_debug_assert(vece != MO_8);
3112        if (type == TCG_TYPE_V256) {
3113            insn |= P_VEXL;
3114        }
3115        tcg_out_vex_modrm(s, insn, sub, a0, a1);
3116        tcg_out8(s, a2);
3117        break;
3118
3119    case INDEX_op_ld_vec:
3120        tcg_out_ld(s, type, a0, a1, a2);
3121        break;
3122    case INDEX_op_st_vec:
3123        tcg_out_st(s, type, a0, a1, a2);
3124        break;
3125    case INDEX_op_dupm_vec:
3126        tcg_out_dupm_vec(s, type, vece, a0, a1, a2);
3127        break;
3128
3129    case INDEX_op_x86_shufps_vec:
3130        insn = OPC_SHUFPS;
3131        sub = args[3];
3132        goto gen_simd_imm8;
3133    case INDEX_op_x86_blend_vec:
3134        if (vece == MO_16) {
3135            insn = OPC_PBLENDW;
3136        } else if (vece == MO_32) {
3137            insn = (have_avx2 ? OPC_VPBLENDD : OPC_BLENDPS);
3138        } else {
3139            g_assert_not_reached();
3140        }
3141        sub = args[3];
3142        goto gen_simd_imm8;
3143    case INDEX_op_x86_vperm2i128_vec:
3144        insn = OPC_VPERM2I128;
3145        sub = args[3];
3146        goto gen_simd_imm8;
3147    case INDEX_op_x86_vpshldi_vec:
3148        insn = vpshldi_insn[vece];
3149        sub = args[3];
3150        goto gen_simd_imm8;
3151
3152    case INDEX_op_not_vec:
3153        insn = OPC_VPTERNLOGQ;
3154        a2 = a1;
3155        sub = 0x33; /* !B */
3156        goto gen_simd_imm8;
3157    case INDEX_op_nor_vec:
3158        insn = OPC_VPTERNLOGQ;
3159        sub = 0x11; /* norCB */
3160        goto gen_simd_imm8;
3161    case INDEX_op_nand_vec:
3162        insn = OPC_VPTERNLOGQ;
3163        sub = 0x77; /* nandCB */
3164        goto gen_simd_imm8;
3165    case INDEX_op_eqv_vec:
3166        insn = OPC_VPTERNLOGQ;
3167        sub = 0x99; /* xnorCB */
3168        goto gen_simd_imm8;
3169    case INDEX_op_orc_vec:
3170        insn = OPC_VPTERNLOGQ;
3171        sub = 0xdd; /* orB!C */
3172        goto gen_simd_imm8;
3173
3174    case INDEX_op_bitsel_vec:
3175        insn = OPC_VPTERNLOGQ;
3176        a3 = args[3];
3177        if (a0 == a1) {
3178            a1 = a2;
3179            a2 = a3;
3180            sub = 0xca; /* A?B:C */
3181        } else if (a0 == a2) {
3182            a2 = a3;
3183            sub = 0xe2; /* B?A:C */
3184        } else {
3185            tcg_out_mov(s, type, a0, a3);
3186            sub = 0xb8; /* B?C:A */
3187        }
3188        goto gen_simd_imm8;
3189
3190    gen_simd_imm8:
3191        tcg_debug_assert(insn != OPC_UD2);
3192        if (type == TCG_TYPE_V256) {
3193            insn |= P_VEXL;
3194        }
3195        tcg_out_vex_modrm(s, insn, a0, a1, a2);
3196        tcg_out8(s, sub);
3197        break;
3198
3199    case INDEX_op_x86_vpblendvb_vec:
3200        insn = OPC_VPBLENDVB;
3201        if (type == TCG_TYPE_V256) {
3202            insn |= P_VEXL;
3203        }
3204        tcg_out_vex_modrm(s, insn, a0, a1, a2);
3205        tcg_out8(s, args[3] << 4);
3206        break;
3207
3208    case INDEX_op_x86_psrldq_vec:
3209        tcg_out_vex_modrm(s, OPC_GRP14, 3, a0, a1);
3210        tcg_out8(s, a2);
3211        break;
3212
3213    case INDEX_op_mov_vec:  /* Always emitted via tcg_out_mov.  */
3214    case INDEX_op_dup_vec:  /* Always emitted via tcg_out_dup_vec.  */
3215    default:
3216        g_assert_not_reached();
3217    }
3218}
3219
3220static TCGConstraintSetIndex tcg_target_op_def(TCGOpcode op)
3221{
3222    switch (op) {
3223    case INDEX_op_goto_ptr:
3224        return C_O0_I1(r);
3225
3226    case INDEX_op_ld8u_i32:
3227    case INDEX_op_ld8u_i64:
3228    case INDEX_op_ld8s_i32:
3229    case INDEX_op_ld8s_i64:
3230    case INDEX_op_ld16u_i32:
3231    case INDEX_op_ld16u_i64:
3232    case INDEX_op_ld16s_i32:
3233    case INDEX_op_ld16s_i64:
3234    case INDEX_op_ld_i32:
3235    case INDEX_op_ld32u_i64:
3236    case INDEX_op_ld32s_i64:
3237    case INDEX_op_ld_i64:
3238        return C_O1_I1(r, r);
3239
3240    case INDEX_op_st8_i32:
3241    case INDEX_op_st8_i64:
3242        return C_O0_I2(qi, r);
3243
3244    case INDEX_op_st16_i32:
3245    case INDEX_op_st16_i64:
3246    case INDEX_op_st_i32:
3247    case INDEX_op_st32_i64:
3248        return C_O0_I2(ri, r);
3249
3250    case INDEX_op_st_i64:
3251        return C_O0_I2(re, r);
3252
3253    case INDEX_op_add_i32:
3254    case INDEX_op_add_i64:
3255        return C_O1_I2(r, r, re);
3256
3257    case INDEX_op_sub_i32:
3258    case INDEX_op_sub_i64:
3259    case INDEX_op_mul_i32:
3260    case INDEX_op_mul_i64:
3261    case INDEX_op_or_i32:
3262    case INDEX_op_or_i64:
3263    case INDEX_op_xor_i32:
3264    case INDEX_op_xor_i64:
3265        return C_O1_I2(r, 0, re);
3266
3267    case INDEX_op_and_i32:
3268    case INDEX_op_and_i64:
3269        return C_O1_I2(r, 0, reZ);
3270
3271    case INDEX_op_andc_i32:
3272    case INDEX_op_andc_i64:
3273        return C_O1_I2(r, r, rI);
3274
3275    case INDEX_op_shl_i32:
3276    case INDEX_op_shl_i64:
3277    case INDEX_op_shr_i32:
3278    case INDEX_op_shr_i64:
3279    case INDEX_op_sar_i32:
3280    case INDEX_op_sar_i64:
3281        return have_bmi2 ? C_O1_I2(r, r, ri) : C_O1_I2(r, 0, ci);
3282
3283    case INDEX_op_rotl_i32:
3284    case INDEX_op_rotl_i64:
3285    case INDEX_op_rotr_i32:
3286    case INDEX_op_rotr_i64:
3287        return C_O1_I2(r, 0, ci);
3288
3289    case INDEX_op_brcond_i32:
3290    case INDEX_op_brcond_i64:
3291        return C_O0_I2(r, re);
3292
3293    case INDEX_op_bswap16_i32:
3294    case INDEX_op_bswap16_i64:
3295    case INDEX_op_bswap32_i32:
3296    case INDEX_op_bswap32_i64:
3297    case INDEX_op_bswap64_i64:
3298    case INDEX_op_neg_i32:
3299    case INDEX_op_neg_i64:
3300    case INDEX_op_not_i32:
3301    case INDEX_op_not_i64:
3302    case INDEX_op_extrh_i64_i32:
3303        return C_O1_I1(r, 0);
3304
3305    case INDEX_op_ext8s_i32:
3306    case INDEX_op_ext8s_i64:
3307    case INDEX_op_ext8u_i32:
3308    case INDEX_op_ext8u_i64:
3309        return C_O1_I1(r, q);
3310
3311    case INDEX_op_ext16s_i32:
3312    case INDEX_op_ext16s_i64:
3313    case INDEX_op_ext16u_i32:
3314    case INDEX_op_ext16u_i64:
3315    case INDEX_op_ext32s_i64:
3316    case INDEX_op_ext32u_i64:
3317    case INDEX_op_ext_i32_i64:
3318    case INDEX_op_extu_i32_i64:
3319    case INDEX_op_extrl_i64_i32:
3320    case INDEX_op_extract_i32:
3321    case INDEX_op_extract_i64:
3322    case INDEX_op_sextract_i32:
3323    case INDEX_op_ctpop_i32:
3324    case INDEX_op_ctpop_i64:
3325        return C_O1_I1(r, r);
3326
3327    case INDEX_op_extract2_i32:
3328    case INDEX_op_extract2_i64:
3329        return C_O1_I2(r, 0, r);
3330
3331    case INDEX_op_deposit_i32:
3332    case INDEX_op_deposit_i64:
3333        return C_O1_I2(Q, 0, Q);
3334
3335    case INDEX_op_setcond_i32:
3336    case INDEX_op_setcond_i64:
3337        return C_O1_I2(q, r, re);
3338
3339    case INDEX_op_movcond_i32:
3340    case INDEX_op_movcond_i64:
3341        return C_O1_I4(r, r, re, r, 0);
3342
3343    case INDEX_op_div2_i32:
3344    case INDEX_op_div2_i64:
3345    case INDEX_op_divu2_i32:
3346    case INDEX_op_divu2_i64:
3347        return C_O2_I3(a, d, 0, 1, r);
3348
3349    case INDEX_op_mulu2_i32:
3350    case INDEX_op_mulu2_i64:
3351    case INDEX_op_muls2_i32:
3352    case INDEX_op_muls2_i64:
3353        return C_O2_I2(a, d, a, r);
3354
3355    case INDEX_op_add2_i32:
3356    case INDEX_op_add2_i64:
3357    case INDEX_op_sub2_i32:
3358    case INDEX_op_sub2_i64:
3359        return C_O2_I4(r, r, 0, 1, re, re);
3360
3361    case INDEX_op_ctz_i32:
3362    case INDEX_op_ctz_i64:
3363        return have_bmi1 ? C_N1_I2(r, r, rW) : C_N1_I2(r, r, r);
3364
3365    case INDEX_op_clz_i32:
3366    case INDEX_op_clz_i64:
3367        return have_lzcnt ? C_N1_I2(r, r, rW) : C_N1_I2(r, r, r);
3368
3369    case INDEX_op_qemu_ld_i32:
3370        return (TARGET_LONG_BITS <= TCG_TARGET_REG_BITS
3371                ? C_O1_I1(r, L) : C_O1_I2(r, L, L));
3372
3373    case INDEX_op_qemu_st_i32:
3374        return (TARGET_LONG_BITS <= TCG_TARGET_REG_BITS
3375                ? C_O0_I2(L, L) : C_O0_I3(L, L, L));
3376    case INDEX_op_qemu_st8_i32:
3377        return (TARGET_LONG_BITS <= TCG_TARGET_REG_BITS
3378                ? C_O0_I2(s, L) : C_O0_I3(s, L, L));
3379
3380    case INDEX_op_qemu_ld_i64:
3381        return (TCG_TARGET_REG_BITS == 64 ? C_O1_I1(r, L)
3382                : TARGET_LONG_BITS <= TCG_TARGET_REG_BITS ? C_O2_I1(r, r, L)
3383                : C_O2_I2(r, r, L, L));
3384
3385    case INDEX_op_qemu_st_i64:
3386        return (TCG_TARGET_REG_BITS == 64 ? C_O0_I2(L, L)
3387                : TARGET_LONG_BITS <= TCG_TARGET_REG_BITS ? C_O0_I3(L, L, L)
3388                : C_O0_I4(L, L, L, L));
3389
3390    case INDEX_op_brcond2_i32:
3391        return C_O0_I4(r, r, ri, ri);
3392
3393    case INDEX_op_setcond2_i32:
3394        return C_O1_I4(r, r, r, ri, ri);
3395
3396    case INDEX_op_ld_vec:
3397    case INDEX_op_dupm_vec:
3398        return C_O1_I1(x, r);
3399
3400    case INDEX_op_st_vec:
3401        return C_O0_I2(x, r);
3402
3403    case INDEX_op_add_vec:
3404    case INDEX_op_sub_vec:
3405    case INDEX_op_mul_vec:
3406    case INDEX_op_and_vec:
3407    case INDEX_op_or_vec:
3408    case INDEX_op_xor_vec:
3409    case INDEX_op_andc_vec:
3410    case INDEX_op_orc_vec:
3411    case INDEX_op_nand_vec:
3412    case INDEX_op_nor_vec:
3413    case INDEX_op_eqv_vec:
3414    case INDEX_op_ssadd_vec:
3415    case INDEX_op_usadd_vec:
3416    case INDEX_op_sssub_vec:
3417    case INDEX_op_ussub_vec:
3418    case INDEX_op_smin_vec:
3419    case INDEX_op_umin_vec:
3420    case INDEX_op_smax_vec:
3421    case INDEX_op_umax_vec:
3422    case INDEX_op_shlv_vec:
3423    case INDEX_op_shrv_vec:
3424    case INDEX_op_sarv_vec:
3425    case INDEX_op_rotlv_vec:
3426    case INDEX_op_rotrv_vec:
3427    case INDEX_op_shls_vec:
3428    case INDEX_op_shrs_vec:
3429    case INDEX_op_sars_vec:
3430    case INDEX_op_cmp_vec:
3431    case INDEX_op_x86_shufps_vec:
3432    case INDEX_op_x86_blend_vec:
3433    case INDEX_op_x86_packss_vec:
3434    case INDEX_op_x86_packus_vec:
3435    case INDEX_op_x86_vperm2i128_vec:
3436    case INDEX_op_x86_punpckl_vec:
3437    case INDEX_op_x86_punpckh_vec:
3438    case INDEX_op_x86_vpshldi_vec:
3439#if TCG_TARGET_REG_BITS == 32
3440    case INDEX_op_dup2_vec:
3441#endif
3442        return C_O1_I2(x, x, x);
3443
3444    case INDEX_op_abs_vec:
3445    case INDEX_op_dup_vec:
3446    case INDEX_op_not_vec:
3447    case INDEX_op_shli_vec:
3448    case INDEX_op_shri_vec:
3449    case INDEX_op_sari_vec:
3450    case INDEX_op_rotli_vec:
3451    case INDEX_op_x86_psrldq_vec:
3452        return C_O1_I1(x, x);
3453
3454    case INDEX_op_x86_vpshldv_vec:
3455    case INDEX_op_x86_vpshrdv_vec:
3456        return C_O1_I3(x, 0, x, x);
3457
3458    case INDEX_op_bitsel_vec:
3459    case INDEX_op_x86_vpblendvb_vec:
3460        return C_O1_I3(x, x, x, x);
3461
3462    default:
3463        g_assert_not_reached();
3464    }
3465}
3466
3467int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
3468{
3469    switch (opc) {
3470    case INDEX_op_add_vec:
3471    case INDEX_op_sub_vec:
3472    case INDEX_op_and_vec:
3473    case INDEX_op_or_vec:
3474    case INDEX_op_xor_vec:
3475    case INDEX_op_andc_vec:
3476    case INDEX_op_orc_vec:
3477    case INDEX_op_nand_vec:
3478    case INDEX_op_nor_vec:
3479    case INDEX_op_eqv_vec:
3480    case INDEX_op_not_vec:
3481    case INDEX_op_bitsel_vec:
3482        return 1;
3483    case INDEX_op_cmp_vec:
3484    case INDEX_op_cmpsel_vec:
3485        return -1;
3486
3487    case INDEX_op_rotli_vec:
3488        return have_avx512vl && vece >= MO_32 ? 1 : -1;
3489
3490    case INDEX_op_shli_vec:
3491    case INDEX_op_shri_vec:
3492        /* We must expand the operation for MO_8.  */
3493        return vece == MO_8 ? -1 : 1;
3494
3495    case INDEX_op_sari_vec:
3496        switch (vece) {
3497        case MO_8:
3498            return -1;
3499        case MO_16:
3500        case MO_32:
3501            return 1;
3502        case MO_64:
3503            if (have_avx512vl) {
3504                return 1;
3505            }
3506            /*
3507             * We can emulate this for MO_64, but it does not pay off
3508             * unless we're producing at least 4 values.
3509             */
3510            return type >= TCG_TYPE_V256 ? -1 : 0;
3511        }
3512        return 0;
3513
3514    case INDEX_op_shls_vec:
3515    case INDEX_op_shrs_vec:
3516        return vece >= MO_16;
3517    case INDEX_op_sars_vec:
3518        switch (vece) {
3519        case MO_16:
3520        case MO_32:
3521            return 1;
3522        case MO_64:
3523            return have_avx512vl;
3524        }
3525        return 0;
3526    case INDEX_op_rotls_vec:
3527        return vece >= MO_16 ? -1 : 0;
3528
3529    case INDEX_op_shlv_vec:
3530    case INDEX_op_shrv_vec:
3531        switch (vece) {
3532        case MO_16:
3533            return have_avx512bw;
3534        case MO_32:
3535        case MO_64:
3536            return have_avx2;
3537        }
3538        return 0;
3539    case INDEX_op_sarv_vec:
3540        switch (vece) {
3541        case MO_16:
3542            return have_avx512bw;
3543        case MO_32:
3544            return have_avx2;
3545        case MO_64:
3546            return have_avx512vl;
3547        }
3548        return 0;
3549    case INDEX_op_rotlv_vec:
3550    case INDEX_op_rotrv_vec:
3551        switch (vece) {
3552        case MO_16:
3553            return have_avx512vbmi2 ? -1 : 0;
3554        case MO_32:
3555        case MO_64:
3556            return have_avx512vl ? 1 : have_avx2 ? -1 : 0;
3557        }
3558        return 0;
3559
3560    case INDEX_op_mul_vec:
3561        switch (vece) {
3562        case MO_8:
3563            return -1;
3564        case MO_64:
3565            return have_avx512dq;
3566        }
3567        return 1;
3568
3569    case INDEX_op_ssadd_vec:
3570    case INDEX_op_usadd_vec:
3571    case INDEX_op_sssub_vec:
3572    case INDEX_op_ussub_vec:
3573        return vece <= MO_16;
3574    case INDEX_op_smin_vec:
3575    case INDEX_op_smax_vec:
3576    case INDEX_op_umin_vec:
3577    case INDEX_op_umax_vec:
3578    case INDEX_op_abs_vec:
3579        return vece <= MO_32 || have_avx512vl;
3580
3581    default:
3582        return 0;
3583    }
3584}
3585
3586static void expand_vec_shi(TCGType type, unsigned vece, TCGOpcode opc,
3587                           TCGv_vec v0, TCGv_vec v1, TCGArg imm)
3588{
3589    TCGv_vec t1, t2;
3590
3591    tcg_debug_assert(vece == MO_8);
3592
3593    t1 = tcg_temp_new_vec(type);
3594    t2 = tcg_temp_new_vec(type);
3595
3596    /*
3597     * Unpack to W, shift, and repack.  Tricky bits:
3598     * (1) Use punpck*bw x,x to produce DDCCBBAA,
3599     *     i.e. duplicate in other half of the 16-bit lane.
3600     * (2) For right-shift, add 8 so that the high half of the lane
3601     *     becomes zero.  For left-shift, and left-rotate, we must
3602     *     shift up and down again.
3603     * (3) Step 2 leaves high half zero such that PACKUSWB
3604     *     (pack with unsigned saturation) does not modify
3605     *     the quantity.
3606     */
3607    vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
3608              tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
3609    vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
3610              tcgv_vec_arg(t2), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
3611
3612    if (opc != INDEX_op_rotli_vec) {
3613        imm += 8;
3614    }
3615    if (opc == INDEX_op_shri_vec) {
3616        tcg_gen_shri_vec(MO_16, t1, t1, imm);
3617        tcg_gen_shri_vec(MO_16, t2, t2, imm);
3618    } else {
3619        tcg_gen_shli_vec(MO_16, t1, t1, imm);
3620        tcg_gen_shli_vec(MO_16, t2, t2, imm);
3621        tcg_gen_shri_vec(MO_16, t1, t1, 8);
3622        tcg_gen_shri_vec(MO_16, t2, t2, 8);
3623    }
3624
3625    vec_gen_3(INDEX_op_x86_packus_vec, type, MO_8,
3626              tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t2));
3627    tcg_temp_free_vec(t1);
3628    tcg_temp_free_vec(t2);
3629}
3630
3631static void expand_vec_sari(TCGType type, unsigned vece,
3632                            TCGv_vec v0, TCGv_vec v1, TCGArg imm)
3633{
3634    TCGv_vec t1, t2;
3635
3636    switch (vece) {
3637    case MO_8:
3638        /* Unpack to W, shift, and repack, as in expand_vec_shi.  */
3639        t1 = tcg_temp_new_vec(type);
3640        t2 = tcg_temp_new_vec(type);
3641        vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
3642                  tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
3643        vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
3644                  tcgv_vec_arg(t2), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
3645        tcg_gen_sari_vec(MO_16, t1, t1, imm + 8);
3646        tcg_gen_sari_vec(MO_16, t2, t2, imm + 8);
3647        vec_gen_3(INDEX_op_x86_packss_vec, type, MO_8,
3648                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t2));
3649        tcg_temp_free_vec(t1);
3650        tcg_temp_free_vec(t2);
3651        break;
3652
3653    case MO_64:
3654        t1 = tcg_temp_new_vec(type);
3655        if (imm <= 32) {
3656            /*
3657             * We can emulate a small sign extend by performing an arithmetic
3658             * 32-bit shift and overwriting the high half of a 64-bit logical
3659             * shift.  Note that the ISA says shift of 32 is valid, but TCG
3660             * does not, so we have to bound the smaller shift -- we get the
3661             * same result in the high half either way.
3662             */
3663            tcg_gen_sari_vec(MO_32, t1, v1, MIN(imm, 31));
3664            tcg_gen_shri_vec(MO_64, v0, v1, imm);
3665            vec_gen_4(INDEX_op_x86_blend_vec, type, MO_32,
3666                      tcgv_vec_arg(v0), tcgv_vec_arg(v0),
3667                      tcgv_vec_arg(t1), 0xaa);
3668        } else {
3669            /* Otherwise we will need to use a compare vs 0 to produce
3670             * the sign-extend, shift and merge.
3671             */
3672            tcg_gen_cmp_vec(TCG_COND_GT, MO_64, t1,
3673                            tcg_constant_vec(type, MO_64, 0), v1);
3674            tcg_gen_shri_vec(MO_64, v0, v1, imm);
3675            tcg_gen_shli_vec(MO_64, t1, t1, 64 - imm);
3676            tcg_gen_or_vec(MO_64, v0, v0, t1);
3677        }
3678        tcg_temp_free_vec(t1);
3679        break;
3680
3681    default:
3682        g_assert_not_reached();
3683    }
3684}
3685
3686static void expand_vec_rotli(TCGType type, unsigned vece,
3687                             TCGv_vec v0, TCGv_vec v1, TCGArg imm)
3688{
3689    TCGv_vec t;
3690
3691    if (vece == MO_8) {
3692        expand_vec_shi(type, vece, INDEX_op_rotli_vec, v0, v1, imm);
3693        return;
3694    }
3695
3696    if (have_avx512vbmi2) {
3697        vec_gen_4(INDEX_op_x86_vpshldi_vec, type, vece,
3698                  tcgv_vec_arg(v0), tcgv_vec_arg(v1), tcgv_vec_arg(v1), imm);
3699        return;
3700    }
3701
3702    t = tcg_temp_new_vec(type);
3703    tcg_gen_shli_vec(vece, t, v1, imm);
3704    tcg_gen_shri_vec(vece, v0, v1, (8 << vece) - imm);
3705    tcg_gen_or_vec(vece, v0, v0, t);
3706    tcg_temp_free_vec(t);
3707}
3708
3709static void expand_vec_rotv(TCGType type, unsigned vece, TCGv_vec v0,
3710                            TCGv_vec v1, TCGv_vec sh, bool right)
3711{
3712    TCGv_vec t;
3713
3714    if (have_avx512vbmi2) {
3715        vec_gen_4(right ? INDEX_op_x86_vpshrdv_vec : INDEX_op_x86_vpshldv_vec,
3716                  type, vece, tcgv_vec_arg(v0), tcgv_vec_arg(v1),
3717                  tcgv_vec_arg(v1), tcgv_vec_arg(sh));
3718        return;
3719    }
3720
3721    t = tcg_temp_new_vec(type);
3722    tcg_gen_dupi_vec(vece, t, 8 << vece);
3723    tcg_gen_sub_vec(vece, t, t, sh);
3724    if (right) {
3725        tcg_gen_shlv_vec(vece, t, v1, t);
3726        tcg_gen_shrv_vec(vece, v0, v1, sh);
3727    } else {
3728        tcg_gen_shrv_vec(vece, t, v1, t);
3729        tcg_gen_shlv_vec(vece, v0, v1, sh);
3730    }
3731    tcg_gen_or_vec(vece, v0, v0, t);
3732    tcg_temp_free_vec(t);
3733}
3734
3735static void expand_vec_rotls(TCGType type, unsigned vece,
3736                             TCGv_vec v0, TCGv_vec v1, TCGv_i32 lsh)
3737{
3738    TCGv_vec t = tcg_temp_new_vec(type);
3739
3740    tcg_debug_assert(vece != MO_8);
3741
3742    if (vece >= MO_32 ? have_avx512vl : have_avx512vbmi2) {
3743        tcg_gen_dup_i32_vec(vece, t, lsh);
3744        if (vece >= MO_32) {
3745            tcg_gen_rotlv_vec(vece, v0, v1, t);
3746        } else {
3747            expand_vec_rotv(type, vece, v0, v1, t, false);
3748        }
3749    } else {
3750        TCGv_i32 rsh = tcg_temp_new_i32();
3751
3752        tcg_gen_neg_i32(rsh, lsh);
3753        tcg_gen_andi_i32(rsh, rsh, (8 << vece) - 1);
3754        tcg_gen_shls_vec(vece, t, v1, lsh);
3755        tcg_gen_shrs_vec(vece, v0, v1, rsh);
3756        tcg_gen_or_vec(vece, v0, v0, t);
3757
3758        tcg_temp_free_i32(rsh);
3759    }
3760
3761    tcg_temp_free_vec(t);
3762}
3763
3764static void expand_vec_mul(TCGType type, unsigned vece,
3765                           TCGv_vec v0, TCGv_vec v1, TCGv_vec v2)
3766{
3767    TCGv_vec t1, t2, t3, t4, zero;
3768
3769    tcg_debug_assert(vece == MO_8);
3770
3771    /*
3772     * Unpack v1 bytes to words, 0 | x.
3773     * Unpack v2 bytes to words, y | 0.
3774     * This leaves the 8-bit result, x * y, with 8 bits of right padding.
3775     * Shift logical right by 8 bits to clear the high 8 bytes before
3776     * using an unsigned saturated pack.
3777     *
3778     * The difference between the V64, V128 and V256 cases is merely how
3779     * we distribute the expansion between temporaries.
3780     */
3781    switch (type) {
3782    case TCG_TYPE_V64:
3783        t1 = tcg_temp_new_vec(TCG_TYPE_V128);
3784        t2 = tcg_temp_new_vec(TCG_TYPE_V128);
3785        zero = tcg_constant_vec(TCG_TYPE_V128, MO_8, 0);
3786        vec_gen_3(INDEX_op_x86_punpckl_vec, TCG_TYPE_V128, MO_8,
3787                  tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
3788        vec_gen_3(INDEX_op_x86_punpckl_vec, TCG_TYPE_V128, MO_8,
3789                  tcgv_vec_arg(t2), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
3790        tcg_gen_mul_vec(MO_16, t1, t1, t2);
3791        tcg_gen_shri_vec(MO_16, t1, t1, 8);
3792        vec_gen_3(INDEX_op_x86_packus_vec, TCG_TYPE_V128, MO_8,
3793                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t1));
3794        tcg_temp_free_vec(t1);
3795        tcg_temp_free_vec(t2);
3796        break;
3797
3798    case TCG_TYPE_V128:
3799    case TCG_TYPE_V256:
3800        t1 = tcg_temp_new_vec(type);
3801        t2 = tcg_temp_new_vec(type);
3802        t3 = tcg_temp_new_vec(type);
3803        t4 = tcg_temp_new_vec(type);
3804        zero = tcg_constant_vec(TCG_TYPE_V128, MO_8, 0);
3805        vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
3806                  tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
3807        vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
3808                  tcgv_vec_arg(t2), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
3809        vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
3810                  tcgv_vec_arg(t3), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
3811        vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
3812                  tcgv_vec_arg(t4), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
3813        tcg_gen_mul_vec(MO_16, t1, t1, t2);
3814        tcg_gen_mul_vec(MO_16, t3, t3, t4);
3815        tcg_gen_shri_vec(MO_16, t1, t1, 8);
3816        tcg_gen_shri_vec(MO_16, t3, t3, 8);
3817        vec_gen_3(INDEX_op_x86_packus_vec, type, MO_8,
3818                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t3));
3819        tcg_temp_free_vec(t1);
3820        tcg_temp_free_vec(t2);
3821        tcg_temp_free_vec(t3);
3822        tcg_temp_free_vec(t4);
3823        break;
3824
3825    default:
3826        g_assert_not_reached();
3827    }
3828}
3829
3830static bool expand_vec_cmp_noinv(TCGType type, unsigned vece, TCGv_vec v0,
3831                                 TCGv_vec v1, TCGv_vec v2, TCGCond cond)
3832{
3833    enum {
3834        NEED_INV  = 1,
3835        NEED_SWAP = 2,
3836        NEED_BIAS = 4,
3837        NEED_UMIN = 8,
3838        NEED_UMAX = 16,
3839    };
3840    TCGv_vec t1, t2, t3;
3841    uint8_t fixup;
3842
3843    switch (cond) {
3844    case TCG_COND_EQ:
3845    case TCG_COND_GT:
3846        fixup = 0;
3847        break;
3848    case TCG_COND_NE:
3849    case TCG_COND_LE:
3850        fixup = NEED_INV;
3851        break;
3852    case TCG_COND_LT:
3853        fixup = NEED_SWAP;
3854        break;
3855    case TCG_COND_GE:
3856        fixup = NEED_SWAP | NEED_INV;
3857        break;
3858    case TCG_COND_LEU:
3859        if (tcg_can_emit_vec_op(INDEX_op_umin_vec, type, vece)) {
3860            fixup = NEED_UMIN;
3861        } else {
3862            fixup = NEED_BIAS | NEED_INV;
3863        }
3864        break;
3865    case TCG_COND_GTU:
3866        if (tcg_can_emit_vec_op(INDEX_op_umin_vec, type, vece)) {
3867            fixup = NEED_UMIN | NEED_INV;
3868        } else {
3869            fixup = NEED_BIAS;
3870        }
3871        break;
3872    case TCG_COND_GEU:
3873        if (tcg_can_emit_vec_op(INDEX_op_umax_vec, type, vece)) {
3874            fixup = NEED_UMAX;
3875        } else {
3876            fixup = NEED_BIAS | NEED_SWAP | NEED_INV;
3877        }
3878        break;
3879    case TCG_COND_LTU:
3880        if (tcg_can_emit_vec_op(INDEX_op_umax_vec, type, vece)) {
3881            fixup = NEED_UMAX | NEED_INV;
3882        } else {
3883            fixup = NEED_BIAS | NEED_SWAP;
3884        }
3885        break;
3886    default:
3887        g_assert_not_reached();
3888    }
3889
3890    if (fixup & NEED_INV) {
3891        cond = tcg_invert_cond(cond);
3892    }
3893    if (fixup & NEED_SWAP) {
3894        t1 = v1, v1 = v2, v2 = t1;
3895        cond = tcg_swap_cond(cond);
3896    }
3897
3898    t1 = t2 = NULL;
3899    if (fixup & (NEED_UMIN | NEED_UMAX)) {
3900        t1 = tcg_temp_new_vec(type);
3901        if (fixup & NEED_UMIN) {
3902            tcg_gen_umin_vec(vece, t1, v1, v2);
3903        } else {
3904            tcg_gen_umax_vec(vece, t1, v1, v2);
3905        }
3906        v2 = t1;
3907        cond = TCG_COND_EQ;
3908    } else if (fixup & NEED_BIAS) {
3909        t1 = tcg_temp_new_vec(type);
3910        t2 = tcg_temp_new_vec(type);
3911        t3 = tcg_constant_vec(type, vece, 1ull << ((8 << vece) - 1));
3912        tcg_gen_sub_vec(vece, t1, v1, t3);
3913        tcg_gen_sub_vec(vece, t2, v2, t3);
3914        v1 = t1;
3915        v2 = t2;
3916        cond = tcg_signed_cond(cond);
3917    }
3918
3919    tcg_debug_assert(cond == TCG_COND_EQ || cond == TCG_COND_GT);
3920    /* Expand directly; do not recurse.  */
3921    vec_gen_4(INDEX_op_cmp_vec, type, vece,
3922              tcgv_vec_arg(v0), tcgv_vec_arg(v1), tcgv_vec_arg(v2), cond);
3923
3924    if (t1) {
3925        tcg_temp_free_vec(t1);
3926        if (t2) {
3927            tcg_temp_free_vec(t2);
3928        }
3929    }
3930    return fixup & NEED_INV;
3931}
3932
3933static void expand_vec_cmp(TCGType type, unsigned vece, TCGv_vec v0,
3934                           TCGv_vec v1, TCGv_vec v2, TCGCond cond)
3935{
3936    if (expand_vec_cmp_noinv(type, vece, v0, v1, v2, cond)) {
3937        tcg_gen_not_vec(vece, v0, v0);
3938    }
3939}
3940
3941static void expand_vec_cmpsel(TCGType type, unsigned vece, TCGv_vec v0,
3942                              TCGv_vec c1, TCGv_vec c2,
3943                              TCGv_vec v3, TCGv_vec v4, TCGCond cond)
3944{
3945    TCGv_vec t = tcg_temp_new_vec(type);
3946
3947    if (expand_vec_cmp_noinv(type, vece, t, c1, c2, cond)) {
3948        /* Invert the sense of the compare by swapping arguments.  */
3949        TCGv_vec x;
3950        x = v3, v3 = v4, v4 = x;
3951    }
3952    vec_gen_4(INDEX_op_x86_vpblendvb_vec, type, vece,
3953              tcgv_vec_arg(v0), tcgv_vec_arg(v4),
3954              tcgv_vec_arg(v3), tcgv_vec_arg(t));
3955    tcg_temp_free_vec(t);
3956}
3957
3958void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
3959                       TCGArg a0, ...)
3960{
3961    va_list va;
3962    TCGArg a2;
3963    TCGv_vec v0, v1, v2, v3, v4;
3964
3965    va_start(va, a0);
3966    v0 = temp_tcgv_vec(arg_temp(a0));
3967    v1 = temp_tcgv_vec(arg_temp(va_arg(va, TCGArg)));
3968    a2 = va_arg(va, TCGArg);
3969
3970    switch (opc) {
3971    case INDEX_op_shli_vec:
3972    case INDEX_op_shri_vec:
3973        expand_vec_shi(type, vece, opc, v0, v1, a2);
3974        break;
3975
3976    case INDEX_op_sari_vec:
3977        expand_vec_sari(type, vece, v0, v1, a2);
3978        break;
3979
3980    case INDEX_op_rotli_vec:
3981        expand_vec_rotli(type, vece, v0, v1, a2);
3982        break;
3983
3984    case INDEX_op_rotls_vec:
3985        expand_vec_rotls(type, vece, v0, v1, temp_tcgv_i32(arg_temp(a2)));
3986        break;
3987
3988    case INDEX_op_rotlv_vec:
3989        v2 = temp_tcgv_vec(arg_temp(a2));
3990        expand_vec_rotv(type, vece, v0, v1, v2, false);
3991        break;
3992    case INDEX_op_rotrv_vec:
3993        v2 = temp_tcgv_vec(arg_temp(a2));
3994        expand_vec_rotv(type, vece, v0, v1, v2, true);
3995        break;
3996
3997    case INDEX_op_mul_vec:
3998        v2 = temp_tcgv_vec(arg_temp(a2));
3999        expand_vec_mul(type, vece, v0, v1, v2);
4000        break;
4001
4002    case INDEX_op_cmp_vec:
4003        v2 = temp_tcgv_vec(arg_temp(a2));
4004        expand_vec_cmp(type, vece, v0, v1, v2, va_arg(va, TCGArg));
4005        break;
4006
4007    case INDEX_op_cmpsel_vec:
4008        v2 = temp_tcgv_vec(arg_temp(a2));
4009        v3 = temp_tcgv_vec(arg_temp(va_arg(va, TCGArg)));
4010        v4 = temp_tcgv_vec(arg_temp(va_arg(va, TCGArg)));
4011        expand_vec_cmpsel(type, vece, v0, v1, v2, v3, v4, va_arg(va, TCGArg));
4012        break;
4013
4014    default:
4015        break;
4016    }
4017
4018    va_end(va);
4019}
4020
4021static const int tcg_target_callee_save_regs[] = {
4022#if TCG_TARGET_REG_BITS == 64
4023    TCG_REG_RBP,
4024    TCG_REG_RBX,
4025#if defined(_WIN64)
4026    TCG_REG_RDI,
4027    TCG_REG_RSI,
4028#endif
4029    TCG_REG_R12,
4030    TCG_REG_R13,
4031    TCG_REG_R14, /* Currently used for the global env. */
4032    TCG_REG_R15,
4033#else
4034    TCG_REG_EBP, /* Currently used for the global env. */
4035    TCG_REG_EBX,
4036    TCG_REG_ESI,
4037    TCG_REG_EDI,
4038#endif
4039};
4040
4041/* Compute frame size via macros, to share between tcg_target_qemu_prologue
4042   and tcg_register_jit.  */
4043
4044#define PUSH_SIZE \
4045    ((1 + ARRAY_SIZE(tcg_target_callee_save_regs)) \
4046     * (TCG_TARGET_REG_BITS / 8))
4047
4048#define FRAME_SIZE \
4049    ((PUSH_SIZE \
4050      + TCG_STATIC_CALL_ARGS_SIZE \
4051      + CPU_TEMP_BUF_NLONGS * sizeof(long) \
4052      + TCG_TARGET_STACK_ALIGN - 1) \
4053     & ~(TCG_TARGET_STACK_ALIGN - 1))
4054
4055/* Generate global QEMU prologue and epilogue code */
4056static void tcg_target_qemu_prologue(TCGContext *s)
4057{
4058    int i, stack_addend;
4059
4060    /* TB prologue */
4061
4062    /* Reserve some stack space, also for TCG temps.  */
4063    stack_addend = FRAME_SIZE - PUSH_SIZE;
4064    tcg_set_frame(s, TCG_REG_CALL_STACK, TCG_STATIC_CALL_ARGS_SIZE,
4065                  CPU_TEMP_BUF_NLONGS * sizeof(long));
4066
4067    /* Save all callee saved registers.  */
4068    for (i = 0; i < ARRAY_SIZE(tcg_target_callee_save_regs); i++) {
4069        tcg_out_push(s, tcg_target_callee_save_regs[i]);
4070    }
4071
4072#if TCG_TARGET_REG_BITS == 32
4073    tcg_out_ld(s, TCG_TYPE_PTR, TCG_AREG0, TCG_REG_ESP,
4074               (ARRAY_SIZE(tcg_target_callee_save_regs) + 1) * 4);
4075    tcg_out_addi(s, TCG_REG_ESP, -stack_addend);
4076    /* jmp *tb.  */
4077    tcg_out_modrm_offset(s, OPC_GRP5, EXT5_JMPN_Ev, TCG_REG_ESP,
4078                         (ARRAY_SIZE(tcg_target_callee_save_regs) + 2) * 4
4079                         + stack_addend);
4080#else
4081# if !defined(CONFIG_SOFTMMU) && TCG_TARGET_REG_BITS == 64
4082    if (guest_base) {
4083        int seg = setup_guest_base_seg();
4084        if (seg != 0) {
4085            x86_guest_base_seg = seg;
4086        } else if (guest_base == (int32_t)guest_base) {
4087            x86_guest_base_offset = guest_base;
4088        } else {
4089            /* Choose R12 because, as a base, it requires a SIB byte. */
4090            x86_guest_base_index = TCG_REG_R12;
4091            tcg_out_movi(s, TCG_TYPE_PTR, x86_guest_base_index, guest_base);
4092            tcg_regset_set_reg(s->reserved_regs, x86_guest_base_index);
4093        }
4094    }
4095# endif
4096    tcg_out_mov(s, TCG_TYPE_PTR, TCG_AREG0, tcg_target_call_iarg_regs[0]);
4097    tcg_out_addi(s, TCG_REG_ESP, -stack_addend);
4098    /* jmp *tb.  */
4099    tcg_out_modrm(s, OPC_GRP5, EXT5_JMPN_Ev, tcg_target_call_iarg_regs[1]);
4100#endif
4101
4102    /*
4103     * Return path for goto_ptr. Set return value to 0, a-la exit_tb,
4104     * and fall through to the rest of the epilogue.
4105     */
4106    tcg_code_gen_epilogue = tcg_splitwx_to_rx(s->code_ptr);
4107    tcg_out_movi(s, TCG_TYPE_REG, TCG_REG_EAX, 0);
4108
4109    /* TB epilogue */
4110    tb_ret_addr = tcg_splitwx_to_rx(s->code_ptr);
4111
4112    tcg_out_addi(s, TCG_REG_CALL_STACK, stack_addend);
4113
4114    if (have_avx2) {
4115        tcg_out_vex_opc(s, OPC_VZEROUPPER, 0, 0, 0, 0);
4116    }
4117    for (i = ARRAY_SIZE(tcg_target_callee_save_regs) - 1; i >= 0; i--) {
4118        tcg_out_pop(s, tcg_target_callee_save_regs[i]);
4119    }
4120    tcg_out_opc(s, OPC_RET, 0, 0, 0);
4121}
4122
4123static void tcg_out_nop_fill(tcg_insn_unit *p, int count)
4124{
4125    memset(p, 0x90, count);
4126}
4127
4128static void tcg_target_init(TCGContext *s)
4129{
4130#ifdef CONFIG_CPUID_H
4131    unsigned a, b, c, d, b7 = 0, c7 = 0;
4132    unsigned max = __get_cpuid_max(0, 0);
4133
4134    if (max >= 7) {
4135        /* BMI1 is available on AMD Piledriver and Intel Haswell CPUs.  */
4136        __cpuid_count(7, 0, a, b7, c7, d);
4137        have_bmi1 = (b7 & bit_BMI) != 0;
4138        have_bmi2 = (b7 & bit_BMI2) != 0;
4139    }
4140
4141    if (max >= 1) {
4142        __cpuid(1, a, b, c, d);
4143#ifndef have_cmov
4144        /* For 32-bit, 99% certainty that we're running on hardware that
4145           supports cmov, but we still need to check.  In case cmov is not
4146           available, we'll use a small forward branch.  */
4147        have_cmov = (d & bit_CMOV) != 0;
4148#endif
4149
4150        /* MOVBE is only available on Intel Atom and Haswell CPUs, so we
4151           need to probe for it.  */
4152        have_movbe = (c & bit_MOVBE) != 0;
4153        have_popcnt = (c & bit_POPCNT) != 0;
4154
4155        /* There are a number of things we must check before we can be
4156           sure of not hitting invalid opcode.  */
4157        if (c & bit_OSXSAVE) {
4158            unsigned bv = xgetbv_low(0);
4159
4160            if ((bv & 6) == 6) {
4161                have_avx1 = (c & bit_AVX) != 0;
4162                have_avx2 = (b7 & bit_AVX2) != 0;
4163
4164                /*
4165                 * There are interesting instructions in AVX512, so long
4166                 * as we have AVX512VL, which indicates support for EVEX
4167                 * on sizes smaller than 512 bits.  We are required to
4168                 * check that OPMASK and all extended ZMM state are enabled
4169                 * even if we're not using them -- the insns will fault.
4170                 */
4171                if ((bv & 0xe0) == 0xe0
4172                    && (b7 & bit_AVX512F)
4173                    && (b7 & bit_AVX512VL)) {
4174                    have_avx512vl = true;
4175                    have_avx512bw = (b7 & bit_AVX512BW) != 0;
4176                    have_avx512dq = (b7 & bit_AVX512DQ) != 0;
4177                    have_avx512vbmi2 = (c7 & bit_AVX512VBMI2) != 0;
4178                }
4179            }
4180        }
4181    }
4182
4183    max = __get_cpuid_max(0x8000000, 0);
4184    if (max >= 1) {
4185        __cpuid(0x80000001, a, b, c, d);
4186        /* LZCNT was introduced with AMD Barcelona and Intel Haswell CPUs.  */
4187        have_lzcnt = (c & bit_LZCNT) != 0;
4188    }
4189#endif /* CONFIG_CPUID_H */
4190
4191    tcg_target_available_regs[TCG_TYPE_I32] = ALL_GENERAL_REGS;
4192    if (TCG_TARGET_REG_BITS == 64) {
4193        tcg_target_available_regs[TCG_TYPE_I64] = ALL_GENERAL_REGS;
4194    }
4195    if (have_avx1) {
4196        tcg_target_available_regs[TCG_TYPE_V64] = ALL_VECTOR_REGS;
4197        tcg_target_available_regs[TCG_TYPE_V128] = ALL_VECTOR_REGS;
4198    }
4199    if (have_avx2) {
4200        tcg_target_available_regs[TCG_TYPE_V256] = ALL_VECTOR_REGS;
4201    }
4202
4203    tcg_target_call_clobber_regs = ALL_VECTOR_REGS;
4204    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_EAX);
4205    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_EDX);
4206    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_ECX);
4207    if (TCG_TARGET_REG_BITS == 64) {
4208#if !defined(_WIN64)
4209        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_RDI);
4210        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_RSI);
4211#endif
4212        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R8);
4213        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R9);
4214        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R10);
4215        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R11);
4216    }
4217
4218    s->reserved_regs = 0;
4219    tcg_regset_set_reg(s->reserved_regs, TCG_REG_CALL_STACK);
4220#ifdef _WIN64
4221    /* These are call saved, and we don't save them, so don't use them. */
4222    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM6);
4223    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM7);
4224    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM8);
4225    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM9);
4226    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM10);
4227    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM11);
4228    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM12);
4229    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM13);
4230    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM14);
4231    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM15);
4232#endif
4233}
4234
4235typedef struct {
4236    DebugFrameHeader h;
4237    uint8_t fde_def_cfa[4];
4238    uint8_t fde_reg_ofs[14];
4239} DebugFrame;
4240
4241/* We're expecting a 2 byte uleb128 encoded value.  */
4242QEMU_BUILD_BUG_ON(FRAME_SIZE >= (1 << 14));
4243
4244#if !defined(__ELF__)
4245    /* Host machine without ELF. */
4246#elif TCG_TARGET_REG_BITS == 64
4247#define ELF_HOST_MACHINE EM_X86_64
4248static const DebugFrame debug_frame = {
4249    .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
4250    .h.cie.id = -1,
4251    .h.cie.version = 1,
4252    .h.cie.code_align = 1,
4253    .h.cie.data_align = 0x78,             /* sleb128 -8 */
4254    .h.cie.return_column = 16,
4255
4256    /* Total FDE size does not include the "len" member.  */
4257    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
4258
4259    .fde_def_cfa = {
4260        12, 7,                          /* DW_CFA_def_cfa %rsp, ... */
4261        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
4262        (FRAME_SIZE >> 7)
4263    },
4264    .fde_reg_ofs = {
4265        0x90, 1,                        /* DW_CFA_offset, %rip, -8 */
4266        /* The following ordering must match tcg_target_callee_save_regs.  */
4267        0x86, 2,                        /* DW_CFA_offset, %rbp, -16 */
4268        0x83, 3,                        /* DW_CFA_offset, %rbx, -24 */
4269        0x8c, 4,                        /* DW_CFA_offset, %r12, -32 */
4270        0x8d, 5,                        /* DW_CFA_offset, %r13, -40 */
4271        0x8e, 6,                        /* DW_CFA_offset, %r14, -48 */
4272        0x8f, 7,                        /* DW_CFA_offset, %r15, -56 */
4273    }
4274};
4275#else
4276#define ELF_HOST_MACHINE EM_386
4277static const DebugFrame debug_frame = {
4278    .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
4279    .h.cie.id = -1,
4280    .h.cie.version = 1,
4281    .h.cie.code_align = 1,
4282    .h.cie.data_align = 0x7c,             /* sleb128 -4 */
4283    .h.cie.return_column = 8,
4284
4285    /* Total FDE size does not include the "len" member.  */
4286    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
4287
4288    .fde_def_cfa = {
4289        12, 4,                          /* DW_CFA_def_cfa %esp, ... */
4290        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
4291        (FRAME_SIZE >> 7)
4292    },
4293    .fde_reg_ofs = {
4294        0x88, 1,                        /* DW_CFA_offset, %eip, -4 */
4295        /* The following ordering must match tcg_target_callee_save_regs.  */
4296        0x85, 2,                        /* DW_CFA_offset, %ebp, -8 */
4297        0x83, 3,                        /* DW_CFA_offset, %ebx, -12 */
4298        0x86, 4,                        /* DW_CFA_offset, %esi, -16 */
4299        0x87, 5,                        /* DW_CFA_offset, %edi, -20 */
4300    }
4301};
4302#endif
4303
4304#if defined(ELF_HOST_MACHINE)
4305void tcg_register_jit(const void *buf, size_t buf_size)
4306{
4307    tcg_register_jit_int(buf, buf_size, &debug_frame, sizeof(debug_frame));
4308}
4309#endif
4310