1 | /* |
2 | ** x86/x64 instruction emitter. |
3 | ** Copyright (C) 2005-2014 Mike Pall. See Copyright Notice in luajit.h |
4 | */ |
5 | |
6 | /* -- Emit basic instructions --------------------------------------------- */ |
7 | |
8 | #define MODRM(mode, r1, r2) ((MCode)((mode)+(((r1)&7)<<3)+((r2)&7))) |
9 | |
10 | #if LJ_64 |
11 | #define REXRB(p, rr, rb) \ |
12 | { MCode rex = 0x40 + (((rr)>>1)&4) + (((rb)>>3)&1); \ |
13 | if (rex != 0x40) *--(p) = rex; } |
14 | #define FORCE_REX 0x200 |
15 | #define REX_64 (FORCE_REX|0x080000) |
16 | #else |
17 | #define REXRB(p, rr, rb) ((void)0) |
18 | #define FORCE_REX 0 |
19 | #define REX_64 0 |
20 | #endif |
21 | |
22 | #define emit_i8(as, i) (*--as->mcp = (MCode)(i)) |
23 | #define emit_i32(as, i) (*(int32_t *)(as->mcp-4) = (i), as->mcp -= 4) |
24 | #define emit_u32(as, u) (*(uint32_t *)(as->mcp-4) = (u), as->mcp -= 4) |
25 | |
26 | #define emit_x87op(as, xo) \ |
27 | (*(uint16_t *)(as->mcp-2) = (uint16_t)(xo), as->mcp -= 2) |
28 | |
29 | /* op */ |
30 | static LJ_AINLINE MCode *emit_op(x86Op xo, Reg rr, Reg rb, Reg rx, |
31 | MCode *p, int delta) |
32 | { |
33 | int n = (int8_t)xo; |
34 | #if defined(__GNUC__) |
35 | if (__builtin_constant_p(xo) && n == -2) |
36 | p[delta-2] = (MCode)(xo >> 24); |
37 | else if (__builtin_constant_p(xo) && n == -3) |
38 | *(uint16_t *)(p+delta-3) = (uint16_t)(xo >> 16); |
39 | else |
40 | #endif |
41 | *(uint32_t *)(p+delta-5) = (uint32_t)xo; |
42 | p += n + delta; |
43 | #if LJ_64 |
44 | { |
45 | uint32_t rex = 0x40 + ((rr>>1)&(4+(FORCE_REX>>1)))+((rx>>2)&2)+((rb>>3)&1); |
46 | if (rex != 0x40) { |
47 | rex |= (rr >> 16); |
48 | if (n == -4) { *p = (MCode)rex; rex = (MCode)(xo >> 8); } |
49 | else if ((xo & 0xffffff) == 0x6600fd) { *p = (MCode)rex; rex = 0x66; } |
50 | *--p = (MCode)rex; |
51 | } |
52 | } |
53 | #else |
54 | UNUSED(rr); UNUSED(rb); UNUSED(rx); |
55 | #endif |
56 | return p; |
57 | } |
58 | |
59 | /* op + modrm */ |
60 | #define emit_opm(xo, mode, rr, rb, p, delta) \ |
61 | (p[(delta)-1] = MODRM((mode), (rr), (rb)), \ |
62 | emit_op((xo), (rr), (rb), 0, (p), (delta))) |
63 | |
64 | /* op + modrm + sib */ |
65 | #define emit_opmx(xo, mode, scale, rr, rb, rx, p) \ |
66 | (p[-1] = MODRM((scale), (rx), (rb)), \ |
67 | p[-2] = MODRM((mode), (rr), RID_ESP), \ |
68 | emit_op((xo), (rr), (rb), (rx), (p), -1)) |
69 | |
70 | /* op r1, r2 */ |
71 | static void emit_rr(ASMState *as, x86Op xo, Reg r1, Reg r2) |
72 | { |
73 | MCode *p = as->mcp; |
74 | as->mcp = emit_opm(xo, XM_REG, r1, r2, p, 0); |
75 | } |
76 | |
77 | #if LJ_64 && defined(LUA_USE_ASSERT) |
78 | /* [addr] is sign-extended in x64 and must be in lower 2G (not 4G). */ |
79 | static int32_t ptr2addr(const void *p) |
80 | { |
81 | lua_assert((uintptr_t)p < (uintptr_t)0x80000000); |
82 | return i32ptr(p); |
83 | } |
84 | #else |
85 | #define ptr2addr(p) (i32ptr((p))) |
86 | #endif |
87 | |
88 | /* op r, [addr] */ |
89 | static void emit_rma(ASMState *as, x86Op xo, Reg rr, const void *addr) |
90 | { |
91 | MCode *p = as->mcp; |
92 | *(int32_t *)(p-4) = ptr2addr(addr); |
93 | #if LJ_64 |
94 | p[-5] = MODRM(XM_SCALE1, RID_ESP, RID_EBP); |
95 | as->mcp = emit_opm(xo, XM_OFS0, rr, RID_ESP, p, -5); |
96 | #else |
97 | as->mcp = emit_opm(xo, XM_OFS0, rr, RID_EBP, p, -4); |
98 | #endif |
99 | } |
100 | |
101 | /* op r, [base+ofs] */ |
102 | static void emit_rmro(ASMState *as, x86Op xo, Reg rr, Reg rb, int32_t ofs) |
103 | { |
104 | MCode *p = as->mcp; |
105 | x86Mode mode; |
106 | if (ra_hasreg(rb)) { |
107 | if (ofs == 0 && (rb&7) != RID_EBP) { |
108 | mode = XM_OFS0; |
109 | } else if (checki8(ofs)) { |
110 | *--p = (MCode)ofs; |
111 | mode = XM_OFS8; |
112 | } else { |
113 | p -= 4; |
114 | *(int32_t *)p = ofs; |
115 | mode = XM_OFS32; |
116 | } |
117 | if ((rb&7) == RID_ESP) |
118 | *--p = MODRM(XM_SCALE1, RID_ESP, RID_ESP); |
119 | } else { |
120 | *(int32_t *)(p-4) = ofs; |
121 | #if LJ_64 |
122 | p[-5] = MODRM(XM_SCALE1, RID_ESP, RID_EBP); |
123 | p -= 5; |
124 | rb = RID_ESP; |
125 | #else |
126 | p -= 4; |
127 | rb = RID_EBP; |
128 | #endif |
129 | mode = XM_OFS0; |
130 | } |
131 | as->mcp = emit_opm(xo, mode, rr, rb, p, 0); |
132 | } |
133 | |
134 | /* op r, [base+idx*scale+ofs] */ |
135 | static void emit_rmrxo(ASMState *as, x86Op xo, Reg rr, Reg rb, Reg rx, |
136 | x86Mode scale, int32_t ofs) |
137 | { |
138 | MCode *p = as->mcp; |
139 | x86Mode mode; |
140 | if (ofs == 0 && (rb&7) != RID_EBP) { |
141 | mode = XM_OFS0; |
142 | } else if (checki8(ofs)) { |
143 | mode = XM_OFS8; |
144 | *--p = (MCode)ofs; |
145 | } else { |
146 | mode = XM_OFS32; |
147 | p -= 4; |
148 | *(int32_t *)p = ofs; |
149 | } |
150 | as->mcp = emit_opmx(xo, mode, scale, rr, rb, rx, p); |
151 | } |
152 | |
153 | /* op r, i */ |
154 | static void emit_gri(ASMState *as, x86Group xg, Reg rb, int32_t i) |
155 | { |
156 | MCode *p = as->mcp; |
157 | x86Op xo; |
158 | if (checki8(i)) { |
159 | *--p = (MCode)i; |
160 | xo = XG_TOXOi8(xg); |
161 | } else { |
162 | p -= 4; |
163 | *(int32_t *)p = i; |
164 | xo = XG_TOXOi(xg); |
165 | } |
166 | as->mcp = emit_opm(xo, XM_REG, (Reg)(xg & 7) | (rb & REX_64), rb, p, 0); |
167 | } |
168 | |
169 | /* op [base+ofs], i */ |
170 | static void emit_gmroi(ASMState *as, x86Group xg, Reg rb, int32_t ofs, |
171 | int32_t i) |
172 | { |
173 | x86Op xo; |
174 | if (checki8(i)) { |
175 | emit_i8(as, i); |
176 | xo = XG_TOXOi8(xg); |
177 | } else { |
178 | emit_i32(as, i); |
179 | xo = XG_TOXOi(xg); |
180 | } |
181 | emit_rmro(as, xo, (Reg)(xg & 7), rb, ofs); |
182 | } |
183 | |
184 | #define emit_shifti(as, xg, r, i) \ |
185 | (emit_i8(as, (i)), emit_rr(as, XO_SHIFTi, (Reg)(xg), (r))) |
186 | |
187 | /* op r, rm/mrm */ |
188 | static void emit_mrm(ASMState *as, x86Op xo, Reg rr, Reg rb) |
189 | { |
190 | MCode *p = as->mcp; |
191 | x86Mode mode = XM_REG; |
192 | if (rb == RID_MRM) { |
193 | rb = as->mrm.base; |
194 | if (rb == RID_NONE) { |
195 | rb = RID_EBP; |
196 | mode = XM_OFS0; |
197 | p -= 4; |
198 | *(int32_t *)p = as->mrm.ofs; |
199 | if (as->mrm.idx != RID_NONE) |
200 | goto mrmidx; |
201 | #if LJ_64 |
202 | *--p = MODRM(XM_SCALE1, RID_ESP, RID_EBP); |
203 | rb = RID_ESP; |
204 | #endif |
205 | } else { |
206 | if (as->mrm.ofs == 0 && (rb&7) != RID_EBP) { |
207 | mode = XM_OFS0; |
208 | } else if (checki8(as->mrm.ofs)) { |
209 | *--p = (MCode)as->mrm.ofs; |
210 | mode = XM_OFS8; |
211 | } else { |
212 | p -= 4; |
213 | *(int32_t *)p = as->mrm.ofs; |
214 | mode = XM_OFS32; |
215 | } |
216 | if (as->mrm.idx != RID_NONE) { |
217 | mrmidx: |
218 | as->mcp = emit_opmx(xo, mode, as->mrm.scale, rr, rb, as->mrm.idx, p); |
219 | return; |
220 | } |
221 | if ((rb&7) == RID_ESP) |
222 | *--p = MODRM(XM_SCALE1, RID_ESP, RID_ESP); |
223 | } |
224 | } |
225 | as->mcp = emit_opm(xo, mode, rr, rb, p, 0); |
226 | } |
227 | |
228 | /* op rm/mrm, i */ |
229 | static void emit_gmrmi(ASMState *as, x86Group xg, Reg rb, int32_t i) |
230 | { |
231 | x86Op xo; |
232 | if (checki8(i)) { |
233 | emit_i8(as, i); |
234 | xo = XG_TOXOi8(xg); |
235 | } else { |
236 | emit_i32(as, i); |
237 | xo = XG_TOXOi(xg); |
238 | } |
239 | emit_mrm(as, xo, (Reg)(xg & 7) | (rb & REX_64), (rb & ~REX_64)); |
240 | } |
241 | |
242 | /* -- Emit loads/stores --------------------------------------------------- */ |
243 | |
244 | /* Instruction selection for XMM moves. */ |
245 | #define XMM_MOVRR(as) ((as->flags & JIT_F_SPLIT_XMM) ? XO_MOVSD : XO_MOVAPS) |
246 | #define XMM_MOVRM(as) ((as->flags & JIT_F_SPLIT_XMM) ? XO_MOVLPD : XO_MOVSD) |
247 | |
248 | /* mov [base+ofs], i */ |
249 | static void emit_movmroi(ASMState *as, Reg base, int32_t ofs, int32_t i) |
250 | { |
251 | emit_i32(as, i); |
252 | emit_rmro(as, XO_MOVmi, 0, base, ofs); |
253 | } |
254 | |
255 | /* mov [base+ofs], r */ |
256 | #define emit_movtomro(as, r, base, ofs) \ |
257 | emit_rmro(as, XO_MOVto, (r), (base), (ofs)) |
258 | |
259 | /* Get/set global_State fields. */ |
260 | #define emit_opgl(as, xo, r, field) \ |
261 | emit_rma(as, (xo), (r), (void *)&J2G(as->J)->field) |
262 | #define emit_getgl(as, r, field) emit_opgl(as, XO_MOV, (r), field) |
263 | #define emit_setgl(as, r, field) emit_opgl(as, XO_MOVto, (r), field) |
264 | |
265 | #define emit_setvmstate(as, i) \ |
266 | (emit_i32(as, i), emit_opgl(as, XO_MOVmi, 0, vmstate)) |
267 | |
268 | /* mov r, i / xor r, r */ |
269 | static void emit_loadi(ASMState *as, Reg r, int32_t i) |
270 | { |
271 | /* XOR r,r is shorter, but modifies the flags. This is bad for HIOP. */ |
272 | if (i == 0 && !(LJ_32 && (IR(as->curins)->o == IR_HIOP || |
273 | (as->curins+1 < as->T->nins && |
274 | IR(as->curins+1)->o == IR_HIOP)))) { |
275 | emit_rr(as, XO_ARITH(XOg_XOR), r, r); |
276 | } else { |
277 | MCode *p = as->mcp; |
278 | *(int32_t *)(p-4) = i; |
279 | p[-5] = (MCode)(XI_MOVri+(r&7)); |
280 | p -= 5; |
281 | REXRB(p, 0, r); |
282 | as->mcp = p; |
283 | } |
284 | } |
285 | |
286 | /* mov r, addr */ |
287 | #define emit_loada(as, r, addr) \ |
288 | emit_loadi(as, (r), ptr2addr((addr))) |
289 | |
290 | #if LJ_64 |
291 | /* mov r, imm64 or shorter 32 bit extended load. */ |
292 | static void emit_loadu64(ASMState *as, Reg r, uint64_t u64) |
293 | { |
294 | if (checku32(u64)) { /* 32 bit load clears upper 32 bits. */ |
295 | emit_loadi(as, r, (int32_t)u64); |
296 | } else if (checki32((int64_t)u64)) { /* Sign-extended 32 bit load. */ |
297 | MCode *p = as->mcp; |
298 | *(int32_t *)(p-4) = (int32_t)u64; |
299 | as->mcp = emit_opm(XO_MOVmi, XM_REG, REX_64, r, p, -4); |
300 | } else { /* Full-size 64 bit load. */ |
301 | MCode *p = as->mcp; |
302 | *(uint64_t *)(p-8) = u64; |
303 | p[-9] = (MCode)(XI_MOVri+(r&7)); |
304 | p[-10] = 0x48 + ((r>>3)&1); |
305 | p -= 10; |
306 | as->mcp = p; |
307 | } |
308 | } |
309 | #endif |
310 | |
311 | /* movsd r, [&tv->n] / xorps r, r */ |
312 | static void emit_loadn(ASMState *as, Reg r, cTValue *tv) |
313 | { |
314 | if (tvispzero(tv)) /* Use xor only for +0. */ |
315 | emit_rr(as, XO_XORPS, r, r); |
316 | else |
317 | emit_rma(as, XMM_MOVRM(as), r, &tv->n); |
318 | } |
319 | |
320 | /* -- Emit control-flow instructions -------------------------------------- */ |
321 | |
322 | /* Label for short jumps. */ |
323 | typedef MCode *MCLabel; |
324 | |
325 | #if LJ_32 && LJ_HASFFI |
326 | /* jmp short target */ |
327 | static void emit_sjmp(ASMState *as, MCLabel target) |
328 | { |
329 | MCode *p = as->mcp; |
330 | ptrdiff_t delta = target - p; |
331 | lua_assert(delta == (int8_t)delta); |
332 | p[-1] = (MCode)(int8_t)delta; |
333 | p[-2] = XI_JMPs; |
334 | as->mcp = p - 2; |
335 | } |
336 | #endif |
337 | |
338 | /* jcc short target */ |
339 | static void emit_sjcc(ASMState *as, int cc, MCLabel target) |
340 | { |
341 | MCode *p = as->mcp; |
342 | ptrdiff_t delta = target - p; |
343 | lua_assert(delta == (int8_t)delta); |
344 | p[-1] = (MCode)(int8_t)delta; |
345 | p[-2] = (MCode)(XI_JCCs+(cc&15)); |
346 | as->mcp = p - 2; |
347 | } |
348 | |
349 | /* jcc short (pending target) */ |
350 | static MCLabel emit_sjcc_label(ASMState *as, int cc) |
351 | { |
352 | MCode *p = as->mcp; |
353 | p[-1] = 0; |
354 | p[-2] = (MCode)(XI_JCCs+(cc&15)); |
355 | as->mcp = p - 2; |
356 | return p; |
357 | } |
358 | |
359 | /* Fixup jcc short target. */ |
360 | static void emit_sfixup(ASMState *as, MCLabel source) |
361 | { |
362 | source[-1] = (MCode)(as->mcp-source); |
363 | } |
364 | |
365 | /* Return label pointing to current PC. */ |
366 | #define emit_label(as) ((as)->mcp) |
367 | |
368 | /* Compute relative 32 bit offset for jump and call instructions. */ |
369 | static LJ_AINLINE int32_t jmprel(MCode *p, MCode *target) |
370 | { |
371 | ptrdiff_t delta = target - p; |
372 | lua_assert(delta == (int32_t)delta); |
373 | return (int32_t)delta; |
374 | } |
375 | |
376 | /* jcc target */ |
377 | static void emit_jcc(ASMState *as, int cc, MCode *target) |
378 | { |
379 | MCode *p = as->mcp; |
380 | *(int32_t *)(p-4) = jmprel(p, target); |
381 | p[-5] = (MCode)(XI_JCCn+(cc&15)); |
382 | p[-6] = 0x0f; |
383 | as->mcp = p - 6; |
384 | } |
385 | |
386 | /* jmp target */ |
387 | static void emit_jmp(ASMState *as, MCode *target) |
388 | { |
389 | MCode *p = as->mcp; |
390 | *(int32_t *)(p-4) = jmprel(p, target); |
391 | p[-5] = XI_JMP; |
392 | as->mcp = p - 5; |
393 | } |
394 | |
395 | /* call target */ |
396 | static void emit_call_(ASMState *as, MCode *target) |
397 | { |
398 | MCode *p = as->mcp; |
399 | #if LJ_64 |
400 | if (target-p != (int32_t)(target-p)) { |
401 | /* Assumes RID_RET is never an argument to calls and always clobbered. */ |
402 | emit_rr(as, XO_GROUP5, XOg_CALL, RID_RET); |
403 | emit_loadu64(as, RID_RET, (uint64_t)target); |
404 | return; |
405 | } |
406 | #endif |
407 | *(int32_t *)(p-4) = jmprel(p, target); |
408 | p[-5] = XI_CALL; |
409 | as->mcp = p - 5; |
410 | } |
411 | |
412 | #define emit_call(as, f) emit_call_(as, (MCode *)(void *)(f)) |
413 | |
414 | /* -- Emit generic operations --------------------------------------------- */ |
415 | |
416 | /* Use 64 bit operations to handle 64 bit IR types. */ |
417 | #if LJ_64 |
418 | #define REX_64IR(ir, r) ((r) + (irt_is64((ir)->t) ? REX_64 : 0)) |
419 | #else |
420 | #define REX_64IR(ir, r) (r) |
421 | #endif |
422 | |
423 | /* Generic move between two regs. */ |
424 | static void emit_movrr(ASMState *as, IRIns *ir, Reg dst, Reg src) |
425 | { |
426 | UNUSED(ir); |
427 | if (dst < RID_MAX_GPR) |
428 | emit_rr(as, XO_MOV, REX_64IR(ir, dst), src); |
429 | else |
430 | emit_rr(as, XMM_MOVRR(as), dst, src); |
431 | } |
432 | |
433 | /* Generic load of register from stack slot. */ |
434 | static void emit_spload(ASMState *as, IRIns *ir, Reg r, int32_t ofs) |
435 | { |
436 | if (r < RID_MAX_GPR) |
437 | emit_rmro(as, XO_MOV, REX_64IR(ir, r), RID_ESP, ofs); |
438 | else |
439 | emit_rmro(as, irt_isnum(ir->t) ? XMM_MOVRM(as) : XO_MOVSS, r, RID_ESP, ofs); |
440 | } |
441 | |
442 | /* Generic store of register to stack slot. */ |
443 | static void emit_spstore(ASMState *as, IRIns *ir, Reg r, int32_t ofs) |
444 | { |
445 | if (r < RID_MAX_GPR) |
446 | emit_rmro(as, XO_MOVto, REX_64IR(ir, r), RID_ESP, ofs); |
447 | else |
448 | emit_rmro(as, irt_isnum(ir->t) ? XO_MOVSDto : XO_MOVSSto, r, RID_ESP, ofs); |
449 | } |
450 | |
451 | /* Add offset to pointer. */ |
452 | static void emit_addptr(ASMState *as, Reg r, int32_t ofs) |
453 | { |
454 | if (ofs) { |
455 | if ((as->flags & JIT_F_LEA_AGU)) |
456 | emit_rmro(as, XO_LEA, r, r, ofs); |
457 | else |
458 | emit_gri(as, XG_ARITHi(XOg_ADD), r, ofs); |
459 | } |
460 | } |
461 | |
462 | #define emit_spsub(as, ofs) emit_addptr(as, RID_ESP|REX_64, -(ofs)) |
463 | |
464 | /* Prefer rematerialization of BASE/L from global_State over spills. */ |
465 | #define emit_canremat(ref) ((ref) <= REF_BASE) |
466 | |
467 | |