PPC: Compile math.sqrt() to fsqrt instruction.
[luajit-2.0/celess22.git] / dynasm / dasm_x86.h
blob7c6dcd3d4707def7670e34be5729108b47aa03f0
1 /*
2 ** DynASM x86 encoding engine.
3 ** Copyright (C) 2005-2012 Mike Pall. All rights reserved.
4 ** Released under the MIT license. See dynasm.lua for full copyright notice.
5 */
7 #include <stddef.h>
8 #include <stdarg.h>
9 #include <string.h>
10 #include <stdlib.h>
12 #define DASM_ARCH "x86"
14 #ifndef DASM_EXTERN
15 #define DASM_EXTERN(a,b,c,d) 0
16 #endif
18 /* Action definitions. DASM_STOP must be 255. */
19 enum {
20 DASM_DISP = 233,
21 DASM_IMM_S, DASM_IMM_B, DASM_IMM_W, DASM_IMM_D, DASM_IMM_WB, DASM_IMM_DB,
22 DASM_VREG, DASM_SPACE, DASM_SETLABEL, DASM_REL_A, DASM_REL_LG, DASM_REL_PC,
23 DASM_IMM_LG, DASM_IMM_PC, DASM_LABEL_LG, DASM_LABEL_PC, DASM_ALIGN,
24 DASM_EXTERN, DASM_ESC, DASM_MARK, DASM_SECTION, DASM_STOP
27 /* Maximum number of section buffer positions for a single dasm_put() call. */
28 #define DASM_MAXSECPOS 25
30 /* DynASM encoder status codes. Action list offset or number are or'ed in. */
31 #define DASM_S_OK 0x00000000
32 #define DASM_S_NOMEM 0x01000000
33 #define DASM_S_PHASE 0x02000000
34 #define DASM_S_MATCH_SEC 0x03000000
35 #define DASM_S_RANGE_I 0x11000000
36 #define DASM_S_RANGE_SEC 0x12000000
37 #define DASM_S_RANGE_LG 0x13000000
38 #define DASM_S_RANGE_PC 0x14000000
39 #define DASM_S_RANGE_VREG 0x15000000
40 #define DASM_S_UNDEF_L 0x21000000
41 #define DASM_S_UNDEF_PC 0x22000000
43 /* Macros to convert positions (8 bit section + 24 bit index). */
44 #define DASM_POS2IDX(pos) ((pos)&0x00ffffff)
45 #define DASM_POS2BIAS(pos) ((pos)&0xff000000)
46 #define DASM_SEC2POS(sec) ((sec)<<24)
47 #define DASM_POS2SEC(pos) ((pos)>>24)
48 #define DASM_POS2PTR(D, pos) (D->sections[DASM_POS2SEC(pos)].rbuf + (pos))
50 /* Action list type. */
51 typedef const unsigned char *dasm_ActList;
53 /* Per-section structure. */
54 typedef struct dasm_Section {
55 int *rbuf; /* Biased buffer pointer (negative section bias). */
56 int *buf; /* True buffer pointer. */
57 size_t bsize; /* Buffer size in bytes. */
58 int pos; /* Biased buffer position. */
59 int epos; /* End of biased buffer position - max single put. */
60 int ofs; /* Byte offset into section. */
61 } dasm_Section;
63 /* Core structure holding the DynASM encoding state. */
64 struct dasm_State {
65 size_t psize; /* Allocated size of this structure. */
66 dasm_ActList actionlist; /* Current actionlist pointer. */
67 int *lglabels; /* Local/global chain/pos ptrs. */
68 size_t lgsize;
69 int *pclabels; /* PC label chains/pos ptrs. */
70 size_t pcsize;
71 void **globals; /* Array of globals (bias -10). */
72 dasm_Section *section; /* Pointer to active section. */
73 size_t codesize; /* Total size of all code sections. */
74 int maxsection; /* 0 <= sectionidx < maxsection. */
75 int status; /* Status code. */
76 dasm_Section sections[1]; /* All sections. Alloc-extended. */
79 /* The size of the core structure depends on the max. number of sections. */
80 #define DASM_PSZ(ms) (sizeof(dasm_State)+(ms-1)*sizeof(dasm_Section))
83 /* Initialize DynASM state. */
84 void dasm_init(Dst_DECL, int maxsection)
86 dasm_State *D;
87 size_t psz = 0;
88 int i;
89 Dst_REF = NULL;
90 DASM_M_GROW(Dst, struct dasm_State, Dst_REF, psz, DASM_PSZ(maxsection));
91 D = Dst_REF;
92 D->psize = psz;
93 D->lglabels = NULL;
94 D->lgsize = 0;
95 D->pclabels = NULL;
96 D->pcsize = 0;
97 D->globals = NULL;
98 D->maxsection = maxsection;
99 for (i = 0; i < maxsection; i++) {
100 D->sections[i].buf = NULL; /* Need this for pass3. */
101 D->sections[i].rbuf = D->sections[i].buf - DASM_SEC2POS(i);
102 D->sections[i].bsize = 0;
103 D->sections[i].epos = 0; /* Wrong, but is recalculated after resize. */
107 /* Free DynASM state. */
108 void dasm_free(Dst_DECL)
110 dasm_State *D = Dst_REF;
111 int i;
112 for (i = 0; i < D->maxsection; i++)
113 if (D->sections[i].buf)
114 DASM_M_FREE(Dst, D->sections[i].buf, D->sections[i].bsize);
115 if (D->pclabels) DASM_M_FREE(Dst, D->pclabels, D->pcsize);
116 if (D->lglabels) DASM_M_FREE(Dst, D->lglabels, D->lgsize);
117 DASM_M_FREE(Dst, D, D->psize);
120 /* Setup global label array. Must be called before dasm_setup(). */
121 void dasm_setupglobal(Dst_DECL, void **gl, unsigned int maxgl)
123 dasm_State *D = Dst_REF;
124 D->globals = gl - 10; /* Negative bias to compensate for locals. */
125 DASM_M_GROW(Dst, int, D->lglabels, D->lgsize, (10+maxgl)*sizeof(int));
128 /* Grow PC label array. Can be called after dasm_setup(), too. */
129 void dasm_growpc(Dst_DECL, unsigned int maxpc)
131 dasm_State *D = Dst_REF;
132 size_t osz = D->pcsize;
133 DASM_M_GROW(Dst, int, D->pclabels, D->pcsize, maxpc*sizeof(int));
134 memset((void *)(((unsigned char *)D->pclabels)+osz), 0, D->pcsize-osz);
137 /* Setup encoder. */
138 void dasm_setup(Dst_DECL, const void *actionlist)
140 dasm_State *D = Dst_REF;
141 int i;
142 D->actionlist = (dasm_ActList)actionlist;
143 D->status = DASM_S_OK;
144 D->section = &D->sections[0];
145 memset((void *)D->lglabels, 0, D->lgsize);
146 if (D->pclabels) memset((void *)D->pclabels, 0, D->pcsize);
147 for (i = 0; i < D->maxsection; i++) {
148 D->sections[i].pos = DASM_SEC2POS(i);
149 D->sections[i].ofs = 0;
154 #ifdef DASM_CHECKS
155 #define CK(x, st) \
156 do { if (!(x)) { \
157 D->status = DASM_S_##st|(int)(p-D->actionlist-1); return; } } while (0)
158 #define CKPL(kind, st) \
159 do { if ((size_t)((char *)pl-(char *)D->kind##labels) >= D->kind##size) { \
160 D->status=DASM_S_RANGE_##st|(int)(p-D->actionlist-1); return; } } while (0)
161 #else
162 #define CK(x, st) ((void)0)
163 #define CKPL(kind, st) ((void)0)
164 #endif
166 /* Pass 1: Store actions and args, link branches/labels, estimate offsets. */
167 void dasm_put(Dst_DECL, int start, ...)
169 va_list ap;
170 dasm_State *D = Dst_REF;
171 dasm_ActList p = D->actionlist + start;
172 dasm_Section *sec = D->section;
173 int pos = sec->pos, ofs = sec->ofs, mrm = 4;
174 int *b;
176 if (pos >= sec->epos) {
177 DASM_M_GROW(Dst, int, sec->buf, sec->bsize,
178 sec->bsize + 2*DASM_MAXSECPOS*sizeof(int));
179 sec->rbuf = sec->buf - DASM_POS2BIAS(pos);
180 sec->epos = (int)sec->bsize/sizeof(int) - DASM_MAXSECPOS+DASM_POS2BIAS(pos);
183 b = sec->rbuf;
184 b[pos++] = start;
186 va_start(ap, start);
187 while (1) {
188 int action = *p++;
189 if (action < DASM_DISP) {
190 ofs++;
191 } else if (action <= DASM_REL_A) {
192 int n = va_arg(ap, int);
193 b[pos++] = n;
194 switch (action) {
195 case DASM_DISP:
196 if (n == 0) { if ((mrm&7) == 4) mrm = p[-2]; if ((mrm&7) != 5) break; }
197 case DASM_IMM_DB: if (((n+128)&-256) == 0) goto ob;
198 case DASM_REL_A: /* Assumes ptrdiff_t is int. !x64 */
199 case DASM_IMM_D: ofs += 4; break;
200 case DASM_IMM_S: CK(((n+128)&-256) == 0, RANGE_I); goto ob;
201 case DASM_IMM_B: CK((n&-256) == 0, RANGE_I); ob: ofs++; break;
202 case DASM_IMM_WB: if (((n+128)&-256) == 0) goto ob;
203 case DASM_IMM_W: CK((n&-65536) == 0, RANGE_I); ofs += 2; break;
204 case DASM_SPACE: p++; ofs += n; break;
205 case DASM_SETLABEL: b[pos-2] = -0x40000000; break; /* Neg. label ofs. */
206 case DASM_VREG: CK((n&-8) == 0 && (n != 4 || (*p&1) == 0), RANGE_VREG);
207 if (*p++ == 1 && *p == DASM_DISP) mrm = n; continue;
209 mrm = 4;
210 } else {
211 int *pl, n;
212 switch (action) {
213 case DASM_REL_LG:
214 case DASM_IMM_LG:
215 n = *p++; pl = D->lglabels + n;
216 if (n <= 246) { CKPL(lg, LG); goto putrel; } /* Bkwd rel or global. */
217 pl -= 246; n = *pl;
218 if (n < 0) n = 0; /* Start new chain for fwd rel if label exists. */
219 goto linkrel;
220 case DASM_REL_PC:
221 case DASM_IMM_PC: pl = D->pclabels + va_arg(ap, int); CKPL(pc, PC);
222 putrel:
223 n = *pl;
224 if (n < 0) { /* Label exists. Get label pos and store it. */
225 b[pos] = -n;
226 } else {
227 linkrel:
228 b[pos] = n; /* Else link to rel chain, anchored at label. */
229 *pl = pos;
231 pos++;
232 ofs += 4; /* Maximum offset needed. */
233 if (action == DASM_REL_LG || action == DASM_REL_PC)
234 b[pos++] = ofs; /* Store pass1 offset estimate. */
235 break;
236 case DASM_LABEL_LG: pl = D->lglabels + *p++; CKPL(lg, LG); goto putlabel;
237 case DASM_LABEL_PC: pl = D->pclabels + va_arg(ap, int); CKPL(pc, PC);
238 putlabel:
239 n = *pl; /* n > 0: Collapse rel chain and replace with label pos. */
240 while (n > 0) { int *pb = DASM_POS2PTR(D, n); n = *pb; *pb = pos; }
241 *pl = -pos; /* Label exists now. */
242 b[pos++] = ofs; /* Store pass1 offset estimate. */
243 break;
244 case DASM_ALIGN:
245 ofs += *p++; /* Maximum alignment needed (arg is 2**n-1). */
246 b[pos++] = ofs; /* Store pass1 offset estimate. */
247 break;
248 case DASM_EXTERN: p += 2; ofs += 4; break;
249 case DASM_ESC: p++; ofs++; break;
250 case DASM_MARK: mrm = p[-2]; break;
251 case DASM_SECTION:
252 n = *p; CK(n < D->maxsection, RANGE_SEC); D->section = &D->sections[n];
253 case DASM_STOP: goto stop;
257 stop:
258 va_end(ap);
259 sec->pos = pos;
260 sec->ofs = ofs;
262 #undef CK
264 /* Pass 2: Link sections, shrink branches/aligns, fix label offsets. */
265 int dasm_link(Dst_DECL, size_t *szp)
267 dasm_State *D = Dst_REF;
268 int secnum;
269 int ofs = 0;
271 #ifdef DASM_CHECKS
272 *szp = 0;
273 if (D->status != DASM_S_OK) return D->status;
275 int pc;
276 for (pc = 0; pc*sizeof(int) < D->pcsize; pc++)
277 if (D->pclabels[pc] > 0) return DASM_S_UNDEF_PC|pc;
279 #endif
281 { /* Handle globals not defined in this translation unit. */
282 int idx;
283 for (idx = 10; idx*sizeof(int) < D->lgsize; idx++) {
284 int n = D->lglabels[idx];
285 /* Undefined label: Collapse rel chain and replace with marker (< 0). */
286 while (n > 0) { int *pb = DASM_POS2PTR(D, n); n = *pb; *pb = -idx; }
290 /* Combine all code sections. No support for data sections (yet). */
291 for (secnum = 0; secnum < D->maxsection; secnum++) {
292 dasm_Section *sec = D->sections + secnum;
293 int *b = sec->rbuf;
294 int pos = DASM_SEC2POS(secnum);
295 int lastpos = sec->pos;
297 while (pos != lastpos) {
298 dasm_ActList p = D->actionlist + b[pos++];
299 while (1) {
300 int op, action = *p++;
301 switch (action) {
302 case DASM_REL_LG: p++; op = p[-3]; goto rel_pc;
303 case DASM_REL_PC: op = p[-2]; rel_pc: {
304 int shrink = op == 0xe9 ? 3 : ((op&0xf0) == 0x80 ? 4 : 0);
305 if (shrink) { /* Shrinkable branch opcode? */
306 int lofs, lpos = b[pos];
307 if (lpos < 0) goto noshrink; /* Ext global? */
308 lofs = *DASM_POS2PTR(D, lpos);
309 if (lpos > pos) { /* Fwd label: add cumulative section offsets. */
310 int i;
311 for (i = secnum; i < DASM_POS2SEC(lpos); i++)
312 lofs += D->sections[i].ofs;
313 } else {
314 lofs -= ofs; /* Bkwd label: unfix offset. */
316 lofs -= b[pos+1]; /* Short branch ok? */
317 if (lofs >= -128-shrink && lofs <= 127) ofs -= shrink; /* Yes. */
318 else { noshrink: shrink = 0; } /* No, cannot shrink op. */
320 b[pos+1] = shrink;
321 pos += 2;
322 break;
324 case DASM_SPACE: case DASM_IMM_LG: case DASM_VREG: p++;
325 case DASM_DISP: case DASM_IMM_S: case DASM_IMM_B: case DASM_IMM_W:
326 case DASM_IMM_D: case DASM_IMM_WB: case DASM_IMM_DB:
327 case DASM_SETLABEL: case DASM_REL_A: case DASM_IMM_PC: pos++; break;
328 case DASM_LABEL_LG: p++;
329 case DASM_LABEL_PC: b[pos++] += ofs; break; /* Fix label offset. */
330 case DASM_ALIGN: ofs -= (b[pos++]+ofs)&*p++; break; /* Adjust ofs. */
331 case DASM_EXTERN: p += 2; break;
332 case DASM_ESC: p++; break;
333 case DASM_MARK: break;
334 case DASM_SECTION: case DASM_STOP: goto stop;
337 stop: (void)0;
339 ofs += sec->ofs; /* Next section starts right after current section. */
342 D->codesize = ofs; /* Total size of all code sections */
343 *szp = ofs;
344 return DASM_S_OK;
347 #define dasmb(x) *cp++ = (unsigned char)(x)
348 #ifndef DASM_ALIGNED_WRITES
349 #define dasmw(x) \
350 do { *((unsigned short *)cp) = (unsigned short)(x); cp+=2; } while (0)
351 #define dasmd(x) \
352 do { *((unsigned int *)cp) = (unsigned int)(x); cp+=4; } while (0)
353 #else
354 #define dasmw(x) do { dasmb(x); dasmb((x)>>8); } while (0)
355 #define dasmd(x) do { dasmw(x); dasmw((x)>>16); } while (0)
356 #endif
358 /* Pass 3: Encode sections. */
359 int dasm_encode(Dst_DECL, void *buffer)
361 dasm_State *D = Dst_REF;
362 unsigned char *base = (unsigned char *)buffer;
363 unsigned char *cp = base;
364 int secnum;
366 /* Encode all code sections. No support for data sections (yet). */
367 for (secnum = 0; secnum < D->maxsection; secnum++) {
368 dasm_Section *sec = D->sections + secnum;
369 int *b = sec->buf;
370 int *endb = sec->rbuf + sec->pos;
372 while (b != endb) {
373 dasm_ActList p = D->actionlist + *b++;
374 unsigned char *mark = NULL;
375 while (1) {
376 int action = *p++;
377 int n = (action >= DASM_DISP && action <= DASM_ALIGN) ? *b++ : 0;
378 switch (action) {
379 case DASM_DISP: if (!mark) mark = cp; {
380 unsigned char *mm = mark;
381 if (*p != DASM_IMM_DB && *p != DASM_IMM_WB) mark = NULL;
382 if (n == 0) { int mrm = mm[-1]&7; if (mrm == 4) mrm = mm[0]&7;
383 if (mrm != 5) { mm[-1] -= 0x80; break; } }
384 if (((n+128) & -256) != 0) goto wd; else mm[-1] -= 0x40;
386 case DASM_IMM_S: case DASM_IMM_B: wb: dasmb(n); break;
387 case DASM_IMM_DB: if (((n+128)&-256) == 0) {
388 db: if (!mark) mark = cp; mark[-2] += 2; mark = NULL; goto wb;
389 } else mark = NULL;
390 case DASM_IMM_D: wd: dasmd(n); break;
391 case DASM_IMM_WB: if (((n+128)&-256) == 0) goto db; else mark = NULL;
392 case DASM_IMM_W: dasmw(n); break;
393 case DASM_VREG: { int t = *p++; if (t >= 2) n<<=3; cp[-1] |= n; break; }
394 case DASM_REL_LG: p++; if (n >= 0) goto rel_pc;
395 b++; n = (int)(ptrdiff_t)D->globals[-n];
396 case DASM_REL_A: rel_a: n -= (int)(ptrdiff_t)(cp+4); goto wd; /* !x64 */
397 case DASM_REL_PC: rel_pc: {
398 int shrink = *b++;
399 int *pb = DASM_POS2PTR(D, n); if (*pb < 0) { n = pb[1]; goto rel_a; }
400 n = *pb - ((int)(cp-base) + 4-shrink);
401 if (shrink == 0) goto wd;
402 if (shrink == 4) { cp--; cp[-1] = *cp-0x10; } else cp[-1] = 0xeb;
403 goto wb;
405 case DASM_IMM_LG:
406 p++; if (n < 0) { n = (int)(ptrdiff_t)D->globals[-n]; goto wd; }
407 case DASM_IMM_PC: {
408 int *pb = DASM_POS2PTR(D, n);
409 n = *pb < 0 ? pb[1] : (*pb + (int)(ptrdiff_t)base);
410 goto wd;
412 case DASM_LABEL_LG: {
413 int idx = *p++;
414 if (idx >= 10)
415 D->globals[idx] = (void *)(base + (*p == DASM_SETLABEL ? *b : n));
416 break;
418 case DASM_LABEL_PC: case DASM_SETLABEL: break;
419 case DASM_SPACE: { int fill = *p++; while (n--) *cp++ = fill; break; }
420 case DASM_ALIGN:
421 n = *p++;
422 while (((cp-base) & n)) *cp++ = 0x90; /* nop */
423 break;
424 case DASM_EXTERN: n = DASM_EXTERN(Dst, cp, p[1], *p); p += 2; goto wd;
425 case DASM_MARK: mark = cp; break;
426 case DASM_ESC: action = *p++;
427 default: *cp++ = action; break;
428 case DASM_SECTION: case DASM_STOP: goto stop;
431 stop: (void)0;
435 if (base + D->codesize != cp) /* Check for phase errors. */
436 return DASM_S_PHASE;
437 return DASM_S_OK;
440 /* Get PC label offset. */
441 int dasm_getpclabel(Dst_DECL, unsigned int pc)
443 dasm_State *D = Dst_REF;
444 if (pc*sizeof(int) < D->pcsize) {
445 int pos = D->pclabels[pc];
446 if (pos < 0) return *DASM_POS2PTR(D, -pos);
447 if (pos > 0) return -1; /* Undefined. */
449 return -2; /* Unused or out of range. */
452 #ifdef DASM_CHECKS
453 /* Optional sanity checker to call between isolated encoding steps. */
454 int dasm_checkstep(Dst_DECL, int secmatch)
456 dasm_State *D = Dst_REF;
457 if (D->status == DASM_S_OK) {
458 int i;
459 for (i = 1; i <= 9; i++) {
460 if (D->lglabels[i] > 0) { D->status = DASM_S_UNDEF_L|i; break; }
461 D->lglabels[i] = 0;
464 if (D->status == DASM_S_OK && secmatch >= 0 &&
465 D->section != &D->sections[secmatch])
466 D->status = DASM_S_MATCH_SEC|(int)(D->section-D->sections);
467 return D->status;
469 #endif