Fix allocation of struct in registers on ARM
[tinycc.git] / arm-gen.c
blobc9d4e5597c4fe3766013f8c846186c72ef626027
1 /*
2 * ARMv4 code generator for TCC
3 *
4 * Copyright (c) 2003 Daniel Glöckner
5 * Copyright (c) 2012 Thomas Preud'homme
7 * Based on i386-gen.c by Fabrice Bellard
9 * This library is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public
11 * License as published by the Free Software Foundation; either
12 * version 2 of the License, or (at your option) any later version.
14 * This library is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with this library; if not, write to the Free Software
21 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
24 #ifdef TARGET_DEFS_ONLY
26 #ifdef TCC_ARM_EABI
27 #ifndef TCC_ARM_VFP /* Avoid useless warning */
28 #define TCC_ARM_VFP
29 #endif
30 #endif
32 /* number of available registers */
33 #ifdef TCC_ARM_VFP
34 #define NB_REGS 13
35 #else
36 #define NB_REGS 9
37 #endif
39 #ifndef TCC_ARM_VERSION
40 # define TCC_ARM_VERSION 5
41 #endif
43 /* a register can belong to several classes. The classes must be
44 sorted from more general to more precise (see gv2() code which does
45 assumptions on it). */
46 #define RC_INT 0x0001 /* generic integer register */
47 #define RC_FLOAT 0x0002 /* generic float register */
48 #define RC_R0 0x0004
49 #define RC_R1 0x0008
50 #define RC_R2 0x0010
51 #define RC_R3 0x0020
52 #define RC_R12 0x0040
53 #define RC_F0 0x0080
54 #define RC_F1 0x0100
55 #define RC_F2 0x0200
56 #define RC_F3 0x0400
57 #ifdef TCC_ARM_VFP
58 #define RC_F4 0x0800
59 #define RC_F5 0x1000
60 #define RC_F6 0x2000
61 #define RC_F7 0x4000
62 #endif
63 #define RC_IRET RC_R0 /* function return: integer register */
64 #define RC_LRET RC_R1 /* function return: second integer register */
65 #define RC_FRET RC_F0 /* function return: float register */
67 /* pretty names for the registers */
68 enum {
69 TREG_R0 = 0,
70 TREG_R1,
71 TREG_R2,
72 TREG_R3,
73 TREG_R12,
74 TREG_F0,
75 TREG_F1,
76 TREG_F2,
77 TREG_F3,
78 #ifdef TCC_ARM_VFP
79 TREG_F4,
80 TREG_F5,
81 TREG_F6,
82 TREG_F7,
83 #endif
86 #ifdef TCC_ARM_VFP
87 #define T2CPR(t) (((t) & VT_BTYPE) != VT_FLOAT ? 0x100 : 0)
88 #endif
90 /* return registers for function */
91 #define REG_IRET TREG_R0 /* single word int return register */
92 #define REG_LRET TREG_R1 /* second word return register (for long long) */
93 #define REG_FRET TREG_F0 /* float return register */
95 #ifdef TCC_ARM_EABI
96 #define TOK___divdi3 TOK___aeabi_ldivmod
97 #define TOK___moddi3 TOK___aeabi_ldivmod
98 #define TOK___udivdi3 TOK___aeabi_uldivmod
99 #define TOK___umoddi3 TOK___aeabi_uldivmod
100 #endif
102 /* defined if function parameters must be evaluated in reverse order */
103 #define INVERT_FUNC_PARAMS
105 /* defined if structures are passed as pointers. Otherwise structures
106 are directly pushed on stack. */
107 /* #define FUNC_STRUCT_PARAM_AS_PTR */
109 /* pointer size, in bytes */
110 #define PTR_SIZE 4
112 /* long double size and alignment, in bytes */
113 #ifdef TCC_ARM_VFP
114 #define LDOUBLE_SIZE 8
115 #endif
117 #ifndef LDOUBLE_SIZE
118 #define LDOUBLE_SIZE 8
119 #endif
121 #ifdef TCC_ARM_EABI
122 #define LDOUBLE_ALIGN 8
123 #else
124 #define LDOUBLE_ALIGN 4
125 #endif
127 /* maximum alignment (for aligned attribute support) */
128 #define MAX_ALIGN 8
130 #define CHAR_IS_UNSIGNED
132 /******************************************************/
133 /* ELF defines */
135 #define EM_TCC_TARGET EM_ARM
137 /* relocation type for 32 bit data relocation */
138 #define R_DATA_32 R_ARM_ABS32
139 #define R_DATA_PTR R_ARM_ABS32
140 #define R_JMP_SLOT R_ARM_JUMP_SLOT
141 #define R_COPY R_ARM_COPY
143 #define ELF_START_ADDR 0x00008000
144 #define ELF_PAGE_SIZE 0x1000
146 /******************************************************/
147 #else /* ! TARGET_DEFS_ONLY */
148 /******************************************************/
149 #include "tcc.h"
151 ST_DATA const int reg_classes[NB_REGS] = {
152 /* r0 */ RC_INT | RC_R0,
153 /* r1 */ RC_INT | RC_R1,
154 /* r2 */ RC_INT | RC_R2,
155 /* r3 */ RC_INT | RC_R3,
156 /* r12 */ RC_INT | RC_R12,
157 /* f0 */ RC_FLOAT | RC_F0,
158 /* f1 */ RC_FLOAT | RC_F1,
159 /* f2 */ RC_FLOAT | RC_F2,
160 /* f3 */ RC_FLOAT | RC_F3,
161 #ifdef TCC_ARM_VFP
162 /* d4/s8 */ RC_FLOAT | RC_F4,
163 /* d5/s10 */ RC_FLOAT | RC_F5,
164 /* d6/s12 */ RC_FLOAT | RC_F6,
165 /* d7/s14 */ RC_FLOAT | RC_F7,
166 #endif
169 static int func_sub_sp_offset, last_itod_magic;
170 static int leaffunc;
172 #if defined(TCC_ARM_EABI) && defined(TCC_ARM_VFP)
173 static CType float_type, double_type, func_float_type, func_double_type;
174 ST_FUNC void arm_init_types(void)
176 float_type.t = VT_FLOAT;
177 double_type.t = VT_DOUBLE;
178 func_float_type.t = VT_FUNC;
179 func_float_type.ref = sym_push(SYM_FIELD, &float_type, FUNC_CDECL, FUNC_OLD);
180 func_double_type.t = VT_FUNC;
181 func_double_type.ref = sym_push(SYM_FIELD, &double_type, FUNC_CDECL, FUNC_OLD);
183 #else
184 #define func_float_type func_old_type
185 #define func_double_type func_old_type
186 #define func_ldouble_type func_old_type
187 ST_FUNC void arm_init_types(void) {}
188 #endif
190 static int two2mask(int a,int b) {
191 return (reg_classes[a]|reg_classes[b])&~(RC_INT|RC_FLOAT);
194 static int regmask(int r) {
195 return reg_classes[r]&~(RC_INT|RC_FLOAT);
198 /******************************************************/
200 void o(uint32_t i)
202 /* this is a good place to start adding big-endian support*/
203 int ind1;
205 ind1 = ind + 4;
206 if (!cur_text_section)
207 tcc_error("compiler error! This happens f.ex. if the compiler\n"
208 "can't evaluate constant expressions outside of a function.");
209 if (ind1 > cur_text_section->data_allocated)
210 section_realloc(cur_text_section, ind1);
211 cur_text_section->data[ind++] = i&255;
212 i>>=8;
213 cur_text_section->data[ind++] = i&255;
214 i>>=8;
215 cur_text_section->data[ind++] = i&255;
216 i>>=8;
217 cur_text_section->data[ind++] = i;
220 static uint32_t stuff_const(uint32_t op, uint32_t c)
222 int try_neg=0;
223 uint32_t nc = 0, negop = 0;
225 switch(op&0x1F00000)
227 case 0x800000: //add
228 case 0x400000: //sub
229 try_neg=1;
230 negop=op^0xC00000;
231 nc=-c;
232 break;
233 case 0x1A00000: //mov
234 case 0x1E00000: //mvn
235 try_neg=1;
236 negop=op^0x400000;
237 nc=~c;
238 break;
239 case 0x200000: //xor
240 if(c==~0)
241 return (op&0xF010F000)|((op>>16)&0xF)|0x1E00000;
242 break;
243 case 0x0: //and
244 if(c==~0)
245 return (op&0xF010F000)|((op>>16)&0xF)|0x1A00000;
246 case 0x1C00000: //bic
247 try_neg=1;
248 negop=op^0x1C00000;
249 nc=~c;
250 break;
251 case 0x1800000: //orr
252 if(c==~0)
253 return (op&0xFFF0FFFF)|0x1E00000;
254 break;
256 do {
257 uint32_t m;
258 int i;
259 if(c<256) /* catch undefined <<32 */
260 return op|c;
261 for(i=2;i<32;i+=2) {
262 m=(0xff>>i)|(0xff<<(32-i));
263 if(!(c&~m))
264 return op|(i<<7)|(c<<i)|(c>>(32-i));
266 op=negop;
267 c=nc;
268 } while(try_neg--);
269 return 0;
273 //only add,sub
274 void stuff_const_harder(uint32_t op, uint32_t v) {
275 uint32_t x;
276 x=stuff_const(op,v);
277 if(x)
278 o(x);
279 else {
280 uint32_t a[16], nv, no, o2, n2;
281 int i,j,k;
282 a[0]=0xff;
283 o2=(op&0xfff0ffff)|((op&0xf000)<<4);;
284 for(i=1;i<16;i++)
285 a[i]=(a[i-1]>>2)|(a[i-1]<<30);
286 for(i=0;i<12;i++)
287 for(j=i<4?i+12:15;j>=i+4;j--)
288 if((v&(a[i]|a[j]))==v) {
289 o(stuff_const(op,v&a[i]));
290 o(stuff_const(o2,v&a[j]));
291 return;
293 no=op^0xC00000;
294 n2=o2^0xC00000;
295 nv=-v;
296 for(i=0;i<12;i++)
297 for(j=i<4?i+12:15;j>=i+4;j--)
298 if((nv&(a[i]|a[j]))==nv) {
299 o(stuff_const(no,nv&a[i]));
300 o(stuff_const(n2,nv&a[j]));
301 return;
303 for(i=0;i<8;i++)
304 for(j=i+4;j<12;j++)
305 for(k=i<4?i+12:15;k>=j+4;k--)
306 if((v&(a[i]|a[j]|a[k]))==v) {
307 o(stuff_const(op,v&a[i]));
308 o(stuff_const(o2,v&a[j]));
309 o(stuff_const(o2,v&a[k]));
310 return;
312 no=op^0xC00000;
313 nv=-v;
314 for(i=0;i<8;i++)
315 for(j=i+4;j<12;j++)
316 for(k=i<4?i+12:15;k>=j+4;k--)
317 if((nv&(a[i]|a[j]|a[k]))==nv) {
318 o(stuff_const(no,nv&a[i]));
319 o(stuff_const(n2,nv&a[j]));
320 o(stuff_const(n2,nv&a[k]));
321 return;
323 o(stuff_const(op,v&a[0]));
324 o(stuff_const(o2,v&a[4]));
325 o(stuff_const(o2,v&a[8]));
326 o(stuff_const(o2,v&a[12]));
330 ST_FUNC uint32_t encbranch(int pos, int addr, int fail)
332 addr-=pos+8;
333 addr/=4;
334 if(addr>=0x1000000 || addr<-0x1000000) {
335 if(fail)
336 tcc_error("FIXME: function bigger than 32MB");
337 return 0;
339 return 0x0A000000|(addr&0xffffff);
342 int decbranch(int pos)
344 int x;
345 x=*(uint32_t *)(cur_text_section->data + pos);
346 x&=0x00ffffff;
347 if(x&0x800000)
348 x-=0x1000000;
349 return x*4+pos+8;
352 /* output a symbol and patch all calls to it */
353 void gsym_addr(int t, int a)
355 uint32_t *x;
356 int lt;
357 while(t) {
358 x=(uint32_t *)(cur_text_section->data + t);
359 t=decbranch(lt=t);
360 if(a==lt+4)
361 *x=0xE1A00000; // nop
362 else {
363 *x &= 0xff000000;
364 *x |= encbranch(lt,a,1);
369 void gsym(int t)
371 gsym_addr(t, ind);
374 #ifdef TCC_ARM_VFP
375 static uint32_t vfpr(int r)
377 if(r<TREG_F0 || r>TREG_F7)
378 tcc_error("compiler error! register %i is no vfp register",r);
379 return r-5;
381 #else
382 static uint32_t fpr(int r)
384 if(r<TREG_F0 || r>TREG_F3)
385 tcc_error("compiler error! register %i is no fpa register",r);
386 return r-5;
388 #endif
390 static uint32_t intr(int r)
392 if(r==4)
393 return 12;
394 if((r<0 || r>4) && r!=14)
395 tcc_error("compiler error! register %i is no int register",r);
396 return r;
399 static void calcaddr(uint32_t *base, int *off, int *sgn, int maxoff, unsigned shift)
401 if(*off>maxoff || *off&((1<<shift)-1)) {
402 uint32_t x, y;
403 x=0xE280E000;
404 if(*sgn)
405 x=0xE240E000;
406 x|=(*base)<<16;
407 *base=14; // lr
408 y=stuff_const(x,*off&~maxoff);
409 if(y) {
410 o(y);
411 *off&=maxoff;
412 return;
414 y=stuff_const(x,(*off+maxoff)&~maxoff);
415 if(y) {
416 o(y);
417 *sgn=!*sgn;
418 *off=((*off+maxoff)&~maxoff)-*off;
419 return;
421 stuff_const_harder(x,*off&~maxoff);
422 *off&=maxoff;
426 static uint32_t mapcc(int cc)
428 switch(cc)
430 case TOK_ULT:
431 return 0x30000000; /* CC/LO */
432 case TOK_UGE:
433 return 0x20000000; /* CS/HS */
434 case TOK_EQ:
435 return 0x00000000; /* EQ */
436 case TOK_NE:
437 return 0x10000000; /* NE */
438 case TOK_ULE:
439 return 0x90000000; /* LS */
440 case TOK_UGT:
441 return 0x80000000; /* HI */
442 case TOK_Nset:
443 return 0x40000000; /* MI */
444 case TOK_Nclear:
445 return 0x50000000; /* PL */
446 case TOK_LT:
447 return 0xB0000000; /* LT */
448 case TOK_GE:
449 return 0xA0000000; /* GE */
450 case TOK_LE:
451 return 0xD0000000; /* LE */
452 case TOK_GT:
453 return 0xC0000000; /* GT */
455 tcc_error("unexpected condition code");
456 return 0xE0000000; /* AL */
459 static int negcc(int cc)
461 switch(cc)
463 case TOK_ULT:
464 return TOK_UGE;
465 case TOK_UGE:
466 return TOK_ULT;
467 case TOK_EQ:
468 return TOK_NE;
469 case TOK_NE:
470 return TOK_EQ;
471 case TOK_ULE:
472 return TOK_UGT;
473 case TOK_UGT:
474 return TOK_ULE;
475 case TOK_Nset:
476 return TOK_Nclear;
477 case TOK_Nclear:
478 return TOK_Nset;
479 case TOK_LT:
480 return TOK_GE;
481 case TOK_GE:
482 return TOK_LT;
483 case TOK_LE:
484 return TOK_GT;
485 case TOK_GT:
486 return TOK_LE;
488 tcc_error("unexpected condition code");
489 return TOK_NE;
492 /* load 'r' from value 'sv' */
493 void load(int r, SValue *sv)
495 int v, ft, fc, fr, sign;
496 uint32_t op;
497 SValue v1;
499 fr = sv->r;
500 ft = sv->type.t;
501 fc = sv->c.ul;
503 if(fc>=0)
504 sign=0;
505 else {
506 sign=1;
507 fc=-fc;
510 v = fr & VT_VALMASK;
511 if (fr & VT_LVAL) {
512 uint32_t base = 0xB; // fp
513 if(v == VT_LLOCAL) {
514 v1.type.t = VT_PTR;
515 v1.r = VT_LOCAL | VT_LVAL;
516 v1.c.ul = sv->c.ul;
517 load(base=14 /* lr */, &v1);
518 fc=sign=0;
519 v=VT_LOCAL;
520 } else if(v == VT_CONST) {
521 v1.type.t = VT_PTR;
522 v1.r = fr&~VT_LVAL;
523 v1.c.ul = sv->c.ul;
524 v1.sym=sv->sym;
525 load(base=14, &v1);
526 fc=sign=0;
527 v=VT_LOCAL;
528 } else if(v < VT_CONST) {
529 base=intr(v);
530 fc=sign=0;
531 v=VT_LOCAL;
533 if(v == VT_LOCAL) {
534 if(is_float(ft)) {
535 calcaddr(&base,&fc,&sign,1020,2);
536 #ifdef TCC_ARM_VFP
537 op=0xED100A00; /* flds */
538 if(!sign)
539 op|=0x800000;
540 if ((ft & VT_BTYPE) != VT_FLOAT)
541 op|=0x100; /* flds -> fldd */
542 o(op|(vfpr(r)<<12)|(fc>>2)|(base<<16));
543 #else
544 op=0xED100100;
545 if(!sign)
546 op|=0x800000;
547 #if LDOUBLE_SIZE == 8
548 if ((ft & VT_BTYPE) != VT_FLOAT)
549 op|=0x8000;
550 #else
551 if ((ft & VT_BTYPE) == VT_DOUBLE)
552 op|=0x8000;
553 else if ((ft & VT_BTYPE) == VT_LDOUBLE)
554 op|=0x400000;
555 #endif
556 o(op|(fpr(r)<<12)|(fc>>2)|(base<<16));
557 #endif
558 } else if((ft & (VT_BTYPE|VT_UNSIGNED)) == VT_BYTE
559 || (ft & VT_BTYPE) == VT_SHORT) {
560 calcaddr(&base,&fc,&sign,255,0);
561 op=0xE1500090;
562 if ((ft & VT_BTYPE) == VT_SHORT)
563 op|=0x20;
564 if ((ft & VT_UNSIGNED) == 0)
565 op|=0x40;
566 if(!sign)
567 op|=0x800000;
568 o(op|(intr(r)<<12)|(base<<16)|((fc&0xf0)<<4)|(fc&0xf));
569 } else {
570 calcaddr(&base,&fc,&sign,4095,0);
571 op=0xE5100000;
572 if(!sign)
573 op|=0x800000;
574 if ((ft & VT_BTYPE) == VT_BYTE || (ft & VT_BTYPE) == VT_BOOL)
575 op|=0x400000;
576 o(op|(intr(r)<<12)|fc|(base<<16));
578 return;
580 } else {
581 if (v == VT_CONST) {
582 op=stuff_const(0xE3A00000|(intr(r)<<12),sv->c.ul);
583 if (fr & VT_SYM || !op) {
584 o(0xE59F0000|(intr(r)<<12));
585 o(0xEA000000);
586 if(fr & VT_SYM)
587 greloc(cur_text_section, sv->sym, ind, R_ARM_ABS32);
588 o(sv->c.ul);
589 } else
590 o(op);
591 return;
592 } else if (v == VT_LOCAL) {
593 op=stuff_const(0xE28B0000|(intr(r)<<12),sv->c.ul);
594 if (fr & VT_SYM || !op) {
595 o(0xE59F0000|(intr(r)<<12));
596 o(0xEA000000);
597 if(fr & VT_SYM) // needed ?
598 greloc(cur_text_section, sv->sym, ind, R_ARM_ABS32);
599 o(sv->c.ul);
600 o(0xE08B0000|(intr(r)<<12)|intr(r));
601 } else
602 o(op);
603 return;
604 } else if(v == VT_CMP) {
605 o(mapcc(sv->c.ul)|0x3A00001|(intr(r)<<12));
606 o(mapcc(negcc(sv->c.ul))|0x3A00000|(intr(r)<<12));
607 return;
608 } else if (v == VT_JMP || v == VT_JMPI) {
609 int t;
610 t = v & 1;
611 o(0xE3A00000|(intr(r)<<12)|t);
612 o(0xEA000000);
613 gsym(sv->c.ul);
614 o(0xE3A00000|(intr(r)<<12)|(t^1));
615 return;
616 } else if (v < VT_CONST) {
617 if(is_float(ft))
618 #ifdef TCC_ARM_VFP
619 o(0xEEB00A40|(vfpr(r)<<12)|vfpr(v)|T2CPR(ft)); /* fcpyX */
620 #else
621 o(0xEE008180|(fpr(r)<<12)|fpr(v));
622 #endif
623 else
624 o(0xE1A00000|(intr(r)<<12)|intr(v));
625 return;
628 tcc_error("load unimplemented!");
631 /* store register 'r' in lvalue 'v' */
632 void store(int r, SValue *sv)
634 SValue v1;
635 int v, ft, fc, fr, sign;
636 uint32_t op;
638 fr = sv->r;
639 ft = sv->type.t;
640 fc = sv->c.ul;
642 if(fc>=0)
643 sign=0;
644 else {
645 sign=1;
646 fc=-fc;
649 v = fr & VT_VALMASK;
650 if (fr & VT_LVAL || fr == VT_LOCAL) {
651 uint32_t base = 0xb;
652 if(v < VT_CONST) {
653 base=intr(v);
654 v=VT_LOCAL;
655 fc=sign=0;
656 } else if(v == VT_CONST) {
657 v1.type.t = ft;
658 v1.r = fr&~VT_LVAL;
659 v1.c.ul = sv->c.ul;
660 v1.sym=sv->sym;
661 load(base=14, &v1);
662 fc=sign=0;
663 v=VT_LOCAL;
665 if(v == VT_LOCAL) {
666 if(is_float(ft)) {
667 calcaddr(&base,&fc,&sign,1020,2);
668 #ifdef TCC_ARM_VFP
669 op=0xED000A00; /* fsts */
670 if(!sign)
671 op|=0x800000;
672 if ((ft & VT_BTYPE) != VT_FLOAT)
673 op|=0x100; /* fsts -> fstd */
674 o(op|(vfpr(r)<<12)|(fc>>2)|(base<<16));
675 #else
676 op=0xED000100;
677 if(!sign)
678 op|=0x800000;
679 #if LDOUBLE_SIZE == 8
680 if ((ft & VT_BTYPE) != VT_FLOAT)
681 op|=0x8000;
682 #else
683 if ((ft & VT_BTYPE) == VT_DOUBLE)
684 op|=0x8000;
685 if ((ft & VT_BTYPE) == VT_LDOUBLE)
686 op|=0x400000;
687 #endif
688 o(op|(fpr(r)<<12)|(fc>>2)|(base<<16));
689 #endif
690 return;
691 } else if((ft & VT_BTYPE) == VT_SHORT) {
692 calcaddr(&base,&fc,&sign,255,0);
693 op=0xE14000B0;
694 if(!sign)
695 op|=0x800000;
696 o(op|(intr(r)<<12)|(base<<16)|((fc&0xf0)<<4)|(fc&0xf));
697 } else {
698 calcaddr(&base,&fc,&sign,4095,0);
699 op=0xE5000000;
700 if(!sign)
701 op|=0x800000;
702 if ((ft & VT_BTYPE) == VT_BYTE || (ft & VT_BTYPE) == VT_BOOL)
703 op|=0x400000;
704 o(op|(intr(r)<<12)|fc|(base<<16));
706 return;
709 tcc_error("store unimplemented");
712 static void gadd_sp(int val)
714 stuff_const_harder(0xE28DD000,val);
717 /* 'is_jmp' is '1' if it is a jump */
718 static void gcall_or_jmp(int is_jmp)
720 int r;
721 if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
722 uint32_t x;
723 /* constant case */
724 x=encbranch(ind,ind+vtop->c.ul,0);
725 if(x) {
726 if (vtop->r & VT_SYM) {
727 /* relocation case */
728 greloc(cur_text_section, vtop->sym, ind, R_ARM_PC24);
729 } else
730 put_elf_reloc(symtab_section, cur_text_section, ind, R_ARM_PC24, 0);
731 o(x|(is_jmp?0xE0000000:0xE1000000));
732 } else {
733 if(!is_jmp)
734 o(0xE28FE004); // add lr,pc,#4
735 o(0xE51FF004); // ldr pc,[pc,#-4]
736 if (vtop->r & VT_SYM)
737 greloc(cur_text_section, vtop->sym, ind, R_ARM_ABS32);
738 o(vtop->c.ul);
740 } else {
741 /* otherwise, indirect call */
742 r = gv(RC_INT);
743 if(!is_jmp)
744 o(0xE1A0E00F); // mov lr,pc
745 o(0xE1A0F000|intr(r)); // mov pc,r
749 #ifdef TCC_ARM_HARDFLOAT
750 static int is_float_hgen_aggr(CType *type)
752 if ((type->t & VT_BTYPE) == VT_STRUCT) {
753 struct Sym *ref;
754 int btype, nb_fields = 0;
756 ref = type->ref;
757 btype = ref->type.t & VT_BTYPE;
758 if (btype == VT_FLOAT || btype == VT_DOUBLE) {
759 for(; ref && btype == (ref->type.t & VT_BTYPE); ref = ref->next, nb_fields++);
760 return !ref && nb_fields <= 4;
763 return 0;
766 struct avail_regs {
767 /* worst case: f(float, double, 3 float struct, double, 3 float struct, double) */
768 signed char avail[3];
769 int first_hole;
770 int last_hole;
771 int first_free_reg;
774 #define AVAIL_REGS_INITIALIZER (struct avail_regs) { { 0, 0, 0}, 0, 0, 0 }
776 /* Assign a register for a CPRC param with correct size and alignment
777 * size and align are in bytes, as returned by type_size */
778 int assign_fpreg(struct avail_regs *avregs, int align, int size)
780 int first_reg = 0;
782 if (avregs->first_free_reg == -1)
783 return -1;
784 if (align >> 3) { // alignment needed (base type: double)
785 first_reg = avregs->first_free_reg;
786 if (first_reg & 1)
787 avregs->avail[avregs->last_hole++] = first_reg++;
788 } else {
789 if (size == 4 && avregs->first_hole != avregs->last_hole)
790 return avregs->avail[avregs->first_hole++];
791 else
792 first_reg = avregs->first_free_reg;
794 if (first_reg + size / 4 <= 16) {
795 avregs->first_free_reg = first_reg + size / 4;
796 return first_reg;
798 avregs->first_free_reg = -1;
799 return -1;
801 #endif
803 /* Return 1 if this function returns via an sret pointer, 0 otherwise */
804 ST_FUNC int gfunc_sret(CType *vt, CType *ret, int *ret_align) {
805 #if TCC_ARM_EABI
806 int size, align;
807 size = type_size(vt, &align);
808 if (size > 4) {
809 return 1;
810 } else {
811 *ret_align = 4;
812 ret->ref = NULL;
813 ret->t = VT_INT;
814 return 0;
816 #else
817 return 1;
818 #endif
821 /* Generate function call. The function address is pushed first, then
822 all the parameters in call order. This functions pops all the
823 parameters and the function address. */
824 void gfunc_call(int nb_args)
826 int size, align, r, args_size, i, ncrn, ncprn, argno, vfp_argno;
827 signed char plan[4][2]={{-1,-1},{-1,-1},{-1,-1},{-1,-1}};
828 SValue *before_stack = NULL; /* SValue before first on stack argument */
829 SValue *before_creg = NULL; /* SValue before first argument of type struct in core register */
830 #ifdef TCC_ARM_HARDFLOAT
831 struct avail_regs avregs = AVAIL_REGS_INITIALIZER;
832 signed char vfp_plan[16];
833 int plan2[4+16];
834 int variadic;
835 #else
836 int plan2[4]={0,0,0,0};
837 #endif
838 int vfp_todo=0;
839 int todo=0, keep;
841 #ifdef TCC_ARM_HARDFLOAT
842 memset(vfp_plan, -1, sizeof(vfp_plan));
843 memset(plan2, 0, sizeof(plan2));
844 variadic = (vtop[-nb_args].type.ref->c == FUNC_ELLIPSIS);
845 #endif
846 r = vtop->r & VT_VALMASK;
847 if (r == VT_CMP || (r & ~1) == VT_JMP)
848 gv(RC_INT);
849 #ifdef TCC_ARM_EABI
850 if((vtop[-nb_args].type.ref->type.t & VT_BTYPE) == VT_STRUCT
851 && type_size(&vtop[-nb_args].type.ref->type, &align) <= 4) {
852 SValue tmp;
853 tmp=vtop[-nb_args];
854 vtop[-nb_args]=vtop[-nb_args+1];
855 vtop[-nb_args+1]=tmp;
856 --nb_args;
859 vpushi(0), nb_args++;
860 vtop->type.t = VT_LLONG;
861 #endif
862 ncrn = ncprn = argno = vfp_argno = args_size = 0;
863 /* Assign argument to registers and stack with alignment.
864 If, considering alignment constraints, enough registers of the correct type
865 (core or VFP) are free for the current argument, assign them to it, else
866 allocate on stack with correct alignment. Whenever a structure is allocated
867 in registers or on stack, it is always put on the stack at this stage. The
868 stack is divided in 3 zones. The zone are, from high addresses to low
869 addresses: structures to be loaded in core registers, structures to be
870 loaded in VFP registers, argument allocated to stack. SValue's representing
871 structures in the first zone are moved just after the SValue pointed by
872 before_stack. SValue's representing structures in the second zone are
873 moved just after the SValue pointer by before_creg. */
874 for(i = nb_args; i-- ;) {
875 int j, assigned_vfpreg = 0;
876 size = type_size(&vtop[-i].type, &align);
877 switch(vtop[-i].type.t & VT_BTYPE) {
878 case VT_STRUCT:
879 case VT_FLOAT:
880 case VT_DOUBLE:
881 case VT_LDOUBLE:
882 #ifdef TCC_ARM_HARDFLOAT
883 if (!variadic) {
884 int hfa = 0; /* Homogeneous float aggregate */
886 if (is_float(vtop[-i].type.t)
887 || (hfa = is_float_hgen_aggr(&vtop[-i].type))) {
888 int end_reg;
890 assigned_vfpreg = assign_fpreg(&avregs, align, size);
891 end_reg = assigned_vfpreg + (size - 1) / 4;
892 if (assigned_vfpreg >= 0) {
893 vfp_plan[vfp_argno++]=TREG_F0 + assigned_vfpreg/2;
894 if (hfa) {
895 /* if before_creg is not set, it means that no parameter has been
896 * allocated in core register. This implied that no argument has
897 * been allocated on stack neither because a VFP was available for
898 * this parameter. */
899 if (before_creg) {
900 /* before_creg already exists and we just update it */
901 vrote(&vtop[-i], &vtop[-i] - before_creg);
902 before_creg++;
904 for (j = assigned_vfpreg; j <= end_reg; j++)
905 vfp_todo|=(1<<j);
907 continue;
908 } else {
909 if (!hfa)
910 vfp_argno++;
911 if (!before_stack)
912 before_stack = &vtop[-i-1];
913 break;
917 #endif
918 ncrn = (ncrn + (align-1)/4) & -(align/4);
919 size = (size + 3) & -4;
920 if (ncrn + size/4 <= 4 || (ncrn < 4 && assigned_vfpreg != -1)) {
921 if (before_stack) {
922 vrote(&vtop[-i], &vtop[-i] - before_stack);
923 before_stack++;
924 /* before_stack can only have been set because all VFP registers are
925 * assigned, so no need to care about before_creg if before_stack is
926 * set since no more argument will be allocated in a VFP register. */
927 } else if (!before_creg)
928 before_creg = &vtop[-i];
929 for (j = ncrn; j < 4 && j < ncrn + size / 4; j++)
930 todo|=(1<<j);
931 ncrn+=size/4;
932 if (ncrn > 4) {
933 args_size = (ncrn - 4) * 4;
934 if (!before_stack)
935 before_stack = &vtop[-i-1];
937 } else {
938 ncrn = 4;
939 /* No need to set before_creg since it has already been set when
940 * assigning argument to core registers */
941 if (!before_stack)
942 before_stack = &vtop[-i-1];
943 break;
945 continue;
946 default:
947 #ifdef TCC_ARM_EABI
948 if (!i) {
949 break;
951 #endif
952 if (ncrn < 4) {
953 int is_long = (vtop[-i].type.t & VT_BTYPE) == VT_LLONG;
955 if (is_long) {
956 ncrn = (ncrn + 1) & -2;
957 if (ncrn == 4) {
958 argno++;
959 break;
962 plan[argno++][0]=ncrn++;
963 if (is_long) {
964 plan[argno-1][1]=ncrn++;
966 continue;
968 argno++;
970 #ifdef TCC_ARM_EABI
971 if(args_size & (align-1)) {
972 vpushi(0);
973 vtop->type.t = VT_VOID; /* padding */
974 vrott(i+2);
975 args_size += 4;
976 nb_args++;
977 argno++;
979 #endif
980 args_size += (size + 3) & -4;
982 #ifdef TCC_ARM_EABI
983 vtop--, nb_args--;
984 #endif
985 args_size = keep = 0;
986 for(i = 0;i < nb_args; i++) {
987 vrotb(keep+1);
988 if ((vtop->type.t & VT_BTYPE) == VT_STRUCT) {
989 size = type_size(&vtop->type, &align);
990 /* align to stack align size */
991 size = (size + 3) & -4;
992 /* allocate the necessary size on stack */
993 gadd_sp(-size);
994 /* generate structure store */
995 r = get_reg(RC_INT);
996 o(0xE1A0000D|(intr(r)<<12));
997 vset(&vtop->type, r | VT_LVAL, 0);
998 vswap();
999 vstore();
1000 vtop--;
1001 args_size += size;
1002 } else if (is_float(vtop->type.t)) {
1003 #ifdef TCC_ARM_HARDFLOAT
1004 if (!variadic && --vfp_argno<16 && vfp_plan[vfp_argno]!=-1) {
1005 plan2[keep++]=vfp_plan[vfp_argno];
1006 continue;
1008 #endif
1009 #ifdef TCC_ARM_VFP
1010 r=vfpr(gv(RC_FLOAT))<<12;
1011 size=4;
1012 if ((vtop->type.t & VT_BTYPE) != VT_FLOAT)
1014 size=8;
1015 r|=0x101; /* fstms -> fstmd */
1017 o(0xED2D0A01+r);
1018 #else
1019 r=fpr(gv(RC_FLOAT))<<12;
1020 if ((vtop->type.t & VT_BTYPE) == VT_FLOAT)
1021 size = 4;
1022 else if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
1023 size = 8;
1024 else
1025 size = LDOUBLE_SIZE;
1027 if (size == 12)
1028 r|=0x400000;
1029 else if(size == 8)
1030 r|=0x8000;
1032 o(0xED2D0100|r|(size>>2));
1033 #endif
1034 vtop--;
1035 args_size += size;
1036 } else {
1037 int s;
1038 /* simple type (currently always same size) */
1039 /* XXX: implicit cast ? */
1040 size=4;
1041 if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
1042 lexpand_nr();
1043 s=-1;
1044 if(--argno<4 && plan[argno][1]!=-1)
1045 s=plan[argno][1];
1046 argno++;
1047 size = 8;
1048 if(s==-1) {
1049 r = gv(RC_INT);
1050 o(0xE52D0004|(intr(r)<<12)); /* str r,[sp,#-4]! */
1051 vtop--;
1052 } else {
1053 size=0;
1054 plan2[keep]=s;
1055 keep++;
1056 vswap();
1059 s=-1;
1060 if(--argno<4 && plan[argno][0]!=-1)
1061 s=plan[argno][0];
1062 #ifdef TCC_ARM_EABI
1063 if(vtop->type.t == VT_VOID) {
1064 if(s == -1)
1065 o(0xE24DD004); /* sub sp,sp,#4 */
1066 vtop--;
1067 } else
1068 #endif
1069 if(s == -1) {
1070 r = gv(RC_INT);
1071 o(0xE52D0004|(intr(r)<<12)); /* str r,[sp,#-4]! */
1072 vtop--;
1073 } else {
1074 size=0;
1075 plan2[keep]=s;
1076 keep++;
1078 args_size += size;
1081 for(i = 0; i < keep; i++) {
1082 vrotb(keep);
1083 gv(regmask(plan2[i]));
1084 #ifdef TCC_ARM_HARDFLOAT
1085 /* arg is in s(2d+1): plan2[i]<plan2[i+1] => alignment occured (ex f,d,f) */
1086 if (i < keep - 1 && is_float(vtop->type.t) && (plan2[i] <= plan2[i + 1])) {
1087 o(0xEEF00A40|(vfpr(plan2[i])<<12)|vfpr(plan2[i]));
1089 #endif
1091 save_regs(keep); /* save used temporary registers */
1092 keep++;
1093 if(vfp_todo) {
1094 int nb_fregs=0;
1096 for(i=0;i<16;i++)
1097 if(vfp_todo&(1<<i)) {
1098 o(0xED9D0A00|(i&1)<<22|(i>>1)<<12|nb_fregs);
1099 vpushi(0);
1100 /* There might be 2 floats in a double VFP reg but that doesn't seem
1101 to matter */
1102 if (!(i%2))
1103 vtop->r=TREG_F0+i/2;
1104 keep++;
1105 nb_fregs++;
1107 if (nb_fregs) {
1108 gadd_sp(nb_fregs*4);
1109 args_size-=nb_fregs*4;
1112 if(ncrn) {
1113 int nb_regs=0;
1114 if (ncrn>4)
1115 ncrn=4;
1116 todo&=((1<<ncrn)-1);
1117 if(todo) {
1118 int i;
1119 o(0xE8BD0000|todo);
1120 for(i=0;i<4;i++)
1121 if(todo&(1<<i)) {
1122 vpushi(0);
1123 vtop->r=i;
1124 keep++;
1125 nb_regs++;
1128 args_size-=nb_regs*4;
1130 vrotb(keep);
1131 gcall_or_jmp(0);
1132 if (args_size)
1133 gadd_sp(args_size);
1134 #ifdef TCC_ARM_EABI
1135 if((vtop->type.ref->type.t & VT_BTYPE) == VT_STRUCT
1136 && type_size(&vtop->type.ref->type, &align) <= 4)
1138 store(REG_IRET,vtop-keep);
1139 ++keep;
1141 #ifdef TCC_ARM_VFP
1142 #ifdef TCC_ARM_HARDFLOAT
1143 else if(variadic && is_float(vtop->type.ref->type.t)) {
1144 #else
1145 else if(is_float(vtop->type.ref->type.t)) {
1146 #endif
1147 if((vtop->type.ref->type.t & VT_BTYPE) == VT_FLOAT) {
1148 o(0xEE000A10); /* fmsr s0,r0 */
1149 } else {
1150 o(0xEE000B10); /* fmdlr d0,r0 */
1151 o(0xEE201B10); /* fmdhr d0,r1 */
1154 #endif
1155 #endif
1156 vtop-=keep;
1157 leaffunc = 0;
1160 /* generate function prolog of type 't' */
1161 void gfunc_prolog(CType *func_type)
1163 Sym *sym,*sym2;
1164 int n,nf,size,align, variadic, struct_ret = 0;
1165 #ifdef TCC_ARM_HARDFLOAT
1166 struct avail_regs avregs = AVAIL_REGS_INITIALIZER;
1167 #endif
1169 sym = func_type->ref;
1170 func_vt = sym->type;
1172 n = nf = 0;
1173 variadic = (func_type->ref->c == FUNC_ELLIPSIS);
1174 if((func_vt.t & VT_BTYPE) == VT_STRUCT
1175 && type_size(&func_vt,&align) > 4)
1177 n++;
1178 struct_ret = 1;
1179 func_vc = 12; /* Offset from fp of the place to store the result */
1181 for(sym2=sym->next;sym2 && (n<4 || nf<16);sym2=sym2->next) {
1182 size = type_size(&sym2->type, &align);
1183 #ifdef TCC_ARM_HARDFLOAT
1184 if (!variadic && (is_float(sym2->type.t)
1185 || is_float_hgen_aggr(&sym2->type))) {
1186 int tmpnf = assign_fpreg(&avregs, align, size) + 1;
1187 nf = (tmpnf > nf) ? tmpnf : nf;
1188 } else
1189 #endif
1190 if (n < 4)
1191 n += (size + 3) / 4;
1193 o(0xE1A0C00D); /* mov ip,sp */
1194 if(variadic)
1195 n=4;
1196 if(n) {
1197 if(n>4)
1198 n=4;
1199 #ifdef TCC_ARM_EABI
1200 n=(n+1)&-2;
1201 #endif
1202 o(0xE92D0000|((1<<n)-1)); /* save r0-r4 on stack if needed */
1204 if (nf) {
1205 if (nf>16)
1206 nf=16;
1207 nf=(nf+1)&-2; /* nf => HARDFLOAT => EABI */
1208 o(0xED2D0A00|nf); /* save s0-s15 on stack if needed */
1210 o(0xE92D5800); /* save fp, ip, lr */
1211 o(0xE1A0B00D); /* mov fp, sp */
1212 func_sub_sp_offset = ind;
1213 o(0xE1A00000); /* nop, leave space for stack adjustment in epilogue */
1215 int addr, pn = struct_ret, sn = 0; /* pn=core, sn=stack */
1217 #ifdef TCC_ARM_HARDFLOAT
1218 func_vc += nf * 4;
1219 avregs = AVAIL_REGS_INITIALIZER;
1220 #endif
1221 while ((sym = sym->next)) {
1222 CType *type;
1223 type = &sym->type;
1224 size = type_size(type, &align);
1225 size = (size + 3) >> 2;
1226 align = (align + 3) & ~3;
1227 #ifdef TCC_ARM_HARDFLOAT
1228 if (!variadic && (is_float(sym->type.t)
1229 || is_float_hgen_aggr(&sym->type))) {
1230 int fpn = assign_fpreg(&avregs, align, size << 2);
1231 if (fpn >= 0) {
1232 addr = fpn * 4;
1233 } else
1234 goto from_stack;
1235 } else
1236 #endif
1237 if (pn < 4) {
1238 #ifdef TCC_ARM_EABI
1239 pn = (pn + (align-1)/4) & -(align/4);
1240 #endif
1241 addr = (nf + pn) * 4;
1242 pn += size;
1243 if (!sn && pn > 4)
1244 sn = (pn - 4);
1245 } else {
1246 #ifdef TCC_ARM_HARDFLOAT
1247 from_stack:
1248 #endif
1249 #ifdef TCC_ARM_EABI
1250 sn = (sn + (align-1)/4) & -(align/4);
1251 #endif
1252 addr = (n + nf + sn) * 4;
1253 sn += size;
1255 sym_push(sym->v & ~SYM_FIELD, type, VT_LOCAL | lvalue_type(type->t), addr+12);
1258 last_itod_magic=0;
1259 leaffunc = 1;
1260 loc = 0;
1263 /* generate function epilog */
1264 void gfunc_epilog(void)
1266 uint32_t x;
1267 int diff;
1268 #ifdef TCC_ARM_EABI
1269 /* Useless but harmless copy of the float result into main register(s) in case
1270 of variadic function in the hardfloat variant */
1271 if(is_float(func_vt.t)) {
1272 if((func_vt.t & VT_BTYPE) == VT_FLOAT)
1273 o(0xEE100A10); /* fmrs r0, s0 */
1274 else {
1275 o(0xEE100B10); /* fmrdl r0, d0 */
1276 o(0xEE301B10); /* fmrdh r1, d0 */
1279 #endif
1280 o(0xE89BA800); /* restore fp, sp, pc */
1281 diff = (-loc + 3) & -4;
1282 #ifdef TCC_ARM_EABI
1283 if(!leaffunc)
1284 diff = ((diff + 11) & -8) - 4;
1285 #endif
1286 if(diff > 0) {
1287 x=stuff_const(0xE24BD000, diff); /* sub sp,fp,# */
1288 if(x)
1289 *(uint32_t *)(cur_text_section->data + func_sub_sp_offset) = x;
1290 else {
1291 int addr;
1292 addr=ind;
1293 o(0xE59FC004); /* ldr ip,[pc+4] */
1294 o(0xE04BD00C); /* sub sp,fp,ip */
1295 o(0xE1A0F00E); /* mov pc,lr */
1296 o(diff);
1297 *(uint32_t *)(cur_text_section->data + func_sub_sp_offset) = 0xE1000000|encbranch(func_sub_sp_offset,addr,1);
1302 /* generate a jump to a label */
1303 int gjmp(int t)
1305 int r;
1306 r=ind;
1307 o(0xE0000000|encbranch(r,t,1));
1308 return r;
1311 /* generate a jump to a fixed address */
1312 void gjmp_addr(int a)
1314 gjmp(a);
1317 /* generate a test. set 'inv' to invert test. Stack entry is popped */
1318 int gtst(int inv, int t)
1320 int v, r;
1321 uint32_t op;
1322 v = vtop->r & VT_VALMASK;
1323 r=ind;
1324 if (v == VT_CMP) {
1325 op=mapcc(inv?negcc(vtop->c.i):vtop->c.i);
1326 op|=encbranch(r,t,1);
1327 o(op);
1328 t=r;
1329 } else if (v == VT_JMP || v == VT_JMPI) {
1330 if ((v & 1) == inv) {
1331 if(!vtop->c.i)
1332 vtop->c.i=t;
1333 else {
1334 uint32_t *x;
1335 int p,lp;
1336 if(t) {
1337 p = vtop->c.i;
1338 do {
1339 p = decbranch(lp=p);
1340 } while(p);
1341 x = (uint32_t *)(cur_text_section->data + lp);
1342 *x &= 0xff000000;
1343 *x |= encbranch(lp,t,1);
1345 t = vtop->c.i;
1347 } else {
1348 t = gjmp(t);
1349 gsym(vtop->c.i);
1351 } else {
1352 if (is_float(vtop->type.t)) {
1353 r=gv(RC_FLOAT);
1354 #ifdef TCC_ARM_VFP
1355 o(0xEEB50A40|(vfpr(r)<<12)|T2CPR(vtop->type.t)); /* fcmpzX */
1356 o(0xEEF1FA10); /* fmstat */
1357 #else
1358 o(0xEE90F118|(fpr(r)<<16));
1359 #endif
1360 vtop->r = VT_CMP;
1361 vtop->c.i = TOK_NE;
1362 return gtst(inv, t);
1363 } else if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
1364 /* constant jmp optimization */
1365 if ((vtop->c.i != 0) != inv)
1366 t = gjmp(t);
1367 } else {
1368 v = gv(RC_INT);
1369 o(0xE3300000|(intr(v)<<16));
1370 vtop->r = VT_CMP;
1371 vtop->c.i = TOK_NE;
1372 return gtst(inv, t);
1375 vtop--;
1376 return t;
1379 /* generate an integer binary operation */
1380 void gen_opi(int op)
1382 int c, func = 0;
1383 uint32_t opc = 0, r, fr;
1384 unsigned short retreg = REG_IRET;
1386 c=0;
1387 switch(op) {
1388 case '+':
1389 opc = 0x8;
1390 c=1;
1391 break;
1392 case TOK_ADDC1: /* add with carry generation */
1393 opc = 0x9;
1394 c=1;
1395 break;
1396 case '-':
1397 opc = 0x4;
1398 c=1;
1399 break;
1400 case TOK_SUBC1: /* sub with carry generation */
1401 opc = 0x5;
1402 c=1;
1403 break;
1404 case TOK_ADDC2: /* add with carry use */
1405 opc = 0xA;
1406 c=1;
1407 break;
1408 case TOK_SUBC2: /* sub with carry use */
1409 opc = 0xC;
1410 c=1;
1411 break;
1412 case '&':
1413 opc = 0x0;
1414 c=1;
1415 break;
1416 case '^':
1417 opc = 0x2;
1418 c=1;
1419 break;
1420 case '|':
1421 opc = 0x18;
1422 c=1;
1423 break;
1424 case '*':
1425 gv2(RC_INT, RC_INT);
1426 r = vtop[-1].r;
1427 fr = vtop[0].r;
1428 vtop--;
1429 o(0xE0000090|(intr(r)<<16)|(intr(r)<<8)|intr(fr));
1430 return;
1431 case TOK_SHL:
1432 opc = 0;
1433 c=2;
1434 break;
1435 case TOK_SHR:
1436 opc = 1;
1437 c=2;
1438 break;
1439 case TOK_SAR:
1440 opc = 2;
1441 c=2;
1442 break;
1443 case '/':
1444 case TOK_PDIV:
1445 func=TOK___divsi3;
1446 c=3;
1447 break;
1448 case TOK_UDIV:
1449 func=TOK___udivsi3;
1450 c=3;
1451 break;
1452 case '%':
1453 #ifdef TCC_ARM_EABI
1454 func=TOK___aeabi_idivmod;
1455 retreg=REG_LRET;
1456 #else
1457 func=TOK___modsi3;
1458 #endif
1459 c=3;
1460 break;
1461 case TOK_UMOD:
1462 #ifdef TCC_ARM_EABI
1463 func=TOK___aeabi_uidivmod;
1464 retreg=REG_LRET;
1465 #else
1466 func=TOK___umodsi3;
1467 #endif
1468 c=3;
1469 break;
1470 case TOK_UMULL:
1471 gv2(RC_INT, RC_INT);
1472 r=intr(vtop[-1].r2=get_reg(RC_INT));
1473 c=vtop[-1].r;
1474 vtop[-1].r=get_reg_ex(RC_INT,regmask(c));
1475 vtop--;
1476 o(0xE0800090|(r<<16)|(intr(vtop->r)<<12)|(intr(c)<<8)|intr(vtop[1].r));
1477 return;
1478 default:
1479 opc = 0x15;
1480 c=1;
1481 break;
1483 switch(c) {
1484 case 1:
1485 if((vtop[-1].r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
1486 if(opc == 4 || opc == 5 || opc == 0xc) {
1487 vswap();
1488 opc|=2; // sub -> rsb
1491 if ((vtop->r & VT_VALMASK) == VT_CMP ||
1492 (vtop->r & (VT_VALMASK & ~1)) == VT_JMP)
1493 gv(RC_INT);
1494 vswap();
1495 c=intr(gv(RC_INT));
1496 vswap();
1497 opc=0xE0000000|(opc<<20)|(c<<16);
1498 if((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
1499 uint32_t x;
1500 x=stuff_const(opc|0x2000000,vtop->c.i);
1501 if(x) {
1502 r=intr(vtop[-1].r=get_reg_ex(RC_INT,regmask(vtop[-1].r)));
1503 o(x|(r<<12));
1504 goto done;
1507 fr=intr(gv(RC_INT));
1508 r=intr(vtop[-1].r=get_reg_ex(RC_INT,two2mask(vtop->r,vtop[-1].r)));
1509 o(opc|(r<<12)|fr);
1510 done:
1511 vtop--;
1512 if (op >= TOK_ULT && op <= TOK_GT) {
1513 vtop->r = VT_CMP;
1514 vtop->c.i = op;
1516 break;
1517 case 2:
1518 opc=0xE1A00000|(opc<<5);
1519 if ((vtop->r & VT_VALMASK) == VT_CMP ||
1520 (vtop->r & (VT_VALMASK & ~1)) == VT_JMP)
1521 gv(RC_INT);
1522 vswap();
1523 r=intr(gv(RC_INT));
1524 vswap();
1525 opc|=r;
1526 if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
1527 fr=intr(vtop[-1].r=get_reg_ex(RC_INT,regmask(vtop[-1].r)));
1528 c = vtop->c.i & 0x1f;
1529 o(opc|(c<<7)|(fr<<12));
1530 } else {
1531 fr=intr(gv(RC_INT));
1532 c=intr(vtop[-1].r=get_reg_ex(RC_INT,two2mask(vtop->r,vtop[-1].r)));
1533 o(opc|(c<<12)|(fr<<8)|0x10);
1535 vtop--;
1536 break;
1537 case 3:
1538 vpush_global_sym(&func_old_type, func);
1539 vrott(3);
1540 gfunc_call(2);
1541 vpushi(0);
1542 vtop->r = retreg;
1543 break;
1544 default:
1545 tcc_error("gen_opi %i unimplemented!",op);
1549 #ifdef TCC_ARM_VFP
1550 static int is_zero(int i)
1552 if((vtop[i].r & (VT_VALMASK | VT_LVAL | VT_SYM)) != VT_CONST)
1553 return 0;
1554 if (vtop[i].type.t == VT_FLOAT)
1555 return (vtop[i].c.f == 0.f);
1556 else if (vtop[i].type.t == VT_DOUBLE)
1557 return (vtop[i].c.d == 0.0);
1558 return (vtop[i].c.ld == 0.l);
1561 /* generate a floating point operation 'v = t1 op t2' instruction. The
1562 * two operands are guaranted to have the same floating point type */
1563 void gen_opf(int op)
1565 uint32_t x;
1566 int fneg=0,r;
1567 x=0xEE000A00|T2CPR(vtop->type.t);
1568 switch(op) {
1569 case '+':
1570 if(is_zero(-1))
1571 vswap();
1572 if(is_zero(0)) {
1573 vtop--;
1574 return;
1576 x|=0x300000;
1577 break;
1578 case '-':
1579 x|=0x300040;
1580 if(is_zero(0)) {
1581 vtop--;
1582 return;
1584 if(is_zero(-1)) {
1585 x|=0x810000; /* fsubX -> fnegX */
1586 vswap();
1587 vtop--;
1588 fneg=1;
1590 break;
1591 case '*':
1592 x|=0x200000;
1593 break;
1594 case '/':
1595 x|=0x800000;
1596 break;
1597 default:
1598 if(op < TOK_ULT || op > TOK_GT) {
1599 tcc_error("unknown fp op %x!",op);
1600 return;
1602 if(is_zero(-1)) {
1603 vswap();
1604 switch(op) {
1605 case TOK_LT: op=TOK_GT; break;
1606 case TOK_GE: op=TOK_ULE; break;
1607 case TOK_LE: op=TOK_GE; break;
1608 case TOK_GT: op=TOK_ULT; break;
1611 x|=0xB40040; /* fcmpX */
1612 if(op!=TOK_EQ && op!=TOK_NE)
1613 x|=0x80; /* fcmpX -> fcmpeX */
1614 if(is_zero(0)) {
1615 vtop--;
1616 o(x|0x10000|(vfpr(gv(RC_FLOAT))<<12)); /* fcmp(e)X -> fcmp(e)zX */
1617 } else {
1618 x|=vfpr(gv(RC_FLOAT));
1619 vswap();
1620 o(x|(vfpr(gv(RC_FLOAT))<<12));
1621 vtop--;
1623 o(0xEEF1FA10); /* fmstat */
1625 switch(op) {
1626 case TOK_LE: op=TOK_ULE; break;
1627 case TOK_LT: op=TOK_ULT; break;
1628 case TOK_UGE: op=TOK_GE; break;
1629 case TOK_UGT: op=TOK_GT; break;
1632 vtop->r = VT_CMP;
1633 vtop->c.i = op;
1634 return;
1636 r=gv(RC_FLOAT);
1637 x|=vfpr(r);
1638 r=regmask(r);
1639 if(!fneg) {
1640 int r2;
1641 vswap();
1642 r2=gv(RC_FLOAT);
1643 x|=vfpr(r2)<<16;
1644 r|=regmask(r2);
1646 vtop->r=get_reg_ex(RC_FLOAT,r);
1647 if(!fneg)
1648 vtop--;
1649 o(x|(vfpr(vtop->r)<<12));
1652 #else
1653 static uint32_t is_fconst()
1655 long double f;
1656 uint32_t r;
1657 if((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) != VT_CONST)
1658 return 0;
1659 if (vtop->type.t == VT_FLOAT)
1660 f = vtop->c.f;
1661 else if (vtop->type.t == VT_DOUBLE)
1662 f = vtop->c.d;
1663 else
1664 f = vtop->c.ld;
1665 if(!ieee_finite(f))
1666 return 0;
1667 r=0x8;
1668 if(f<0.0) {
1669 r=0x18;
1670 f=-f;
1672 if(f==0.0)
1673 return r;
1674 if(f==1.0)
1675 return r|1;
1676 if(f==2.0)
1677 return r|2;
1678 if(f==3.0)
1679 return r|3;
1680 if(f==4.0)
1681 return r|4;
1682 if(f==5.0)
1683 return r|5;
1684 if(f==0.5)
1685 return r|6;
1686 if(f==10.0)
1687 return r|7;
1688 return 0;
1691 /* generate a floating point operation 'v = t1 op t2' instruction. The
1692 two operands are guaranted to have the same floating point type */
1693 void gen_opf(int op)
1695 uint32_t x, r, r2, c1, c2;
1696 //fputs("gen_opf\n",stderr);
1697 vswap();
1698 c1 = is_fconst();
1699 vswap();
1700 c2 = is_fconst();
1701 x=0xEE000100;
1702 #if LDOUBLE_SIZE == 8
1703 if ((vtop->type.t & VT_BTYPE) != VT_FLOAT)
1704 x|=0x80;
1705 #else
1706 if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
1707 x|=0x80;
1708 else if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE)
1709 x|=0x80000;
1710 #endif
1711 switch(op)
1713 case '+':
1714 if(!c2) {
1715 vswap();
1716 c2=c1;
1718 vswap();
1719 r=fpr(gv(RC_FLOAT));
1720 vswap();
1721 if(c2) {
1722 if(c2>0xf)
1723 x|=0x200000; // suf
1724 r2=c2&0xf;
1725 } else {
1726 r2=fpr(gv(RC_FLOAT));
1728 break;
1729 case '-':
1730 if(c2) {
1731 if(c2<=0xf)
1732 x|=0x200000; // suf
1733 r2=c2&0xf;
1734 vswap();
1735 r=fpr(gv(RC_FLOAT));
1736 vswap();
1737 } else if(c1 && c1<=0xf) {
1738 x|=0x300000; // rsf
1739 r2=c1;
1740 r=fpr(gv(RC_FLOAT));
1741 vswap();
1742 } else {
1743 x|=0x200000; // suf
1744 vswap();
1745 r=fpr(gv(RC_FLOAT));
1746 vswap();
1747 r2=fpr(gv(RC_FLOAT));
1749 break;
1750 case '*':
1751 if(!c2 || c2>0xf) {
1752 vswap();
1753 c2=c1;
1755 vswap();
1756 r=fpr(gv(RC_FLOAT));
1757 vswap();
1758 if(c2 && c2<=0xf)
1759 r2=c2;
1760 else
1761 r2=fpr(gv(RC_FLOAT));
1762 x|=0x100000; // muf
1763 break;
1764 case '/':
1765 if(c2 && c2<=0xf) {
1766 x|=0x400000; // dvf
1767 r2=c2;
1768 vswap();
1769 r=fpr(gv(RC_FLOAT));
1770 vswap();
1771 } else if(c1 && c1<=0xf) {
1772 x|=0x500000; // rdf
1773 r2=c1;
1774 r=fpr(gv(RC_FLOAT));
1775 vswap();
1776 } else {
1777 x|=0x400000; // dvf
1778 vswap();
1779 r=fpr(gv(RC_FLOAT));
1780 vswap();
1781 r2=fpr(gv(RC_FLOAT));
1783 break;
1784 default:
1785 if(op >= TOK_ULT && op <= TOK_GT) {
1786 x|=0xd0f110; // cmfe
1787 /* bug (intention?) in Linux FPU emulator
1788 doesn't set carry if equal */
1789 switch(op) {
1790 case TOK_ULT:
1791 case TOK_UGE:
1792 case TOK_ULE:
1793 case TOK_UGT:
1794 tcc_error("unsigned comparision on floats?");
1795 break;
1796 case TOK_LT:
1797 op=TOK_Nset;
1798 break;
1799 case TOK_LE:
1800 op=TOK_ULE; /* correct in unordered case only if AC bit in FPSR set */
1801 break;
1802 case TOK_EQ:
1803 case TOK_NE:
1804 x&=~0x400000; // cmfe -> cmf
1805 break;
1807 if(c1 && !c2) {
1808 c2=c1;
1809 vswap();
1810 switch(op) {
1811 case TOK_Nset:
1812 op=TOK_GT;
1813 break;
1814 case TOK_GE:
1815 op=TOK_ULE;
1816 break;
1817 case TOK_ULE:
1818 op=TOK_GE;
1819 break;
1820 case TOK_GT:
1821 op=TOK_Nset;
1822 break;
1825 vswap();
1826 r=fpr(gv(RC_FLOAT));
1827 vswap();
1828 if(c2) {
1829 if(c2>0xf)
1830 x|=0x200000;
1831 r2=c2&0xf;
1832 } else {
1833 r2=fpr(gv(RC_FLOAT));
1835 vtop[-1].r = VT_CMP;
1836 vtop[-1].c.i = op;
1837 } else {
1838 tcc_error("unknown fp op %x!",op);
1839 return;
1842 if(vtop[-1].r == VT_CMP)
1843 c1=15;
1844 else {
1845 c1=vtop->r;
1846 if(r2&0x8)
1847 c1=vtop[-1].r;
1848 vtop[-1].r=get_reg_ex(RC_FLOAT,two2mask(vtop[-1].r,c1));
1849 c1=fpr(vtop[-1].r);
1851 vtop--;
1852 o(x|(r<<16)|(c1<<12)|r2);
1854 #endif
1856 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
1857 and 'long long' cases. */
1858 ST_FUNC void gen_cvt_itof1(int t)
1860 uint32_t r, r2;
1861 int bt;
1862 bt=vtop->type.t & VT_BTYPE;
1863 if(bt == VT_INT || bt == VT_SHORT || bt == VT_BYTE) {
1864 #ifndef TCC_ARM_VFP
1865 uint32_t dsize = 0;
1866 #endif
1867 r=intr(gv(RC_INT));
1868 #ifdef TCC_ARM_VFP
1869 r2=vfpr(vtop->r=get_reg(RC_FLOAT));
1870 o(0xEE000A10|(r<<12)|(r2<<16)); /* fmsr */
1871 r2|=r2<<12;
1872 if(!(vtop->type.t & VT_UNSIGNED))
1873 r2|=0x80; /* fuitoX -> fsituX */
1874 o(0xEEB80A40|r2|T2CPR(t)); /* fYitoX*/
1875 #else
1876 r2=fpr(vtop->r=get_reg(RC_FLOAT));
1877 if((t & VT_BTYPE) != VT_FLOAT)
1878 dsize=0x80; /* flts -> fltd */
1879 o(0xEE000110|dsize|(r2<<16)|(r<<12)); /* flts */
1880 if((vtop->type.t & (VT_UNSIGNED|VT_BTYPE)) == (VT_UNSIGNED|VT_INT)) {
1881 uint32_t off = 0;
1882 o(0xE3500000|(r<<12)); /* cmp */
1883 r=fpr(get_reg(RC_FLOAT));
1884 if(last_itod_magic) {
1885 off=ind+8-last_itod_magic;
1886 off/=4;
1887 if(off>255)
1888 off=0;
1890 o(0xBD1F0100|(r<<12)|off); /* ldflts */
1891 if(!off) {
1892 o(0xEA000000); /* b */
1893 last_itod_magic=ind;
1894 o(0x4F800000); /* 4294967296.0f */
1896 o(0xBE000100|dsize|(r2<<16)|(r2<<12)|r); /* adflt */
1898 #endif
1899 return;
1900 } else if(bt == VT_LLONG) {
1901 int func;
1902 CType *func_type = 0;
1903 if((t & VT_BTYPE) == VT_FLOAT) {
1904 func_type = &func_float_type;
1905 if(vtop->type.t & VT_UNSIGNED)
1906 func=TOK___floatundisf;
1907 else
1908 func=TOK___floatdisf;
1909 #if LDOUBLE_SIZE != 8
1910 } else if((t & VT_BTYPE) == VT_LDOUBLE) {
1911 func_type = &func_ldouble_type;
1912 if(vtop->type.t & VT_UNSIGNED)
1913 func=TOK___floatundixf;
1914 else
1915 func=TOK___floatdixf;
1916 } else if((t & VT_BTYPE) == VT_DOUBLE) {
1917 #else
1918 } else if((t & VT_BTYPE) == VT_DOUBLE || (t & VT_BTYPE) == VT_LDOUBLE) {
1919 #endif
1920 func_type = &func_double_type;
1921 if(vtop->type.t & VT_UNSIGNED)
1922 func=TOK___floatundidf;
1923 else
1924 func=TOK___floatdidf;
1926 if(func_type) {
1927 vpush_global_sym(func_type, func);
1928 vswap();
1929 gfunc_call(1);
1930 vpushi(0);
1931 vtop->r=TREG_F0;
1932 return;
1935 tcc_error("unimplemented gen_cvt_itof %x!",vtop->type.t);
1938 /* convert fp to int 't' type */
1939 void gen_cvt_ftoi(int t)
1941 uint32_t r, r2;
1942 int u, func = 0;
1943 u=t&VT_UNSIGNED;
1944 t&=VT_BTYPE;
1945 r2=vtop->type.t & VT_BTYPE;
1946 if(t==VT_INT) {
1947 #ifdef TCC_ARM_VFP
1948 r=vfpr(gv(RC_FLOAT));
1949 u=u?0:0x10000;
1950 o(0xEEBC0AC0|(r<<12)|r|T2CPR(r2)|u); /* ftoXizY */
1951 r2=intr(vtop->r=get_reg(RC_INT));
1952 o(0xEE100A10|(r<<16)|(r2<<12));
1953 return;
1954 #else
1955 if(u) {
1956 if(r2 == VT_FLOAT)
1957 func=TOK___fixunssfsi;
1958 #if LDOUBLE_SIZE != 8
1959 else if(r2 == VT_LDOUBLE)
1960 func=TOK___fixunsxfsi;
1961 else if(r2 == VT_DOUBLE)
1962 #else
1963 else if(r2 == VT_LDOUBLE || r2 == VT_DOUBLE)
1964 #endif
1965 func=TOK___fixunsdfsi;
1966 } else {
1967 r=fpr(gv(RC_FLOAT));
1968 r2=intr(vtop->r=get_reg(RC_INT));
1969 o(0xEE100170|(r2<<12)|r);
1970 return;
1972 #endif
1973 } else if(t == VT_LLONG) { // unsigned handled in gen_cvt_ftoi1
1974 if(r2 == VT_FLOAT)
1975 func=TOK___fixsfdi;
1976 #if LDOUBLE_SIZE != 8
1977 else if(r2 == VT_LDOUBLE)
1978 func=TOK___fixxfdi;
1979 else if(r2 == VT_DOUBLE)
1980 #else
1981 else if(r2 == VT_LDOUBLE || r2 == VT_DOUBLE)
1982 #endif
1983 func=TOK___fixdfdi;
1985 if(func) {
1986 vpush_global_sym(&func_old_type, func);
1987 vswap();
1988 gfunc_call(1);
1989 vpushi(0);
1990 if(t == VT_LLONG)
1991 vtop->r2 = REG_LRET;
1992 vtop->r = REG_IRET;
1993 return;
1995 tcc_error("unimplemented gen_cvt_ftoi!");
1998 /* convert from one floating point type to another */
1999 void gen_cvt_ftof(int t)
2001 #ifdef TCC_ARM_VFP
2002 if(((vtop->type.t & VT_BTYPE) == VT_FLOAT) != ((t & VT_BTYPE) == VT_FLOAT)) {
2003 uint32_t r = vfpr(gv(RC_FLOAT));
2004 o(0xEEB70AC0|(r<<12)|r|T2CPR(vtop->type.t));
2006 #else
2007 /* all we have to do on i386 and FPA ARM is to put the float in a register */
2008 gv(RC_FLOAT);
2009 #endif
2012 /* computed goto support */
2013 void ggoto(void)
2015 gcall_or_jmp(1);
2016 vtop--;
2019 /* Save the stack pointer onto the stack and return the location of its address */
2020 ST_FUNC void gen_vla_sp_save(int addr) {
2021 tcc_error("variable length arrays unsupported for this target");
2024 /* Restore the SP from a location on the stack */
2025 ST_FUNC void gen_vla_sp_restore(int addr) {
2026 tcc_error("variable length arrays unsupported for this target");
2029 /* Subtract from the stack pointer, and push the resulting value onto the stack */
2030 ST_FUNC void gen_vla_alloc(CType *type, int align) {
2031 tcc_error("variable length arrays unsupported for this target");
2034 /* end of ARM code generator */
2035 /*************************************************************/
2036 #endif
2037 /*************************************************************/