Fix type_to_str test for unsigned int
[tinycc.git] / arm-gen.c
blob372c468c01c13b334f8775f53864ce43d3a328dd
1 /*
2 * ARMv4 code generator for TCC
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 #if defined(TCC_ARM_EABI) && !defined(TCC_ARM_VFP)
27 #error "Currently TinyCC only supports float computation with VFP instructions"
28 #endif
30 /* number of available registers */
31 #ifdef TCC_ARM_VFP
32 #define NB_REGS 13
33 #else
34 #define NB_REGS 9
35 #endif
37 #ifndef TCC_ARM_VERSION
38 # define TCC_ARM_VERSION 5
39 #endif
41 /* a register can belong to several classes. The classes must be
42 sorted from more general to more precise (see gv2() code which does
43 assumptions on it). */
44 #define RC_INT 0x0001 /* generic integer register */
45 #define RC_FLOAT 0x0002 /* generic float register */
46 #define RC_R0 0x0004
47 #define RC_R1 0x0008
48 #define RC_R2 0x0010
49 #define RC_R3 0x0020
50 #define RC_R12 0x0040
51 #define RC_F0 0x0080
52 #define RC_F1 0x0100
53 #define RC_F2 0x0200
54 #define RC_F3 0x0400
55 #ifdef TCC_ARM_VFP
56 #define RC_F4 0x0800
57 #define RC_F5 0x1000
58 #define RC_F6 0x2000
59 #define RC_F7 0x4000
60 #endif
61 #define RC_IRET RC_R0 /* function return: integer register */
62 #define RC_LRET RC_R1 /* function return: second integer register */
63 #define RC_FRET RC_F0 /* function return: float register */
65 /* pretty names for the registers */
66 enum {
67 TREG_R0 = 0,
68 TREG_R1,
69 TREG_R2,
70 TREG_R3,
71 TREG_R12,
72 TREG_F0,
73 TREG_F1,
74 TREG_F2,
75 TREG_F3,
76 #ifdef TCC_ARM_VFP
77 TREG_F4,
78 TREG_F5,
79 TREG_F6,
80 TREG_F7,
81 #endif
84 #ifdef TCC_ARM_VFP
85 #define T2CPR(t) (((t) & VT_BTYPE) != VT_FLOAT ? 0x100 : 0)
86 #endif
88 /* return registers for function */
89 #define REG_IRET TREG_R0 /* single word int return register */
90 #define REG_LRET TREG_R1 /* second word return register (for long long) */
91 #define REG_FRET TREG_F0 /* float return register */
93 #ifdef TCC_ARM_EABI
94 #define TOK___divdi3 TOK___aeabi_ldivmod
95 #define TOK___moddi3 TOK___aeabi_ldivmod
96 #define TOK___udivdi3 TOK___aeabi_uldivmod
97 #define TOK___umoddi3 TOK___aeabi_uldivmod
98 #endif
100 /* defined if function parameters must be evaluated in reverse order */
101 #define INVERT_FUNC_PARAMS
103 /* defined if structures are passed as pointers. Otherwise structures
104 are directly pushed on stack. */
105 /* #define FUNC_STRUCT_PARAM_AS_PTR */
107 /* pointer size, in bytes */
108 #define PTR_SIZE 4
110 /* long double size and alignment, in bytes */
111 #ifdef TCC_ARM_VFP
112 #define LDOUBLE_SIZE 8
113 #endif
115 #ifndef LDOUBLE_SIZE
116 #define LDOUBLE_SIZE 8
117 #endif
119 #ifdef TCC_ARM_EABI
120 #define LDOUBLE_ALIGN 8
121 #else
122 #define LDOUBLE_ALIGN 4
123 #endif
125 /* maximum alignment (for aligned attribute support) */
126 #define MAX_ALIGN 8
128 #define CHAR_IS_UNSIGNED
130 /******************************************************/
131 /* ELF defines */
133 #define EM_TCC_TARGET EM_ARM
135 /* relocation type for 32 bit data relocation */
136 #define R_DATA_32 R_ARM_ABS32
137 #define R_DATA_PTR R_ARM_ABS32
138 #define R_JMP_SLOT R_ARM_JUMP_SLOT
139 #define R_COPY R_ARM_COPY
141 #define ELF_START_ADDR 0x00008000
142 #define ELF_PAGE_SIZE 0x1000
144 enum float_abi {
145 ARM_SOFTFP_FLOAT,
146 ARM_HARD_FLOAT,
149 /******************************************************/
150 #else /* ! TARGET_DEFS_ONLY */
151 /******************************************************/
152 #include "tcc.h"
154 enum float_abi float_abi;
156 ST_DATA const int reg_classes[NB_REGS] = {
157 /* r0 */ RC_INT | RC_R0,
158 /* r1 */ RC_INT | RC_R1,
159 /* r2 */ RC_INT | RC_R2,
160 /* r3 */ RC_INT | RC_R3,
161 /* r12 */ RC_INT | RC_R12,
162 /* f0 */ RC_FLOAT | RC_F0,
163 /* f1 */ RC_FLOAT | RC_F1,
164 /* f2 */ RC_FLOAT | RC_F2,
165 /* f3 */ RC_FLOAT | RC_F3,
166 #ifdef TCC_ARM_VFP
167 /* d4/s8 */ RC_FLOAT | RC_F4,
168 /* d5/s10 */ RC_FLOAT | RC_F5,
169 /* d6/s12 */ RC_FLOAT | RC_F6,
170 /* d7/s14 */ RC_FLOAT | RC_F7,
171 #endif
174 static int func_sub_sp_offset, last_itod_magic;
175 static int leaffunc;
177 #if defined(TCC_ARM_EABI) && defined(TCC_ARM_VFP)
178 static CType float_type, double_type, func_float_type, func_double_type;
179 ST_FUNC void arm_init(struct TCCState *s)
181 float_type.t = VT_FLOAT;
182 double_type.t = VT_DOUBLE;
183 func_float_type.t = VT_FUNC;
184 func_float_type.ref = sym_push(SYM_FIELD, &float_type, FUNC_CDECL, FUNC_OLD);
185 func_double_type.t = VT_FUNC;
186 func_double_type.ref = sym_push(SYM_FIELD, &double_type, FUNC_CDECL, FUNC_OLD);
188 float_abi = s->float_abi;
190 #else
191 #define func_float_type func_old_type
192 #define func_double_type func_old_type
193 #define func_ldouble_type func_old_type
194 ST_FUNC void arm_init(struct TCCState *s) {}
195 #endif
197 static int two2mask(int a,int b) {
198 return (reg_classes[a]|reg_classes[b])&~(RC_INT|RC_FLOAT);
201 static int regmask(int r) {
202 return reg_classes[r]&~(RC_INT|RC_FLOAT);
205 /******************************************************/
207 #ifdef TCC_ARM_EABI
208 char *default_elfinterp(struct TCCState *s)
210 if (s->float_abi == ARM_HARD_FLOAT)
211 return "/lib/ld-linux-armhf.so.3";
212 else
213 return "/lib/ld-linux.so.3";
215 #endif
217 void o(uint32_t i)
219 /* this is a good place to start adding big-endian support*/
220 int ind1;
222 ind1 = ind + 4;
223 if (!cur_text_section)
224 tcc_error("compiler error! This happens f.ex. if the compiler\n"
225 "can't evaluate constant expressions outside of a function.");
226 if (ind1 > cur_text_section->data_allocated)
227 section_realloc(cur_text_section, ind1);
228 cur_text_section->data[ind++] = i&255;
229 i>>=8;
230 cur_text_section->data[ind++] = i&255;
231 i>>=8;
232 cur_text_section->data[ind++] = i&255;
233 i>>=8;
234 cur_text_section->data[ind++] = i;
237 static uint32_t stuff_const(uint32_t op, uint32_t c)
239 int try_neg=0;
240 uint32_t nc = 0, negop = 0;
242 switch(op&0x1F00000)
244 case 0x800000: //add
245 case 0x400000: //sub
246 try_neg=1;
247 negop=op^0xC00000;
248 nc=-c;
249 break;
250 case 0x1A00000: //mov
251 case 0x1E00000: //mvn
252 try_neg=1;
253 negop=op^0x400000;
254 nc=~c;
255 break;
256 case 0x200000: //xor
257 if(c==~0)
258 return (op&0xF010F000)|((op>>16)&0xF)|0x1E00000;
259 break;
260 case 0x0: //and
261 if(c==~0)
262 return (op&0xF010F000)|((op>>16)&0xF)|0x1A00000;
263 case 0x1C00000: //bic
264 try_neg=1;
265 negop=op^0x1C00000;
266 nc=~c;
267 break;
268 case 0x1800000: //orr
269 if(c==~0)
270 return (op&0xFFF0FFFF)|0x1E00000;
271 break;
273 do {
274 uint32_t m;
275 int i;
276 if(c<256) /* catch undefined <<32 */
277 return op|c;
278 for(i=2;i<32;i+=2) {
279 m=(0xff>>i)|(0xff<<(32-i));
280 if(!(c&~m))
281 return op|(i<<7)|(c<<i)|(c>>(32-i));
283 op=negop;
284 c=nc;
285 } while(try_neg--);
286 return 0;
290 //only add,sub
291 void stuff_const_harder(uint32_t op, uint32_t v) {
292 uint32_t x;
293 x=stuff_const(op,v);
294 if(x)
295 o(x);
296 else {
297 uint32_t a[16], nv, no, o2, n2;
298 int i,j,k;
299 a[0]=0xff;
300 o2=(op&0xfff0ffff)|((op&0xf000)<<4);;
301 for(i=1;i<16;i++)
302 a[i]=(a[i-1]>>2)|(a[i-1]<<30);
303 for(i=0;i<12;i++)
304 for(j=i<4?i+12:15;j>=i+4;j--)
305 if((v&(a[i]|a[j]))==v) {
306 o(stuff_const(op,v&a[i]));
307 o(stuff_const(o2,v&a[j]));
308 return;
310 no=op^0xC00000;
311 n2=o2^0xC00000;
312 nv=-v;
313 for(i=0;i<12;i++)
314 for(j=i<4?i+12:15;j>=i+4;j--)
315 if((nv&(a[i]|a[j]))==nv) {
316 o(stuff_const(no,nv&a[i]));
317 o(stuff_const(n2,nv&a[j]));
318 return;
320 for(i=0;i<8;i++)
321 for(j=i+4;j<12;j++)
322 for(k=i<4?i+12:15;k>=j+4;k--)
323 if((v&(a[i]|a[j]|a[k]))==v) {
324 o(stuff_const(op,v&a[i]));
325 o(stuff_const(o2,v&a[j]));
326 o(stuff_const(o2,v&a[k]));
327 return;
329 no=op^0xC00000;
330 nv=-v;
331 for(i=0;i<8;i++)
332 for(j=i+4;j<12;j++)
333 for(k=i<4?i+12:15;k>=j+4;k--)
334 if((nv&(a[i]|a[j]|a[k]))==nv) {
335 o(stuff_const(no,nv&a[i]));
336 o(stuff_const(n2,nv&a[j]));
337 o(stuff_const(n2,nv&a[k]));
338 return;
340 o(stuff_const(op,v&a[0]));
341 o(stuff_const(o2,v&a[4]));
342 o(stuff_const(o2,v&a[8]));
343 o(stuff_const(o2,v&a[12]));
347 ST_FUNC uint32_t encbranch(int pos, int addr, int fail)
349 addr-=pos+8;
350 addr/=4;
351 if(addr>=0x1000000 || addr<-0x1000000) {
352 if(fail)
353 tcc_error("FIXME: function bigger than 32MB");
354 return 0;
356 return 0x0A000000|(addr&0xffffff);
359 int decbranch(int pos)
361 int x;
362 x=*(uint32_t *)(cur_text_section->data + pos);
363 x&=0x00ffffff;
364 if(x&0x800000)
365 x-=0x1000000;
366 return x*4+pos+8;
369 /* output a symbol and patch all calls to it */
370 void gsym_addr(int t, int a)
372 uint32_t *x;
373 int lt;
374 while(t) {
375 x=(uint32_t *)(cur_text_section->data + t);
376 t=decbranch(lt=t);
377 if(a==lt+4)
378 *x=0xE1A00000; // nop
379 else {
380 *x &= 0xff000000;
381 *x |= encbranch(lt,a,1);
386 void gsym(int t)
388 gsym_addr(t, ind);
391 #ifdef TCC_ARM_VFP
392 static uint32_t vfpr(int r)
394 if(r<TREG_F0 || r>TREG_F7)
395 tcc_error("compiler error! register %i is no vfp register",r);
396 return r-5;
398 #else
399 static uint32_t fpr(int r)
401 if(r<TREG_F0 || r>TREG_F3)
402 tcc_error("compiler error! register %i is no fpa register",r);
403 return r-5;
405 #endif
407 static uint32_t intr(int r)
409 if(r==4)
410 return 12;
411 if((r<0 || r>4) && r!=14)
412 tcc_error("compiler error! register %i is no int register",r);
413 return r;
416 static void calcaddr(uint32_t *base, int *off, int *sgn, int maxoff, unsigned shift)
418 if(*off>maxoff || *off&((1<<shift)-1)) {
419 uint32_t x, y;
420 x=0xE280E000;
421 if(*sgn)
422 x=0xE240E000;
423 x|=(*base)<<16;
424 *base=14; // lr
425 y=stuff_const(x,*off&~maxoff);
426 if(y) {
427 o(y);
428 *off&=maxoff;
429 return;
431 y=stuff_const(x,(*off+maxoff)&~maxoff);
432 if(y) {
433 o(y);
434 *sgn=!*sgn;
435 *off=((*off+maxoff)&~maxoff)-*off;
436 return;
438 stuff_const_harder(x,*off&~maxoff);
439 *off&=maxoff;
443 static uint32_t mapcc(int cc)
445 switch(cc)
447 case TOK_ULT:
448 return 0x30000000; /* CC/LO */
449 case TOK_UGE:
450 return 0x20000000; /* CS/HS */
451 case TOK_EQ:
452 return 0x00000000; /* EQ */
453 case TOK_NE:
454 return 0x10000000; /* NE */
455 case TOK_ULE:
456 return 0x90000000; /* LS */
457 case TOK_UGT:
458 return 0x80000000; /* HI */
459 case TOK_Nset:
460 return 0x40000000; /* MI */
461 case TOK_Nclear:
462 return 0x50000000; /* PL */
463 case TOK_LT:
464 return 0xB0000000; /* LT */
465 case TOK_GE:
466 return 0xA0000000; /* GE */
467 case TOK_LE:
468 return 0xD0000000; /* LE */
469 case TOK_GT:
470 return 0xC0000000; /* GT */
472 tcc_error("unexpected condition code");
473 return 0xE0000000; /* AL */
476 static int negcc(int cc)
478 switch(cc)
480 case TOK_ULT:
481 return TOK_UGE;
482 case TOK_UGE:
483 return TOK_ULT;
484 case TOK_EQ:
485 return TOK_NE;
486 case TOK_NE:
487 return TOK_EQ;
488 case TOK_ULE:
489 return TOK_UGT;
490 case TOK_UGT:
491 return TOK_ULE;
492 case TOK_Nset:
493 return TOK_Nclear;
494 case TOK_Nclear:
495 return TOK_Nset;
496 case TOK_LT:
497 return TOK_GE;
498 case TOK_GE:
499 return TOK_LT;
500 case TOK_LE:
501 return TOK_GT;
502 case TOK_GT:
503 return TOK_LE;
505 tcc_error("unexpected condition code");
506 return TOK_NE;
509 /* load 'r' from value 'sv' */
510 void load(int r, SValue *sv)
512 int v, ft, fc, fr, sign;
513 uint32_t op;
514 SValue v1;
516 fr = sv->r;
517 ft = sv->type.t;
518 fc = sv->c.ul;
520 if(fc>=0)
521 sign=0;
522 else {
523 sign=1;
524 fc=-fc;
527 v = fr & VT_VALMASK;
528 if (fr & VT_LVAL) {
529 uint32_t base = 0xB; // fp
530 if(v == VT_LLOCAL) {
531 v1.type.t = VT_PTR;
532 v1.r = VT_LOCAL | VT_LVAL;
533 v1.c.ul = sv->c.ul;
534 load(base=14 /* lr */, &v1);
535 fc=sign=0;
536 v=VT_LOCAL;
537 } else if(v == VT_CONST) {
538 v1.type.t = VT_PTR;
539 v1.r = fr&~VT_LVAL;
540 v1.c.ul = sv->c.ul;
541 v1.sym=sv->sym;
542 load(base=14, &v1);
543 fc=sign=0;
544 v=VT_LOCAL;
545 } else if(v < VT_CONST) {
546 base=intr(v);
547 fc=sign=0;
548 v=VT_LOCAL;
550 if(v == VT_LOCAL) {
551 if(is_float(ft)) {
552 calcaddr(&base,&fc,&sign,1020,2);
553 #ifdef TCC_ARM_VFP
554 op=0xED100A00; /* flds */
555 if(!sign)
556 op|=0x800000;
557 if ((ft & VT_BTYPE) != VT_FLOAT)
558 op|=0x100; /* flds -> fldd */
559 o(op|(vfpr(r)<<12)|(fc>>2)|(base<<16));
560 #else
561 op=0xED100100;
562 if(!sign)
563 op|=0x800000;
564 #if LDOUBLE_SIZE == 8
565 if ((ft & VT_BTYPE) != VT_FLOAT)
566 op|=0x8000;
567 #else
568 if ((ft & VT_BTYPE) == VT_DOUBLE)
569 op|=0x8000;
570 else if ((ft & VT_BTYPE) == VT_LDOUBLE)
571 op|=0x400000;
572 #endif
573 o(op|(fpr(r)<<12)|(fc>>2)|(base<<16));
574 #endif
575 } else if((ft & (VT_BTYPE|VT_UNSIGNED)) == VT_BYTE
576 || (ft & VT_BTYPE) == VT_SHORT) {
577 calcaddr(&base,&fc,&sign,255,0);
578 op=0xE1500090;
579 if ((ft & VT_BTYPE) == VT_SHORT)
580 op|=0x20;
581 if ((ft & VT_UNSIGNED) == 0)
582 op|=0x40;
583 if(!sign)
584 op|=0x800000;
585 o(op|(intr(r)<<12)|(base<<16)|((fc&0xf0)<<4)|(fc&0xf));
586 } else {
587 calcaddr(&base,&fc,&sign,4095,0);
588 op=0xE5100000;
589 if(!sign)
590 op|=0x800000;
591 if ((ft & VT_BTYPE) == VT_BYTE || (ft & VT_BTYPE) == VT_BOOL)
592 op|=0x400000;
593 o(op|(intr(r)<<12)|fc|(base<<16));
595 return;
597 } else {
598 if (v == VT_CONST) {
599 op=stuff_const(0xE3A00000|(intr(r)<<12),sv->c.ul);
600 if (fr & VT_SYM || !op) {
601 o(0xE59F0000|(intr(r)<<12));
602 o(0xEA000000);
603 if(fr & VT_SYM)
604 greloc(cur_text_section, sv->sym, ind, R_ARM_ABS32);
605 o(sv->c.ul);
606 } else
607 o(op);
608 return;
609 } else if (v == VT_LOCAL) {
610 op=stuff_const(0xE28B0000|(intr(r)<<12),sv->c.ul);
611 if (fr & VT_SYM || !op) {
612 o(0xE59F0000|(intr(r)<<12));
613 o(0xEA000000);
614 if(fr & VT_SYM) // needed ?
615 greloc(cur_text_section, sv->sym, ind, R_ARM_ABS32);
616 o(sv->c.ul);
617 o(0xE08B0000|(intr(r)<<12)|intr(r));
618 } else
619 o(op);
620 return;
621 } else if(v == VT_CMP) {
622 o(mapcc(sv->c.ul)|0x3A00001|(intr(r)<<12));
623 o(mapcc(negcc(sv->c.ul))|0x3A00000|(intr(r)<<12));
624 return;
625 } else if (v == VT_JMP || v == VT_JMPI) {
626 int t;
627 t = v & 1;
628 o(0xE3A00000|(intr(r)<<12)|t);
629 o(0xEA000000);
630 gsym(sv->c.ul);
631 o(0xE3A00000|(intr(r)<<12)|(t^1));
632 return;
633 } else if (v < VT_CONST) {
634 if(is_float(ft))
635 #ifdef TCC_ARM_VFP
636 o(0xEEB00A40|(vfpr(r)<<12)|vfpr(v)|T2CPR(ft)); /* fcpyX */
637 #else
638 o(0xEE008180|(fpr(r)<<12)|fpr(v));
639 #endif
640 else
641 o(0xE1A00000|(intr(r)<<12)|intr(v));
642 return;
645 tcc_error("load unimplemented!");
648 /* store register 'r' in lvalue 'v' */
649 void store(int r, SValue *sv)
651 SValue v1;
652 int v, ft, fc, fr, sign;
653 uint32_t op;
655 fr = sv->r;
656 ft = sv->type.t;
657 fc = sv->c.ul;
659 if(fc>=0)
660 sign=0;
661 else {
662 sign=1;
663 fc=-fc;
666 v = fr & VT_VALMASK;
667 if (fr & VT_LVAL || fr == VT_LOCAL) {
668 uint32_t base = 0xb;
669 if(v < VT_CONST) {
670 base=intr(v);
671 v=VT_LOCAL;
672 fc=sign=0;
673 } else if(v == VT_CONST) {
674 v1.type.t = ft;
675 v1.r = fr&~VT_LVAL;
676 v1.c.ul = sv->c.ul;
677 v1.sym=sv->sym;
678 load(base=14, &v1);
679 fc=sign=0;
680 v=VT_LOCAL;
682 if(v == VT_LOCAL) {
683 if(is_float(ft)) {
684 calcaddr(&base,&fc,&sign,1020,2);
685 #ifdef TCC_ARM_VFP
686 op=0xED000A00; /* fsts */
687 if(!sign)
688 op|=0x800000;
689 if ((ft & VT_BTYPE) != VT_FLOAT)
690 op|=0x100; /* fsts -> fstd */
691 o(op|(vfpr(r)<<12)|(fc>>2)|(base<<16));
692 #else
693 op=0xED000100;
694 if(!sign)
695 op|=0x800000;
696 #if LDOUBLE_SIZE == 8
697 if ((ft & VT_BTYPE) != VT_FLOAT)
698 op|=0x8000;
699 #else
700 if ((ft & VT_BTYPE) == VT_DOUBLE)
701 op|=0x8000;
702 if ((ft & VT_BTYPE) == VT_LDOUBLE)
703 op|=0x400000;
704 #endif
705 o(op|(fpr(r)<<12)|(fc>>2)|(base<<16));
706 #endif
707 return;
708 } else if((ft & VT_BTYPE) == VT_SHORT) {
709 calcaddr(&base,&fc,&sign,255,0);
710 op=0xE14000B0;
711 if(!sign)
712 op|=0x800000;
713 o(op|(intr(r)<<12)|(base<<16)|((fc&0xf0)<<4)|(fc&0xf));
714 } else {
715 calcaddr(&base,&fc,&sign,4095,0);
716 op=0xE5000000;
717 if(!sign)
718 op|=0x800000;
719 if ((ft & VT_BTYPE) == VT_BYTE || (ft & VT_BTYPE) == VT_BOOL)
720 op|=0x400000;
721 o(op|(intr(r)<<12)|fc|(base<<16));
723 return;
726 tcc_error("store unimplemented");
729 static void gadd_sp(int val)
731 stuff_const_harder(0xE28DD000,val);
734 /* 'is_jmp' is '1' if it is a jump */
735 static void gcall_or_jmp(int is_jmp)
737 int r;
738 if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
739 uint32_t x;
740 /* constant case */
741 x=encbranch(ind,ind+vtop->c.ul,0);
742 if(x) {
743 if (vtop->r & VT_SYM) {
744 /* relocation case */
745 greloc(cur_text_section, vtop->sym, ind, R_ARM_PC24);
746 } else
747 put_elf_reloc(symtab_section, cur_text_section, ind, R_ARM_PC24, 0);
748 o(x|(is_jmp?0xE0000000:0xE1000000));
749 } else {
750 if(!is_jmp)
751 o(0xE28FE004); // add lr,pc,#4
752 o(0xE51FF004); // ldr pc,[pc,#-4]
753 if (vtop->r & VT_SYM)
754 greloc(cur_text_section, vtop->sym, ind, R_ARM_ABS32);
755 o(vtop->c.ul);
757 } else {
758 /* otherwise, indirect call */
759 r = gv(RC_INT);
760 if(!is_jmp)
761 o(0xE1A0E00F); // mov lr,pc
762 o(0xE1A0F000|intr(r)); // mov pc,r
766 /* Return whether a structure is an homogeneous float aggregate or not.
767 The answer is true if all the elements of the structure are of the same
768 primitive float type and there is less than 4 elements.
770 type: the type corresponding to the structure to be tested */
771 static int is_hgen_float_aggr(CType *type)
773 if ((type->t & VT_BTYPE) == VT_STRUCT) {
774 struct Sym *ref;
775 int btype, nb_fields = 0;
777 ref = type->ref->next;
778 btype = ref->type.t & VT_BTYPE;
779 if (btype == VT_FLOAT || btype == VT_DOUBLE) {
780 for(; ref && btype == (ref->type.t & VT_BTYPE); ref = ref->next, nb_fields++);
781 return !ref && nb_fields <= 4;
784 return 0;
787 struct avail_regs {
788 signed char avail[3]; /* 3 holes max with only float and double alignments */
789 int first_hole; /* first available hole */
790 int last_hole; /* last available hole (none if equal to first_hole) */
791 int first_free_reg; /* next free register in the sequence, hole excluded */
794 #define AVAIL_REGS_INITIALIZER (struct avail_regs) { { 0, 0, 0}, 0, 0, 0 }
796 /* Find suitable registers for a VFP Co-Processor Register Candidate (VFP CPRC
797 param) according to the rules described in the procedure call standard for
798 the ARM architecture (AAPCS). If found, the registers are assigned to this
799 VFP CPRC parameter. Registers are allocated in sequence unless a hole exists
800 and the parameter is a single float.
802 avregs: opaque structure to keep track of available VFP co-processor regs
803 align: alignment contraints for the param, as returned by type_size()
804 size: size of the parameter, as returned by type_size() */
805 int assign_vfpreg(struct avail_regs *avregs, int align, int size)
807 int first_reg = 0;
809 if (avregs->first_free_reg == -1)
810 return -1;
811 if (align >> 3) { /* double alignment */
812 first_reg = avregs->first_free_reg;
813 /* alignment contraint not respected so use next reg and record hole */
814 if (first_reg & 1)
815 avregs->avail[avregs->last_hole++] = first_reg++;
816 } else { /* no special alignment (float or array of float) */
817 /* if single float and a hole is available, assign the param to it */
818 if (size == 4 && avregs->first_hole != avregs->last_hole)
819 return avregs->avail[avregs->first_hole++];
820 else
821 first_reg = avregs->first_free_reg;
823 if (first_reg + size / 4 <= 16) {
824 avregs->first_free_reg = first_reg + size / 4;
825 return first_reg;
827 avregs->first_free_reg = -1;
828 return -1;
831 /* Returns whether all params need to be passed in core registers or not.
832 This is the case for function part of the runtime ABI. */
833 int floats_in_core_regs(SValue *sval)
835 if (!sval->sym)
836 return 0;
838 switch (sval->sym->v) {
839 case TOK___floatundisf:
840 case TOK___floatundidf:
841 case TOK___fixunssfdi:
842 case TOK___fixunsdfdi:
843 #ifndef TCC_ARM_VFP
844 case TOK___fixunsxfdi:
845 #endif
846 case TOK___floatdisf:
847 case TOK___floatdidf:
848 case TOK___fixsfdi:
849 case TOK___fixdfdi:
850 return 1;
852 default:
853 return 0;
857 /* Return the number of registers needed to return the struct, or 0 if
858 returning via struct pointer. */
859 ST_FUNC int gfunc_sret(CType *vt, int variadic, CType *ret, int *ret_align) {
860 #ifdef TCC_ARM_EABI
861 int size, align;
862 size = type_size(vt, &align);
863 if (float_abi == ARM_HARD_FLOAT && !variadic &&
864 (is_float(vt->t) || is_hgen_float_aggr(vt))) {
865 *ret_align = 8;
866 ret->ref = NULL;
867 ret->t = VT_DOUBLE;
868 return (size + 7) >> 3;
869 } else if (size <= 4) {
870 *ret_align = 4;
871 ret->ref = NULL;
872 ret->t = VT_INT;
873 return 1;
874 } else
875 return 0;
876 #else
877 return 0;
878 #endif
881 /* Parameters are classified according to how they are copied to their final
882 destination for the function call. Because the copying is performed class
883 after class according to the order in the union below, it is important that
884 some constraints about the order of the members of this union are respected:
885 - CORE_STRUCT_CLASS must come after STACK_CLASS;
886 - CORE_CLASS must come after STACK_CLASS, CORE_STRUCT_CLASS and
887 VFP_STRUCT_CLASS;
888 - VFP_STRUCT_CLASS must come after VFP_CLASS.
889 See the comment for the main loop in copy_params() for the reason. */
890 enum reg_class {
891 STACK_CLASS = 0,
892 CORE_STRUCT_CLASS,
893 VFP_CLASS,
894 VFP_STRUCT_CLASS,
895 CORE_CLASS,
896 NB_CLASSES
899 struct param_plan {
900 int start; /* first reg or addr used depending on the class */
901 int end; /* last reg used or next free addr depending on the class */
902 SValue *sval; /* pointer to SValue on the value stack */
903 struct param_plan *prev; /* previous element in this class */
906 struct plan {
907 struct param_plan *pplans; /* array of all the param plans */
908 struct param_plan *clsplans[NB_CLASSES]; /* per class lists of param plans */
911 #define add_param_plan(plan,pplan,class) \
912 do { \
913 pplan.prev = plan->clsplans[class]; \
914 plan->pplans[plan ## _nb] = pplan; \
915 plan->clsplans[class] = &plan->pplans[plan ## _nb++]; \
916 } while(0)
918 /* Assign parameters to registers and stack with alignment according to the
919 rules in the procedure call standard for the ARM architecture (AAPCS).
920 The overall assignment is recorded in an array of per parameter structures
921 called parameter plans. The parameter plans are also further organized in a
922 number of linked lists, one per class of parameter (see the comment for the
923 definition of union reg_class).
925 nb_args: number of parameters of the function for which a call is generated
926 float_abi: float ABI in use for this function call
927 plan: the structure where the overall assignment is recorded
928 todo: a bitmap that record which core registers hold a parameter
930 Returns the amount of stack space needed for parameter passing
932 Note: this function allocated an array in plan->pplans with tcc_malloc. It
933 is the responsability of the caller to free this array once used (ie not
934 before copy_params). */
935 static int assign_regs(int nb_args, int float_abi, struct plan *plan, int *todo)
937 int i, size, align;
938 int ncrn /* next core register number */, nsaa /* next stacked argument address*/;
939 int plan_nb = 0;
940 struct param_plan pplan;
941 struct avail_regs avregs = AVAIL_REGS_INITIALIZER;
943 ncrn = nsaa = 0;
944 *todo = 0;
945 plan->pplans = tcc_malloc(nb_args * sizeof(*plan->pplans));
946 memset(plan->clsplans, 0, sizeof(plan->clsplans));
947 for(i = nb_args; i-- ;) {
948 int j, start_vfpreg = 0;
949 size = type_size(&vtop[-i].type, &align);
950 size = (size + 3) & ~3;
951 align = (align + 3) & ~3;
952 switch(vtop[-i].type.t & VT_BTYPE) {
953 case VT_STRUCT:
954 case VT_FLOAT:
955 case VT_DOUBLE:
956 case VT_LDOUBLE:
957 if (float_abi == ARM_HARD_FLOAT) {
958 int is_hfa = 0; /* Homogeneous float aggregate */
960 if (is_float(vtop[-i].type.t)
961 || (is_hfa = is_hgen_float_aggr(&vtop[-i].type))) {
962 int end_vfpreg;
964 start_vfpreg = assign_vfpreg(&avregs, align, size);
965 end_vfpreg = start_vfpreg + ((size - 1) >> 2);
966 if (start_vfpreg >= 0) {
967 pplan = (struct param_plan) {start_vfpreg, end_vfpreg, &vtop[-i]};
968 if (is_hfa)
969 add_param_plan(plan, pplan, VFP_STRUCT_CLASS);
970 else
971 add_param_plan(plan, pplan, VFP_CLASS);
972 continue;
973 } else
974 break;
977 ncrn = (ncrn + (align-1)/4) & ~((align/4) - 1);
978 if (ncrn + size/4 <= 4 || (ncrn < 4 && start_vfpreg != -1)) {
979 /* The parameter is allocated both in core register and on stack. As
980 * such, it can be of either class: it would either be the last of
981 * CORE_STRUCT_CLASS or the first of STACK_CLASS. */
982 for (j = ncrn; j < 4 && j < ncrn + size / 4; j++)
983 *todo|=(1<<j);
984 pplan = (struct param_plan) {ncrn, j, &vtop[-i]};
985 add_param_plan(plan, pplan, CORE_STRUCT_CLASS);
986 ncrn += size/4;
987 if (ncrn > 4)
988 nsaa = (ncrn - 4) * 4;
989 } else {
990 ncrn = 4;
991 break;
993 continue;
994 default:
995 if (ncrn < 4) {
996 int is_long = (vtop[-i].type.t & VT_BTYPE) == VT_LLONG;
998 if (is_long) {
999 ncrn = (ncrn + 1) & -2;
1000 if (ncrn == 4)
1001 break;
1003 pplan = (struct param_plan) {ncrn, ncrn, &vtop[-i]};
1004 ncrn++;
1005 if (is_long)
1006 pplan.end = ncrn++;
1007 add_param_plan(plan, pplan, CORE_CLASS);
1008 continue;
1011 nsaa = (nsaa + (align - 1)) & ~(align - 1);
1012 pplan = (struct param_plan) {nsaa, nsaa + size, &vtop[-i]};
1013 add_param_plan(plan, pplan, STACK_CLASS);
1014 nsaa += size; /* size already rounded up before */
1016 return nsaa;
1019 #undef add_param_plan
1021 /* Copy parameters to their final destination (core reg, VFP reg or stack) for
1022 function call.
1024 nb_args: number of parameters the function take
1025 plan: the overall assignment plan for parameters
1026 todo: a bitmap indicating what core reg will hold a parameter
1028 Returns the number of SValue added by this function on the value stack */
1029 static int copy_params(int nb_args, struct plan *plan, int todo)
1031 int size, align, r, i, nb_extra_sval = 0;
1032 struct param_plan *pplan;
1034 /* Several constraints require parameters to be copied in a specific order:
1035 - structures are copied to the stack before being loaded in a reg;
1036 - floats loaded to an odd numbered VFP reg are first copied to the
1037 preceding even numbered VFP reg and then moved to the next VFP reg.
1039 It is thus important that:
1040 - structures assigned to core regs must be copied after parameters
1041 assigned to the stack but before structures assigned to VFP regs because
1042 a structure can lie partly in core registers and partly on the stack;
1043 - parameters assigned to the stack and all structures be copied before
1044 parameters assigned to a core reg since copying a parameter to the stack
1045 require using a core reg;
1046 - parameters assigned to VFP regs be copied before structures assigned to
1047 VFP regs as the copy might use an even numbered VFP reg that already
1048 holds part of a structure. */
1049 for(i = 0; i < NB_CLASSES; i++) {
1050 for(pplan = plan->clsplans[i]; pplan; pplan = pplan->prev) {
1051 vpushv(pplan->sval);
1052 pplan->sval->r = pplan->sval->r2 = VT_CONST; /* disable entry */
1053 switch(i) {
1054 case STACK_CLASS:
1055 case CORE_STRUCT_CLASS:
1056 case VFP_STRUCT_CLASS:
1057 if ((pplan->sval->type.t & VT_BTYPE) == VT_STRUCT) {
1058 int padding = 0;
1059 size = type_size(&pplan->sval->type, &align);
1060 /* align to stack align size */
1061 size = (size + 3) & ~3;
1062 if (i == STACK_CLASS && pplan->prev)
1063 padding = pplan->start - pplan->prev->end;
1064 size += padding; /* Add padding if any */
1065 /* allocate the necessary size on stack */
1066 gadd_sp(-size);
1067 /* generate structure store */
1068 r = get_reg(RC_INT);
1069 o(0xE28D0000|(intr(r)<<12)|padding); /* add r, sp, padding */
1070 vset(&vtop->type, r | VT_LVAL, 0);
1071 vswap();
1072 vstore(); /* memcpy to current sp + potential padding */
1074 /* Homogeneous float aggregate are loaded to VFP registers
1075 immediately since there is no way of loading data in multiple
1076 non consecutive VFP registers as what is done for other
1077 structures (see the use of todo). */
1078 if (i == VFP_STRUCT_CLASS) {
1079 int first = pplan->start, nb = pplan->end - first + 1;
1080 /* vpop.32 {pplan->start, ..., pplan->end} */
1081 o(0xECBD0A00|(first&1)<<22|(first>>1)<<12|nb);
1082 /* No need to write the register used to a SValue since VFP regs
1083 cannot be used for gcall_or_jmp */
1085 } else {
1086 if (is_float(pplan->sval->type.t)) {
1087 #ifdef TCC_ARM_VFP
1088 r = vfpr(gv(RC_FLOAT)) << 12;
1089 if ((pplan->sval->type.t & VT_BTYPE) == VT_FLOAT)
1090 size = 4;
1091 else {
1092 size = 8;
1093 r |= 0x101; /* vpush.32 -> vpush.64 */
1095 o(0xED2D0A01 + r); /* vpush */
1096 #else
1097 r = fpr(gv(RC_FLOAT)) << 12;
1098 if ((pplan->sval->type.t & VT_BTYPE) == VT_FLOAT)
1099 size = 4;
1100 else if ((pplan->sval->type.t & VT_BTYPE) == VT_DOUBLE)
1101 size = 8;
1102 else
1103 size = LDOUBLE_SIZE;
1105 if (size == 12)
1106 r |= 0x400000;
1107 else if(size == 8)
1108 r|=0x8000;
1110 o(0xED2D0100|r|(size>>2)); /* some kind of vpush for FPA */
1111 #endif
1112 } else {
1113 /* simple type (currently always same size) */
1114 /* XXX: implicit cast ? */
1115 size=4;
1116 if ((pplan->sval->type.t & VT_BTYPE) == VT_LLONG) {
1117 lexpand_nr();
1118 size = 8;
1119 r = gv(RC_INT);
1120 o(0xE52D0004|(intr(r)<<12)); /* push r */
1121 vtop--;
1123 r = gv(RC_INT);
1124 o(0xE52D0004|(intr(r)<<12)); /* push r */
1126 if (i == STACK_CLASS && pplan->prev)
1127 gadd_sp(pplan->prev->end - pplan->start); /* Add padding if any */
1129 break;
1131 case VFP_CLASS:
1132 gv(regmask(TREG_F0 + (pplan->start >> 1)));
1133 if (pplan->start & 1) { /* Must be in upper part of double register */
1134 o(0xEEF00A40|((pplan->start>>1)<<12)|(pplan->start>>1)); /* vmov.f32 s(n+1), sn */
1135 vtop->r = VT_CONST; /* avoid being saved on stack by gv for next float */
1137 break;
1139 case CORE_CLASS:
1140 if ((pplan->sval->type.t & VT_BTYPE) == VT_LLONG) {
1141 lexpand_nr();
1142 gv(regmask(pplan->end));
1143 pplan->sval->r2 = vtop->r;
1144 vtop--;
1146 gv(regmask(pplan->start));
1147 /* Mark register as used so that gcall_or_jmp use another one
1148 (regs >=4 are free as never used to pass parameters) */
1149 pplan->sval->r = vtop->r;
1150 break;
1152 vtop--;
1156 /* Manually free remaining registers since next parameters are loaded
1157 * manually, without the help of gv(int). */
1158 save_regs(nb_args);
1160 if(todo) {
1161 o(0xE8BD0000|todo); /* pop {todo} */
1162 for(pplan = plan->clsplans[CORE_STRUCT_CLASS]; pplan; pplan = pplan->prev) {
1163 int r;
1164 pplan->sval->r = pplan->start;
1165 /* An SValue can only pin 2 registers at best (r and r2) but a structure
1166 can occupy more than 2 registers. Thus, we need to push on the value
1167 stack some fake parameter to have on SValue for each registers used
1168 by a structure (r2 is not used). */
1169 for (r = pplan->start + 1; r <= pplan->end; r++) {
1170 if (todo & (1 << r)) {
1171 nb_extra_sval++;
1172 vpushi(0);
1173 vtop->r = r;
1178 return nb_extra_sval;
1181 /* Generate function call. The function address is pushed first, then
1182 all the parameters in call order. This functions pops all the
1183 parameters and the function address. */
1184 void gfunc_call(int nb_args)
1186 int r, args_size;
1187 int variadic, def_float_abi = float_abi;
1188 int todo;
1189 struct plan plan;
1191 #ifdef TCC_ARM_EABI
1192 if (float_abi == ARM_HARD_FLOAT) {
1193 variadic = (vtop[-nb_args].type.ref->c == FUNC_ELLIPSIS);
1194 if (variadic || floats_in_core_regs(&vtop[-nb_args]))
1195 float_abi = ARM_SOFTFP_FLOAT;
1197 #endif
1198 /* cannot let cpu flags if other instruction are generated. Also avoid leaving
1199 VT_JMP anywhere except on the top of the stack because it would complicate
1200 the code generator. */
1201 r = vtop->r & VT_VALMASK;
1202 if (r == VT_CMP || (r & ~1) == VT_JMP)
1203 gv(RC_INT);
1205 args_size = assign_regs(nb_args, float_abi, &plan, &todo);
1207 #ifdef TCC_ARM_EABI
1208 if (args_size & 7) { /* Stack must be 8 byte aligned at fct call for EABI */
1209 args_size = (args_size + 7) & ~7;
1210 o(0xE24DD004); /* sub sp, sp, #4 */
1212 #endif
1214 nb_args += copy_params(nb_args, &plan, todo);
1215 tcc_free(plan.pplans);
1217 /* Move fct SValue on top as required by gcall_or_jmp */
1218 vrotb(nb_args + 1);
1219 gcall_or_jmp(0);
1220 if (args_size)
1221 gadd_sp(args_size); /* pop all parameters passed on the stack */
1222 #if defined(TCC_ARM_EABI) && defined(TCC_ARM_VFP)
1223 if(float_abi == ARM_SOFTFP_FLOAT && is_float(vtop->type.ref->type.t)) {
1224 if((vtop->type.ref->type.t & VT_BTYPE) == VT_FLOAT) {
1225 o(0xEE000A10); /*vmov s0, r0 */
1226 } else {
1227 o(0xEE000B10); /* vmov.32 d0[0], r0 */
1228 o(0xEE201B10); /* vmov.32 d0[1], r1 */
1231 #endif
1232 vtop -= nb_args + 1; /* Pop all params and fct address from value stack */
1233 leaffunc = 0; /* we are calling a function, so we aren't in a leaf function */
1234 float_abi = def_float_abi;
1237 /* generate function prolog of type 't' */
1238 void gfunc_prolog(CType *func_type)
1240 Sym *sym,*sym2;
1241 int n, nf, size, align, struct_ret = 0;
1242 int addr, pn, sn; /* pn=core, sn=stack */
1243 struct avail_regs avregs = AVAIL_REGS_INITIALIZER;
1244 CType ret_type;
1246 sym = func_type->ref;
1247 func_vt = sym->type;
1248 func_var = (func_type->ref->c == FUNC_ELLIPSIS);
1250 n = nf = 0;
1251 if ((func_vt.t & VT_BTYPE) == VT_STRUCT &&
1252 !gfunc_sret(&func_vt, func_var, &ret_type, &align))
1254 n++;
1255 struct_ret = 1;
1256 func_vc = 12; /* Offset from fp of the place to store the result */
1258 for(sym2 = sym->next; sym2 && (n < 4 || nf < 16); sym2 = sym2->next) {
1259 size = type_size(&sym2->type, &align);
1260 #ifdef TCC_ARM_EABI
1261 if (float_abi == ARM_HARD_FLOAT && !func_var &&
1262 (is_float(sym2->type.t) || is_hgen_float_aggr(&sym2->type))) {
1263 int tmpnf = assign_vfpreg(&avregs, align, size);
1264 tmpnf += (size + 3) / 4;
1265 nf = (tmpnf > nf) ? tmpnf : nf;
1266 } else
1267 #endif
1268 if (n < 4)
1269 n += (size + 3) / 4;
1271 o(0xE1A0C00D); /* mov ip,sp */
1272 if (func_var)
1273 n=4;
1274 if (n) {
1275 if(n>4)
1276 n=4;
1277 #ifdef TCC_ARM_EABI
1278 n=(n+1)&-2;
1279 #endif
1280 o(0xE92D0000|((1<<n)-1)); /* save r0-r4 on stack if needed */
1282 if (nf) {
1283 if (nf>16)
1284 nf=16;
1285 nf=(nf+1)&-2; /* nf => HARDFLOAT => EABI */
1286 o(0xED2D0A00|nf); /* save s0-s15 on stack if needed */
1288 o(0xE92D5800); /* save fp, ip, lr */
1289 o(0xE1A0B00D); /* mov fp, sp */
1290 func_sub_sp_offset = ind;
1291 o(0xE1A00000); /* nop, leave space for stack adjustment in epilog */
1293 #ifdef TCC_ARM_EABI
1294 if (float_abi == ARM_HARD_FLOAT) {
1295 func_vc += nf * 4;
1296 avregs = AVAIL_REGS_INITIALIZER;
1298 #endif
1299 pn = struct_ret, sn = 0;
1300 while ((sym = sym->next)) {
1301 CType *type;
1302 type = &sym->type;
1303 size = type_size(type, &align);
1304 size = (size + 3) >> 2;
1305 align = (align + 3) & ~3;
1306 #ifdef TCC_ARM_EABI
1307 if (float_abi == ARM_HARD_FLOAT && !func_var && (is_float(sym->type.t)
1308 || is_hgen_float_aggr(&sym->type))) {
1309 int fpn = assign_vfpreg(&avregs, align, size << 2);
1310 if (fpn >= 0)
1311 addr = fpn * 4;
1312 else
1313 goto from_stack;
1314 } else
1315 #endif
1316 if (pn < 4) {
1317 #ifdef TCC_ARM_EABI
1318 pn = (pn + (align-1)/4) & -(align/4);
1319 #endif
1320 addr = (nf + pn) * 4;
1321 pn += size;
1322 if (!sn && pn > 4)
1323 sn = (pn - 4);
1324 } else {
1325 from_stack:
1326 #ifdef TCC_ARM_EABI
1327 sn = (sn + (align-1)/4) & -(align/4);
1328 #endif
1329 addr = (n + nf + sn) * 4;
1330 sn += size;
1332 sym_push(sym->v & ~SYM_FIELD, type, VT_LOCAL | lvalue_type(type->t),
1333 addr + 12);
1335 last_itod_magic=0;
1336 leaffunc = 1;
1337 loc = 0;
1340 /* generate function epilog */
1341 void gfunc_epilog(void)
1343 uint32_t x;
1344 int diff;
1345 /* Copy float return value to core register if base standard is used and
1346 float computation is made with VFP */
1347 #if defined(TCC_ARM_EABI) && defined(TCC_ARM_VFP)
1348 if ((float_abi == ARM_SOFTFP_FLOAT || func_var) && is_float(func_vt.t)) {
1349 if((func_vt.t & VT_BTYPE) == VT_FLOAT)
1350 o(0xEE100A10); /* fmrs r0, s0 */
1351 else {
1352 o(0xEE100B10); /* fmrdl r0, d0 */
1353 o(0xEE301B10); /* fmrdh r1, d0 */
1356 #endif
1357 o(0xE89BA800); /* restore fp, sp, pc */
1358 diff = (-loc + 3) & -4;
1359 #ifdef TCC_ARM_EABI
1360 if(!leaffunc)
1361 diff = ((diff + 11) & -8) - 4;
1362 #endif
1363 if(diff > 0) {
1364 x=stuff_const(0xE24BD000, diff); /* sub sp,fp,# */
1365 if(x)
1366 *(uint32_t *)(cur_text_section->data + func_sub_sp_offset) = x;
1367 else {
1368 int addr;
1369 addr=ind;
1370 o(0xE59FC004); /* ldr ip,[pc+4] */
1371 o(0xE04BD00C); /* sub sp,fp,ip */
1372 o(0xE1A0F00E); /* mov pc,lr */
1373 o(diff);
1374 *(uint32_t *)(cur_text_section->data + func_sub_sp_offset) = 0xE1000000|encbranch(func_sub_sp_offset,addr,1);
1379 /* generate a jump to a label */
1380 int gjmp(int t)
1382 int r;
1383 r=ind;
1384 o(0xE0000000|encbranch(r,t,1));
1385 return r;
1388 /* generate a jump to a fixed address */
1389 void gjmp_addr(int a)
1391 gjmp(a);
1394 /* generate a test. set 'inv' to invert test. Stack entry is popped */
1395 int gtst(int inv, int t)
1397 int v, r;
1398 uint32_t op;
1399 v = vtop->r & VT_VALMASK;
1400 r=ind;
1401 if (v == VT_CMP) {
1402 op=mapcc(inv?negcc(vtop->c.i):vtop->c.i);
1403 op|=encbranch(r,t,1);
1404 o(op);
1405 t=r;
1406 } else { /* VT_JMP || VT_JMPI */
1407 if ((v & 1) == inv) {
1408 if(!vtop->c.i)
1409 vtop->c.i=t;
1410 else {
1411 uint32_t *x;
1412 int p,lp;
1413 if(t) {
1414 p = vtop->c.i;
1415 do {
1416 p = decbranch(lp=p);
1417 } while(p);
1418 x = (uint32_t *)(cur_text_section->data + lp);
1419 *x &= 0xff000000;
1420 *x |= encbranch(lp,t,1);
1422 t = vtop->c.i;
1424 } else {
1425 t = gjmp(t);
1426 gsym(vtop->c.i);
1429 vtop--;
1430 return t;
1433 /* generate an integer binary operation */
1434 void gen_opi(int op)
1436 int c, func = 0;
1437 uint32_t opc = 0, r, fr;
1438 unsigned short retreg = REG_IRET;
1440 c=0;
1441 switch(op) {
1442 case '+':
1443 opc = 0x8;
1444 c=1;
1445 break;
1446 case TOK_ADDC1: /* add with carry generation */
1447 opc = 0x9;
1448 c=1;
1449 break;
1450 case '-':
1451 opc = 0x4;
1452 c=1;
1453 break;
1454 case TOK_SUBC1: /* sub with carry generation */
1455 opc = 0x5;
1456 c=1;
1457 break;
1458 case TOK_ADDC2: /* add with carry use */
1459 opc = 0xA;
1460 c=1;
1461 break;
1462 case TOK_SUBC2: /* sub with carry use */
1463 opc = 0xC;
1464 c=1;
1465 break;
1466 case '&':
1467 opc = 0x0;
1468 c=1;
1469 break;
1470 case '^':
1471 opc = 0x2;
1472 c=1;
1473 break;
1474 case '|':
1475 opc = 0x18;
1476 c=1;
1477 break;
1478 case '*':
1479 gv2(RC_INT, RC_INT);
1480 r = vtop[-1].r;
1481 fr = vtop[0].r;
1482 vtop--;
1483 o(0xE0000090|(intr(r)<<16)|(intr(r)<<8)|intr(fr));
1484 return;
1485 case TOK_SHL:
1486 opc = 0;
1487 c=2;
1488 break;
1489 case TOK_SHR:
1490 opc = 1;
1491 c=2;
1492 break;
1493 case TOK_SAR:
1494 opc = 2;
1495 c=2;
1496 break;
1497 case '/':
1498 case TOK_PDIV:
1499 func=TOK___divsi3;
1500 c=3;
1501 break;
1502 case TOK_UDIV:
1503 func=TOK___udivsi3;
1504 c=3;
1505 break;
1506 case '%':
1507 #ifdef TCC_ARM_EABI
1508 func=TOK___aeabi_idivmod;
1509 retreg=REG_LRET;
1510 #else
1511 func=TOK___modsi3;
1512 #endif
1513 c=3;
1514 break;
1515 case TOK_UMOD:
1516 #ifdef TCC_ARM_EABI
1517 func=TOK___aeabi_uidivmod;
1518 retreg=REG_LRET;
1519 #else
1520 func=TOK___umodsi3;
1521 #endif
1522 c=3;
1523 break;
1524 case TOK_UMULL:
1525 gv2(RC_INT, RC_INT);
1526 r=intr(vtop[-1].r2=get_reg(RC_INT));
1527 c=vtop[-1].r;
1528 vtop[-1].r=get_reg_ex(RC_INT,regmask(c));
1529 vtop--;
1530 o(0xE0800090|(r<<16)|(intr(vtop->r)<<12)|(intr(c)<<8)|intr(vtop[1].r));
1531 return;
1532 default:
1533 opc = 0x15;
1534 c=1;
1535 break;
1537 switch(c) {
1538 case 1:
1539 if((vtop[-1].r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
1540 if(opc == 4 || opc == 5 || opc == 0xc) {
1541 vswap();
1542 opc|=2; // sub -> rsb
1545 if ((vtop->r & VT_VALMASK) == VT_CMP ||
1546 (vtop->r & (VT_VALMASK & ~1)) == VT_JMP)
1547 gv(RC_INT);
1548 vswap();
1549 c=intr(gv(RC_INT));
1550 vswap();
1551 opc=0xE0000000|(opc<<20)|(c<<16);
1552 if((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
1553 uint32_t x;
1554 x=stuff_const(opc|0x2000000,vtop->c.i);
1555 if(x) {
1556 r=intr(vtop[-1].r=get_reg_ex(RC_INT,regmask(vtop[-1].r)));
1557 o(x|(r<<12));
1558 goto done;
1561 fr=intr(gv(RC_INT));
1562 r=intr(vtop[-1].r=get_reg_ex(RC_INT,two2mask(vtop->r,vtop[-1].r)));
1563 o(opc|(r<<12)|fr);
1564 done:
1565 vtop--;
1566 if (op >= TOK_ULT && op <= TOK_GT) {
1567 vtop->r = VT_CMP;
1568 vtop->c.i = op;
1570 break;
1571 case 2:
1572 opc=0xE1A00000|(opc<<5);
1573 if ((vtop->r & VT_VALMASK) == VT_CMP ||
1574 (vtop->r & (VT_VALMASK & ~1)) == VT_JMP)
1575 gv(RC_INT);
1576 vswap();
1577 r=intr(gv(RC_INT));
1578 vswap();
1579 opc|=r;
1580 if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
1581 fr=intr(vtop[-1].r=get_reg_ex(RC_INT,regmask(vtop[-1].r)));
1582 c = vtop->c.i & 0x1f;
1583 o(opc|(c<<7)|(fr<<12));
1584 } else {
1585 fr=intr(gv(RC_INT));
1586 c=intr(vtop[-1].r=get_reg_ex(RC_INT,two2mask(vtop->r,vtop[-1].r)));
1587 o(opc|(c<<12)|(fr<<8)|0x10);
1589 vtop--;
1590 break;
1591 case 3:
1592 vpush_global_sym(&func_old_type, func);
1593 vrott(3);
1594 gfunc_call(2);
1595 vpushi(0);
1596 vtop->r = retreg;
1597 break;
1598 default:
1599 tcc_error("gen_opi %i unimplemented!",op);
1603 #ifdef TCC_ARM_VFP
1604 static int is_zero(int i)
1606 if((vtop[i].r & (VT_VALMASK | VT_LVAL | VT_SYM)) != VT_CONST)
1607 return 0;
1608 if (vtop[i].type.t == VT_FLOAT)
1609 return (vtop[i].c.f == 0.f);
1610 else if (vtop[i].type.t == VT_DOUBLE)
1611 return (vtop[i].c.d == 0.0);
1612 return (vtop[i].c.ld == 0.l);
1615 /* generate a floating point operation 'v = t1 op t2' instruction. The
1616 * two operands are guaranted to have the same floating point type */
1617 void gen_opf(int op)
1619 uint32_t x;
1620 int fneg=0,r;
1621 x=0xEE000A00|T2CPR(vtop->type.t);
1622 switch(op) {
1623 case '+':
1624 if(is_zero(-1))
1625 vswap();
1626 if(is_zero(0)) {
1627 vtop--;
1628 return;
1630 x|=0x300000;
1631 break;
1632 case '-':
1633 x|=0x300040;
1634 if(is_zero(0)) {
1635 vtop--;
1636 return;
1638 if(is_zero(-1)) {
1639 x|=0x810000; /* fsubX -> fnegX */
1640 vswap();
1641 vtop--;
1642 fneg=1;
1644 break;
1645 case '*':
1646 x|=0x200000;
1647 break;
1648 case '/':
1649 x|=0x800000;
1650 break;
1651 default:
1652 if(op < TOK_ULT || op > TOK_GT) {
1653 tcc_error("unknown fp op %x!",op);
1654 return;
1656 if(is_zero(-1)) {
1657 vswap();
1658 switch(op) {
1659 case TOK_LT: op=TOK_GT; break;
1660 case TOK_GE: op=TOK_ULE; break;
1661 case TOK_LE: op=TOK_GE; break;
1662 case TOK_GT: op=TOK_ULT; break;
1665 x|=0xB40040; /* fcmpX */
1666 if(op!=TOK_EQ && op!=TOK_NE)
1667 x|=0x80; /* fcmpX -> fcmpeX */
1668 if(is_zero(0)) {
1669 vtop--;
1670 o(x|0x10000|(vfpr(gv(RC_FLOAT))<<12)); /* fcmp(e)X -> fcmp(e)zX */
1671 } else {
1672 x|=vfpr(gv(RC_FLOAT));
1673 vswap();
1674 o(x|(vfpr(gv(RC_FLOAT))<<12));
1675 vtop--;
1677 o(0xEEF1FA10); /* fmstat */
1679 switch(op) {
1680 case TOK_LE: op=TOK_ULE; break;
1681 case TOK_LT: op=TOK_ULT; break;
1682 case TOK_UGE: op=TOK_GE; break;
1683 case TOK_UGT: op=TOK_GT; break;
1686 vtop->r = VT_CMP;
1687 vtop->c.i = op;
1688 return;
1690 r=gv(RC_FLOAT);
1691 x|=vfpr(r);
1692 r=regmask(r);
1693 if(!fneg) {
1694 int r2;
1695 vswap();
1696 r2=gv(RC_FLOAT);
1697 x|=vfpr(r2)<<16;
1698 r|=regmask(r2);
1700 vtop->r=get_reg_ex(RC_FLOAT,r);
1701 if(!fneg)
1702 vtop--;
1703 o(x|(vfpr(vtop->r)<<12));
1706 #else
1707 static uint32_t is_fconst()
1709 long double f;
1710 uint32_t r;
1711 if((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) != VT_CONST)
1712 return 0;
1713 if (vtop->type.t == VT_FLOAT)
1714 f = vtop->c.f;
1715 else if (vtop->type.t == VT_DOUBLE)
1716 f = vtop->c.d;
1717 else
1718 f = vtop->c.ld;
1719 if(!ieee_finite(f))
1720 return 0;
1721 r=0x8;
1722 if(f<0.0) {
1723 r=0x18;
1724 f=-f;
1726 if(f==0.0)
1727 return r;
1728 if(f==1.0)
1729 return r|1;
1730 if(f==2.0)
1731 return r|2;
1732 if(f==3.0)
1733 return r|3;
1734 if(f==4.0)
1735 return r|4;
1736 if(f==5.0)
1737 return r|5;
1738 if(f==0.5)
1739 return r|6;
1740 if(f==10.0)
1741 return r|7;
1742 return 0;
1745 /* generate a floating point operation 'v = t1 op t2' instruction. The
1746 two operands are guaranted to have the same floating point type */
1747 void gen_opf(int op)
1749 uint32_t x, r, r2, c1, c2;
1750 //fputs("gen_opf\n",stderr);
1751 vswap();
1752 c1 = is_fconst();
1753 vswap();
1754 c2 = is_fconst();
1755 x=0xEE000100;
1756 #if LDOUBLE_SIZE == 8
1757 if ((vtop->type.t & VT_BTYPE) != VT_FLOAT)
1758 x|=0x80;
1759 #else
1760 if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
1761 x|=0x80;
1762 else if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE)
1763 x|=0x80000;
1764 #endif
1765 switch(op)
1767 case '+':
1768 if(!c2) {
1769 vswap();
1770 c2=c1;
1772 vswap();
1773 r=fpr(gv(RC_FLOAT));
1774 vswap();
1775 if(c2) {
1776 if(c2>0xf)
1777 x|=0x200000; // suf
1778 r2=c2&0xf;
1779 } else {
1780 r2=fpr(gv(RC_FLOAT));
1782 break;
1783 case '-':
1784 if(c2) {
1785 if(c2<=0xf)
1786 x|=0x200000; // suf
1787 r2=c2&0xf;
1788 vswap();
1789 r=fpr(gv(RC_FLOAT));
1790 vswap();
1791 } else if(c1 && c1<=0xf) {
1792 x|=0x300000; // rsf
1793 r2=c1;
1794 r=fpr(gv(RC_FLOAT));
1795 vswap();
1796 } else {
1797 x|=0x200000; // suf
1798 vswap();
1799 r=fpr(gv(RC_FLOAT));
1800 vswap();
1801 r2=fpr(gv(RC_FLOAT));
1803 break;
1804 case '*':
1805 if(!c2 || c2>0xf) {
1806 vswap();
1807 c2=c1;
1809 vswap();
1810 r=fpr(gv(RC_FLOAT));
1811 vswap();
1812 if(c2 && c2<=0xf)
1813 r2=c2;
1814 else
1815 r2=fpr(gv(RC_FLOAT));
1816 x|=0x100000; // muf
1817 break;
1818 case '/':
1819 if(c2 && c2<=0xf) {
1820 x|=0x400000; // dvf
1821 r2=c2;
1822 vswap();
1823 r=fpr(gv(RC_FLOAT));
1824 vswap();
1825 } else if(c1 && c1<=0xf) {
1826 x|=0x500000; // rdf
1827 r2=c1;
1828 r=fpr(gv(RC_FLOAT));
1829 vswap();
1830 } else {
1831 x|=0x400000; // dvf
1832 vswap();
1833 r=fpr(gv(RC_FLOAT));
1834 vswap();
1835 r2=fpr(gv(RC_FLOAT));
1837 break;
1838 default:
1839 if(op >= TOK_ULT && op <= TOK_GT) {
1840 x|=0xd0f110; // cmfe
1841 /* bug (intention?) in Linux FPU emulator
1842 doesn't set carry if equal */
1843 switch(op) {
1844 case TOK_ULT:
1845 case TOK_UGE:
1846 case TOK_ULE:
1847 case TOK_UGT:
1848 tcc_error("unsigned comparision on floats?");
1849 break;
1850 case TOK_LT:
1851 op=TOK_Nset;
1852 break;
1853 case TOK_LE:
1854 op=TOK_ULE; /* correct in unordered case only if AC bit in FPSR set */
1855 break;
1856 case TOK_EQ:
1857 case TOK_NE:
1858 x&=~0x400000; // cmfe -> cmf
1859 break;
1861 if(c1 && !c2) {
1862 c2=c1;
1863 vswap();
1864 switch(op) {
1865 case TOK_Nset:
1866 op=TOK_GT;
1867 break;
1868 case TOK_GE:
1869 op=TOK_ULE;
1870 break;
1871 case TOK_ULE:
1872 op=TOK_GE;
1873 break;
1874 case TOK_GT:
1875 op=TOK_Nset;
1876 break;
1879 vswap();
1880 r=fpr(gv(RC_FLOAT));
1881 vswap();
1882 if(c2) {
1883 if(c2>0xf)
1884 x|=0x200000;
1885 r2=c2&0xf;
1886 } else {
1887 r2=fpr(gv(RC_FLOAT));
1889 vtop[-1].r = VT_CMP;
1890 vtop[-1].c.i = op;
1891 } else {
1892 tcc_error("unknown fp op %x!",op);
1893 return;
1896 if(vtop[-1].r == VT_CMP)
1897 c1=15;
1898 else {
1899 c1=vtop->r;
1900 if(r2&0x8)
1901 c1=vtop[-1].r;
1902 vtop[-1].r=get_reg_ex(RC_FLOAT,two2mask(vtop[-1].r,c1));
1903 c1=fpr(vtop[-1].r);
1905 vtop--;
1906 o(x|(r<<16)|(c1<<12)|r2);
1908 #endif
1910 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
1911 and 'long long' cases. */
1912 ST_FUNC void gen_cvt_itof1(int t)
1914 uint32_t r, r2;
1915 int bt;
1916 bt=vtop->type.t & VT_BTYPE;
1917 if(bt == VT_INT || bt == VT_SHORT || bt == VT_BYTE) {
1918 #ifndef TCC_ARM_VFP
1919 uint32_t dsize = 0;
1920 #endif
1921 r=intr(gv(RC_INT));
1922 #ifdef TCC_ARM_VFP
1923 r2=vfpr(vtop->r=get_reg(RC_FLOAT));
1924 o(0xEE000A10|(r<<12)|(r2<<16)); /* fmsr */
1925 r2|=r2<<12;
1926 if(!(vtop->type.t & VT_UNSIGNED))
1927 r2|=0x80; /* fuitoX -> fsituX */
1928 o(0xEEB80A40|r2|T2CPR(t)); /* fYitoX*/
1929 #else
1930 r2=fpr(vtop->r=get_reg(RC_FLOAT));
1931 if((t & VT_BTYPE) != VT_FLOAT)
1932 dsize=0x80; /* flts -> fltd */
1933 o(0xEE000110|dsize|(r2<<16)|(r<<12)); /* flts */
1934 if((vtop->type.t & (VT_UNSIGNED|VT_BTYPE)) == (VT_UNSIGNED|VT_INT)) {
1935 uint32_t off = 0;
1936 o(0xE3500000|(r<<12)); /* cmp */
1937 r=fpr(get_reg(RC_FLOAT));
1938 if(last_itod_magic) {
1939 off=ind+8-last_itod_magic;
1940 off/=4;
1941 if(off>255)
1942 off=0;
1944 o(0xBD1F0100|(r<<12)|off); /* ldflts */
1945 if(!off) {
1946 o(0xEA000000); /* b */
1947 last_itod_magic=ind;
1948 o(0x4F800000); /* 4294967296.0f */
1950 o(0xBE000100|dsize|(r2<<16)|(r2<<12)|r); /* adflt */
1952 #endif
1953 return;
1954 } else if(bt == VT_LLONG) {
1955 int func;
1956 CType *func_type = 0;
1957 if((t & VT_BTYPE) == VT_FLOAT) {
1958 func_type = &func_float_type;
1959 if(vtop->type.t & VT_UNSIGNED)
1960 func=TOK___floatundisf;
1961 else
1962 func=TOK___floatdisf;
1963 #if LDOUBLE_SIZE != 8
1964 } else if((t & VT_BTYPE) == VT_LDOUBLE) {
1965 func_type = &func_ldouble_type;
1966 if(vtop->type.t & VT_UNSIGNED)
1967 func=TOK___floatundixf;
1968 else
1969 func=TOK___floatdixf;
1970 } else if((t & VT_BTYPE) == VT_DOUBLE) {
1971 #else
1972 } else if((t & VT_BTYPE) == VT_DOUBLE || (t & VT_BTYPE) == VT_LDOUBLE) {
1973 #endif
1974 func_type = &func_double_type;
1975 if(vtop->type.t & VT_UNSIGNED)
1976 func=TOK___floatundidf;
1977 else
1978 func=TOK___floatdidf;
1980 if(func_type) {
1981 vpush_global_sym(func_type, func);
1982 vswap();
1983 gfunc_call(1);
1984 vpushi(0);
1985 vtop->r=TREG_F0;
1986 return;
1989 tcc_error("unimplemented gen_cvt_itof %x!",vtop->type.t);
1992 /* convert fp to int 't' type */
1993 void gen_cvt_ftoi(int t)
1995 uint32_t r, r2;
1996 int u, func = 0;
1997 u=t&VT_UNSIGNED;
1998 t&=VT_BTYPE;
1999 r2=vtop->type.t & VT_BTYPE;
2000 if(t==VT_INT) {
2001 #ifdef TCC_ARM_VFP
2002 r=vfpr(gv(RC_FLOAT));
2003 u=u?0:0x10000;
2004 o(0xEEBC0AC0|(r<<12)|r|T2CPR(r2)|u); /* ftoXizY */
2005 r2=intr(vtop->r=get_reg(RC_INT));
2006 o(0xEE100A10|(r<<16)|(r2<<12));
2007 return;
2008 #else
2009 if(u) {
2010 if(r2 == VT_FLOAT)
2011 func=TOK___fixunssfsi;
2012 #if LDOUBLE_SIZE != 8
2013 else if(r2 == VT_LDOUBLE)
2014 func=TOK___fixunsxfsi;
2015 else if(r2 == VT_DOUBLE)
2016 #else
2017 else if(r2 == VT_LDOUBLE || r2 == VT_DOUBLE)
2018 #endif
2019 func=TOK___fixunsdfsi;
2020 } else {
2021 r=fpr(gv(RC_FLOAT));
2022 r2=intr(vtop->r=get_reg(RC_INT));
2023 o(0xEE100170|(r2<<12)|r);
2024 return;
2026 #endif
2027 } else if(t == VT_LLONG) { // unsigned handled in gen_cvt_ftoi1
2028 if(r2 == VT_FLOAT)
2029 func=TOK___fixsfdi;
2030 #if LDOUBLE_SIZE != 8
2031 else if(r2 == VT_LDOUBLE)
2032 func=TOK___fixxfdi;
2033 else if(r2 == VT_DOUBLE)
2034 #else
2035 else if(r2 == VT_LDOUBLE || r2 == VT_DOUBLE)
2036 #endif
2037 func=TOK___fixdfdi;
2039 if(func) {
2040 vpush_global_sym(&func_old_type, func);
2041 vswap();
2042 gfunc_call(1);
2043 vpushi(0);
2044 if(t == VT_LLONG)
2045 vtop->r2 = REG_LRET;
2046 vtop->r = REG_IRET;
2047 return;
2049 tcc_error("unimplemented gen_cvt_ftoi!");
2052 /* convert from one floating point type to another */
2053 void gen_cvt_ftof(int t)
2055 #ifdef TCC_ARM_VFP
2056 if(((vtop->type.t & VT_BTYPE) == VT_FLOAT) != ((t & VT_BTYPE) == VT_FLOAT)) {
2057 uint32_t r = vfpr(gv(RC_FLOAT));
2058 o(0xEEB70AC0|(r<<12)|r|T2CPR(vtop->type.t));
2060 #else
2061 /* all we have to do on i386 and FPA ARM is to put the float in a register */
2062 gv(RC_FLOAT);
2063 #endif
2066 /* computed goto support */
2067 void ggoto(void)
2069 gcall_or_jmp(1);
2070 vtop--;
2073 /* Save the stack pointer onto the stack and return the location of its address */
2074 ST_FUNC void gen_vla_sp_save(int addr) {
2075 tcc_error("variable length arrays unsupported for this target");
2078 /* Restore the SP from a location on the stack */
2079 ST_FUNC void gen_vla_sp_restore(int addr) {
2080 tcc_error("variable length arrays unsupported for this target");
2083 /* Subtract from the stack pointer, and push the resulting value onto the stack */
2084 ST_FUNC void gen_vla_alloc(CType *type, int align) {
2085 tcc_error("variable length arrays unsupported for this target");
2088 /* end of ARM code generator */
2089 /*************************************************************/
2090 #endif
2091 /*************************************************************/