arm-gen.c

   1 /*
   2  *  ARMv4 code generator for TCC
   3  *
   4  *  Copyright (c) 2003 Daniel Glöckner
   5  *  Copyright (c) 2012 Thomas Preud'homme
   6  *
   7  *  Based on i386-gen.c by Fabrice Bellard
   8  *
   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.
  13  *
  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.
  18  *
  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
  22  */
  23
  24 #ifdef TARGET_DEFS_ONLY
  25
  26 #ifdef TCC_ARM_EABI
  27 #ifndef TCC_ARM_VFP /* Avoid useless warning */
  28 #define TCC_ARM_VFP
  29 #endif
  30 #endif
  31
  32 /* number of available registers */
  33 #ifdef TCC_ARM_VFP
  34 #define NB_REGS            13
  35 #else
  36 #define NB_REGS             9
  37 #endif
  38
  39 #ifndef TCC_ARM_VERSION
  40 # define TCC_ARM_VERSION 5
  41 #endif
  42
  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 */
  66
  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
  84 };
  85
  86 #ifdef TCC_ARM_VFP
  87 #define T2CPR(t) (((t) & VT_BTYPE) != VT_FLOAT ? 0x100 : 0)
  88 #endif
  89
  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 */
  94
  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
 101
 102 /* defined if function parameters must be evaluated in reverse order */
 103 #define INVERT_FUNC_PARAMS
 104
 105 /* defined if structures are passed as pointers. Otherwise structures
 106    are directly pushed on stack. */
 107 /* #define FUNC_STRUCT_PARAM_AS_PTR */
 108
 109 /* pointer size, in bytes */
 110 #define PTR_SIZE 4
 111
 112 /* long double size and alignment, in bytes */
 113 #ifdef TCC_ARM_VFP
 114 #define LDOUBLE_SIZE  8
 115 #endif
 116
 117 #ifndef LDOUBLE_SIZE
 118 #define LDOUBLE_SIZE  8
 119 #endif
 120
 121 #ifdef TCC_ARM_EABI
 122 #define LDOUBLE_ALIGN 8
 123 #else
 124 #define LDOUBLE_ALIGN 4
 125 #endif
 126
 127 /* maximum alignment (for aligned attribute support) */
 128 #define MAX_ALIGN     8
 129
 130 #define CHAR_IS_UNSIGNED
 131
 132 /******************************************************/
 133 /* ELF defines */
 134
 135 #define EM_TCC_TARGET EM_ARM
 136
 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
 142
 143 #define ELF_START_ADDR 0x00008000
 144 #define ELF_PAGE_SIZE  0x1000
 145
 146 /******************************************************/
 147 #else /* ! TARGET_DEFS_ONLY */
 148 /******************************************************/
 149 #include "tcc.h"
 150
 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
 167 };
 168
 169 static int func_sub_sp_offset, last_itod_magic;
 170 static int leaffunc;
 171
 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)
 175 {
 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);
 182 }
 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
 189
 190 static int two2mask(int a,int b) {
 191   return (reg_classes[a]|reg_classes[b])&~(RC_INT|RC_FLOAT);
 192 }
 193
 194 static int regmask(int r) {
 195   return reg_classes[r]&~(RC_INT|RC_FLOAT);
 196 }
 197
 198 /******************************************************/
 199
 200 void o(uint32_t i)
 201 {
 202   /* this is a good place to start adding big-endian support*/
 203   int ind1;
 204
 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;
 218 }
 219
 220 static uint32_t stuff_const(uint32_t op, uint32_t c)
 221 {
 222   int try_neg=0;
 223   uint32_t nc = 0, negop = 0;
 224
 225   switch(op&0x1F00000)
 226   {
 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;
 255   }
 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));
 265     }
 266     op=negop;
 267     c=nc;
 268   } while(try_neg--);
 269   return 0;
 270 }
 271
 272
 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;
 292         }
 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;
 302         }
 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;
 311           }
 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;
 322           }
 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]));
 327   }
 328 }
 329
 330 ST_FUNC uint32_t encbranch(int pos, int addr, int fail)
 331 {
 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;
 338   }
 339   return 0x0A000000|(addr&0xffffff);
 340 }
 341
 342 int decbranch(int pos)
 343 {
 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;
 350 }
 351
 352 /* output a symbol and patch all calls to it */
 353 void gsym_addr(int t, int a)
 354 {
 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);
 365     }
 366   }
 367 }
 368
 369 void gsym(int t)
 370 {
 371   gsym_addr(t, ind);
 372 }
 373
 374 #ifdef TCC_ARM_VFP
 375 static uint32_t vfpr(int r)
 376 {
 377   if(r<TREG_F0 || r>TREG_F7)
 378     tcc_error("compiler error! register %i is no vfp register",r);
 379   return r-5;
 380 }
 381 #else
 382 static uint32_t fpr(int r)
 383 {
 384   if(r<TREG_F0 || r>TREG_F3)
 385     tcc_error("compiler error! register %i is no fpa register",r);
 386   return r-5;
 387 }
 388 #endif
 389
 390 static uint32_t intr(int r)
 391 {
 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;
 397 }
 398
 399 static void calcaddr(uint32_t *base, int *off, int *sgn, int maxoff, unsigned shift)
 400 {
 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;
 413     }
 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;
 420     }
 421     stuff_const_harder(x,*off&~maxoff);
 422     *off&=maxoff;
 423   }
 424 }
 425
 426 static uint32_t mapcc(int cc)
 427 {
 428   switch(cc)
 429   {
 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 */
 454   }
 455   tcc_error("unexpected condition code");
 456   return 0xE0000000; /* AL */
 457 }
 458
 459 static int negcc(int cc)
 460 {
 461   switch(cc)
 462   {
 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;
 487   }
 488   tcc_error("unexpected condition code");
 489   return TOK_NE;
 490 }
 491
 492 /* load 'r' from value 'sv' */
 493 void load(int r, SValue *sv)
 494 {
 495   int v, ft, fc, fr, sign;
 496   uint32_t op;
 497   SValue v1;
 498
 499   fr = sv->r;
 500   ft = sv->type.t;
 501   fc = sv->c.ul;
 502
 503   if(fc>=0)
 504     sign=0;
 505   else {
 506     sign=1;
 507     fc=-fc;
 508   }
 509
 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;
 532     }
 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));
 577       }
 578       return;
 579     }
 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;
 626     }
 627   }
 628   tcc_error("load unimplemented!");
 629 }
 630
 631 /* store register 'r' in lvalue 'v' */
 632 void store(int r, SValue *sv)
 633 {
 634   SValue v1;
 635   int v, ft, fc, fr, sign;
 636   uint32_t op;
 637
 638   fr = sv->r;
 639   ft = sv->type.t;
 640   fc = sv->c.ul;
 641
 642   if(fc>=0)
 643     sign=0;
 644   else {
 645     sign=1;
 646     fc=-fc;
 647   }
 648
 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;
 664     }
 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));
 705       }
 706       return;
 707     }
 708   }
 709   tcc_error("store unimplemented");
 710 }
 711
 712 static void gadd_sp(int val)
 713 {
 714   stuff_const_harder(0xE28DD000,val);
 715 }
 716
 717 /* 'is_jmp' is '1' if it is a jump */
 718 static void gcall_or_jmp(int is_jmp)
 719 {
 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);
 739     }
 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
 746   }
 747 }
 748
 749 /* Return 1 if this function returns via an sret pointer, 0 otherwise */
 750 ST_FUNC int gfunc_sret(CType *vt, CType *ret, int *ret_align) {
 751 #ifdef TCC_ARM_EABI
 752     int size, align;
 753     size = type_size(vt, &align);
 754     if (size > 4) {
 755         return 1;
 756     } else {
 757         *ret_align = 4;
 758         ret->ref = NULL;
 759         ret->t = VT_INT;
 760         return 0;
 761     }
 762 #else
 763     return 1;
 764 #endif
 765 }
 766
 767 #ifdef TCC_ARM_HARDFLOAT
 768 /* Return whether a structure is an homogeneous float aggregate or not.
 769    The answer is true if all the elements of the structure are of the same
 770    primitive float type and there is less than 4 elements.
 771
 772    type: the type corresponding to the structure to be tested */
 773 static int is_hgen_float_aggr(CType *type)
 774 {
 775   if ((type->t & VT_BTYPE) == VT_STRUCT) {
 776     struct Sym *ref;
 777     int btype, nb_fields = 0;
 778
 779     ref = type->ref;
 780     btype = ref->type.t & VT_BTYPE;
 781     if (btype == VT_FLOAT || btype == VT_DOUBLE) {
 782       for(; ref && btype == (ref->type.t & VT_BTYPE); ref = ref->next, nb_fields++);
 783       return !ref && nb_fields <= 4;
 784     }
 785   }
 786   return 0;
 787 }
 788
 789 struct avail_regs {
 790   signed char avail[3]; /* 3 holes max with only float and double alignments */
 791   int first_hole; /* first available hole */
 792   int last_hole; /* last available hole (none if equal to first_hole) */
 793   int first_free_reg; /* next free register in the sequence, hole excluded */
 794 };
 795
 796 #define AVAIL_REGS_INITIALIZER (struct avail_regs) { { 0, 0, 0}, 0, 0, 0 }
 797
 798 /* Find suitable registers for a VFP Co-Processor Register Candidate (VFP CPRC
 799    param) according to the rules described in the procedure call standard for
 800    the ARM architecture (AAPCS). If found, the registers are assigned to this
 801    VFP CPRC parameter. Registers are allocated in sequence unless a hole exists
 802    and the parameter is a single float.
 803
 804    avregs: opaque structure to keep track of available VFP co-processor regs
 805    align: alignment contraints for the param, as returned by type_size()
 806    size: size of the parameter, as returned by type_size() */
 807 int assign_vfpreg(struct avail_regs *avregs, int align, int size)
 808 {
 809   int first_reg = 0;
 810
 811   if (avregs->first_free_reg == -1)
 812     return -1;
 813   if (align >> 3) { /* double alignment */
 814     first_reg = avregs->first_free_reg;
 815     /* alignment contraint not respected so use next reg and record hole */
 816     if (first_reg & 1)
 817       avregs->avail[avregs->last_hole++] = first_reg++;
 818   } else { /* no special alignment (float or array of float) */
 819     /* if single float and a hole is available, assign the param to it */
 820     if (size == 4 && avregs->first_hole != avregs->last_hole)
 821       return avregs->avail[avregs->first_hole++];
 822     else
 823       first_reg = avregs->first_free_reg;
 824   }
 825   if (first_reg + size / 4 <= 16) {
 826     avregs->first_free_reg = first_reg + size / 4;
 827     return first_reg;
 828   }
 829   avregs->first_free_reg = -1;
 830   return -1;
 831 }
 832 #endif
 833
 834 /* Parameters are classified according to how they are copied to their final
 835    destination for the function call. Because the copying is performed class
 836    after class according to the order in the union below, it is important that
 837    some constraints about the order of the members of this union are respected:
 838    - CORE_STRUCT_CLASS must come after STACK_CLASS;
 839    - CORE_CLASS must come after STACK_CLASS, CORE_STRUCT_CLASS and
 840      VFP_STRUCT_CLASS;
 841    - VFP_STRUCT_CLASS must come after VFP_CLASS.
 842    See the comment for the main loop in copy_params() for the reason. */
 843 enum reg_class {
 844         STACK_CLASS = 0,
 845         CORE_STRUCT_CLASS,
 846         VFP_CLASS,
 847         VFP_STRUCT_CLASS,
 848         CORE_CLASS,
 849         NB_CLASSES
 850 };
 851
 852 struct param_plan {
 853     int start; /* first reg or addr used depending on the class */
 854     int end; /* last reg used or next free addr depending on the class */
 855     SValue *sval; /* pointer to SValue on the value stack */
 856     struct param_plan *prev; /*  previous element in this class */
 857 };
 858
 859 struct plan {
 860     struct param_plan *pplans; /* array of all the param plans */
 861     struct param_plan *clsplans[NB_CLASSES]; /* per class lists of param plans */
 862 };
 863
 864 #define add_param_plan(plan,pplan,class)                        \
 865     do {                                                        \
 866         pplan.prev = plan->clsplans[class];                     \
 867         plan->pplans[plan ## _nb] = pplan;                      \
 868         plan->clsplans[class] = &plan->pplans[plan ## _nb++];   \
 869     } while(0)
 870
 871 /* Assign parameters to registers and stack with alignment according to the
 872    rules in the procedure call standard for the ARM architecture (AAPCS).
 873    The overall assignment is recorded in an array of per parameter structures
 874    called parameter plans. The parameter plans are also further organized in a
 875    number of linked lists, one per class of parameter (see the comment for the
 876    definition of union reg_class).
 877
 878    nb_args: number of parameters of the function for which a call is generated
 879    variadic: whether the function is a variadic function or not
 880    plan: the structure where the overall assignment is recorded
 881    todo: a bitmap that record which core registers hold a parameter
 882
 883    Returns the amount of stack space needed for parameter passing
 884
 885    Note: this function allocated an array in plan->pplans with tcc_malloc. It
 886    is the responsability of the caller to free this array once used (ie not
 887    before copy_params). */
 888 static int assign_regs(int nb_args, int variadic, struct plan *plan, int *todo)
 889 {
 890   int i, size, align;
 891   int ncrn /* next core register number */, nsaa /* next stacked argument address*/;
 892   int plan_nb = 0;
 893   struct param_plan pplan;
 894 #ifdef TCC_ARM_HARDFLOAT
 895   struct avail_regs avregs = AVAIL_REGS_INITIALIZER;
 896 #endif
 897
 898   ncrn = nsaa = 0;
 899   *todo = 0;
 900   plan->pplans = tcc_malloc(nb_args * sizeof(*plan->pplans));
 901   memset(plan->clsplans, 0, sizeof(plan->clsplans));
 902   for(i = nb_args; i-- ;) {
 903     int j, start_vfpreg = 0;
 904     size = type_size(&vtop[-i].type, &align);
 905     switch(vtop[-i].type.t & VT_BTYPE) {
 906       case VT_STRUCT:
 907       case VT_FLOAT:
 908       case VT_DOUBLE:
 909       case VT_LDOUBLE:
 910 #ifdef TCC_ARM_HARDFLOAT
 911       if (!variadic) {
 912         int is_hfa = 0; /* Homogeneous float aggregate */
 913
 914         if (is_float(vtop[-i].type.t)
 915             || (is_hfa = is_hgen_float_aggr(&vtop[-i].type))) {
 916           int end_vfpreg;
 917
 918           start_vfpreg = assign_vfpreg(&avregs, align, size);
 919           end_vfpreg = start_vfpreg + ((size - 1) >> 2);
 920           if (start_vfpreg >= 0) {
 921             pplan = (struct param_plan) {start_vfpreg, end_vfpreg, &vtop[-i]};
 922             if (is_hfa)
 923               add_param_plan(plan, pplan, VFP_STRUCT_CLASS);
 924             else
 925               add_param_plan(plan, pplan, VFP_CLASS);
 926             continue;
 927           } else
 928             break;
 929         }
 930       }
 931 #endif
 932       ncrn = (ncrn + (align-1)/4) & -(align/4);
 933       size = (size + 3) & -4;
 934       if (ncrn + size/4 <= 4 || (ncrn < 4 && start_vfpreg != -1)) {
 935         /* The parameter is allocated both in core register and on stack. As
 936          * such, it can be of either class: it would either be the last of
 937          * CORE_STRUCT_CLASS or the first of STACK_CLASS. */
 938         for (j = ncrn; j < 4 && j < ncrn + size / 4; j++)
 939           *todo|=(1<<j);
 940         pplan = (struct param_plan) {ncrn, j, &vtop[-i]};
 941         add_param_plan(plan, pplan, CORE_STRUCT_CLASS);
 942         ncrn += size/4;
 943         if (ncrn > 4)
 944           nsaa = (ncrn - 4) * 4;
 945       } else {
 946         ncrn = 4;
 947         break;
 948       }
 949       continue;
 950       default:
 951       if (ncrn < 4) {
 952         int is_long = (vtop[-i].type.t & VT_BTYPE) == VT_LLONG;
 953
 954         if (is_long) {
 955           ncrn = (ncrn + 1) & -2;
 956           if (ncrn == 4)
 957             break;
 958         }
 959         pplan = (struct param_plan) {ncrn, ncrn, &vtop[-i]};
 960         ncrn++;
 961         if (is_long)
 962           pplan.end = ncrn++;
 963         add_param_plan(plan, pplan, CORE_CLASS);
 964         continue;
 965       }
 966     }
 967     nsaa = (nsaa + (align - 1)) & ~(align - 1);
 968     pplan = (struct param_plan) {nsaa, nsaa + size, &vtop[-i]};
 969     add_param_plan(plan, pplan, STACK_CLASS);
 970     nsaa += size; /* size already rounded up before */
 971   }
 972   return nsaa;
 973 }
 974
 975 #undef add_param_plan
 976
 977 /* Copy parameters to their final destination (core reg, VFP reg or stack) for
 978    function call.
 979
 980    nb_args: number of parameters the function take
 981    plan: the overall assignment plan for parameters
 982    todo: a bitmap indicating what core reg will hold a parameter */
 983 static void copy_params(int nb_args, struct plan *plan, int todo)
 984 {
 985   int size, align, r, i;
 986   struct param_plan *pplan;
 987
 988   /* Put argument on stack (structure are put on stack no matter how they are
 989    * passed via register or the stack). */
 990 #ifdef TCC_ARM_EABI
 991   if ((pplan = plan->clsplans[STACK_CLASS]) && pplan->end & 7)
 992     o(0xE24DD004); /* sub sp, sp, #4 */
 993 #endif
 994    /* Several constraints require parameters to be copied in a specific order:
 995       - structures are copied to the stack before being loaded in a reg;
 996       - floats loaded to an odd numbered VFP reg are first copied to the
 997         preceding even numbered VFP reg and then moved to the next VFP reg.
 998
 999       It is thus important that:
1000       - structures assigned to core regs must be copied after parameters
1001         assigned to the stack but before structures assigned to VFP regs because
1002         a structure can lie partly in core registers and partly on the stack;
1003       - parameters assigned to the stack and all structures be copied before
1004         parameters assigned to a core reg since copying a parameter to the stack
1005         require using a core reg;
1006       - parameters assigned to VFP regs be copied before structures assigned to
1007         VFP regs as the copy might use an even numbered VFP reg that already
1008         holds part of a structure. */
1009   for(i = 0; i < NB_CLASSES; i++) {
1010     for(pplan = plan->clsplans[i]; pplan; pplan = pplan->prev) {
1011       vpushv(pplan->sval);
1012       pplan->sval->r = pplan->sval->r2 = VT_CONST; /* disable entry */
1013       switch(i) {
1014         case STACK_CLASS:
1015         case CORE_STRUCT_CLASS:
1016         case VFP_STRUCT_CLASS:
1017           if ((pplan->sval->type.t & VT_BTYPE) == VT_STRUCT) {
1018             size = type_size(&pplan->sval->type, &align);
1019             /* align to stack align size */
1020             size = (size + 3) & ~3;
1021             if (i == STACK_CLASS && pplan->prev)
1022               size += pplan->start - pplan->prev->end; /* Add padding if any */
1023             /* allocate the necessary size on stack */
1024             gadd_sp(-size);
1025             /* generate structure store */
1026             r = get_reg(RC_INT);
1027             o(0xE1A0000D|(intr(r)<<12)); /* mov r, sp */
1028             vset(&vtop->type, r | VT_LVAL, 0);
1029             vswap();
1030             vstore(); /* memcpy to current sp */
1031             /* Homogeneous float aggregate are loaded to VFP registers
1032                immediately since there is no way of loading data in multiple
1033                non consecutive VFP registers as what is done for other
1034                structures (see the use of todo). */
1035             if (i == VFP_STRUCT_CLASS) {
1036               int first = pplan->start, nb = pplan->end - first + 1;
1037               /* vpop.32 {pplan->start, ..., pplan->end} */
1038               o(0xECBD0A00|(first&1)<<22|(first>>1)<<12|nb);
1039               /* No need to write the register used to a SValue since VFP regs
1040                  cannot be used for gcall_or_jmp */
1041             }
1042           } else {
1043             if (is_float(pplan->sval->type.t)) {
1044 #ifdef TCC_ARM_VFP
1045               r = vfpr(gv(RC_FLOAT)) << 12;
1046               if ((pplan->sval->type.t & VT_BTYPE) == VT_FLOAT)
1047                 size = 4;
1048               else {
1049                 size = 8;
1050                 r |= 0x101; /* vpush.32 -> vpush.64 */
1051               }
1052               o(0xED2D0A01 + r); /* vpush */
1053 #else
1054               r = fpr(gv(RC_FLOAT)) << 12;
1055               if ((pplan->sval->type.t & VT_BTYPE) == VT_FLOAT)
1056                 size = 4;
1057               else if ((pplan->sval->type.t & VT_BTYPE) == VT_DOUBLE)
1058                 size = 8;
1059               else
1060                 size = LDOUBLE_SIZE;
1061
1062               if (size == 12)
1063                 r |= 0x400000;
1064               else if(size == 8)
1065                 r|=0x8000;
1066
1067               o(0xED2D0100|r|(size>>2)); /* some kind of vpush for FPA */
1068 #endif
1069             } else {
1070               /* simple type (currently always same size) */
1071               /* XXX: implicit cast ? */
1072               size=4;
1073               if ((pplan->sval->type.t & VT_BTYPE) == VT_LLONG) {
1074                 lexpand_nr();
1075                 size = 8;
1076                 r = gv(RC_INT);
1077                 o(0xE52D0004|(intr(r)<<12)); /* push r */
1078                 vtop--;
1079               }
1080               r = gv(RC_INT);
1081               o(0xE52D0004|(intr(r)<<12)); /* push r */
1082             }
1083             if (i == STACK_CLASS && pplan->prev)
1084               gadd_sp(pplan->prev->end - pplan->start); /* Add padding if any */
1085           }
1086           break;
1087
1088         case VFP_CLASS:
1089           gv(regmask(TREG_F0 + (pplan->start >> 1)));
1090           if (pplan->start & 1) { /* Must be in upper part of double register */
1091             o(0xEEF00A40|((pplan->start>>1)<<12)|(pplan->start>>1)); /* vmov.f32 s(n+1), sn */
1092             vtop->r = VT_CONST; /* avoid being saved on stack by gv for next float */
1093           }
1094           break;
1095
1096         case CORE_CLASS:
1097           if ((pplan->sval->type.t & VT_BTYPE) == VT_LLONG) {
1098             lexpand_nr();
1099             gv(regmask(pplan->end));
1100             pplan->sval->r2 = vtop->r;
1101             vtop--;
1102           }
1103           gv(regmask(pplan->start));
1104           /* Mark register as used so that gcall_or_jmp use another one
1105              (regs >=4 are free as never used to pass parameters) */
1106           pplan->sval->r = vtop->r;
1107           break;
1108       }
1109       vtop--;
1110     }
1111   }
1112
1113   /* Manually free remaining registers since next parameters are loaded
1114    * manually, without the help of gv(int). */
1115   save_regs(nb_args);
1116
1117   if(todo) {
1118     o(0xE8BD0000|todo); /* pop {todo} */
1119     for(pplan = plan->clsplans[CORE_STRUCT_CLASS]; pplan; pplan = pplan->prev) {
1120       pplan->sval->r = pplan->start;
1121       if ((pplan->sval->type.t & VT_BTYPE) == VT_LLONG)
1122         pplan->sval->r2 = pplan->end;
1123     }
1124   }
1125 }
1126
1127 /* Generate function call. The function address is pushed first, then
1128    all the parameters in call order. This functions pops all the
1129    parameters and the function address. */
1130 void gfunc_call(int nb_args)
1131 {
1132   int align, r, args_size;
1133   int variadic;
1134   int todo;
1135   struct plan plan;
1136
1137   variadic = (vtop[-nb_args].type.ref->c == FUNC_ELLIPSIS);
1138   /* cannot let cpu flags if other instruction are generated. Also avoid leaving
1139      VT_JMP anywhere except on the top of the stack because it would complicate
1140      the code generator. */
1141   r = vtop->r & VT_VALMASK;
1142   if (r == VT_CMP || (r & ~1) == VT_JMP)
1143     gv(RC_INT);
1144 #ifdef TCC_ARM_EABI
1145   /* return type is a struct so caller of gfunc_call (unary(void) in tccgen.c)
1146      assumed it had to be passed by a pointer. Since it's less than 4 bytes, we
1147      can actually pass it directly in a register. */
1148   if((vtop[-nb_args].type.ref->type.t & VT_BTYPE) == VT_STRUCT
1149      && type_size(&vtop[-nb_args].type.ref->type, &align) <= 4) {
1150     SValue tmp;
1151     tmp=vtop[-nb_args];
1152     vtop[-nb_args]=vtop[-nb_args+1];
1153     vtop[-nb_args+1]=tmp;
1154     --nb_args;
1155   }
1156 #endif
1157
1158   args_size = assign_regs(nb_args, variadic, &plan, &todo);
1159   copy_params(nb_args, &plan, todo);
1160   tcc_free(plan.pplans);
1161
1162   /* Move fct SValue on top as required by gcall_or_jmp */
1163   vrotb(nb_args + 1);
1164   gcall_or_jmp(0);
1165   if (args_size)
1166       gadd_sp(args_size); /* pop all parameters passed on the stack */
1167 #ifdef TCC_ARM_EABI
1168   if((vtop->type.ref->type.t & VT_BTYPE) == VT_STRUCT
1169      && type_size(&vtop->type.ref->type, &align) <= 4) {
1170     store(REG_IRET,vtop-nb_args-1);
1171     nb_args++;
1172   }
1173 #ifdef TCC_ARM_VFP
1174 #ifdef TCC_ARM_HARDFLOAT
1175   else if(variadic && is_float(vtop->type.ref->type.t)) {
1176 #else
1177   else if(is_float(vtop->type.ref->type.t)) {
1178 #endif
1179     if((vtop->type.ref->type.t & VT_BTYPE) == VT_FLOAT) {
1180       o(0xEE000A10); /*vmov s0, r0 */
1181     } else {
1182       o(0xEE000B10); /* vmov.32 d0[0], r0 */
1183       o(0xEE201B10); /* vmov.32 d0[1], r1 */
1184     }
1185   }
1186 #endif
1187 #endif
1188   vtop -= nb_args + 1; /* Pop all params and fct address from value stack */
1189   leaffunc = 0; /* we are calling a function, so we aren't in a leaf function */
1190 }
1191
1192 /* generate function prolog of type 't' */
1193 void gfunc_prolog(CType *func_type)
1194 {
1195   Sym *sym,*sym2;
1196   int n,nf,size,align, variadic, struct_ret = 0;
1197 #ifdef TCC_ARM_HARDFLOAT
1198   struct avail_regs avregs = AVAIL_REGS_INITIALIZER;
1199 #endif
1200
1201   sym = func_type->ref;
1202   func_vt = sym->type;
1203
1204   n = nf = 0;
1205   variadic = (func_type->ref->c == FUNC_ELLIPSIS);
1206   if((func_vt.t & VT_BTYPE) == VT_STRUCT
1207      && type_size(&func_vt,&align) > 4)
1208   {
1209     n++;
1210     struct_ret = 1;
1211     func_vc = 12; /* Offset from fp of the place to store the result */
1212   }
1213   for(sym2=sym->next;sym2 && (n<4 || nf<16);sym2=sym2->next) {
1214     size = type_size(&sym2->type, &align);
1215 #ifdef TCC_ARM_HARDFLOAT
1216     if (!variadic && (is_float(sym2->type.t)
1217         || is_hgen_float_aggr(&sym2->type))) {
1218       int tmpnf = assign_vfpreg(&avregs, align, size) + 1;
1219       nf = (tmpnf > nf) ? tmpnf : nf;
1220     } else
1221 #endif
1222     if (n < 4)
1223       n += (size + 3) / 4;
1224   }
1225   o(0xE1A0C00D); /* mov ip,sp */
1226   if(variadic)
1227     n=4;
1228   if(n) {
1229     if(n>4)
1230       n=4;
1231 #ifdef TCC_ARM_EABI
1232     n=(n+1)&-2;
1233 #endif
1234     o(0xE92D0000|((1<<n)-1)); /* save r0-r4 on stack if needed */
1235   }
1236   if (nf) {
1237     if (nf>16)
1238       nf=16;
1239     nf=(nf+1)&-2; /* nf => HARDFLOAT => EABI */
1240     o(0xED2D0A00|nf); /* save s0-s15 on stack if needed */
1241   }
1242   o(0xE92D5800); /* save fp, ip, lr */
1243   o(0xE1A0B00D); /* mov fp, sp */
1244   func_sub_sp_offset = ind;
1245   o(0xE1A00000); /* nop, leave space for stack adjustment in epilogue */
1246   {
1247     int addr, pn = struct_ret, sn = 0; /* pn=core, sn=stack */
1248
1249 #ifdef TCC_ARM_HARDFLOAT
1250     func_vc += nf * 4;
1251     avregs = AVAIL_REGS_INITIALIZER;
1252 #endif
1253     while ((sym = sym->next)) {
1254       CType *type;
1255       type = &sym->type;
1256       size = type_size(type, &align);
1257       size = (size + 3) >> 2;
1258       align = (align + 3) & ~3;
1259 #ifdef TCC_ARM_HARDFLOAT
1260       if (!variadic && (is_float(sym->type.t)
1261           || is_hgen_float_aggr(&sym->type))) {
1262         int fpn = assign_vfpreg(&avregs, align, size << 2);
1263         if (fpn >= 0) {
1264           addr = fpn * 4;
1265         } else
1266           goto from_stack;
1267       } else
1268 #endif
1269       if (pn < 4) {
1270 #ifdef TCC_ARM_EABI
1271         pn = (pn + (align-1)/4) & -(align/4);
1272 #endif
1273         addr = (nf + pn) * 4;
1274         pn += size;
1275         if (!sn && pn > 4)
1276           sn = (pn - 4);
1277       } else {
1278 #ifdef TCC_ARM_HARDFLOAT
1279 from_stack:
1280 #endif
1281 #ifdef TCC_ARM_EABI
1282         sn = (sn + (align-1)/4) & -(align/4);
1283 #endif
1284         addr = (n + nf + sn) * 4;
1285         sn += size;
1286       }
1287       sym_push(sym->v & ~SYM_FIELD, type, VT_LOCAL | lvalue_type(type->t), addr+12);
1288     }
1289   }
1290   last_itod_magic=0;
1291   leaffunc = 1;
1292   loc = 0;
1293 }
1294
1295 /* generate function epilog */
1296 void gfunc_epilog(void)
1297 {
1298   uint32_t x;
1299   int diff;
1300 #ifdef TCC_ARM_EABI
1301   /* Useless but harmless copy of the float result into main register(s) in case
1302      of variadic function in the hardfloat variant */
1303   if(is_float(func_vt.t)) {
1304     if((func_vt.t & VT_BTYPE) == VT_FLOAT)
1305       o(0xEE100A10); /* fmrs r0, s0 */
1306     else {
1307       o(0xEE100B10); /* fmrdl r0, d0 */
1308       o(0xEE301B10); /* fmrdh r1, d0 */
1309     }
1310   }
1311 #endif
1312   o(0xE89BA800); /* restore fp, sp, pc */
1313   diff = (-loc + 3) & -4;
1314 #ifdef TCC_ARM_EABI
1315   if(!leaffunc)
1316     diff = ((diff + 11) & -8) - 4;
1317 #endif
1318   if(diff > 0) {
1319     x=stuff_const(0xE24BD000, diff); /* sub sp,fp,# */
1320     if(x)
1321       *(uint32_t *)(cur_text_section->data + func_sub_sp_offset) = x;
1322     else {
1323       int addr;
1324       addr=ind;
1325       o(0xE59FC004); /* ldr ip,[pc+4] */
1326       o(0xE04BD00C); /* sub sp,fp,ip  */
1327       o(0xE1A0F00E); /* mov pc,lr */
1328       o(diff);
1329       *(uint32_t *)(cur_text_section->data + func_sub_sp_offset) = 0xE1000000|encbranch(func_sub_sp_offset,addr,1);
1330     }
1331   }
1332 }
1333
1334 /* generate a jump to a label */
1335 int gjmp(int t)
1336 {
1337   int r;
1338   r=ind;
1339   o(0xE0000000|encbranch(r,t,1));
1340   return r;
1341 }
1342
1343 /* generate a jump to a fixed address */
1344 void gjmp_addr(int a)
1345 {
1346   gjmp(a);
1347 }
1348
1349 /* generate a test. set 'inv' to invert test. Stack entry is popped */
1350 int gtst(int inv, int t)
1351 {
1352   int v, r;
1353   uint32_t op;
1354   v = vtop->r & VT_VALMASK;
1355   r=ind;
1356   if (v == VT_CMP) {
1357     op=mapcc(inv?negcc(vtop->c.i):vtop->c.i);
1358     op|=encbranch(r,t,1);
1359     o(op);
1360     t=r;
1361   } else if (v == VT_JMP || v == VT_JMPI) {
1362     if ((v & 1) == inv) {
1363       if(!vtop->c.i)
1364         vtop->c.i=t;
1365       else {
1366         uint32_t *x;
1367         int p,lp;
1368         if(t) {
1369           p = vtop->c.i;
1370           do {
1371             p = decbranch(lp=p);
1372           } while(p);
1373           x = (uint32_t *)(cur_text_section->data + lp);
1374           *x &= 0xff000000;
1375           *x |= encbranch(lp,t,1);
1376         }
1377         t = vtop->c.i;
1378       }
1379     } else {
1380       t = gjmp(t);
1381       gsym(vtop->c.i);
1382     }
1383   } else {
1384     if (is_float(vtop->type.t)) {
1385       r=gv(RC_FLOAT);
1386 #ifdef TCC_ARM_VFP
1387       o(0xEEB50A40|(vfpr(r)<<12)|T2CPR(vtop->type.t)); /* fcmpzX */
1388       o(0xEEF1FA10); /* fmstat */
1389 #else
1390       o(0xEE90F118|(fpr(r)<<16));
1391 #endif
1392       vtop->r = VT_CMP;
1393       vtop->c.i = TOK_NE;
1394       return gtst(inv, t);
1395     } else if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
1396       /* constant jmp optimization */
1397       if ((vtop->c.i != 0) != inv)
1398         t = gjmp(t);
1399     } else {
1400       v = gv(RC_INT);
1401       o(0xE3300000|(intr(v)<<16));
1402       vtop->r = VT_CMP;
1403       vtop->c.i = TOK_NE;
1404       return gtst(inv, t);
1405     }
1406   }
1407   vtop--;
1408   return t;
1409 }
1410
1411 /* generate an integer binary operation */
1412 void gen_opi(int op)
1413 {
1414   int c, func = 0;
1415   uint32_t opc = 0, r, fr;
1416   unsigned short retreg = REG_IRET;
1417
1418   c=0;
1419   switch(op) {
1420     case '+':
1421       opc = 0x8;
1422       c=1;
1423       break;
1424     case TOK_ADDC1: /* add with carry generation */
1425       opc = 0x9;
1426       c=1;
1427       break;
1428     case '-':
1429       opc = 0x4;
1430       c=1;
1431       break;
1432     case TOK_SUBC1: /* sub with carry generation */
1433       opc = 0x5;
1434       c=1;
1435       break;
1436     case TOK_ADDC2: /* add with carry use */
1437       opc = 0xA;
1438       c=1;
1439       break;
1440     case TOK_SUBC2: /* sub with carry use */
1441       opc = 0xC;
1442       c=1;
1443       break;
1444     case '&':
1445       opc = 0x0;
1446       c=1;
1447       break;
1448     case '^':
1449       opc = 0x2;
1450       c=1;
1451       break;
1452     case '|':
1453       opc = 0x18;
1454       c=1;
1455       break;
1456     case '*':
1457       gv2(RC_INT, RC_INT);
1458       r = vtop[-1].r;
1459       fr = vtop[0].r;
1460       vtop--;
1461       o(0xE0000090|(intr(r)<<16)|(intr(r)<<8)|intr(fr));
1462       return;
1463     case TOK_SHL:
1464       opc = 0;
1465       c=2;
1466       break;
1467     case TOK_SHR:
1468       opc = 1;
1469       c=2;
1470       break;
1471     case TOK_SAR:
1472       opc = 2;
1473       c=2;
1474       break;
1475     case '/':
1476     case TOK_PDIV:
1477       func=TOK___divsi3;
1478       c=3;
1479       break;
1480     case TOK_UDIV:
1481       func=TOK___udivsi3;
1482       c=3;
1483       break;
1484     case '%':
1485 #ifdef TCC_ARM_EABI
1486       func=TOK___aeabi_idivmod;
1487       retreg=REG_LRET;
1488 #else
1489       func=TOK___modsi3;
1490 #endif
1491       c=3;
1492       break;
1493     case TOK_UMOD:
1494 #ifdef TCC_ARM_EABI
1495       func=TOK___aeabi_uidivmod;
1496       retreg=REG_LRET;
1497 #else
1498       func=TOK___umodsi3;
1499 #endif
1500       c=3;
1501       break;
1502     case TOK_UMULL:
1503       gv2(RC_INT, RC_INT);
1504       r=intr(vtop[-1].r2=get_reg(RC_INT));
1505       c=vtop[-1].r;
1506       vtop[-1].r=get_reg_ex(RC_INT,regmask(c));
1507       vtop--;
1508       o(0xE0800090|(r<<16)|(intr(vtop->r)<<12)|(intr(c)<<8)|intr(vtop[1].r));
1509       return;
1510     default:
1511       opc = 0x15;
1512       c=1;
1513       break;
1514   }
1515   switch(c) {
1516     case 1:
1517       if((vtop[-1].r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
1518         if(opc == 4 || opc == 5 || opc == 0xc) {
1519           vswap();
1520           opc|=2; // sub -> rsb
1521         }
1522       }
1523       if ((vtop->r & VT_VALMASK) == VT_CMP ||
1524           (vtop->r & (VT_VALMASK & ~1)) == VT_JMP)
1525         gv(RC_INT);
1526       vswap();
1527       c=intr(gv(RC_INT));
1528       vswap();
1529       opc=0xE0000000|(opc<<20)|(c<<16);
1530       if((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
1531         uint32_t x;
1532         x=stuff_const(opc|0x2000000,vtop->c.i);
1533         if(x) {
1534           r=intr(vtop[-1].r=get_reg_ex(RC_INT,regmask(vtop[-1].r)));
1535           o(x|(r<<12));
1536           goto done;
1537         }
1538       }
1539       fr=intr(gv(RC_INT));
1540       r=intr(vtop[-1].r=get_reg_ex(RC_INT,two2mask(vtop->r,vtop[-1].r)));
1541       o(opc|(r<<12)|fr);
1542 done:
1543       vtop--;
1544       if (op >= TOK_ULT && op <= TOK_GT) {
1545         vtop->r = VT_CMP;
1546         vtop->c.i = op;
1547       }
1548       break;
1549     case 2:
1550       opc=0xE1A00000|(opc<<5);
1551       if ((vtop->r & VT_VALMASK) == VT_CMP ||
1552           (vtop->r & (VT_VALMASK & ~1)) == VT_JMP)
1553         gv(RC_INT);
1554       vswap();
1555       r=intr(gv(RC_INT));
1556       vswap();
1557       opc|=r;
1558       if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
1559         fr=intr(vtop[-1].r=get_reg_ex(RC_INT,regmask(vtop[-1].r)));
1560         c = vtop->c.i & 0x1f;
1561         o(opc|(c<<7)|(fr<<12));
1562       } else {
1563         fr=intr(gv(RC_INT));
1564         c=intr(vtop[-1].r=get_reg_ex(RC_INT,two2mask(vtop->r,vtop[-1].r)));
1565         o(opc|(c<<12)|(fr<<8)|0x10);
1566       }
1567       vtop--;
1568       break;
1569     case 3:
1570       vpush_global_sym(&func_old_type, func);
1571       vrott(3);
1572       gfunc_call(2);
1573       vpushi(0);
1574       vtop->r = retreg;
1575       break;
1576     default:
1577       tcc_error("gen_opi %i unimplemented!",op);
1578   }
1579 }
1580
1581 #ifdef TCC_ARM_VFP
1582 static int is_zero(int i)
1583 {
1584   if((vtop[i].r & (VT_VALMASK | VT_LVAL | VT_SYM)) != VT_CONST)
1585     return 0;
1586   if (vtop[i].type.t == VT_FLOAT)
1587     return (vtop[i].c.f == 0.f);
1588   else if (vtop[i].type.t == VT_DOUBLE)
1589     return (vtop[i].c.d == 0.0);
1590   return (vtop[i].c.ld == 0.l);
1591 }
1592
1593 /* generate a floating point operation 'v = t1 op t2' instruction. The
1594  *    two operands are guaranted to have the same floating point type */
1595 void gen_opf(int op)
1596 {
1597   uint32_t x;
1598   int fneg=0,r;
1599   x=0xEE000A00|T2CPR(vtop->type.t);
1600   switch(op) {
1601     case '+':
1602       if(is_zero(-1))
1603         vswap();
1604       if(is_zero(0)) {
1605         vtop--;
1606         return;
1607       }
1608       x|=0x300000;
1609       break;
1610     case '-':
1611       x|=0x300040;
1612       if(is_zero(0)) {
1613         vtop--;
1614         return;
1615       }
1616       if(is_zero(-1)) {
1617         x|=0x810000; /* fsubX -> fnegX */
1618         vswap();
1619         vtop--;
1620         fneg=1;
1621       }
1622       break;
1623     case '*':
1624       x|=0x200000;
1625       break;
1626     case '/':
1627       x|=0x800000;
1628       break;
1629     default:
1630       if(op < TOK_ULT || op > TOK_GT) {
1631         tcc_error("unknown fp op %x!",op);
1632         return;
1633       }
1634       if(is_zero(-1)) {
1635         vswap();
1636         switch(op) {
1637           case TOK_LT: op=TOK_GT; break;
1638           case TOK_GE: op=TOK_ULE; break;
1639           case TOK_LE: op=TOK_GE; break;
1640           case TOK_GT: op=TOK_ULT; break;
1641         }
1642       }
1643       x|=0xB40040; /* fcmpX */
1644       if(op!=TOK_EQ && op!=TOK_NE)
1645         x|=0x80; /* fcmpX -> fcmpeX */
1646       if(is_zero(0)) {
1647         vtop--;
1648         o(x|0x10000|(vfpr(gv(RC_FLOAT))<<12)); /* fcmp(e)X -> fcmp(e)zX */
1649       } else {
1650         x|=vfpr(gv(RC_FLOAT));
1651         vswap();
1652         o(x|(vfpr(gv(RC_FLOAT))<<12));
1653         vtop--;
1654       }
1655       o(0xEEF1FA10); /* fmstat */
1656
1657       switch(op) {
1658         case TOK_LE: op=TOK_ULE; break;
1659         case TOK_LT: op=TOK_ULT; break;
1660         case TOK_UGE: op=TOK_GE; break;
1661         case TOK_UGT: op=TOK_GT; break;
1662       }
1663
1664       vtop->r = VT_CMP;
1665       vtop->c.i = op;
1666       return;
1667   }
1668   r=gv(RC_FLOAT);
1669   x|=vfpr(r);
1670   r=regmask(r);
1671   if(!fneg) {
1672     int r2;
1673     vswap();
1674     r2=gv(RC_FLOAT);
1675     x|=vfpr(r2)<<16;
1676     r|=regmask(r2);
1677   }
1678   vtop->r=get_reg_ex(RC_FLOAT,r);
1679   if(!fneg)
1680     vtop--;
1681   o(x|(vfpr(vtop->r)<<12));
1682 }
1683
1684 #else
1685 static uint32_t is_fconst()
1686 {
1687   long double f;
1688   uint32_t r;
1689   if((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) != VT_CONST)
1690     return 0;
1691   if (vtop->type.t == VT_FLOAT)
1692     f = vtop->c.f;
1693   else if (vtop->type.t == VT_DOUBLE)
1694     f = vtop->c.d;
1695   else
1696     f = vtop->c.ld;
1697   if(!ieee_finite(f))
1698     return 0;
1699   r=0x8;
1700   if(f<0.0) {
1701     r=0x18;
1702     f=-f;
1703   }
1704   if(f==0.0)
1705     return r;
1706   if(f==1.0)
1707     return r|1;
1708   if(f==2.0)
1709     return r|2;
1710   if(f==3.0)
1711     return r|3;
1712   if(f==4.0)
1713     return r|4;
1714   if(f==5.0)
1715     return r|5;
1716   if(f==0.5)
1717     return r|6;
1718   if(f==10.0)
1719     return r|7;
1720   return 0;
1721 }
1722
1723 /* generate a floating point operation 'v = t1 op t2' instruction. The
1724    two operands are guaranted to have the same floating point type */
1725 void gen_opf(int op)
1726 {
1727   uint32_t x, r, r2, c1, c2;
1728   //fputs("gen_opf\n",stderr);
1729   vswap();
1730   c1 = is_fconst();
1731   vswap();
1732   c2 = is_fconst();
1733   x=0xEE000100;
1734 #if LDOUBLE_SIZE == 8
1735   if ((vtop->type.t & VT_BTYPE) != VT_FLOAT)
1736     x|=0x80;
1737 #else
1738   if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
1739     x|=0x80;
1740   else if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE)
1741     x|=0x80000;
1742 #endif
1743   switch(op)
1744   {
1745     case '+':
1746       if(!c2) {
1747         vswap();
1748         c2=c1;
1749       }
1750       vswap();
1751       r=fpr(gv(RC_FLOAT));
1752       vswap();
1753       if(c2) {
1754         if(c2>0xf)
1755           x|=0x200000; // suf
1756         r2=c2&0xf;
1757       } else {
1758         r2=fpr(gv(RC_FLOAT));
1759       }
1760       break;
1761     case '-':
1762       if(c2) {
1763         if(c2<=0xf)
1764           x|=0x200000; // suf
1765         r2=c2&0xf;
1766         vswap();
1767         r=fpr(gv(RC_FLOAT));
1768         vswap();
1769       } else if(c1 && c1<=0xf) {
1770         x|=0x300000; // rsf
1771         r2=c1;
1772         r=fpr(gv(RC_FLOAT));
1773         vswap();
1774       } else {
1775         x|=0x200000; // suf
1776         vswap();
1777         r=fpr(gv(RC_FLOAT));
1778         vswap();
1779         r2=fpr(gv(RC_FLOAT));
1780       }
1781       break;
1782     case '*':
1783       if(!c2 || c2>0xf) {
1784         vswap();
1785         c2=c1;
1786       }
1787       vswap();
1788       r=fpr(gv(RC_FLOAT));
1789       vswap();
1790       if(c2 && c2<=0xf)
1791         r2=c2;
1792       else
1793         r2=fpr(gv(RC_FLOAT));
1794       x|=0x100000; // muf
1795       break;
1796     case '/':
1797       if(c2 && c2<=0xf) {
1798         x|=0x400000; // dvf
1799         r2=c2;
1800         vswap();
1801         r=fpr(gv(RC_FLOAT));
1802         vswap();
1803       } else if(c1 && c1<=0xf) {
1804         x|=0x500000; // rdf
1805         r2=c1;
1806         r=fpr(gv(RC_FLOAT));
1807         vswap();
1808       } else {
1809         x|=0x400000; // dvf
1810         vswap();
1811         r=fpr(gv(RC_FLOAT));
1812         vswap();
1813         r2=fpr(gv(RC_FLOAT));
1814       }
1815       break;
1816     default:
1817       if(op >= TOK_ULT && op <= TOK_GT) {
1818         x|=0xd0f110; // cmfe
1819 /* bug (intention?) in Linux FPU emulator
1820    doesn't set carry if equal */
1821         switch(op) {
1822           case TOK_ULT:
1823           case TOK_UGE:
1824           case TOK_ULE:
1825           case TOK_UGT:
1826             tcc_error("unsigned comparision on floats?");
1827             break;
1828           case TOK_LT:
1829             op=TOK_Nset;
1830             break;
1831           case TOK_LE:
1832             op=TOK_ULE; /* correct in unordered case only if AC bit in FPSR set */
1833             break;
1834           case TOK_EQ:
1835           case TOK_NE:
1836             x&=~0x400000; // cmfe -> cmf
1837             break;
1838         }
1839         if(c1 && !c2) {
1840           c2=c1;
1841           vswap();
1842           switch(op) {
1843             case TOK_Nset:
1844               op=TOK_GT;
1845               break;
1846             case TOK_GE:
1847               op=TOK_ULE;
1848               break;
1849             case TOK_ULE:
1850               op=TOK_GE;
1851               break;
1852             case TOK_GT:
1853               op=TOK_Nset;
1854               break;
1855           }
1856         }
1857         vswap();
1858         r=fpr(gv(RC_FLOAT));
1859         vswap();
1860         if(c2) {
1861           if(c2>0xf)
1862             x|=0x200000;
1863           r2=c2&0xf;
1864         } else {
1865           r2=fpr(gv(RC_FLOAT));
1866         }
1867         vtop[-1].r = VT_CMP;
1868         vtop[-1].c.i = op;
1869       } else {
1870         tcc_error("unknown fp op %x!",op);
1871         return;
1872       }
1873   }
1874   if(vtop[-1].r == VT_CMP)
1875     c1=15;
1876   else {
1877     c1=vtop->r;
1878     if(r2&0x8)
1879       c1=vtop[-1].r;
1880     vtop[-1].r=get_reg_ex(RC_FLOAT,two2mask(vtop[-1].r,c1));
1881     c1=fpr(vtop[-1].r);
1882   }
1883   vtop--;
1884   o(x|(r<<16)|(c1<<12)|r2);
1885 }
1886 #endif
1887
1888 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
1889    and 'long long' cases. */
1890 ST_FUNC void gen_cvt_itof1(int t)
1891 {
1892   uint32_t r, r2;
1893   int bt;
1894   bt=vtop->type.t & VT_BTYPE;
1895   if(bt == VT_INT || bt == VT_SHORT || bt == VT_BYTE) {
1896 #ifndef TCC_ARM_VFP
1897     uint32_t dsize = 0;
1898 #endif
1899     r=intr(gv(RC_INT));
1900 #ifdef TCC_ARM_VFP
1901     r2=vfpr(vtop->r=get_reg(RC_FLOAT));
1902     o(0xEE000A10|(r<<12)|(r2<<16)); /* fmsr */
1903     r2|=r2<<12;
1904     if(!(vtop->type.t & VT_UNSIGNED))
1905       r2|=0x80;                /* fuitoX -> fsituX */
1906     o(0xEEB80A40|r2|T2CPR(t)); /* fYitoX*/
1907 #else
1908     r2=fpr(vtop->r=get_reg(RC_FLOAT));
1909     if((t & VT_BTYPE) != VT_FLOAT)
1910       dsize=0x80;    /* flts -> fltd */
1911     o(0xEE000110|dsize|(r2<<16)|(r<<12)); /* flts */
1912     if((vtop->type.t & (VT_UNSIGNED|VT_BTYPE)) == (VT_UNSIGNED|VT_INT)) {
1913       uint32_t off = 0;
1914       o(0xE3500000|(r<<12));        /* cmp */
1915       r=fpr(get_reg(RC_FLOAT));
1916       if(last_itod_magic) {
1917         off=ind+8-last_itod_magic;
1918         off/=4;
1919         if(off>255)
1920           off=0;
1921       }
1922       o(0xBD1F0100|(r<<12)|off);    /* ldflts */
1923       if(!off) {
1924         o(0xEA000000);              /* b */
1925         last_itod_magic=ind;
1926         o(0x4F800000);              /* 4294967296.0f */
1927       }
1928       o(0xBE000100|dsize|(r2<<16)|(r2<<12)|r); /* adflt */
1929     }
1930 #endif
1931     return;
1932   } else if(bt == VT_LLONG) {
1933     int func;
1934     CType *func_type = 0;
1935     if((t & VT_BTYPE) == VT_FLOAT) {
1936       func_type = &func_float_type;
1937       if(vtop->type.t & VT_UNSIGNED)
1938         func=TOK___floatundisf;
1939       else
1940         func=TOK___floatdisf;
1941 #if LDOUBLE_SIZE != 8
1942     } else if((t & VT_BTYPE) == VT_LDOUBLE) {
1943       func_type = &func_ldouble_type;
1944       if(vtop->type.t & VT_UNSIGNED)
1945         func=TOK___floatundixf;
1946       else
1947         func=TOK___floatdixf;
1948     } else if((t & VT_BTYPE) == VT_DOUBLE) {
1949 #else
1950     } else if((t & VT_BTYPE) == VT_DOUBLE || (t & VT_BTYPE) == VT_LDOUBLE) {
1951 #endif
1952       func_type = &func_double_type;
1953       if(vtop->type.t & VT_UNSIGNED)
1954         func=TOK___floatundidf;
1955       else
1956         func=TOK___floatdidf;
1957     }
1958     if(func_type) {
1959       vpush_global_sym(func_type, func);
1960       vswap();
1961       gfunc_call(1);
1962       vpushi(0);
1963       vtop->r=TREG_F0;
1964       return;
1965     }
1966   }
1967   tcc_error("unimplemented gen_cvt_itof %x!",vtop->type.t);
1968 }
1969
1970 /* convert fp to int 't' type */
1971 void gen_cvt_ftoi(int t)
1972 {
1973   uint32_t r, r2;
1974   int u, func = 0;
1975   u=t&VT_UNSIGNED;
1976   t&=VT_BTYPE;
1977   r2=vtop->type.t & VT_BTYPE;
1978   if(t==VT_INT) {
1979 #ifdef TCC_ARM_VFP
1980     r=vfpr(gv(RC_FLOAT));
1981     u=u?0:0x10000;
1982     o(0xEEBC0AC0|(r<<12)|r|T2CPR(r2)|u); /* ftoXizY */
1983     r2=intr(vtop->r=get_reg(RC_INT));
1984     o(0xEE100A10|(r<<16)|(r2<<12));
1985     return;
1986 #else
1987     if(u) {
1988       if(r2 == VT_FLOAT)
1989         func=TOK___fixunssfsi;
1990 #if LDOUBLE_SIZE != 8
1991       else if(r2 == VT_LDOUBLE)
1992         func=TOK___fixunsxfsi;
1993       else if(r2 == VT_DOUBLE)
1994 #else
1995       else if(r2 == VT_LDOUBLE || r2 == VT_DOUBLE)
1996 #endif
1997         func=TOK___fixunsdfsi;
1998     } else {
1999       r=fpr(gv(RC_FLOAT));
2000       r2=intr(vtop->r=get_reg(RC_INT));
2001       o(0xEE100170|(r2<<12)|r);
2002       return;
2003     }
2004 #endif
2005   } else if(t == VT_LLONG) { // unsigned handled in gen_cvt_ftoi1
2006     if(r2 == VT_FLOAT)
2007       func=TOK___fixsfdi;
2008 #if LDOUBLE_SIZE != 8
2009     else if(r2 == VT_LDOUBLE)
2010       func=TOK___fixxfdi;
2011     else if(r2 == VT_DOUBLE)
2012 #else
2013     else if(r2 == VT_LDOUBLE || r2 == VT_DOUBLE)
2014 #endif
2015       func=TOK___fixdfdi;
2016   }
2017   if(func) {
2018     vpush_global_sym(&func_old_type, func);
2019     vswap();
2020     gfunc_call(1);
2021     vpushi(0);
2022     if(t == VT_LLONG)
2023       vtop->r2 = REG_LRET;
2024     vtop->r = REG_IRET;
2025     return;
2026   }
2027   tcc_error("unimplemented gen_cvt_ftoi!");
2028 }
2029
2030 /* convert from one floating point type to another */
2031 void gen_cvt_ftof(int t)
2032 {
2033 #ifdef TCC_ARM_VFP
2034   if(((vtop->type.t & VT_BTYPE) == VT_FLOAT) != ((t & VT_BTYPE) == VT_FLOAT)) {
2035     uint32_t r = vfpr(gv(RC_FLOAT));
2036     o(0xEEB70AC0|(r<<12)|r|T2CPR(vtop->type.t));
2037   }
2038 #else
2039   /* all we have to do on i386 and FPA ARM is to put the float in a register */
2040   gv(RC_FLOAT);
2041 #endif
2042 }
2043
2044 /* computed goto support */
2045 void ggoto(void)
2046 {
2047   gcall_or_jmp(1);
2048   vtop--;
2049 }
2050
2051 /* Save the stack pointer onto the stack and return the location of its address */
2052 ST_FUNC void gen_vla_sp_save(int addr) {
2053     tcc_error("variable length arrays unsupported for this target");
2054 }
2055
2056 /* Restore the SP from a location on the stack */
2057 ST_FUNC void gen_vla_sp_restore(int addr) {
2058     tcc_error("variable length arrays unsupported for this target");
2059 }
2060
2061 /* Subtract from the stack pointer, and push the resulting value onto the stack */
2062 ST_FUNC void gen_vla_alloc(CType *type, int align) {
2063     tcc_error("variable length arrays unsupported for this target");
2064 }
2065
2066 /* end of ARM code generator */
2067 /*************************************************************/
2068 #endif
2069 /*************************************************************/