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 #ifdef TCC_ARM_HARDFLOAT
 750 /* Return whether a structure is an homogeneous float aggregate or not.
 751    The answer is true if all the elements of the structure are of the same
 752    primitive float type and there is less than 4 elements.
 753
 754    type: the type corresponding to the structure to be tested */
 755 static int is_hgen_float_aggr(CType *type)
 756 {
 757   if ((type->t & VT_BTYPE) == VT_STRUCT) {
 758     struct Sym *ref;
 759     int btype, nb_fields = 0;
 760
 761     ref = type->ref->next;
 762     btype = ref->type.t & VT_BTYPE;
 763     if (btype == VT_FLOAT || btype == VT_DOUBLE) {
 764       for(; ref && btype == (ref->type.t & VT_BTYPE); ref = ref->next, nb_fields++);
 765       return !ref && nb_fields <= 4;
 766     }
 767   }
 768   return 0;
 769 }
 770
 771 struct avail_regs {
 772   signed char avail[3]; /* 3 holes max with only float and double alignments */
 773   int first_hole; /* first available hole */
 774   int last_hole; /* last available hole (none if equal to first_hole) */
 775   int first_free_reg; /* next free register in the sequence, hole excluded */
 776 };
 777
 778 #define AVAIL_REGS_INITIALIZER (struct avail_regs) { { 0, 0, 0}, 0, 0, 0 }
 779
 780 /* Find suitable registers for a VFP Co-Processor Register Candidate (VFP CPRC
 781    param) according to the rules described in the procedure call standard for
 782    the ARM architecture (AAPCS). If found, the registers are assigned to this
 783    VFP CPRC parameter. Registers are allocated in sequence unless a hole exists
 784    and the parameter is a single float.
 785
 786    avregs: opaque structure to keep track of available VFP co-processor regs
 787    align: alignment contraints for the param, as returned by type_size()
 788    size: size of the parameter, as returned by type_size() */
 789 int assign_vfpreg(struct avail_regs *avregs, int align, int size)
 790 {
 791   int first_reg = 0;
 792
 793   if (avregs->first_free_reg == -1)
 794     return -1;
 795   if (align >> 3) { /* double alignment */
 796     first_reg = avregs->first_free_reg;
 797     /* alignment contraint not respected so use next reg and record hole */
 798     if (first_reg & 1)
 799       avregs->avail[avregs->last_hole++] = first_reg++;
 800   } else { /* no special alignment (float or array of float) */
 801     /* if single float and a hole is available, assign the param to it */
 802     if (size == 4 && avregs->first_hole != avregs->last_hole)
 803       return avregs->avail[avregs->first_hole++];
 804     else
 805       first_reg = avregs->first_free_reg;
 806   }
 807   if (first_reg + size / 4 <= 16) {
 808     avregs->first_free_reg = first_reg + size / 4;
 809     return first_reg;
 810   }
 811   avregs->first_free_reg = -1;
 812   return -1;
 813 }
 814 #endif
 815
 816 /* Return the number of registers needed to return the struct, or 0 if
 817    returning via struct pointer. */
 818 ST_FUNC int gfunc_sret(CType *vt, CType *ret, int *ret_align) {
 819 #ifdef TCC_ARM_EABI
 820     int size, align;
 821     size = type_size(vt, &align);
 822 #ifdef TCC_ARM_HARDFLOAT
 823     if (is_float(vt->t) || is_hgen_float_aggr(vt)) {
 824         *ret_align = 8;
 825         ret->ref = NULL;
 826         ret->t = VT_DOUBLE;
 827         return (size + 7) >> 3;
 828     } else
 829 #endif
 830     if (size > 4) {
 831         return 0;
 832     } else {
 833         *ret_align = 4;
 834         ret->ref = NULL;
 835         ret->t = VT_INT;
 836         return 1;
 837     }
 838 #else
 839     return 0;
 840 #endif
 841 }
 842
 843 /* Parameters are classified according to how they are copied to their final
 844    destination for the function call. Because the copying is performed class
 845    after class according to the order in the union below, it is important that
 846    some constraints about the order of the members of this union are respected:
 847    - CORE_STRUCT_CLASS must come after STACK_CLASS;
 848    - CORE_CLASS must come after STACK_CLASS, CORE_STRUCT_CLASS and
 849      VFP_STRUCT_CLASS;
 850    - VFP_STRUCT_CLASS must come after VFP_CLASS.
 851    See the comment for the main loop in copy_params() for the reason. */
 852 enum reg_class {
 853         STACK_CLASS = 0,
 854         CORE_STRUCT_CLASS,
 855         VFP_CLASS,
 856         VFP_STRUCT_CLASS,
 857         CORE_CLASS,
 858         NB_CLASSES
 859 };
 860
 861 struct param_plan {
 862     int start; /* first reg or addr used depending on the class */
 863     int end; /* last reg used or next free addr depending on the class */
 864     SValue *sval; /* pointer to SValue on the value stack */
 865     struct param_plan *prev; /*  previous element in this class */
 866 };
 867
 868 struct plan {
 869     struct param_plan *pplans; /* array of all the param plans */
 870     struct param_plan *clsplans[NB_CLASSES]; /* per class lists of param plans */
 871 };
 872
 873 #define add_param_plan(plan,pplan,class)                        \
 874     do {                                                        \
 875         pplan.prev = plan->clsplans[class];                     \
 876         plan->pplans[plan ## _nb] = pplan;                      \
 877         plan->clsplans[class] = &plan->pplans[plan ## _nb++];   \
 878     } while(0)
 879
 880 /* Assign parameters to registers and stack with alignment according to the
 881    rules in the procedure call standard for the ARM architecture (AAPCS).
 882    The overall assignment is recorded in an array of per parameter structures
 883    called parameter plans. The parameter plans are also further organized in a
 884    number of linked lists, one per class of parameter (see the comment for the
 885    definition of union reg_class).
 886
 887    nb_args: number of parameters of the function for which a call is generated
 888    variadic: whether the function is a variadic function or not
 889    plan: the structure where the overall assignment is recorded
 890    todo: a bitmap that record which core registers hold a parameter
 891
 892    Returns the amount of stack space needed for parameter passing
 893
 894    Note: this function allocated an array in plan->pplans with tcc_malloc. It
 895    is the responsability of the caller to free this array once used (ie not
 896    before copy_params). */
 897 static int assign_regs(int nb_args, int variadic, struct plan *plan, int *todo)
 898 {
 899   int i, size, align;
 900   int ncrn /* next core register number */, nsaa /* next stacked argument address*/;
 901   int plan_nb = 0;
 902   struct param_plan pplan;
 903 #ifdef TCC_ARM_HARDFLOAT
 904   struct avail_regs avregs = AVAIL_REGS_INITIALIZER;
 905 #endif
 906
 907   ncrn = nsaa = 0;
 908   *todo = 0;
 909   plan->pplans = tcc_malloc(nb_args * sizeof(*plan->pplans));
 910   memset(plan->clsplans, 0, sizeof(plan->clsplans));
 911   for(i = nb_args; i-- ;) {
 912     int j, start_vfpreg = 0;
 913     size = type_size(&vtop[-i].type, &align);
 914     switch(vtop[-i].type.t & VT_BTYPE) {
 915       case VT_STRUCT:
 916       case VT_FLOAT:
 917       case VT_DOUBLE:
 918       case VT_LDOUBLE:
 919 #ifdef TCC_ARM_HARDFLOAT
 920       if (!variadic) {
 921         int is_hfa = 0; /* Homogeneous float aggregate */
 922
 923         if (is_float(vtop[-i].type.t)
 924             || (is_hfa = is_hgen_float_aggr(&vtop[-i].type))) {
 925           int end_vfpreg;
 926
 927           start_vfpreg = assign_vfpreg(&avregs, align, size);
 928           end_vfpreg = start_vfpreg + ((size - 1) >> 2);
 929           if (start_vfpreg >= 0) {
 930             pplan = (struct param_plan) {start_vfpreg, end_vfpreg, &vtop[-i]};
 931             if (is_hfa)
 932               add_param_plan(plan, pplan, VFP_STRUCT_CLASS);
 933             else
 934               add_param_plan(plan, pplan, VFP_CLASS);
 935             continue;
 936           } else
 937             break;
 938         }
 939       }
 940 #endif
 941       ncrn = (ncrn + (align-1)/4) & -(align/4);
 942       size = (size + 3) & -4;
 943       if (ncrn + size/4 <= 4 || (ncrn < 4 && start_vfpreg != -1)) {
 944         /* The parameter is allocated both in core register and on stack. As
 945          * such, it can be of either class: it would either be the last of
 946          * CORE_STRUCT_CLASS or the first of STACK_CLASS. */
 947         for (j = ncrn; j < 4 && j < ncrn + size / 4; j++)
 948           *todo|=(1<<j);
 949         pplan = (struct param_plan) {ncrn, j, &vtop[-i]};
 950         add_param_plan(plan, pplan, CORE_STRUCT_CLASS);
 951         ncrn += size/4;
 952         if (ncrn > 4)
 953           nsaa = (ncrn - 4) * 4;
 954       } else {
 955         ncrn = 4;
 956         break;
 957       }
 958       continue;
 959       default:
 960       if (ncrn < 4) {
 961         int is_long = (vtop[-i].type.t & VT_BTYPE) == VT_LLONG;
 962
 963         if (is_long) {
 964           ncrn = (ncrn + 1) & -2;
 965           if (ncrn == 4)
 966             break;
 967         }
 968         pplan = (struct param_plan) {ncrn, ncrn, &vtop[-i]};
 969         ncrn++;
 970         if (is_long)
 971           pplan.end = ncrn++;
 972         add_param_plan(plan, pplan, CORE_CLASS);
 973         continue;
 974       }
 975     }
 976     nsaa = (nsaa + (align - 1)) & ~(align - 1);
 977     pplan = (struct param_plan) {nsaa, nsaa + size, &vtop[-i]};
 978     add_param_plan(plan, pplan, STACK_CLASS);
 979     nsaa += size; /* size already rounded up before */
 980   }
 981   return nsaa;
 982 }
 983
 984 #undef add_param_plan
 985
 986 /* Copy parameters to their final destination (core reg, VFP reg or stack) for
 987    function call.
 988
 989    nb_args: number of parameters the function take
 990    plan: the overall assignment plan for parameters
 991    todo: a bitmap indicating what core reg will hold a parameter
 992
 993    Returns the number of SValue added by this function on the value stack */
 994 static int copy_params(int nb_args, struct plan *plan, int todo)
 995 {
 996   int size, align, r, i, nb_extra_sval = 0;
 997   struct param_plan *pplan;
 998
 999    /* Several constraints require parameters to be copied in a specific order:
1000       - structures are copied to the stack before being loaded in a reg;
1001       - floats loaded to an odd numbered VFP reg are first copied to the
1002         preceding even numbered VFP reg and then moved to the next VFP reg.
1003
1004       It is thus important that:
1005       - structures assigned to core regs must be copied after parameters
1006         assigned to the stack but before structures assigned to VFP regs because
1007         a structure can lie partly in core registers and partly on the stack;
1008       - parameters assigned to the stack and all structures be copied before
1009         parameters assigned to a core reg since copying a parameter to the stack
1010         require using a core reg;
1011       - parameters assigned to VFP regs be copied before structures assigned to
1012         VFP regs as the copy might use an even numbered VFP reg that already
1013         holds part of a structure. */
1014   for(i = 0; i < NB_CLASSES; i++) {
1015     for(pplan = plan->clsplans[i]; pplan; pplan = pplan->prev) {
1016       vpushv(pplan->sval);
1017       pplan->sval->r = pplan->sval->r2 = VT_CONST; /* disable entry */
1018       switch(i) {
1019         case STACK_CLASS:
1020         case CORE_STRUCT_CLASS:
1021         case VFP_STRUCT_CLASS:
1022           if ((pplan->sval->type.t & VT_BTYPE) == VT_STRUCT) {
1023             size = type_size(&pplan->sval->type, &align);
1024             /* align to stack align size */
1025             size = (size + 3) & ~3;
1026             if (i == STACK_CLASS && pplan->prev)
1027               size += pplan->start - pplan->prev->end; /* Add padding if any */
1028             /* allocate the necessary size on stack */
1029             gadd_sp(-size);
1030             /* generate structure store */
1031             r = get_reg(RC_INT);
1032             o(0xE1A0000D|(intr(r)<<12)); /* mov r, sp */
1033             vset(&vtop->type, r | VT_LVAL, 0);
1034             vswap();
1035             vstore(); /* memcpy to current sp */
1036             /* Homogeneous float aggregate are loaded to VFP registers
1037                immediately since there is no way of loading data in multiple
1038                non consecutive VFP registers as what is done for other
1039                structures (see the use of todo). */
1040             if (i == VFP_STRUCT_CLASS) {
1041               int first = pplan->start, nb = pplan->end - first + 1;
1042               /* vpop.32 {pplan->start, ..., pplan->end} */
1043               o(0xECBD0A00|(first&1)<<22|(first>>1)<<12|nb);
1044               /* No need to write the register used to a SValue since VFP regs
1045                  cannot be used for gcall_or_jmp */
1046             }
1047           } else {
1048             if (is_float(pplan->sval->type.t)) {
1049 #ifdef TCC_ARM_VFP
1050               r = vfpr(gv(RC_FLOAT)) << 12;
1051               if ((pplan->sval->type.t & VT_BTYPE) == VT_FLOAT)
1052                 size = 4;
1053               else {
1054                 size = 8;
1055                 r |= 0x101; /* vpush.32 -> vpush.64 */
1056               }
1057               o(0xED2D0A01 + r); /* vpush */
1058 #else
1059               r = fpr(gv(RC_FLOAT)) << 12;
1060               if ((pplan->sval->type.t & VT_BTYPE) == VT_FLOAT)
1061                 size = 4;
1062               else if ((pplan->sval->type.t & VT_BTYPE) == VT_DOUBLE)
1063                 size = 8;
1064               else
1065                 size = LDOUBLE_SIZE;
1066
1067               if (size == 12)
1068                 r |= 0x400000;
1069               else if(size == 8)
1070                 r|=0x8000;
1071
1072               o(0xED2D0100|r|(size>>2)); /* some kind of vpush for FPA */
1073 #endif
1074             } else {
1075               /* simple type (currently always same size) */
1076               /* XXX: implicit cast ? */
1077               size=4;
1078               if ((pplan->sval->type.t & VT_BTYPE) == VT_LLONG) {
1079                 lexpand_nr();
1080                 size = 8;
1081                 r = gv(RC_INT);
1082                 o(0xE52D0004|(intr(r)<<12)); /* push r */
1083                 vtop--;
1084               }
1085               r = gv(RC_INT);
1086               o(0xE52D0004|(intr(r)<<12)); /* push r */
1087             }
1088             if (i == STACK_CLASS && pplan->prev)
1089               gadd_sp(pplan->prev->end - pplan->start); /* Add padding if any */
1090           }
1091           break;
1092
1093         case VFP_CLASS:
1094           gv(regmask(TREG_F0 + (pplan->start >> 1)));
1095           if (pplan->start & 1) { /* Must be in upper part of double register */
1096             o(0xEEF00A40|((pplan->start>>1)<<12)|(pplan->start>>1)); /* vmov.f32 s(n+1), sn */
1097             vtop->r = VT_CONST; /* avoid being saved on stack by gv for next float */
1098           }
1099           break;
1100
1101         case CORE_CLASS:
1102           if ((pplan->sval->type.t & VT_BTYPE) == VT_LLONG) {
1103             lexpand_nr();
1104             gv(regmask(pplan->end));
1105             pplan->sval->r2 = vtop->r;
1106             vtop--;
1107           }
1108           gv(regmask(pplan->start));
1109           /* Mark register as used so that gcall_or_jmp use another one
1110              (regs >=4 are free as never used to pass parameters) */
1111           pplan->sval->r = vtop->r;
1112           break;
1113       }
1114       vtop--;
1115     }
1116   }
1117
1118   /* Manually free remaining registers since next parameters are loaded
1119    * manually, without the help of gv(int). */
1120   save_regs(nb_args);
1121
1122   if(todo) {
1123     o(0xE8BD0000|todo); /* pop {todo} */
1124     for(pplan = plan->clsplans[CORE_STRUCT_CLASS]; pplan; pplan = pplan->prev) {
1125       int r;
1126       pplan->sval->r = pplan->start;
1127       /* TODO: why adding fake param */
1128       for (r = pplan->start + 1; r <= pplan->end; r++) {
1129         if (todo & (1 << r)) {
1130           nb_extra_sval++;
1131           vpushi(0);
1132           vtop->r = r;
1133         }
1134       }
1135     }
1136   }
1137   return nb_extra_sval;
1138 }
1139
1140 /* Generate function call. The function address is pushed first, then
1141    all the parameters in call order. This functions pops all the
1142    parameters and the function address. */
1143 void gfunc_call(int nb_args)
1144 {
1145   int r, args_size;
1146   int variadic;
1147   int todo;
1148   struct plan plan;
1149
1150   variadic = (vtop[-nb_args].type.ref->c == FUNC_ELLIPSIS);
1151   /* cannot let cpu flags if other instruction are generated. Also avoid leaving
1152      VT_JMP anywhere except on the top of the stack because it would complicate
1153      the code generator. */
1154   r = vtop->r & VT_VALMASK;
1155   if (r == VT_CMP || (r & ~1) == VT_JMP)
1156     gv(RC_INT);
1157
1158   args_size = assign_regs(nb_args, variadic, &plan, &todo);
1159
1160 #ifdef TCC_ARM_EABI
1161   if (args_size & 7) { /* Stack must be 8 byte aligned at fct call for EABI */
1162     args_size = (args_size + 7) & ~7;
1163     o(0xE24DD004); /* sub sp, sp, #4 */
1164   }
1165 #endif
1166
1167   nb_args += copy_params(nb_args, &plan, todo);
1168   tcc_free(plan.pplans);
1169
1170   /* Move fct SValue on top as required by gcall_or_jmp */
1171   vrotb(nb_args + 1);
1172   gcall_or_jmp(0);
1173   if (args_size)
1174       gadd_sp(args_size); /* pop all parameters passed on the stack */
1175 #ifdef TCC_ARM_EABI
1176 #ifdef TCC_ARM_VFP
1177 #ifdef TCC_ARM_HARDFLOAT
1178   if(variadic && is_float(vtop->type.ref->type.t)) {
1179 #else
1180   rf(is_float(vtop->type.ref->type.t)) {
1181 #endif
1182     if((vtop->type.ref->type.t & VT_BTYPE) == VT_FLOAT) {
1183       o(0xEE000A10); /*vmov s0, r0 */
1184     } else {
1185       o(0xEE000B10); /* vmov.32 d0[0], r0 */
1186       o(0xEE201B10); /* vmov.32 d0[1], r1 */
1187     }
1188   }
1189 #endif
1190 #endif
1191   vtop -= nb_args + 1; /* Pop all params and fct address from value stack */
1192   leaffunc = 0; /* we are calling a function, so we aren't in a leaf function */
1193 }
1194
1195 /* generate function prolog of type 't' */
1196 void gfunc_prolog(CType *func_type)
1197 {
1198   Sym *sym,*sym2;
1199   int n,nf,size,align, variadic, struct_ret = 0;
1200 #ifdef TCC_ARM_HARDFLOAT
1201   struct avail_regs avregs = AVAIL_REGS_INITIALIZER;
1202 #endif
1203
1204   sym = func_type->ref;
1205   func_vt = sym->type;
1206
1207   n = nf = 0;
1208   variadic = (func_type->ref->c == FUNC_ELLIPSIS);
1209   if((func_vt.t & VT_BTYPE) == VT_STRUCT
1210 #ifdef TCC_ARM_HARDFLOAT
1211      && (variadic || !is_hgen_float_aggr(&func_vt))
1212 #endif
1213      && type_size(&func_vt,&align) > 4)
1214   {
1215     n++;
1216     struct_ret = 1;
1217     func_vc = 12; /* Offset from fp of the place to store the result */
1218   }
1219   for(sym2=sym->next;sym2 && (n<4 || nf<16);sym2=sym2->next) {
1220     size = type_size(&sym2->type, &align);
1221 #ifdef TCC_ARM_HARDFLOAT
1222     if (!variadic && (is_float(sym2->type.t)
1223         || is_hgen_float_aggr(&sym2->type))) {
1224       int tmpnf = assign_vfpreg(&avregs, align, size);
1225       tmpnf += (size + 3) / 4;
1226       nf = (tmpnf > nf) ? tmpnf : nf;
1227     } else
1228 #endif
1229     if (n < 4)
1230       n += (size + 3) / 4;
1231   }
1232   o(0xE1A0C00D); /* mov ip,sp */
1233   if(variadic)
1234     n=4;
1235   if(n) {
1236     if(n>4)
1237       n=4;
1238 #ifdef TCC_ARM_EABI
1239     n=(n+1)&-2;
1240 #endif
1241     o(0xE92D0000|((1<<n)-1)); /* save r0-r4 on stack if needed */
1242   }
1243   if (nf) {
1244     if (nf>16)
1245       nf=16;
1246     nf=(nf+1)&-2; /* nf => HARDFLOAT => EABI */
1247     o(0xED2D0A00|nf); /* save s0-s15 on stack if needed */
1248   }
1249   o(0xE92D5800); /* save fp, ip, lr */
1250   o(0xE1A0B00D); /* mov fp, sp */
1251   func_sub_sp_offset = ind;
1252   o(0xE1A00000); /* nop, leave space for stack adjustment in epilogue */
1253   {
1254     int addr, pn = struct_ret, sn = 0; /* pn=core, sn=stack */
1255
1256 #ifdef TCC_ARM_HARDFLOAT
1257     func_vc += nf * 4;
1258     avregs = AVAIL_REGS_INITIALIZER;
1259 #endif
1260     while ((sym = sym->next)) {
1261       CType *type;
1262       type = &sym->type;
1263       size = type_size(type, &align);
1264       size = (size + 3) >> 2;
1265       align = (align + 3) & ~3;
1266 #ifdef TCC_ARM_HARDFLOAT
1267       if (!variadic && (is_float(sym->type.t)
1268           || is_hgen_float_aggr(&sym->type))) {
1269         int fpn = assign_vfpreg(&avregs, align, size << 2);
1270         if (fpn >= 0) {
1271           addr = fpn * 4;
1272         } else
1273           goto from_stack;
1274       } else
1275 #endif
1276       if (pn < 4) {
1277 #ifdef TCC_ARM_EABI
1278         pn = (pn + (align-1)/4) & -(align/4);
1279 #endif
1280         addr = (nf + pn) * 4;
1281         pn += size;
1282         if (!sn && pn > 4)
1283           sn = (pn - 4);
1284       } else {
1285 #ifdef TCC_ARM_HARDFLOAT
1286 from_stack:
1287 #endif
1288 #ifdef TCC_ARM_EABI
1289         sn = (sn + (align-1)/4) & -(align/4);
1290 #endif
1291         addr = (n + nf + sn) * 4;
1292         sn += size;
1293       }
1294       sym_push(sym->v & ~SYM_FIELD, type, VT_LOCAL | lvalue_type(type->t), addr+12);
1295     }
1296   }
1297   last_itod_magic=0;
1298   leaffunc = 1;
1299   loc = 0;
1300 }
1301
1302 /* generate function epilog */
1303 void gfunc_epilog(void)
1304 {
1305   uint32_t x;
1306   int diff;
1307 #ifdef TCC_ARM_EABI
1308   /* Useless but harmless copy of the float result into main register(s) in case
1309      of variadic function in the hardfloat variant */
1310   if(is_float(func_vt.t)) {
1311     if((func_vt.t & VT_BTYPE) == VT_FLOAT)
1312       o(0xEE100A10); /* fmrs r0, s0 */
1313     else {
1314       o(0xEE100B10); /* fmrdl r0, d0 */
1315       o(0xEE301B10); /* fmrdh r1, d0 */
1316     }
1317   }
1318 #endif
1319   o(0xE89BA800); /* restore fp, sp, pc */
1320   diff = (-loc + 3) & -4;
1321 #ifdef TCC_ARM_EABI
1322   if(!leaffunc)
1323     diff = ((diff + 11) & -8) - 4;
1324 #endif
1325   if(diff > 0) {
1326     x=stuff_const(0xE24BD000, diff); /* sub sp,fp,# */
1327     if(x)
1328       *(uint32_t *)(cur_text_section->data + func_sub_sp_offset) = x;
1329     else {
1330       int addr;
1331       addr=ind;
1332       o(0xE59FC004); /* ldr ip,[pc+4] */
1333       o(0xE04BD00C); /* sub sp,fp,ip  */
1334       o(0xE1A0F00E); /* mov pc,lr */
1335       o(diff);
1336       *(uint32_t *)(cur_text_section->data + func_sub_sp_offset) = 0xE1000000|encbranch(func_sub_sp_offset,addr,1);
1337     }
1338   }
1339 }
1340
1341 /* generate a jump to a label */
1342 int gjmp(int t)
1343 {
1344   int r;
1345   r=ind;
1346   o(0xE0000000|encbranch(r,t,1));
1347   return r;
1348 }
1349
1350 /* generate a jump to a fixed address */
1351 void gjmp_addr(int a)
1352 {
1353   gjmp(a);
1354 }
1355
1356 /* generate a test. set 'inv' to invert test. Stack entry is popped */
1357 int gtst(int inv, int t)
1358 {
1359   int v, r;
1360   uint32_t op;
1361   v = vtop->r & VT_VALMASK;
1362   r=ind;
1363   if (v == VT_CMP) {
1364     op=mapcc(inv?negcc(vtop->c.i):vtop->c.i);
1365     op|=encbranch(r,t,1);
1366     o(op);
1367     t=r;
1368   } else if (v == VT_JMP || v == VT_JMPI) {
1369     if ((v & 1) == inv) {
1370       if(!vtop->c.i)
1371         vtop->c.i=t;
1372       else {
1373         uint32_t *x;
1374         int p,lp;
1375         if(t) {
1376           p = vtop->c.i;
1377           do {
1378             p = decbranch(lp=p);
1379           } while(p);
1380           x = (uint32_t *)(cur_text_section->data + lp);
1381           *x &= 0xff000000;
1382           *x |= encbranch(lp,t,1);
1383         }
1384         t = vtop->c.i;
1385       }
1386     } else {
1387       t = gjmp(t);
1388       gsym(vtop->c.i);
1389     }
1390   } else {
1391     if (is_float(vtop->type.t)) {
1392       r=gv(RC_FLOAT);
1393 #ifdef TCC_ARM_VFP
1394       o(0xEEB50A40|(vfpr(r)<<12)|T2CPR(vtop->type.t)); /* fcmpzX */
1395       o(0xEEF1FA10); /* fmstat */
1396 #else
1397       o(0xEE90F118|(fpr(r)<<16));
1398 #endif
1399       vtop->r = VT_CMP;
1400       vtop->c.i = TOK_NE;
1401       return gtst(inv, t);
1402     } else if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
1403       /* constant jmp optimization */
1404       if ((vtop->c.i != 0) != inv)
1405         t = gjmp(t);
1406     } else {
1407       v = gv(RC_INT);
1408       o(0xE3300000|(intr(v)<<16));
1409       vtop->r = VT_CMP;
1410       vtop->c.i = TOK_NE;
1411       return gtst(inv, t);
1412     }
1413   }
1414   vtop--;
1415   return t;
1416 }
1417
1418 /* generate an integer binary operation */
1419 void gen_opi(int op)
1420 {
1421   int c, func = 0;
1422   uint32_t opc = 0, r, fr;
1423   unsigned short retreg = REG_IRET;
1424
1425   c=0;
1426   switch(op) {
1427     case '+':
1428       opc = 0x8;
1429       c=1;
1430       break;
1431     case TOK_ADDC1: /* add with carry generation */
1432       opc = 0x9;
1433       c=1;
1434       break;
1435     case '-':
1436       opc = 0x4;
1437       c=1;
1438       break;
1439     case TOK_SUBC1: /* sub with carry generation */
1440       opc = 0x5;
1441       c=1;
1442       break;
1443     case TOK_ADDC2: /* add with carry use */
1444       opc = 0xA;
1445       c=1;
1446       break;
1447     case TOK_SUBC2: /* sub with carry use */
1448       opc = 0xC;
1449       c=1;
1450       break;
1451     case '&':
1452       opc = 0x0;
1453       c=1;
1454       break;
1455     case '^':
1456       opc = 0x2;
1457       c=1;
1458       break;
1459     case '|':
1460       opc = 0x18;
1461       c=1;
1462       break;
1463     case '*':
1464       gv2(RC_INT, RC_INT);
1465       r = vtop[-1].r;
1466       fr = vtop[0].r;
1467       vtop--;
1468       o(0xE0000090|(intr(r)<<16)|(intr(r)<<8)|intr(fr));
1469       return;
1470     case TOK_SHL:
1471       opc = 0;
1472       c=2;
1473       break;
1474     case TOK_SHR:
1475       opc = 1;
1476       c=2;
1477       break;
1478     case TOK_SAR:
1479       opc = 2;
1480       c=2;
1481       break;
1482     case '/':
1483     case TOK_PDIV:
1484       func=TOK___divsi3;
1485       c=3;
1486       break;
1487     case TOK_UDIV:
1488       func=TOK___udivsi3;
1489       c=3;
1490       break;
1491     case '%':
1492 #ifdef TCC_ARM_EABI
1493       func=TOK___aeabi_idivmod;
1494       retreg=REG_LRET;
1495 #else
1496       func=TOK___modsi3;
1497 #endif
1498       c=3;
1499       break;
1500     case TOK_UMOD:
1501 #ifdef TCC_ARM_EABI
1502       func=TOK___aeabi_uidivmod;
1503       retreg=REG_LRET;
1504 #else
1505       func=TOK___umodsi3;
1506 #endif
1507       c=3;
1508       break;
1509     case TOK_UMULL:
1510       gv2(RC_INT, RC_INT);
1511       r=intr(vtop[-1].r2=get_reg(RC_INT));
1512       c=vtop[-1].r;
1513       vtop[-1].r=get_reg_ex(RC_INT,regmask(c));
1514       vtop--;
1515       o(0xE0800090|(r<<16)|(intr(vtop->r)<<12)|(intr(c)<<8)|intr(vtop[1].r));
1516       return;
1517     default:
1518       opc = 0x15;
1519       c=1;
1520       break;
1521   }
1522   switch(c) {
1523     case 1:
1524       if((vtop[-1].r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
1525         if(opc == 4 || opc == 5 || opc == 0xc) {
1526           vswap();
1527           opc|=2; // sub -> rsb
1528         }
1529       }
1530       if ((vtop->r & VT_VALMASK) == VT_CMP ||
1531           (vtop->r & (VT_VALMASK & ~1)) == VT_JMP)
1532         gv(RC_INT);
1533       vswap();
1534       c=intr(gv(RC_INT));
1535       vswap();
1536       opc=0xE0000000|(opc<<20)|(c<<16);
1537       if((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
1538         uint32_t x;
1539         x=stuff_const(opc|0x2000000,vtop->c.i);
1540         if(x) {
1541           r=intr(vtop[-1].r=get_reg_ex(RC_INT,regmask(vtop[-1].r)));
1542           o(x|(r<<12));
1543           goto done;
1544         }
1545       }
1546       fr=intr(gv(RC_INT));
1547       r=intr(vtop[-1].r=get_reg_ex(RC_INT,two2mask(vtop->r,vtop[-1].r)));
1548       o(opc|(r<<12)|fr);
1549 done:
1550       vtop--;
1551       if (op >= TOK_ULT && op <= TOK_GT) {
1552         vtop->r = VT_CMP;
1553         vtop->c.i = op;
1554       }
1555       break;
1556     case 2:
1557       opc=0xE1A00000|(opc<<5);
1558       if ((vtop->r & VT_VALMASK) == VT_CMP ||
1559           (vtop->r & (VT_VALMASK & ~1)) == VT_JMP)
1560         gv(RC_INT);
1561       vswap();
1562       r=intr(gv(RC_INT));
1563       vswap();
1564       opc|=r;
1565       if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
1566         fr=intr(vtop[-1].r=get_reg_ex(RC_INT,regmask(vtop[-1].r)));
1567         c = vtop->c.i & 0x1f;
1568         o(opc|(c<<7)|(fr<<12));
1569       } else {
1570         fr=intr(gv(RC_INT));
1571         c=intr(vtop[-1].r=get_reg_ex(RC_INT,two2mask(vtop->r,vtop[-1].r)));
1572         o(opc|(c<<12)|(fr<<8)|0x10);
1573       }
1574       vtop--;
1575       break;
1576     case 3:
1577       vpush_global_sym(&func_old_type, func);
1578       vrott(3);
1579       gfunc_call(2);
1580       vpushi(0);
1581       vtop->r = retreg;
1582       break;
1583     default:
1584       tcc_error("gen_opi %i unimplemented!",op);
1585   }
1586 }
1587
1588 #ifdef TCC_ARM_VFP
1589 static int is_zero(int i)
1590 {
1591   if((vtop[i].r & (VT_VALMASK | VT_LVAL | VT_SYM)) != VT_CONST)
1592     return 0;
1593   if (vtop[i].type.t == VT_FLOAT)
1594     return (vtop[i].c.f == 0.f);
1595   else if (vtop[i].type.t == VT_DOUBLE)
1596     return (vtop[i].c.d == 0.0);
1597   return (vtop[i].c.ld == 0.l);
1598 }
1599
1600 /* generate a floating point operation 'v = t1 op t2' instruction. The
1601  *    two operands are guaranted to have the same floating point type */
1602 void gen_opf(int op)
1603 {
1604   uint32_t x;
1605   int fneg=0,r;
1606   x=0xEE000A00|T2CPR(vtop->type.t);
1607   switch(op) {
1608     case '+':
1609       if(is_zero(-1))
1610         vswap();
1611       if(is_zero(0)) {
1612         vtop--;
1613         return;
1614       }
1615       x|=0x300000;
1616       break;
1617     case '-':
1618       x|=0x300040;
1619       if(is_zero(0)) {
1620         vtop--;
1621         return;
1622       }
1623       if(is_zero(-1)) {
1624         x|=0x810000; /* fsubX -> fnegX */
1625         vswap();
1626         vtop--;
1627         fneg=1;
1628       }
1629       break;
1630     case '*':
1631       x|=0x200000;
1632       break;
1633     case '/':
1634       x|=0x800000;
1635       break;
1636     default:
1637       if(op < TOK_ULT || op > TOK_GT) {
1638         tcc_error("unknown fp op %x!",op);
1639         return;
1640       }
1641       if(is_zero(-1)) {
1642         vswap();
1643         switch(op) {
1644           case TOK_LT: op=TOK_GT; break;
1645           case TOK_GE: op=TOK_ULE; break;
1646           case TOK_LE: op=TOK_GE; break;
1647           case TOK_GT: op=TOK_ULT; break;
1648         }
1649       }
1650       x|=0xB40040; /* fcmpX */
1651       if(op!=TOK_EQ && op!=TOK_NE)
1652         x|=0x80; /* fcmpX -> fcmpeX */
1653       if(is_zero(0)) {
1654         vtop--;
1655         o(x|0x10000|(vfpr(gv(RC_FLOAT))<<12)); /* fcmp(e)X -> fcmp(e)zX */
1656       } else {
1657         x|=vfpr(gv(RC_FLOAT));
1658         vswap();
1659         o(x|(vfpr(gv(RC_FLOAT))<<12));
1660         vtop--;
1661       }
1662       o(0xEEF1FA10); /* fmstat */
1663
1664       switch(op) {
1665         case TOK_LE: op=TOK_ULE; break;
1666         case TOK_LT: op=TOK_ULT; break;
1667         case TOK_UGE: op=TOK_GE; break;
1668         case TOK_UGT: op=TOK_GT; break;
1669       }
1670
1671       vtop->r = VT_CMP;
1672       vtop->c.i = op;
1673       return;
1674   }
1675   r=gv(RC_FLOAT);
1676   x|=vfpr(r);
1677   r=regmask(r);
1678   if(!fneg) {
1679     int r2;
1680     vswap();
1681     r2=gv(RC_FLOAT);
1682     x|=vfpr(r2)<<16;
1683     r|=regmask(r2);
1684   }
1685   vtop->r=get_reg_ex(RC_FLOAT,r);
1686   if(!fneg)
1687     vtop--;
1688   o(x|(vfpr(vtop->r)<<12));
1689 }
1690
1691 #else
1692 static uint32_t is_fconst()
1693 {
1694   long double f;
1695   uint32_t r;
1696   if((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) != VT_CONST)
1697     return 0;
1698   if (vtop->type.t == VT_FLOAT)
1699     f = vtop->c.f;
1700   else if (vtop->type.t == VT_DOUBLE)
1701     f = vtop->c.d;
1702   else
1703     f = vtop->c.ld;
1704   if(!ieee_finite(f))
1705     return 0;
1706   r=0x8;
1707   if(f<0.0) {
1708     r=0x18;
1709     f=-f;
1710   }
1711   if(f==0.0)
1712     return r;
1713   if(f==1.0)
1714     return r|1;
1715   if(f==2.0)
1716     return r|2;
1717   if(f==3.0)
1718     return r|3;
1719   if(f==4.0)
1720     return r|4;
1721   if(f==5.0)
1722     return r|5;
1723   if(f==0.5)
1724     return r|6;
1725   if(f==10.0)
1726     return r|7;
1727   return 0;
1728 }
1729
1730 /* generate a floating point operation 'v = t1 op t2' instruction. The
1731    two operands are guaranted to have the same floating point type */
1732 void gen_opf(int op)
1733 {
1734   uint32_t x, r, r2, c1, c2;
1735   //fputs("gen_opf\n",stderr);
1736   vswap();
1737   c1 = is_fconst();
1738   vswap();
1739   c2 = is_fconst();
1740   x=0xEE000100;
1741 #if LDOUBLE_SIZE == 8
1742   if ((vtop->type.t & VT_BTYPE) != VT_FLOAT)
1743     x|=0x80;
1744 #else
1745   if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
1746     x|=0x80;
1747   else if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE)
1748     x|=0x80000;
1749 #endif
1750   switch(op)
1751   {
1752     case '+':
1753       if(!c2) {
1754         vswap();
1755         c2=c1;
1756       }
1757       vswap();
1758       r=fpr(gv(RC_FLOAT));
1759       vswap();
1760       if(c2) {
1761         if(c2>0xf)
1762           x|=0x200000; // suf
1763         r2=c2&0xf;
1764       } else {
1765         r2=fpr(gv(RC_FLOAT));
1766       }
1767       break;
1768     case '-':
1769       if(c2) {
1770         if(c2<=0xf)
1771           x|=0x200000; // suf
1772         r2=c2&0xf;
1773         vswap();
1774         r=fpr(gv(RC_FLOAT));
1775         vswap();
1776       } else if(c1 && c1<=0xf) {
1777         x|=0x300000; // rsf
1778         r2=c1;
1779         r=fpr(gv(RC_FLOAT));
1780         vswap();
1781       } else {
1782         x|=0x200000; // suf
1783         vswap();
1784         r=fpr(gv(RC_FLOAT));
1785         vswap();
1786         r2=fpr(gv(RC_FLOAT));
1787       }
1788       break;
1789     case '*':
1790       if(!c2 || c2>0xf) {
1791         vswap();
1792         c2=c1;
1793       }
1794       vswap();
1795       r=fpr(gv(RC_FLOAT));
1796       vswap();
1797       if(c2 && c2<=0xf)
1798         r2=c2;
1799       else
1800         r2=fpr(gv(RC_FLOAT));
1801       x|=0x100000; // muf
1802       break;
1803     case '/':
1804       if(c2 && c2<=0xf) {
1805         x|=0x400000; // dvf
1806         r2=c2;
1807         vswap();
1808         r=fpr(gv(RC_FLOAT));
1809         vswap();
1810       } else if(c1 && c1<=0xf) {
1811         x|=0x500000; // rdf
1812         r2=c1;
1813         r=fpr(gv(RC_FLOAT));
1814         vswap();
1815       } else {
1816         x|=0x400000; // dvf
1817         vswap();
1818         r=fpr(gv(RC_FLOAT));
1819         vswap();
1820         r2=fpr(gv(RC_FLOAT));
1821       }
1822       break;
1823     default:
1824       if(op >= TOK_ULT && op <= TOK_GT) {
1825         x|=0xd0f110; // cmfe
1826 /* bug (intention?) in Linux FPU emulator
1827    doesn't set carry if equal */
1828         switch(op) {
1829           case TOK_ULT:
1830           case TOK_UGE:
1831           case TOK_ULE:
1832           case TOK_UGT:
1833             tcc_error("unsigned comparision on floats?");
1834             break;
1835           case TOK_LT:
1836             op=TOK_Nset;
1837             break;
1838           case TOK_LE:
1839             op=TOK_ULE; /* correct in unordered case only if AC bit in FPSR set */
1840             break;
1841           case TOK_EQ:
1842           case TOK_NE:
1843             x&=~0x400000; // cmfe -> cmf
1844             break;
1845         }
1846         if(c1 && !c2) {
1847           c2=c1;
1848           vswap();
1849           switch(op) {
1850             case TOK_Nset:
1851               op=TOK_GT;
1852               break;
1853             case TOK_GE:
1854               op=TOK_ULE;
1855               break;
1856             case TOK_ULE:
1857               op=TOK_GE;
1858               break;
1859             case TOK_GT:
1860               op=TOK_Nset;
1861               break;
1862           }
1863         }
1864         vswap();
1865         r=fpr(gv(RC_FLOAT));
1866         vswap();
1867         if(c2) {
1868           if(c2>0xf)
1869             x|=0x200000;
1870           r2=c2&0xf;
1871         } else {
1872           r2=fpr(gv(RC_FLOAT));
1873         }
1874         vtop[-1].r = VT_CMP;
1875         vtop[-1].c.i = op;
1876       } else {
1877         tcc_error("unknown fp op %x!",op);
1878         return;
1879       }
1880   }
1881   if(vtop[-1].r == VT_CMP)
1882     c1=15;
1883   else {
1884     c1=vtop->r;
1885     if(r2&0x8)
1886       c1=vtop[-1].r;
1887     vtop[-1].r=get_reg_ex(RC_FLOAT,two2mask(vtop[-1].r,c1));
1888     c1=fpr(vtop[-1].r);
1889   }
1890   vtop--;
1891   o(x|(r<<16)|(c1<<12)|r2);
1892 }
1893 #endif
1894
1895 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
1896    and 'long long' cases. */
1897 ST_FUNC void gen_cvt_itof1(int t)
1898 {
1899   uint32_t r, r2;
1900   int bt;
1901   bt=vtop->type.t & VT_BTYPE;
1902   if(bt == VT_INT || bt == VT_SHORT || bt == VT_BYTE) {
1903 #ifndef TCC_ARM_VFP
1904     uint32_t dsize = 0;
1905 #endif
1906     r=intr(gv(RC_INT));
1907 #ifdef TCC_ARM_VFP
1908     r2=vfpr(vtop->r=get_reg(RC_FLOAT));
1909     o(0xEE000A10|(r<<12)|(r2<<16)); /* fmsr */
1910     r2|=r2<<12;
1911     if(!(vtop->type.t & VT_UNSIGNED))
1912       r2|=0x80;                /* fuitoX -> fsituX */
1913     o(0xEEB80A40|r2|T2CPR(t)); /* fYitoX*/
1914 #else
1915     r2=fpr(vtop->r=get_reg(RC_FLOAT));
1916     if((t & VT_BTYPE) != VT_FLOAT)
1917       dsize=0x80;    /* flts -> fltd */
1918     o(0xEE000110|dsize|(r2<<16)|(r<<12)); /* flts */
1919     if((vtop->type.t & (VT_UNSIGNED|VT_BTYPE)) == (VT_UNSIGNED|VT_INT)) {
1920       uint32_t off = 0;
1921       o(0xE3500000|(r<<12));        /* cmp */
1922       r=fpr(get_reg(RC_FLOAT));
1923       if(last_itod_magic) {
1924         off=ind+8-last_itod_magic;
1925         off/=4;
1926         if(off>255)
1927           off=0;
1928       }
1929       o(0xBD1F0100|(r<<12)|off);    /* ldflts */
1930       if(!off) {
1931         o(0xEA000000);              /* b */
1932         last_itod_magic=ind;
1933         o(0x4F800000);              /* 4294967296.0f */
1934       }
1935       o(0xBE000100|dsize|(r2<<16)|(r2<<12)|r); /* adflt */
1936     }
1937 #endif
1938     return;
1939   } else if(bt == VT_LLONG) {
1940     int func;
1941     CType *func_type = 0;
1942     if((t & VT_BTYPE) == VT_FLOAT) {
1943       func_type = &func_float_type;
1944       if(vtop->type.t & VT_UNSIGNED)
1945         func=TOK___floatundisf;
1946       else
1947         func=TOK___floatdisf;
1948 #if LDOUBLE_SIZE != 8
1949     } else if((t & VT_BTYPE) == VT_LDOUBLE) {
1950       func_type = &func_ldouble_type;
1951       if(vtop->type.t & VT_UNSIGNED)
1952         func=TOK___floatundixf;
1953       else
1954         func=TOK___floatdixf;
1955     } else if((t & VT_BTYPE) == VT_DOUBLE) {
1956 #else
1957     } else if((t & VT_BTYPE) == VT_DOUBLE || (t & VT_BTYPE) == VT_LDOUBLE) {
1958 #endif
1959       func_type = &func_double_type;
1960       if(vtop->type.t & VT_UNSIGNED)
1961         func=TOK___floatundidf;
1962       else
1963         func=TOK___floatdidf;
1964     }
1965     if(func_type) {
1966       vpush_global_sym(func_type, func);
1967       vswap();
1968       gfunc_call(1);
1969       vpushi(0);
1970       vtop->r=TREG_F0;
1971       return;
1972     }
1973   }
1974   tcc_error("unimplemented gen_cvt_itof %x!",vtop->type.t);
1975 }
1976
1977 /* convert fp to int 't' type */
1978 void gen_cvt_ftoi(int t)
1979 {
1980   uint32_t r, r2;
1981   int u, func = 0;
1982   u=t&VT_UNSIGNED;
1983   t&=VT_BTYPE;
1984   r2=vtop->type.t & VT_BTYPE;
1985   if(t==VT_INT) {
1986 #ifdef TCC_ARM_VFP
1987     r=vfpr(gv(RC_FLOAT));
1988     u=u?0:0x10000;
1989     o(0xEEBC0AC0|(r<<12)|r|T2CPR(r2)|u); /* ftoXizY */
1990     r2=intr(vtop->r=get_reg(RC_INT));
1991     o(0xEE100A10|(r<<16)|(r2<<12));
1992     return;
1993 #else
1994     if(u) {
1995       if(r2 == VT_FLOAT)
1996         func=TOK___fixunssfsi;
1997 #if LDOUBLE_SIZE != 8
1998       else if(r2 == VT_LDOUBLE)
1999         func=TOK___fixunsxfsi;
2000       else if(r2 == VT_DOUBLE)
2001 #else
2002       else if(r2 == VT_LDOUBLE || r2 == VT_DOUBLE)
2003 #endif
2004         func=TOK___fixunsdfsi;
2005     } else {
2006       r=fpr(gv(RC_FLOAT));
2007       r2=intr(vtop->r=get_reg(RC_INT));
2008       o(0xEE100170|(r2<<12)|r);
2009       return;
2010     }
2011 #endif
2012   } else if(t == VT_LLONG) { // unsigned handled in gen_cvt_ftoi1
2013     if(r2 == VT_FLOAT)
2014       func=TOK___fixsfdi;
2015 #if LDOUBLE_SIZE != 8
2016     else if(r2 == VT_LDOUBLE)
2017       func=TOK___fixxfdi;
2018     else if(r2 == VT_DOUBLE)
2019 #else
2020     else if(r2 == VT_LDOUBLE || r2 == VT_DOUBLE)
2021 #endif
2022       func=TOK___fixdfdi;
2023   }
2024   if(func) {
2025     vpush_global_sym(&func_old_type, func);
2026     vswap();
2027     gfunc_call(1);
2028     vpushi(0);
2029     if(t == VT_LLONG)
2030       vtop->r2 = REG_LRET;
2031     vtop->r = REG_IRET;
2032     return;
2033   }
2034   tcc_error("unimplemented gen_cvt_ftoi!");
2035 }
2036
2037 /* convert from one floating point type to another */
2038 void gen_cvt_ftof(int t)
2039 {
2040 #ifdef TCC_ARM_VFP
2041   if(((vtop->type.t & VT_BTYPE) == VT_FLOAT) != ((t & VT_BTYPE) == VT_FLOAT)) {
2042     uint32_t r = vfpr(gv(RC_FLOAT));
2043     o(0xEEB70AC0|(r<<12)|r|T2CPR(vtop->type.t));
2044   }
2045 #else
2046   /* all we have to do on i386 and FPA ARM is to put the float in a register */
2047   gv(RC_FLOAT);
2048 #endif
2049 }
2050
2051 /* computed goto support */
2052 void ggoto(void)
2053 {
2054   gcall_or_jmp(1);
2055   vtop--;
2056 }
2057
2058 /* Save the stack pointer onto the stack and return the location of its address */
2059 ST_FUNC void gen_vla_sp_save(int addr) {
2060     tcc_error("variable length arrays unsupported for this target");
2061 }
2062
2063 /* Restore the SP from a location on the stack */
2064 ST_FUNC void gen_vla_sp_restore(int addr) {
2065     tcc_error("variable length arrays unsupported for this target");
2066 }
2067
2068 /* Subtract from the stack pointer, and push the resulting value onto the stack */
2069 ST_FUNC void gen_vla_alloc(CType *type, int align) {
2070     tcc_error("variable length arrays unsupported for this target");
2071 }
2072
2073 /* end of ARM code generator */
2074 /*************************************************************/
2075 #endif
2076 /*************************************************************/