lang/luajit/patches/050-ppc-softfloat.patch

   1 From 71b7bc88341945f13f3951e2bb5fd247b639ff7a Mon Sep 17 00:00:00 2001
   2 From: Mike Pall <mike>
   3 Date: Sun, 3 Sep 2017 23:20:53 +0200
   4 Subject: [PATCH] PPC: Add soft-float support to JIT compiler backend.
   5
   6 Contributed by Djordje Kovacevic and Stefan Pejic from RT-RK.com.
   7 Sponsored by Cisco Systems, Inc.
   8 ---
   9  src/lj_arch.h    |   1 -
  10  src/lj_asm_ppc.h | 321 ++++++++++++++++++++++++++++++++++++++++-------
  11  2 files changed, 278 insertions(+), 44 deletions(-)
  12
  13 --- a/src/lj_arch.h
  14 +++ b/src/lj_arch.h
  15 @@ -273,7 +273,6 @@
  16  #endif
  17
  18  #if LJ_ABI_SOFTFP
  19 -#define LJ_ARCH_NOJIT          1  /* NYI */
  20  #define LJ_ARCH_NUMMODE                LJ_NUMMODE_DUAL
  21  #else
  22  #define LJ_ARCH_NUMMODE                LJ_NUMMODE_DUAL_SINGLE
  23 --- a/src/lj_asm_ppc.h
  24 +++ b/src/lj_asm_ppc.h
  25 @@ -226,6 +226,7 @@ static void asm_fusexrefx(ASMState *as,
  26    emit_tab(as, pi, rt, left, right);
  27  }
  28
  29 +#if !LJ_SOFTFP
  30  /* Fuse to multiply-add/sub instruction. */
  31  static int asm_fusemadd(ASMState *as, IRIns *ir, PPCIns pi, PPCIns pir)
  32  {
  33 @@ -245,6 +246,7 @@ static int asm_fusemadd(ASMState *as, IR
  34    }
  35    return 0;
  36  }
  37 +#endif
  38
  39  /* -- Calls --------------------------------------------------------------- */
  40
  41 @@ -253,13 +255,17 @@ static void asm_gencall(ASMState *as, co
  42  {
  43    uint32_t n, nargs = CCI_XNARGS(ci);
  44    int32_t ofs = 8;
  45 -  Reg gpr = REGARG_FIRSTGPR, fpr = REGARG_FIRSTFPR;
  46 +  Reg gpr = REGARG_FIRSTGPR;
  47 +#if !LJ_SOFTFP
  48 +  Reg fpr = REGARG_FIRSTFPR;
  49 +#endif
  50    if ((void *)ci->func)
  51      emit_call(as, (void *)ci->func);
  52    for (n = 0; n < nargs; n++) {  /* Setup args. */
  53      IRRef ref = args[n];
  54      if (ref) {
  55        IRIns *ir = IR(ref);
  56 +#if !LJ_SOFTFP
  57        if (irt_isfp(ir->t)) {
  58         if (fpr <= REGARG_LASTFPR) {
  59           lua_assert(rset_test(as->freeset, fpr));  /* Already evicted. */
  60 @@ -271,7 +277,9 @@ static void asm_gencall(ASMState *as, co
  61           emit_spstore(as, ir, r, ofs);
  62           ofs += irt_isnum(ir->t) ? 8 : 4;
  63         }
  64 -      } else {
  65 +      } else
  66 +#endif
  67 +      {
  68         if (gpr <= REGARG_LASTGPR) {
  69           lua_assert(rset_test(as->freeset, gpr));  /* Already evicted. */
  70           ra_leftov(as, gpr, ref);
  71 @@ -290,8 +298,10 @@ static void asm_gencall(ASMState *as, co
  72      }
  73      checkmclim(as);
  74    }
  75 +#if !LJ_SOFTFP
  76    if ((ci->flags & CCI_VARARG))  /* Vararg calls need to know about FPR use. */
  77      emit_tab(as, fpr == REGARG_FIRSTFPR ? PPCI_CRXOR : PPCI_CREQV, 6, 6, 6);
  78 +#endif
  79  }
  80
  81  /* Setup result reg/sp for call. Evict scratch regs. */
  82 @@ -299,8 +309,10 @@ static void asm_setupresult(ASMState *as
  83  {
  84    RegSet drop = RSET_SCRATCH;
  85    int hiop = ((ir+1)->o == IR_HIOP && !irt_isnil((ir+1)->t));
  86 +#if !LJ_SOFTFP
  87    if ((ci->flags & CCI_NOFPRCLOBBER))
  88      drop &= ~RSET_FPR;
  89 +#endif
  90    if (ra_hasreg(ir->r))
  91      rset_clear(drop, ir->r);  /* Dest reg handled below. */
  92    if (hiop && ra_hasreg((ir+1)->r))
  93 @@ -308,7 +320,7 @@ static void asm_setupresult(ASMState *as
  94    ra_evictset(as, drop);  /* Evictions must be performed first. */
  95    if (ra_used(ir)) {
  96      lua_assert(!irt_ispri(ir->t));
  97 -    if (irt_isfp(ir->t)) {
  98 +    if (!LJ_SOFTFP && irt_isfp(ir->t)) {
  99        if ((ci->flags & CCI_CASTU64)) {
 100         /* Use spill slot or temp slots. */
 101         int32_t ofs = ir->s ? sps_scale(ir->s) : SPOFS_TMP;
 102 @@ -377,6 +389,7 @@ static void asm_retf(ASMState *as, IRIns
 103
 104  /* -- Type conversions ---------------------------------------------------- */
 105
 106 +#if !LJ_SOFTFP
 107  static void asm_tointg(ASMState *as, IRIns *ir, Reg left)
 108  {
 109    RegSet allow = RSET_FPR;
 110 @@ -409,15 +422,23 @@ static void asm_tobit(ASMState *as, IRIn
 111    emit_fai(as, PPCI_STFD, tmp, RID_SP, SPOFS_TMP);
 112    emit_fab(as, PPCI_FADD, tmp, left, right);
 113  }
 114 +#endif
 115
 116  static void asm_conv(ASMState *as, IRIns *ir)
 117  {
 118    IRType st = (IRType)(ir->op2 & IRCONV_SRCMASK);
 119 +#if !LJ_SOFTFP
 120    int stfp = (st == IRT_NUM || st == IRT_FLOAT);
 121 +#endif
 122    IRRef lref = ir->op1;
 123 -  lua_assert(irt_type(ir->t) != st);
 124    lua_assert(!(irt_isint64(ir->t) ||
 125                (st == IRT_I64 || st == IRT_U64))); /* Handled by SPLIT. */
 126 +#if LJ_SOFTFP
 127 +  /* FP conversions are handled by SPLIT. */
 128 +  lua_assert(!irt_isfp(ir->t) && !(st == IRT_NUM || st == IRT_FLOAT));
 129 +  /* Can't check for same types: SPLIT uses CONV int.int + BXOR for sfp NEG. */
 130 +#else
 131 +  lua_assert(irt_type(ir->t) != st);
 132    if (irt_isfp(ir->t)) {
 133      Reg dest = ra_dest(as, ir, RSET_FPR);
 134      if (stfp) {  /* FP to FP conversion. */
 135 @@ -476,7 +497,9 @@ static void asm_conv(ASMState *as, IRIns
 136         emit_fb(as, PPCI_FCTIWZ, tmp, left);
 137        }
 138      }
 139 -  } else {
 140 +  } else
 141 +#endif
 142 +  {
 143      Reg dest = ra_dest(as, ir, RSET_GPR);
 144      if (st >= IRT_I8 && st <= IRT_U16) {  /* Extend to 32 bit integer. */
 145        Reg left = ra_alloc1(as, ir->op1, RSET_GPR);
 146 @@ -496,17 +519,41 @@ static void asm_strto(ASMState *as, IRIn
 147  {
 148    const CCallInfo *ci = &lj_ir_callinfo[IRCALL_lj_strscan_num];
 149    IRRef args[2];
 150 -  int32_t ofs;
 151 +  int32_t ofs = SPOFS_TMP;
 152 +#if LJ_SOFTFP
 153 +  ra_evictset(as, RSET_SCRATCH);
 154 +  if (ra_used(ir)) {
 155 +    if (ra_hasspill(ir->s) && ra_hasspill((ir+1)->s) &&
 156 +       (ir->s & 1) == LJ_BE && (ir->s ^ 1) == (ir+1)->s) {
 157 +      int i;
 158 +      for (i = 0; i < 2; i++) {
 159 +       Reg r = (ir+i)->r;
 160 +       if (ra_hasreg(r)) {
 161 +         ra_free(as, r);
 162 +         ra_modified(as, r);
 163 +         emit_spload(as, ir+i, r, sps_scale((ir+i)->s));
 164 +       }
 165 +      }
 166 +      ofs = sps_scale(ir->s & ~1);
 167 +    } else {
 168 +      Reg rhi = ra_dest(as, ir+1, RSET_GPR);
 169 +      Reg rlo = ra_dest(as, ir, rset_exclude(RSET_GPR, rhi));
 170 +      emit_tai(as, PPCI_LWZ, rhi, RID_SP, ofs);
 171 +      emit_tai(as, PPCI_LWZ, rlo, RID_SP, ofs+4);
 172 +    }
 173 +  }
 174 +#else
 175    RegSet drop = RSET_SCRATCH;
 176    if (ra_hasreg(ir->r)) rset_set(drop, ir->r);  /* Spill dest reg (if any). */
 177    ra_evictset(as, drop);
 178 +  if (ir->s) ofs = sps_scale(ir->s);
 179 +#endif
 180    asm_guardcc(as, CC_EQ);
 181    emit_ai(as, PPCI_CMPWI, RID_RET, 0);  /* Test return status. */
 182    args[0] = ir->op1;      /* GCstr *str */
 183    args[1] = ASMREF_TMP1;  /* TValue *n  */
 184    asm_gencall(as, ci, args);
 185    /* Store the result to the spill slot or temp slots. */
 186 -  ofs = ir->s ? sps_scale(ir->s) : SPOFS_TMP;
 187    emit_tai(as, PPCI_ADDI, ra_releasetmp(as, ASMREF_TMP1), RID_SP, ofs);
 188  }
 189
 190 @@ -530,7 +577,10 @@ static void asm_tvptr(ASMState *as, Reg
 191        Reg src = ra_alloc1(as, ref, allow);
 192        emit_setgl(as, src, tmptv.gcr);
 193      }
 194 -    type = ra_allock(as, irt_toitype(ir->t), allow);
 195 +    if (LJ_SOFTFP && (ir+1)->o == IR_HIOP)
 196 +      type = ra_alloc1(as, ref+1, allow);
 197 +    else
 198 +      type = ra_allock(as, irt_toitype(ir->t), allow);
 199      emit_setgl(as, type, tmptv.it);
 200    }
 201  }
 202 @@ -574,11 +624,27 @@ static void asm_href(ASMState *as, IRIns
 203    Reg tisnum = RID_NONE, tmpnum = RID_NONE;
 204    IRRef refkey = ir->op2;
 205    IRIns *irkey = IR(refkey);
 206 +  int isk = irref_isk(refkey);
 207    IRType1 kt = irkey->t;
 208    uint32_t khash;
 209    MCLabel l_end, l_loop, l_next;
 210
 211    rset_clear(allow, tab);
 212 +#if LJ_SOFTFP
 213 +  if (!isk) {
 214 +    key = ra_alloc1(as, refkey, allow);
 215 +    rset_clear(allow, key);
 216 +    if (irkey[1].o == IR_HIOP) {
 217 +      if (ra_hasreg((irkey+1)->r)) {
 218 +       tmpnum = (irkey+1)->r;
 219 +       ra_noweak(as, tmpnum);
 220 +      } else {
 221 +       tmpnum = ra_allocref(as, refkey+1, allow);
 222 +      }
 223 +      rset_clear(allow, tmpnum);
 224 +    }
 225 +  }
 226 +#else
 227    if (irt_isnum(kt)) {
 228      key = ra_alloc1(as, refkey, RSET_FPR);
 229      tmpnum = ra_scratch(as, rset_exclude(RSET_FPR, key));
 230 @@ -588,6 +654,7 @@ static void asm_href(ASMState *as, IRIns
 231      key = ra_alloc1(as, refkey, allow);
 232      rset_clear(allow, key);
 233    }
 234 +#endif
 235    tmp2 = ra_scratch(as, allow);
 236    rset_clear(allow, tmp2);
 237
 238 @@ -610,7 +677,7 @@ static void asm_href(ASMState *as, IRIns
 239      asm_guardcc(as, CC_EQ);
 240    else
 241      emit_condbranch(as, PPCI_BC|PPCF_Y, CC_EQ, l_end);
 242 -  if (irt_isnum(kt)) {
 243 +  if (!LJ_SOFTFP && irt_isnum(kt)) {
 244      emit_fab(as, PPCI_FCMPU, 0, tmpnum, key);
 245      emit_condbranch(as, PPCI_BC, CC_GE, l_next);
 246      emit_ab(as, PPCI_CMPLW, tmp1, tisnum);
 247 @@ -620,7 +687,10 @@ static void asm_href(ASMState *as, IRIns
 248        emit_ab(as, PPCI_CMPW, tmp2, key);
 249        emit_condbranch(as, PPCI_BC, CC_NE, l_next);
 250      }
 251 -    emit_ai(as, PPCI_CMPWI, tmp1, irt_toitype(irkey->t));
 252 +    if (LJ_SOFTFP && ra_hasreg(tmpnum))
 253 +      emit_ab(as, PPCI_CMPW, tmp1, tmpnum);
 254 +    else
 255 +      emit_ai(as, PPCI_CMPWI, tmp1, irt_toitype(irkey->t));
 256      if (!irt_ispri(kt))
 257        emit_tai(as, PPCI_LWZ, tmp2, dest, (int32_t)offsetof(Node, key.gcr));
 258    }
 259 @@ -629,19 +699,19 @@ static void asm_href(ASMState *as, IRIns
 260             (((char *)as->mcp-(char *)l_loop) & 0xffffu);
 261
 262    /* Load main position relative to tab->node into dest. */
 263 -  khash = irref_isk(refkey) ? ir_khash(irkey) : 1;
 264 +  khash = isk ? ir_khash(irkey) : 1;
 265    if (khash == 0) {
 266      emit_tai(as, PPCI_LWZ, dest, tab, (int32_t)offsetof(GCtab, node));
 267    } else {
 268      Reg tmphash = tmp1;
 269 -    if (irref_isk(refkey))
 270 +    if (isk)
 271        tmphash = ra_allock(as, khash, allow);
 272      emit_tab(as, PPCI_ADD, dest, dest, tmp1);
 273      emit_tai(as, PPCI_MULLI, tmp1, tmp1, sizeof(Node));
 274      emit_asb(as, PPCI_AND, tmp1, tmp2, tmphash);
 275      emit_tai(as, PPCI_LWZ, dest, tab, (int32_t)offsetof(GCtab, node));
 276      emit_tai(as, PPCI_LWZ, tmp2, tab, (int32_t)offsetof(GCtab, hmask));
 277 -    if (irref_isk(refkey)) {
 278 +    if (isk) {
 279        /* Nothing to do. */
 280      } else if (irt_isstr(kt)) {
 281        emit_tai(as, PPCI_LWZ, tmp1, key, (int32_t)offsetof(GCstr, hash));
 282 @@ -651,13 +721,19 @@ static void asm_href(ASMState *as, IRIns
 283        emit_asb(as, PPCI_XOR, tmp1, tmp1, tmp2);
 284        emit_rotlwi(as, tmp1, tmp1, (HASH_ROT2+HASH_ROT1)&31);
 285        emit_tab(as, PPCI_SUBF, tmp2, dest, tmp2);
 286 -      if (irt_isnum(kt)) {
 287 +      if (LJ_SOFTFP ? (irkey[1].o == IR_HIOP) : irt_isnum(kt)) {
 288 +#if LJ_SOFTFP
 289 +       emit_asb(as, PPCI_XOR, tmp2, key, tmp1);
 290 +       emit_rotlwi(as, dest, tmp1, HASH_ROT1);
 291 +       emit_tab(as, PPCI_ADD, tmp1, tmpnum, tmpnum);
 292 +#else
 293         int32_t ofs = ra_spill(as, irkey);
 294         emit_asb(as, PPCI_XOR, tmp2, tmp2, tmp1);
 295         emit_rotlwi(as, dest, tmp1, HASH_ROT1);
 296         emit_tab(as, PPCI_ADD, tmp1, tmp1, tmp1);
 297         emit_tai(as, PPCI_LWZ, tmp2, RID_SP, ofs+4);
 298         emit_tai(as, PPCI_LWZ, tmp1, RID_SP, ofs);
 299 +#endif
 300        } else {
 301         emit_asb(as, PPCI_XOR, tmp2, key, tmp1);
 302         emit_rotlwi(as, dest, tmp1, HASH_ROT1);
 303 @@ -784,8 +860,8 @@ static PPCIns asm_fxloadins(IRIns *ir)
 304    case IRT_U8: return PPCI_LBZ;
 305    case IRT_I16: return PPCI_LHA;
 306    case IRT_U16: return PPCI_LHZ;
 307 -  case IRT_NUM: return PPCI_LFD;
 308 -  case IRT_FLOAT: return PPCI_LFS;
 309 +  case IRT_NUM: lua_assert(!LJ_SOFTFP); return PPCI_LFD;
 310 +  case IRT_FLOAT: if (!LJ_SOFTFP) return PPCI_LFS;
 311    default: return PPCI_LWZ;
 312    }
 313  }
 314 @@ -795,8 +871,8 @@ static PPCIns asm_fxstoreins(IRIns *ir)
 315    switch (irt_type(ir->t)) {
 316    case IRT_I8: case IRT_U8: return PPCI_STB;
 317    case IRT_I16: case IRT_U16: return PPCI_STH;
 318 -  case IRT_NUM: return PPCI_STFD;
 319 -  case IRT_FLOAT: return PPCI_STFS;
 320 +  case IRT_NUM: lua_assert(!LJ_SOFTFP); return PPCI_STFD;
 321 +  case IRT_FLOAT: if (!LJ_SOFTFP) return PPCI_STFS;
 322    default: return PPCI_STW;
 323    }
 324  }
 325 @@ -839,7 +915,8 @@ static void asm_fstore(ASMState *as, IRI
 326
 327  static void asm_xload(ASMState *as, IRIns *ir)
 328  {
 329 -  Reg dest = ra_dest(as, ir, irt_isfp(ir->t) ? RSET_FPR : RSET_GPR);
 330 +  Reg dest = ra_dest(as, ir,
 331 +    (!LJ_SOFTFP && irt_isfp(ir->t)) ? RSET_FPR : RSET_GPR);
 332    lua_assert(!(ir->op2 & IRXLOAD_UNALIGNED));
 333    if (irt_isi8(ir->t))
 334      emit_as(as, PPCI_EXTSB, dest, dest);
 335 @@ -857,7 +934,8 @@ static void asm_xstore_(ASMState *as, IR
 336      Reg src = ra_alloc1(as, irb->op1, RSET_GPR);
 337      asm_fusexrefx(as, PPCI_STWBRX, src, ir->op1, rset_exclude(RSET_GPR, src));
 338    } else {
 339 -    Reg src = ra_alloc1(as, ir->op2, irt_isfp(ir->t) ? RSET_FPR : RSET_GPR);
 340 +    Reg src = ra_alloc1(as, ir->op2,
 341 +      (!LJ_SOFTFP && irt_isfp(ir->t)) ? RSET_FPR : RSET_GPR);
 342      asm_fusexref(as, asm_fxstoreins(ir), src, ir->op1,
 343                  rset_exclude(RSET_GPR, src), ofs);
 344    }
 345 @@ -871,10 +949,19 @@ static void asm_ahuvload(ASMState *as, I
 346    Reg dest = RID_NONE, type = RID_TMP, tmp = RID_TMP, idx;
 347    RegSet allow = RSET_GPR;
 348    int32_t ofs = AHUREF_LSX;
 349 +  if (LJ_SOFTFP && (ir+1)->o == IR_HIOP) {
 350 +    t.irt = IRT_NUM;
 351 +    if (ra_used(ir+1)) {
 352 +      type = ra_dest(as, ir+1, allow);
 353 +      rset_clear(allow, type);
 354 +    }
 355 +    ofs = 0;
 356 +  }
 357    if (ra_used(ir)) {
 358 -    lua_assert(irt_isnum(t) || irt_isint(t) || irt_isaddr(t));
 359 -    if (!irt_isnum(t)) ofs = 0;
 360 -    dest = ra_dest(as, ir, irt_isnum(t) ? RSET_FPR : RSET_GPR);
 361 +    lua_assert((LJ_SOFTFP ? 0 : irt_isnum(ir->t)) ||
 362 +              irt_isint(ir->t) || irt_isaddr(ir->t));
 363 +    if (LJ_SOFTFP || !irt_isnum(t)) ofs = 0;
 364 +    dest = ra_dest(as, ir, (!LJ_SOFTFP && irt_isnum(t)) ? RSET_FPR : allow);
 365      rset_clear(allow, dest);
 366    }
 367    idx = asm_fuseahuref(as, ir->op1, &ofs, allow);
 368 @@ -883,12 +970,13 @@ static void asm_ahuvload(ASMState *as, I
 369      asm_guardcc(as, CC_GE);
 370      emit_ab(as, PPCI_CMPLW, type, tisnum);
 371      if (ra_hasreg(dest)) {
 372 -      if (ofs == AHUREF_LSX) {
 373 +      if (!LJ_SOFTFP && ofs == AHUREF_LSX) {
 374         tmp = ra_scratch(as, rset_exclude(rset_exclude(RSET_GPR,
 375                                                        (idx&255)), (idx>>8)));
 376         emit_fab(as, PPCI_LFDX, dest, (idx&255), tmp);
 377        } else {
 378 -       emit_fai(as, PPCI_LFD, dest, idx, ofs);
 379 +       emit_fai(as, LJ_SOFTFP ? PPCI_LWZ : PPCI_LFD, dest, idx,
 380 +                ofs+4*LJ_SOFTFP);
 381        }
 382      }
 383    } else {
 384 @@ -911,7 +999,7 @@ static void asm_ahustore(ASMState *as, I
 385    int32_t ofs = AHUREF_LSX;
 386    if (ir->r == RID_SINK)
 387      return;
 388 -  if (irt_isnum(ir->t)) {
 389 +  if (!LJ_SOFTFP && irt_isnum(ir->t)) {
 390      src = ra_alloc1(as, ir->op2, RSET_FPR);
 391    } else {
 392      if (!irt_ispri(ir->t)) {
 393 @@ -919,11 +1007,14 @@ static void asm_ahustore(ASMState *as, I
 394        rset_clear(allow, src);
 395        ofs = 0;
 396      }
 397 -    type = ra_allock(as, (int32_t)irt_toitype(ir->t), allow);
 398 +    if (LJ_SOFTFP && (ir+1)->o == IR_HIOP)
 399 +      type = ra_alloc1(as, (ir+1)->op2, allow);
 400 +    else
 401 +      type = ra_allock(as, (int32_t)irt_toitype(ir->t), allow);
 402      rset_clear(allow, type);
 403    }
 404    idx = asm_fuseahuref(as, ir->op1, &ofs, allow);
 405 -  if (irt_isnum(ir->t)) {
 406 +  if (!LJ_SOFTFP && irt_isnum(ir->t)) {
 407      if (ofs == AHUREF_LSX) {
 408        emit_fab(as, PPCI_STFDX, src, (idx&255), RID_TMP);
 409        emit_slwi(as, RID_TMP, (idx>>8), 3);
 410 @@ -948,21 +1039,33 @@ static void asm_sload(ASMState *as, IRIn
 411    IRType1 t = ir->t;
 412    Reg dest = RID_NONE, type = RID_NONE, base;
 413    RegSet allow = RSET_GPR;
 414 +  int hiop = (LJ_SOFTFP && (ir+1)->o == IR_HIOP);
 415 +  if (hiop)
 416 +    t.irt = IRT_NUM;
 417    lua_assert(!(ir->op2 & IRSLOAD_PARENT));  /* Handled by asm_head_side(). */
 418 -  lua_assert(irt_isguard(t) || !(ir->op2 & IRSLOAD_TYPECHECK));
 419 +  lua_assert(irt_isguard(ir->t) || !(ir->op2 & IRSLOAD_TYPECHECK));
 420    lua_assert(LJ_DUALNUM ||
 421              !irt_isint(t) || (ir->op2 & (IRSLOAD_CONVERT|IRSLOAD_FRAME)));
 422 +#if LJ_SOFTFP
 423 +  lua_assert(!(ir->op2 & IRSLOAD_CONVERT));  /* Handled by LJ_SOFTFP SPLIT. */
 424 +  if (hiop && ra_used(ir+1)) {
 425 +    type = ra_dest(as, ir+1, allow);
 426 +    rset_clear(allow, type);
 427 +  }
 428 +#else
 429    if ((ir->op2 & IRSLOAD_CONVERT) && irt_isguard(t) && irt_isint(t)) {
 430      dest = ra_scratch(as, RSET_FPR);
 431      asm_tointg(as, ir, dest);
 432      t.irt = IRT_NUM;  /* Continue with a regular number type check. */
 433 -  } else if (ra_used(ir)) {
 434 +  } else
 435 +#endif
 436 +  if (ra_used(ir)) {
 437      lua_assert(irt_isnum(t) || irt_isint(t) || irt_isaddr(t));
 438 -    dest = ra_dest(as, ir, irt_isnum(t) ? RSET_FPR : RSET_GPR);
 439 +    dest = ra_dest(as, ir, (!LJ_SOFTFP && irt_isnum(t)) ? RSET_FPR : allow);
 440      rset_clear(allow, dest);
 441      base = ra_alloc1(as, REF_BASE, allow);
 442      rset_clear(allow, base);
 443 -    if ((ir->op2 & IRSLOAD_CONVERT)) {
 444 +    if (!LJ_SOFTFP && (ir->op2 & IRSLOAD_CONVERT)) {
 445        if (irt_isint(t)) {
 446         emit_tai(as, PPCI_LWZ, dest, RID_SP, SPOFS_TMPLO);
 447         dest = ra_scratch(as, RSET_FPR);
 448 @@ -994,10 +1097,13 @@ dotypecheck:
 449      if ((ir->op2 & IRSLOAD_TYPECHECK)) {
 450        Reg tisnum = ra_allock(as, (int32_t)LJ_TISNUM, allow);
 451        asm_guardcc(as, CC_GE);
 452 -      emit_ab(as, PPCI_CMPLW, RID_TMP, tisnum);
 453 +#if !LJ_SOFTFP
 454        type = RID_TMP;
 455 +#endif
 456 +      emit_ab(as, PPCI_CMPLW, type, tisnum);
 457      }
 458 -    if (ra_hasreg(dest)) emit_fai(as, PPCI_LFD, dest, base, ofs-4);
 459 +    if (ra_hasreg(dest)) emit_fai(as, LJ_SOFTFP ? PPCI_LWZ : PPCI_LFD, dest,
 460 +                                 base, ofs-(LJ_SOFTFP?0:4));
 461    } else {
 462      if ((ir->op2 & IRSLOAD_TYPECHECK)) {
 463        asm_guardcc(as, CC_NE);
 464 @@ -1119,6 +1225,7 @@ static void asm_obar(ASMState *as, IRIns
 465
 466  /* -- Arithmetic and logic operations ------------------------------------- */
 467
 468 +#if !LJ_SOFTFP
 469  static void asm_fparith(ASMState *as, IRIns *ir, PPCIns pi)
 470  {
 471    Reg dest = ra_dest(as, ir, RSET_FPR);
 472 @@ -1146,13 +1253,17 @@ static void asm_fpmath(ASMState *as, IRI
 473    else
 474      asm_callid(as, ir, IRCALL_lj_vm_floor + ir->op2);
 475  }
 476 +#endif
 477
 478  static void asm_add(ASMState *as, IRIns *ir)
 479  {
 480 +#if !LJ_SOFTFP
 481    if (irt_isnum(ir->t)) {
 482      if (!asm_fusemadd(as, ir, PPCI_FMADD, PPCI_FMADD))
 483        asm_fparith(as, ir, PPCI_FADD);
 484 -  } else {
 485 +  } else
 486 +#endif
 487 +  {
 488      Reg dest = ra_dest(as, ir, RSET_GPR);
 489      Reg right, left = ra_hintalloc(as, ir->op1, dest, RSET_GPR);
 490      PPCIns pi;
 491 @@ -1191,10 +1302,13 @@ static void asm_add(ASMState *as, IRIns
 492
 493  static void asm_sub(ASMState *as, IRIns *ir)
 494  {
 495 +#if !LJ_SOFTFP
 496    if (irt_isnum(ir->t)) {
 497      if (!asm_fusemadd(as, ir, PPCI_FMSUB, PPCI_FNMSUB))
 498        asm_fparith(as, ir, PPCI_FSUB);
 499 -  } else {
 500 +  } else
 501 +#endif
 502 +  {
 503      PPCIns pi = PPCI_SUBF;
 504      Reg dest = ra_dest(as, ir, RSET_GPR);
 505      Reg left, right;
 506 @@ -1220,9 +1334,12 @@ static void asm_sub(ASMState *as, IRIns
 507
 508  static void asm_mul(ASMState *as, IRIns *ir)
 509  {
 510 +#if !LJ_SOFTFP
 511    if (irt_isnum(ir->t)) {
 512      asm_fparith(as, ir, PPCI_FMUL);
 513 -  } else {
 514 +  } else
 515 +#endif
 516 +  {
 517      PPCIns pi = PPCI_MULLW;
 518      Reg dest = ra_dest(as, ir, RSET_GPR);
 519      Reg right, left = ra_hintalloc(as, ir->op1, dest, RSET_GPR);
 520 @@ -1250,9 +1367,12 @@ static void asm_mul(ASMState *as, IRIns
 521
 522  static void asm_neg(ASMState *as, IRIns *ir)
 523  {
 524 +#if !LJ_SOFTFP
 525    if (irt_isnum(ir->t)) {
 526      asm_fpunary(as, ir, PPCI_FNEG);
 527 -  } else {
 528 +  } else
 529 +#endif
 530 +  {
 531      Reg dest, left;
 532      PPCIns pi = PPCI_NEG;
 533      if (as->flagmcp == as->mcp) {
 534 @@ -1563,9 +1683,40 @@ static void asm_bitshift(ASMState *as, I
 535                        PPCI_RLWINM|PPCF_MB(0)|PPCF_ME(31))
 536  #define asm_bror(as, ir)       lua_assert(0)
 537
 538 +#if LJ_SOFTFP
 539 +static void asm_sfpmin_max(ASMState *as, IRIns *ir)
 540 +{
 541 +  CCallInfo ci = lj_ir_callinfo[IRCALL_softfp_cmp];
 542 +  IRRef args[4];
 543 +  MCLabel l_right, l_end;
 544 +  Reg desthi = ra_dest(as, ir, RSET_GPR), destlo = ra_dest(as, ir+1, RSET_GPR);
 545 +  Reg righthi, lefthi = ra_alloc2(as, ir, RSET_GPR);
 546 +  Reg rightlo, leftlo = ra_alloc2(as, ir+1, RSET_GPR);
 547 +  PPCCC cond = (IROp)ir->o == IR_MIN ? CC_EQ : CC_NE;
 548 +  righthi = (lefthi >> 8); lefthi &= 255;
 549 +  rightlo = (leftlo >> 8); leftlo &= 255;
 550 +  args[0^LJ_BE] = ir->op1; args[1^LJ_BE] = (ir+1)->op1;
 551 +  args[2^LJ_BE] = ir->op2; args[3^LJ_BE] = (ir+1)->op2;
 552 +  l_end = emit_label(as);
 553 +  if (desthi != righthi) emit_mr(as, desthi, righthi);
 554 +  if (destlo != rightlo) emit_mr(as, destlo, rightlo);
 555 +  l_right = emit_label(as);
 556 +  if (l_end != l_right) emit_jmp(as, l_end);
 557 +  if (desthi != lefthi) emit_mr(as, desthi, lefthi);
 558 +  if (destlo != leftlo) emit_mr(as, destlo, leftlo);
 559 +  if (l_right == as->mcp+1) {
 560 +    cond ^= 4; l_right = l_end; ++as->mcp;
 561 +  }
 562 +  emit_condbranch(as, PPCI_BC, cond, l_right);
 563 +  ra_evictset(as, RSET_SCRATCH);
 564 +  emit_cmpi(as, RID_RET, 1);
 565 +  asm_gencall(as, &ci, args);
 566 +}
 567 +#endif
 568 +
 569  static void asm_min_max(ASMState *as, IRIns *ir, int ismax)
 570  {
 571 -  if (irt_isnum(ir->t)) {
 572 +  if (!LJ_SOFTFP && irt_isnum(ir->t)) {
 573      Reg dest = ra_dest(as, ir, RSET_FPR);
 574      Reg tmp = dest;
 575      Reg right, left = ra_alloc2(as, ir, RSET_FPR);
 576 @@ -1653,7 +1804,7 @@ static void asm_intcomp_(ASMState *as, I
 577  static void asm_comp(ASMState *as, IRIns *ir)
 578  {
 579    PPCCC cc = asm_compmap[ir->o];
 580 -  if (irt_isnum(ir->t)) {
 581 +  if (!LJ_SOFTFP && irt_isnum(ir->t)) {
 582      Reg right, left = ra_alloc2(as, ir, RSET_FPR);
 583      right = (left >> 8); left &= 255;
 584      asm_guardcc(as, (cc >> 4));
 585 @@ -1674,6 +1825,44 @@ static void asm_comp(ASMState *as, IRIns
 586
 587  #define asm_equal(as, ir)      asm_comp(as, ir)
 588
 589 +#if LJ_SOFTFP
 590 +/* SFP comparisons. */
 591 +static void asm_sfpcomp(ASMState *as, IRIns *ir)
 592 +{
 593 +  const CCallInfo *ci = &lj_ir_callinfo[IRCALL_softfp_cmp];
 594 +  RegSet drop = RSET_SCRATCH;
 595 +  Reg r;
 596 +  IRRef args[4];
 597 +  args[0^LJ_BE] = ir->op1; args[1^LJ_BE] = (ir+1)->op1;
 598 +  args[2^LJ_BE] = ir->op2; args[3^LJ_BE] = (ir+1)->op2;
 599 +
 600 +  for (r = REGARG_FIRSTGPR; r <= REGARG_FIRSTGPR+3; r++) {
 601 +    if (!rset_test(as->freeset, r) &&
 602 +       regcost_ref(as->cost[r]) == args[r-REGARG_FIRSTGPR])
 603 +      rset_clear(drop, r);
 604 +  }
 605 +  ra_evictset(as, drop);
 606 +  asm_setupresult(as, ir, ci);
 607 +  switch ((IROp)ir->o) {
 608 +  case IR_ULT:
 609 +    asm_guardcc(as, CC_EQ);
 610 +    emit_ai(as, PPCI_CMPWI, RID_RET, 0);
 611 +  case IR_ULE:
 612 +    asm_guardcc(as, CC_EQ);
 613 +    emit_ai(as, PPCI_CMPWI, RID_RET, 1);
 614 +    break;
 615 +  case IR_GE: case IR_GT:
 616 +    asm_guardcc(as, CC_EQ);
 617 +    emit_ai(as, PPCI_CMPWI, RID_RET, 2);
 618 +  default:
 619 +    asm_guardcc(as, (asm_compmap[ir->o] & 0xf));
 620 +    emit_ai(as, PPCI_CMPWI, RID_RET, 0);
 621 +    break;
 622 +  }
 623 +  asm_gencall(as, ci, args);
 624 +}
 625 +#endif
 626 +
 627  #if LJ_HASFFI
 628  /* 64 bit integer comparisons. */
 629  static void asm_comp64(ASMState *as, IRIns *ir)
 630 @@ -1703,19 +1892,36 @@ static void asm_comp64(ASMState *as, IRI
 631  /* Hiword op of a split 64 bit op. Previous op must be the loword op. */
 632  static void asm_hiop(ASMState *as, IRIns *ir)
 633  {
 634 -#if LJ_HASFFI
 635 +#if LJ_HASFFI || LJ_SOFTFP
 636    /* HIOP is marked as a store because it needs its own DCE logic. */
 637    int uselo = ra_used(ir-1), usehi = ra_used(ir);  /* Loword/hiword used? */
 638    if (LJ_UNLIKELY(!(as->flags & JIT_F_OPT_DCE))) uselo = usehi = 1;
 639    if ((ir-1)->o == IR_CONV) {  /* Conversions to/from 64 bit. */
 640      as->curins--;  /* Always skip the CONV. */
 641 +#if LJ_HASFFI && !LJ_SOFTFP
 642      if (usehi || uselo)
 643        asm_conv64(as, ir);
 644      return;
 645 +#endif
 646    } else if ((ir-1)->o <= IR_NE) {  /* 64 bit integer comparisons. ORDER IR. */
 647      as->curins--;  /* Always skip the loword comparison. */
 648 +#if LJ_SOFTFP
 649 +    if (!irt_isint(ir->t)) {
 650 +      asm_sfpcomp(as, ir-1);
 651 +      return;
 652 +    }
 653 +#endif
 654 +#if LJ_HASFFI
 655      asm_comp64(as, ir);
 656 +#endif
 657 +    return;
 658 +#if LJ_SOFTFP
 659 +  } else if ((ir-1)->o == IR_MIN || (ir-1)->o == IR_MAX) {
 660 +      as->curins--;  /* Always skip the loword min/max. */
 661 +    if (uselo || usehi)
 662 +      asm_sfpmin_max(as, ir-1);
 663      return;
 664 +#endif
 665    } else if ((ir-1)->o == IR_XSTORE) {
 666      as->curins--;  /* Handle both stores here. */
 667      if ((ir-1)->r != RID_SINK) {
 668 @@ -1726,14 +1932,27 @@ static void asm_hiop(ASMState *as, IRIns
 669    }
 670    if (!usehi) return;  /* Skip unused hiword op for all remaining ops. */
 671    switch ((ir-1)->o) {
 672 +#if LJ_HASFFI
 673    case IR_ADD: as->curins--; asm_add64(as, ir); break;
 674    case IR_SUB: as->curins--; asm_sub64(as, ir); break;
 675    case IR_NEG: as->curins--; asm_neg64(as, ir); break;
 676 +#endif
 677 +#if LJ_SOFTFP
 678 +  case IR_SLOAD: case IR_ALOAD: case IR_HLOAD: case IR_ULOAD: case IR_VLOAD:
 679 +  case IR_STRTO:
 680 +    if (!uselo)
 681 +      ra_allocref(as, ir->op1, RSET_GPR);  /* Mark lo op as used. */
 682 +    break;
 683 +#endif
 684    case IR_CALLN:
 685 +  case IR_CALLS:
 686    case IR_CALLXS:
 687      if (!uselo)
 688        ra_allocref(as, ir->op1, RID2RSET(RID_RETLO));  /* Mark lo op as used. */
 689      break;
 690 +#if LJ_SOFTFP
 691 +  case IR_ASTORE: case IR_HSTORE: case IR_USTORE: case IR_TOSTR:
 692 +#endif
 693    case IR_CNEWI:
 694      /* Nothing to do here. Handled by lo op itself. */
 695      break;
 696 @@ -1797,8 +2016,19 @@ static void asm_stack_restore(ASMState *
 697      if ((sn & SNAP_NORESTORE))
 698        continue;
 699      if (irt_isnum(ir->t)) {
 700 +#if LJ_SOFTFP
 701 +      Reg tmp;
 702 +      RegSet allow = rset_exclude(RSET_GPR, RID_BASE);
 703 +      lua_assert(irref_isk(ref));  /* LJ_SOFTFP: must be a number constant. */
 704 +      tmp = ra_allock(as, (int32_t)ir_knum(ir)->u32.lo, allow);
 705 +      emit_tai(as, PPCI_STW, tmp, RID_BASE, ofs+(LJ_BE?4:0));
 706 +      if (rset_test(as->freeset, tmp+1)) allow = RID2RSET(tmp+1);
 707 +      tmp = ra_allock(as, (int32_t)ir_knum(ir)->u32.hi, allow);
 708 +      emit_tai(as, PPCI_STW, tmp, RID_BASE, ofs+(LJ_BE?0:4));
 709 +#else
 710        Reg src = ra_alloc1(as, ref, RSET_FPR);
 711        emit_fai(as, PPCI_STFD, src, RID_BASE, ofs);
 712 +#endif
 713      } else {
 714        Reg type;
 715        RegSet allow = rset_exclude(RSET_GPR, RID_BASE);
 716 @@ -1811,6 +2041,10 @@ static void asm_stack_restore(ASMState *
 717        if ((sn & (SNAP_CONT|SNAP_FRAME))) {
 718         if (s == 0) continue;  /* Do not overwrite link to previous frame. */
 719         type = ra_allock(as, (int32_t)(*flinks--), allow);
 720 +#if LJ_SOFTFP
 721 +      } else if ((sn & SNAP_SOFTFPNUM)) {
 722 +       type = ra_alloc1(as, ref+1, rset_exclude(RSET_GPR, RID_BASE));
 723 +#endif
 724        } else {
 725         type = ra_allock(as, (int32_t)irt_toitype(ir->t), allow);
 726        }
 727 @@ -1947,14 +2181,15 @@ static Reg asm_setup_call_slots(ASMState
 728    int nslots = 2, ngpr = REGARG_NUMGPR, nfpr = REGARG_NUMFPR;
 729    asm_collectargs(as, ir, ci, args);
 730    for (i = 0; i < nargs; i++)
 731 -    if (args[i] && irt_isfp(IR(args[i])->t)) {
 732 +    if (!LJ_SOFTFP && args[i] && irt_isfp(IR(args[i])->t)) {
 733        if (nfpr > 0) nfpr--; else nslots = (nslots+3) & ~1;
 734      } else {
 735        if (ngpr > 0) ngpr--; else nslots++;
 736      }
 737    if (nslots > as->evenspill)  /* Leave room for args in stack slots. */
 738      as->evenspill = nslots;
 739 -  return irt_isfp(ir->t) ? REGSP_HINT(RID_FPRET) : REGSP_HINT(RID_RET);
 740 +  return (!LJ_SOFTFP && irt_isfp(ir->t)) ? REGSP_HINT(RID_FPRET) :
 741 +                                          REGSP_HINT(RID_RET);
 742  }
 743
 744  static void asm_setup_target(ASMState *as)