src/lj_emit_arm.h

   1 /*
   2 ** ARM instruction emitter.
   3 ** Copyright (C) 2005-2011 Mike Pall. See Copyright Notice in luajit.h
   4 */
   5
   6 /* -- Constant encoding --------------------------------------------------- */
   7
   8 static uint8_t emit_invai[16] = {
   9   /* AND */ (ARMI_AND^ARMI_BIC) >> 21,
  10   /* EOR */ 0,
  11   /* SUB */ (ARMI_SUB^ARMI_ADD) >> 21,
  12   /* RSB */ 0,
  13   /* ADD */ (ARMI_ADD^ARMI_SUB) >> 21,
  14   /* ADC */ (ARMI_ADC^ARMI_SBC) >> 21,
  15   /* SBC */ (ARMI_SBC^ARMI_ADC) >> 21,
  16   /* RSC */ 0,
  17   /* TST */ 0,
  18   /* TEQ */ 0,
  19   /* CMP */ (ARMI_CMP^ARMI_CMN) >> 21,
  20   /* CMN */ (ARMI_CMN^ARMI_CMP) >> 21,
  21   /* ORR */ 0,
  22   /* MOV */ (ARMI_MOV^ARMI_MVN) >> 21,
  23   /* BIC */ (ARMI_BIC^ARMI_AND) >> 21,
  24   /* MVN */ (ARMI_MVN^ARMI_MOV) >> 21
  25 };
  26
  27 /* Encode constant in K12 format for data processing instructions. */
  28 static uint32_t emit_isk12(ARMIns ai, int32_t n)
  29 {
  30   uint32_t invai, i, m = (uint32_t)n;
  31   /* K12: unsigned 8 bit value, rotated in steps of two bits. */
  32   for (i = 0; i < 4096; i += 256, m = lj_rol(m, 2))
  33     if (m <= 255) return ARMI_K12|m|i;
  34   /* Otherwise try negation/complement with the inverse instruction. */
  35   invai = emit_invai[((ai >> 21) & 15)];
  36   if (!invai) return 0;  /* Failed. No inverse instruction. */
  37   m = ~(uint32_t)n;
  38   if (invai == ((ARMI_SUB^ARMI_ADD) >> 21) ||
  39       invai == (ARMI_CMP^ARMI_CMN) >> 21) m++;
  40   for (i = 0; i < 4096; i += 256, m = lj_rol(m, 2))
  41     if (m <= 255) return ARMI_K12|(invai<<21)|m|i;
  42   return 0;  /* Failed. */
  43 }
  44
  45 /* -- Emit basic instructions --------------------------------------------- */
  46
  47 static void emit_dnm(ASMState *as, ARMIns ai, Reg rd, Reg rn, Reg rm)
  48 {
  49   *--as->mcp = ai | ARMF_D(rd) | ARMF_N(rn) | ARMF_M(rm);
  50 }
  51
  52 static void emit_dm(ASMState *as, ARMIns ai, Reg rd, Reg rm)
  53 {
  54   *--as->mcp = ai | ARMF_D(rd) | ARMF_M(rm);
  55 }
  56
  57 static void emit_dn(ASMState *as, ARMIns ai, Reg rd, Reg rn)
  58 {
  59   *--as->mcp = ai | ARMF_D(rd) | ARMF_N(rn);
  60 }
  61
  62 static void emit_nm(ASMState *as, ARMIns ai, Reg rn, Reg rm)
  63 {
  64   *--as->mcp = ai | ARMF_N(rn) | ARMF_M(rm);
  65 }
  66
  67 static void emit_d(ASMState *as, ARMIns ai, Reg rd)
  68 {
  69   *--as->mcp = ai | ARMF_D(rd);
  70 }
  71
  72 static void emit_n(ASMState *as, ARMIns ai, Reg rn)
  73 {
  74   *--as->mcp = ai | ARMF_N(rn);
  75 }
  76
  77 static void emit_m(ASMState *as, ARMIns ai, Reg rm)
  78 {
  79   *--as->mcp = ai | ARMF_M(rm);
  80 }
  81
  82 static void emit_lsox(ASMState *as, ARMIns ai, Reg rd, Reg rn, int32_t ofs)
  83 {
  84   lua_assert(ofs >= -255 && ofs <= 255);
  85   if (ofs < 0) ofs = -ofs; else ai |= ARMI_LS_U;
  86   *--as->mcp = ai | ARMI_LS_P | ARMI_LSX_I | ARMF_D(rd) | ARMF_N(rn) |
  87                ((ofs & 0xf0) << 4) | (ofs & 0x0f);
  88 }
  89
  90 static void emit_lso(ASMState *as, ARMIns ai, Reg rd, Reg rn, int32_t ofs)
  91 {
  92   lua_assert(ofs >= -4095 && ofs <= 4095);
  93   /* Combine LDR/STR pairs to LDRD/STRD. */
  94   if (*as->mcp == (ai|ARMI_LS_P|ARMI_LS_U|ARMF_D(rd^1)|ARMF_N(rn)|(ofs^4)) &&
  95       (ai & ~(ARMI_LDR^ARMI_STR)) == ARMI_STR && rd != rn &&
  96       (uint32_t)ofs <= 252 && !(ofs & 3) && !((rd ^ (ofs >>2)) & 1) &&
  97       as->mcp != as->mcloop) {
  98     as->mcp++;
  99     emit_lsox(as, ai == ARMI_LDR ? ARMI_LDRD : ARMI_STRD, rd&~1, rn, ofs&~4);
 100     return;
 101   }
 102   if (ofs < 0) ofs = -ofs; else ai |= ARMI_LS_U;
 103   *--as->mcp = ai | ARMI_LS_P | ARMF_D(rd) | ARMF_N(rn) | ofs;
 104 }
 105
 106 /* -- Emit loads/stores --------------------------------------------------- */
 107
 108 /* Prefer spills of BASE/L. */
 109 #define emit_canremat(ref)      ((ref) < ASMREF_L)
 110
 111 /* Try to find a one step delta relative to another constant. */
 112 static int emit_kdelta1(ASMState *as, Reg d, int32_t i)
 113 {
 114   RegSet work = ~as->freeset & RSET_GPR;
 115   while (work) {
 116     Reg r = rset_picktop(work);
 117     IRRef ref = regcost_ref(as->cost[r]);
 118     lua_assert(r != d);
 119     if (emit_canremat(ref)) {
 120       int32_t delta = i - (ra_iskref(ref) ? ra_krefk(as, ref) : IR(ref)->i);
 121       uint32_t k = emit_isk12(ARMI_ADD, delta);
 122       if (k) {
 123         if (k == ARMI_K12)
 124           emit_dm(as, ARMI_MOV, d, r);
 125         else
 126           emit_dn(as, ARMI_ADD^k, d, r);
 127         return 1;
 128       }
 129     }
 130     rset_clear(work, r);
 131   }
 132   return 0;  /* Failed. */
 133 }
 134
 135 /* Try to find a two step delta relative to another constant. */
 136 static int emit_kdelta2(ASMState *as, Reg d, int32_t i)
 137 {
 138   RegSet work = ~as->freeset & RSET_GPR;
 139   while (work) {
 140     Reg r = rset_picktop(work);
 141     IRRef ref = regcost_ref(as->cost[r]);
 142     lua_assert(r != d);
 143     if (emit_canremat(ref)) {
 144       int32_t delta = i - (ra_iskref(ref) ? ra_krefk(as, ref) : IR(ref)->i);
 145       uint32_t sh, inv = 0, k2, k;
 146       if (delta < 0) { delta = -delta; inv = ARMI_ADD^ARMI_SUB; }
 147       sh = lj_ffs(delta) & ~1;
 148       k2 = emit_isk12(0, delta & (255 << sh));
 149       k = emit_isk12(0, delta & ~(255 << sh));
 150       if (k) {
 151         emit_dn(as, ARMI_ADD^k2^inv, d, d);
 152         emit_dn(as, ARMI_ADD^k^inv, d, r);
 153         return 1;
 154       }
 155     }
 156     rset_clear(work, r);
 157   }
 158   return 0;  /* Failed. */
 159 }
 160
 161 /* Load a 32 bit constant into a GPR. */
 162 static void emit_loadi(ASMState *as, Reg r, int32_t i)
 163 {
 164   uint32_t k = emit_isk12(ARMI_MOV, i);
 165   lua_assert(rset_test(as->freeset, r) || r == RID_TMP);
 166   if (k) {
 167     /* Standard K12 constant. */
 168     emit_d(as, ARMI_MOV^k, r);
 169   } else if ((as->flags & JIT_F_ARMV6T2) && (uint32_t)i < 0x00010000u) {
 170     /* 16 bit loword constant for ARMv6T2. */
 171     emit_d(as, ARMI_MOVW|(i & 0x0fff)|((i & 0xf000)<<4), r);
 172   } else if (emit_kdelta1(as, r, i)) {
 173     /* One step delta relative to another constant. */
 174   } else if ((as->flags & JIT_F_ARMV6T2)) {
 175     /* 32 bit hiword/loword constant for ARMv6T2. */
 176     emit_d(as, ARMI_MOVT|((i>>16) & 0x0fff)|(((i>>16) & 0xf000)<<4), r);
 177     emit_d(as, ARMI_MOVW|(i & 0x0fff)|((i & 0xf000)<<4), r);
 178   } else if (emit_kdelta2(as, r, i)) {
 179     /* Two step delta relative to another constant. */
 180   } else {
 181     /* Otherwise construct the constant with up to 4 instructions. */
 182     /* NYI: use mvn+bic, use pc-relative loads. */
 183     for (;;) {
 184       uint32_t sh = lj_ffs(i) & ~1;
 185       int32_t m = i & (255 << sh);
 186       i &= ~(255 << sh);
 187       if (i == 0) {
 188         emit_d(as, ARMI_MOV ^ emit_isk12(0, m), r);
 189         break;
 190       }
 191       emit_dn(as, ARMI_ORR ^ emit_isk12(0, m), r, r);
 192     }
 193   }
 194 }
 195
 196 #define emit_loada(as, r, addr)         emit_loadi(as, (r), i32ptr((addr)))
 197
 198 static Reg ra_allock(ASMState *as, int32_t k, RegSet allow);
 199
 200 /* Get/set from constant pointer. */
 201 static void emit_lsptr(ASMState *as, ARMIns ai, Reg r, void *p)
 202 {
 203   int32_t i = i32ptr(p);
 204   emit_lso(as, ai, r, ra_allock(as, (i & ~4095), rset_exclude(RSET_GPR, r)),
 205            (i & 4095));
 206 }
 207
 208 /* Get/set global_State fields. */
 209 #define emit_getgl(as, r, field) \
 210   emit_lsptr(as, ARMI_LDR, (r), (void *)&J2G(as->J)->field)
 211 #define emit_setgl(as, r, field) \
 212   emit_lsptr(as, ARMI_STR, (r), (void *)&J2G(as->J)->field)
 213
 214 /* Trace number is determined from pc of exit instruction. */
 215 #define emit_setvmstate(as, i)          UNUSED(i)
 216
 217 /* -- Emit control-flow instructions -------------------------------------- */
 218
 219 /* Label for internal jumps. */
 220 typedef MCode *MCLabel;
 221
 222 /* Return label pointing to current PC. */
 223 #define emit_label(as)          ((as)->mcp)
 224
 225 static void emit_branch(ASMState *as, ARMIns ai, MCode *target)
 226 {
 227   MCode *p = as->mcp;
 228   ptrdiff_t delta = (target - p) - 1;
 229   lua_assert(((delta + 0x00800000) >> 24) == 0);
 230   *--p = ai | ((uint32_t)delta & 0x00ffffffu);
 231   as->mcp = p;
 232 }
 233
 234 static void emit_call(ASMState *as, void *target)
 235 {
 236   MCode *p = --as->mcp;
 237   ptrdiff_t delta = ((char *)target - (char *)p) - 8;
 238   if ((((delta>>2) + 0x00800000) >> 24) == 0) {
 239     if ((delta & 1))
 240       *p = ARMI_BLX | ((uint32_t)(delta>>2) & 0x00ffffffu) | ((delta&2) << 27);
 241     else
 242       *p = ARMI_BL | ((uint32_t)(delta>>2) & 0x00ffffffu);
 243   } else {  /* Target out of range: need indirect call. But don't use R0-R3. */
 244     Reg r = ra_allock(as, i32ptr(target), RSET_RANGE(RID_R4, RID_R12+1));
 245     *p = ARMI_BLXr | ARMF_M(r);
 246   }
 247 }
 248
 249 /* -- Emit generic operations --------------------------------------------- */
 250
 251 /* Generic move between two regs. */
 252 static void emit_movrr(ASMState *as, IRIns *ir, Reg dst, Reg src)
 253 {
 254   lua_assert(!irt_isnum(ir->t)); UNUSED(ir);
 255   if (as->mcp != as->mcloop) {  /* Swap early registers for loads/stores. */
 256     MCode ins = *as->mcp, swp = (src^dst);
 257     if ((ins & 0x0c000000) == 0x04000000 && (ins & 0x02000010) != 0x02000010) {
 258       if (!((ins ^ (dst << 16)) & 0x000f0000))
 259         *as->mcp = ins ^ (swp << 16);  /* Swap N in load/store. */
 260       if (!(ins & 0x00100000) && !((ins ^ (dst << 12)) & 0x0000f000))
 261         *as->mcp = ins ^ (swp << 12);  /* Swap D in store. */
 262     }
 263   }
 264   emit_dm(as, ARMI_MOV, dst, src);
 265 }
 266
 267 /* Generic load of register from stack slot. */
 268 static void emit_spload(ASMState *as, IRIns *ir, Reg r, int32_t ofs)
 269 {
 270   lua_assert(!irt_isnum(ir->t)); UNUSED(ir);
 271   emit_lso(as, ARMI_LDR, r, RID_SP, ofs);
 272 }
 273
 274 /* Generic store of register to stack slot. */
 275 static void emit_spstore(ASMState *as, IRIns *ir, Reg r, int32_t ofs)
 276 {
 277   lua_assert(!irt_isnum(ir->t)); UNUSED(ir);
 278   emit_lso(as, ARMI_STR, r, RID_SP, ofs);
 279 }
 280
 281 /* Emit an arithmetic/logic operation with a constant operand. */
 282 static void emit_opk(ASMState *as, ARMIns ai, Reg dest, Reg src,
 283                      int32_t i, RegSet allow)
 284 {
 285   uint32_t k = emit_isk12(ai, i);
 286   if (k)
 287     emit_dn(as, ai^k, dest, src);
 288   else
 289     emit_dnm(as, ai, dest, src, ra_allock(as, i, allow));
 290 }
 291
 292 /* Add offset to pointer. */
 293 static void emit_addptr(ASMState *as, Reg r, int32_t ofs)
 294 {
 295   if (ofs)
 296     emit_opk(as, ARMI_ADD, r, r, ofs, rset_exclude(RSET_GPR, r));
 297 }
 298
 299 #define emit_spsub(as, ofs)     emit_addptr(as, RID_SP, -(ofs))
 300