ljx

FORK: LuaJIT with native 5.2 and 5.3 support
git clone https://git.neptards.moe/neptards/ljx.git
Log | Files | Refs | README

lj_emit_arm.h (10866B)


      1 /*
      2 ** ARM instruction emitter.
      3 ** Copyright (C) 2005-2016 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 #if !LJ_SOFTFP
    107 static void emit_vlso(ASMState *as, ARMIns ai, Reg rd, Reg rn, int32_t ofs)
    108 {
    109   lua_assert(ofs >= -1020 && ofs <= 1020 && (ofs&3) == 0);
    110   if (ofs < 0) ofs = -ofs; else ai |= ARMI_LS_U;
    111   *--as->mcp = ai | ARMI_LS_P | ARMF_D(rd & 15) | ARMF_N(rn) | (ofs >> 2);
    112 }
    113 #endif
    114 
    115 /* -- Emit loads/stores --------------------------------------------------- */
    116 
    117 /* Prefer spills of BASE/L. */
    118 #define emit_canremat(ref)	((ref) < ASMREF_L)
    119 
    120 /* Try to find a one step delta relative to another constant. */
    121 static int emit_kdelta1(ASMState *as, Reg d, int32_t i)
    122 {
    123   RegSet work = ~as->freeset & RSET_GPR;
    124   while (work) {
    125     Reg r = rset_picktop(work);
    126     IRRef ref = regcost_ref(as->cost[r]);
    127     lua_assert(r != d);
    128     if (emit_canremat(ref)) {
    129       int32_t delta = i - (ra_iskref(ref) ? ra_krefk(as, ref) : IR(ref)->i);
    130       uint32_t k = emit_isk12(ARMI_ADD, delta);
    131       if (k) {
    132 	if (k == ARMI_K12)
    133 	  emit_dm(as, ARMI_MOV, d, r);
    134 	else
    135 	  emit_dn(as, ARMI_ADD^k, d, r);
    136 	return 1;
    137       }
    138     }
    139     rset_clear(work, r);
    140   }
    141   return 0;  /* Failed. */
    142 }
    143 
    144 /* Try to find a two step delta relative to another constant. */
    145 static int emit_kdelta2(ASMState *as, Reg d, int32_t i)
    146 {
    147   RegSet work = ~as->freeset & RSET_GPR;
    148   while (work) {
    149     Reg r = rset_picktop(work);
    150     IRRef ref = regcost_ref(as->cost[r]);
    151     lua_assert(r != d);
    152     if (emit_canremat(ref)) {
    153       int32_t other = ra_iskref(ref) ? ra_krefk(as, ref) : IR(ref)->i;
    154       if (other) {
    155 	int32_t delta = i - other;
    156 	uint32_t sh, inv = 0, k2, k;
    157 	if (delta < 0) { delta = -delta; inv = ARMI_ADD^ARMI_SUB; }
    158 	sh = lj_ffs(delta) & ~1;
    159 	k2 = emit_isk12(0, delta & (255 << sh));
    160 	k = emit_isk12(0, delta & ~(255 << sh));
    161 	if (k) {
    162 	  emit_dn(as, ARMI_ADD^k2^inv, d, d);
    163 	  emit_dn(as, ARMI_ADD^k^inv, d, r);
    164 	  return 1;
    165 	}
    166       }
    167     }
    168     rset_clear(work, r);
    169   }
    170   return 0;  /* Failed. */
    171 }
    172 
    173 /* Load a 32 bit constant into a GPR. */
    174 static void emit_loadi(ASMState *as, Reg r, int32_t i)
    175 {
    176   uint32_t k = emit_isk12(ARMI_MOV, i);
    177   lua_assert(rset_test(as->freeset, r) || r == RID_TMP);
    178   if (k) {
    179     /* Standard K12 constant. */
    180     emit_d(as, ARMI_MOV^k, r);
    181   } else if ((as->flags & JIT_F_ARMV6T2) && (uint32_t)i < 0x00010000u) {
    182     /* 16 bit loword constant for ARMv6T2. */
    183     emit_d(as, ARMI_MOVW|(i & 0x0fff)|((i & 0xf000)<<4), r);
    184   } else if (emit_kdelta1(as, r, i)) {
    185     /* One step delta relative to another constant. */
    186   } else if ((as->flags & JIT_F_ARMV6T2)) {
    187     /* 32 bit hiword/loword constant for ARMv6T2. */
    188     emit_d(as, ARMI_MOVT|((i>>16) & 0x0fff)|(((i>>16) & 0xf000)<<4), r);
    189     emit_d(as, ARMI_MOVW|(i & 0x0fff)|((i & 0xf000)<<4), r);
    190   } else if (emit_kdelta2(as, r, i)) {
    191     /* Two step delta relative to another constant. */
    192   } else {
    193     /* Otherwise construct the constant with up to 4 instructions. */
    194     /* NYI: use mvn+bic, use pc-relative loads. */
    195     for (;;) {
    196       uint32_t sh = lj_ffs(i) & ~1;
    197       int32_t m = i & (255 << sh);
    198       i &= ~(255 << sh);
    199       if (i == 0) {
    200 	emit_d(as, ARMI_MOV ^ emit_isk12(0, m), r);
    201 	break;
    202       }
    203       emit_dn(as, ARMI_ORR ^ emit_isk12(0, m), r, r);
    204     }
    205   }
    206 }
    207 
    208 #define emit_loada(as, r, addr)		emit_loadi(as, (r), i32ptr((addr)))
    209 
    210 static Reg ra_allock(ASMState *as, int32_t k, RegSet allow);
    211 
    212 /* Get/set from constant pointer. */
    213 static void emit_lsptr(ASMState *as, ARMIns ai, Reg r, void *p)
    214 {
    215   int32_t i = i32ptr(p);
    216   emit_lso(as, ai, r, ra_allock(as, (i & ~4095), rset_exclude(RSET_GPR, r)),
    217 	   (i & 4095));
    218 }
    219 
    220 #if !LJ_SOFTFP
    221 /* Load a number constant into an FPR. */
    222 static void emit_loadk64(ASMState *as, Reg r, IRIns *ir)
    223 {
    224   cTValue *tv = ir_knum(ir);
    225   int32_t i;
    226   if ((as->flags & JIT_F_VFPV3) && !tv->u32.lo) {
    227     uint32_t hi = tv->u32.hi;
    228     uint32_t b = ((hi >> 22) & 0x1ff);
    229     if (!(hi & 0xffff) && (b == 0x100 || b == 0x0ff)) {
    230       *--as->mcp = ARMI_VMOVI_D | ARMF_D(r & 15) |
    231 		   ((tv->u32.hi >> 12) & 0x00080000) |
    232 		   ((tv->u32.hi >> 4) & 0x00070000) |
    233 		   ((tv->u32.hi >> 16) & 0x0000000f);
    234       return;
    235     }
    236   }
    237   i = i32ptr(tv);
    238   emit_vlso(as, ARMI_VLDR_D, r,
    239 	    ra_allock(as, (i & ~1020), RSET_GPR), (i & 1020));
    240 }
    241 #endif
    242 
    243 /* Get/set global_State fields. */
    244 #define emit_getgl(as, r, field) \
    245   emit_lsptr(as, ARMI_LDR, (r), (void *)&J2G(as->J)->field)
    246 #define emit_setgl(as, r, field) \
    247   emit_lsptr(as, ARMI_STR, (r), (void *)&J2G(as->J)->field)
    248 
    249 /* Trace number is determined from pc of exit instruction. */
    250 #define emit_setvmstate(as, i)		UNUSED(i)
    251 
    252 /* -- Emit control-flow instructions -------------------------------------- */
    253 
    254 /* Label for internal jumps. */
    255 typedef MCode *MCLabel;
    256 
    257 /* Return label pointing to current PC. */
    258 #define emit_label(as)		((as)->mcp)
    259 
    260 static void emit_branch(ASMState *as, ARMIns ai, MCode *target)
    261 {
    262   MCode *p = as->mcp;
    263   ptrdiff_t delta = (target - p) - 1;
    264   lua_assert(((delta + 0x00800000) >> 24) == 0);
    265   *--p = ai | ((uint32_t)delta & 0x00ffffffu);
    266   as->mcp = p;
    267 }
    268 
    269 #define emit_jmp(as, target) emit_branch(as, ARMI_B, (target))
    270 
    271 static void emit_call(ASMState *as, void *target)
    272 {
    273   MCode *p = --as->mcp;
    274   ptrdiff_t delta = ((char *)target - (char *)p) - 8;
    275   if ((((delta>>2) + 0x00800000) >> 24) == 0) {
    276     if ((delta & 1))
    277       *p = ARMI_BLX | ((uint32_t)(delta>>2) & 0x00ffffffu) | ((delta&2) << 27);
    278     else
    279       *p = ARMI_BL | ((uint32_t)(delta>>2) & 0x00ffffffu);
    280   } else {  /* Target out of range: need indirect call. But don't use R0-R3. */
    281     Reg r = ra_allock(as, i32ptr(target), RSET_RANGE(RID_R4, RID_R12+1));
    282     *p = ARMI_BLXr | ARMF_M(r);
    283   }
    284 }
    285 
    286 /* -- Emit generic operations --------------------------------------------- */
    287 
    288 /* Generic move between two regs. */
    289 static void emit_movrr(ASMState *as, IRIns *ir, Reg dst, Reg src)
    290 {
    291 #if LJ_SOFTFP
    292   lua_assert(!irt_isnum(ir->t)); UNUSED(ir);
    293 #else
    294   if (dst >= RID_MAX_GPR) {
    295     emit_dm(as, irt_isnum(ir->t) ? ARMI_VMOV_D : ARMI_VMOV_S,
    296 	    (dst & 15), (src & 15));
    297     return;
    298   }
    299 #endif
    300   if (as->mcp != as->mcloop) {  /* Swap early registers for loads/stores. */
    301     MCode ins = *as->mcp, swp = (src^dst);
    302     if ((ins & 0x0c000000) == 0x04000000 && (ins & 0x02000010) != 0x02000010) {
    303       if (!((ins ^ (dst << 16)) & 0x000f0000))
    304 	*as->mcp = ins ^ (swp << 16);  /* Swap N in load/store. */
    305       if (!(ins & 0x00100000) && !((ins ^ (dst << 12)) & 0x0000f000))
    306 	*as->mcp = ins ^ (swp << 12);  /* Swap D in store. */
    307     }
    308   }
    309   emit_dm(as, ARMI_MOV, dst, src);
    310 }
    311 
    312 /* Generic load of register with base and (small) offset address. */
    313 static void emit_loadofs(ASMState *as, IRIns *ir, Reg r, Reg base, int32_t ofs)
    314 {
    315 #if LJ_SOFTFP
    316   lua_assert(!irt_isnum(ir->t)); UNUSED(ir);
    317 #else
    318   if (r >= RID_MAX_GPR)
    319     emit_vlso(as, irt_isnum(ir->t) ? ARMI_VLDR_D : ARMI_VLDR_S, r, base, ofs);
    320   else
    321 #endif
    322     emit_lso(as, ARMI_LDR, r, base, ofs);
    323 }
    324 
    325 /* Generic store of register with base and (small) offset address. */
    326 static void emit_storeofs(ASMState *as, IRIns *ir, Reg r, Reg base, int32_t ofs)
    327 {
    328 #if LJ_SOFTFP
    329   lua_assert(!irt_isnum(ir->t)); UNUSED(ir);
    330 #else
    331   if (r >= RID_MAX_GPR)
    332     emit_vlso(as, irt_isnum(ir->t) ? ARMI_VSTR_D : ARMI_VSTR_S, r, base, ofs);
    333   else
    334 #endif
    335     emit_lso(as, ARMI_STR, r, base, ofs);
    336 }
    337 
    338 /* Emit an arithmetic/logic operation with a constant operand. */
    339 static void emit_opk(ASMState *as, ARMIns ai, Reg dest, Reg src,
    340 		     int32_t i, RegSet allow)
    341 {
    342   uint32_t k = emit_isk12(ai, i);
    343   if (k)
    344     emit_dn(as, ai^k, dest, src);
    345   else
    346     emit_dnm(as, ai, dest, src, ra_allock(as, i, allow));
    347 }
    348 
    349 /* Add offset to pointer. */
    350 static void emit_addptr(ASMState *as, Reg r, int32_t ofs)
    351 {
    352   if (ofs)
    353     emit_opk(as, ARMI_ADD, r, r, ofs, rset_exclude(RSET_GPR, r));
    354 }
    355 
    356 #define emit_spsub(as, ofs)	emit_addptr(as, RID_SP, -(ofs))
    357