src/lj_opt_narrow.c

   1 /*
   2 ** NARROW: Narrowing of numbers to integers (double to int32_t).
   3 ** STRIPOV: Stripping of overflow checks.
   4 ** Copyright (C) 2005-2014 Mike Pall. See Copyright Notice in luajit.h
   5 */
   6
   7 #define lj_opt_narrow_c
   8 #define LUA_CORE
   9
  10 #include "lj_obj.h"
  11
  12 #if LJ_HASJIT
  13
  14 #include "lj_bc.h"
  15 #include "lj_ir.h"
  16 #include "lj_jit.h"
  17 #include "lj_iropt.h"
  18 #include "lj_trace.h"
  19 #include "lj_vm.h"
  20 #include "lj_strscan.h"
  21
  22 /* Rationale for narrowing optimizations:
  23 **
  24 ** Lua has only a single number type and this is a FP double by default.
  25 ** Narrowing doubles to integers does not pay off for the interpreter on a
  26 ** current-generation x86/x64 machine. Most FP operations need the same
  27 ** amount of execution resources as their integer counterparts, except
  28 ** with slightly longer latencies. Longer latencies are a non-issue for
  29 ** the interpreter, since they are usually hidden by other overhead.
  30 **
  31 ** The total CPU execution bandwidth is the sum of the bandwidth of the FP
  32 ** and the integer units, because they execute in parallel. The FP units
  33 ** have an equal or higher bandwidth than the integer units. Not using
  34 ** them means losing execution bandwidth. Moving work away from them to
  35 ** the already quite busy integer units is a losing proposition.
  36 **
  37 ** The situation for JIT-compiled code is a bit different: the higher code
  38 ** density makes the extra latencies much more visible. Tight loops expose
  39 ** the latencies for updating the induction variables. Array indexing
  40 ** requires narrowing conversions with high latencies and additional
  41 ** guards (to check that the index is really an integer). And many common
  42 ** optimizations only work on integers.
  43 **
  44 ** One solution would be speculative, eager narrowing of all number loads.
  45 ** This causes many problems, like losing -0 or the need to resolve type
  46 ** mismatches between traces. It also effectively forces the integer type
  47 ** to have overflow-checking semantics. This impedes many basic
  48 ** optimizations and requires adding overflow checks to all integer
  49 ** arithmetic operations (whereas FP arithmetics can do without).
  50 **
  51 ** Always replacing an FP op with an integer op plus an overflow check is
  52 ** counter-productive on a current-generation super-scalar CPU. Although
  53 ** the overflow check branches are highly predictable, they will clog the
  54 ** execution port for the branch unit and tie up reorder buffers. This is
  55 ** turning a pure data-flow dependency into a different data-flow
  56 ** dependency (with slightly lower latency) *plus* a control dependency.
  57 ** In general, you don't want to do this since latencies due to data-flow
  58 ** dependencies can be well hidden by out-of-order execution.
  59 **
  60 ** A better solution is to keep all numbers as FP values and only narrow
  61 ** when it's beneficial to do so. LuaJIT uses predictive narrowing for
  62 ** induction variables and demand-driven narrowing for index expressions,
  63 ** integer arguments and bit operations. Additionally it can eliminate or
  64 ** hoist most of the resulting overflow checks. Regular arithmetic
  65 ** computations are never narrowed to integers.
  66 **
  67 ** The integer type in the IR has convenient wrap-around semantics and
  68 ** ignores overflow. Extra operations have been added for
  69 ** overflow-checking arithmetic (ADDOV/SUBOV) instead of an extra type.
  70 ** Apart from reducing overall complexity of the compiler, this also
  71 ** nicely solves the problem where you want to apply algebraic
  72 ** simplifications to ADD, but not to ADDOV. And the x86/x64 assembler can
  73 ** use lea instead of an add for integer ADD, but not for ADDOV (lea does
  74 ** not affect the flags, but it helps to avoid register moves).
  75 **
  76 **
  77 ** All of the above has to be reconsidered for architectures with slow FP
  78 ** operations or without a hardware FPU. The dual-number mode of LuaJIT
  79 ** addresses this issue. Arithmetic operations are performed on integers
  80 ** as far as possible and overflow checks are added as needed.
  81 **
  82 ** This implies that narrowing for integer arguments and bit operations
  83 ** should also strip overflow checks, e.g. replace ADDOV with ADD. The
  84 ** original overflow guards are weak and can be eliminated by DCE, if
  85 ** there's no other use.
  86 **
  87 ** A slight twist is that it's usually beneficial to use overflow-checked
  88 ** integer arithmetics if all inputs are already integers. This is the only
  89 ** change that affects the single-number mode, too.
  90 */
  91
  92 /* Some local macros to save typing. Undef'd at the end. */
  93 #define IR(ref)                 (&J->cur.ir[(ref)])
  94 #define fins                    (&J->fold.ins)
  95
  96 /* Pass IR on to next optimization in chain (FOLD). */
  97 #define emitir(ot, a, b)        (lj_ir_set(J, (ot), (a), (b)), lj_opt_fold(J))
  98
  99 #define emitir_raw(ot, a, b)    (lj_ir_set(J, (ot), (a), (b)), lj_ir_emit(J))
 100
 101 /* -- Elimination of narrowing type conversions --------------------------- */
 102
 103 /* Narrowing of index expressions and bit operations is demand-driven. The
 104 ** trace recorder emits a narrowing type conversion (CONV.int.num or TOBIT)
 105 ** in all of these cases (e.g. array indexing or string indexing). FOLD
 106 ** already takes care of eliminating simple redundant conversions like
 107 ** CONV.int.num(CONV.num.int(x)) ==> x.
 108 **
 109 ** But the surrounding code is FP-heavy and arithmetic operations are
 110 ** performed on FP numbers (for the single-number mode). Consider a common
 111 ** example such as 'x=t[i+1]', with 'i' already an integer (due to induction
 112 ** variable narrowing). The index expression would be recorded as
 113 **   CONV.int.num(ADD(CONV.num.int(i), 1))
 114 ** which is clearly suboptimal.
 115 **
 116 ** One can do better by recursively backpropagating the narrowing type
 117 ** conversion across FP arithmetic operations. This turns FP ops into
 118 ** their corresponding integer counterparts. Depending on the semantics of
 119 ** the conversion they also need to check for overflow. Currently only ADD
 120 ** and SUB are supported.
 121 **
 122 ** The above example can be rewritten as
 123 **   ADDOV(CONV.int.num(CONV.num.int(i)), 1)
 124 ** and then into ADDOV(i, 1) after folding of the conversions. The original
 125 ** FP ops remain in the IR and are eliminated by DCE since all references to
 126 ** them are gone.
 127 **
 128 ** [In dual-number mode the trace recorder already emits ADDOV etc., but
 129 ** this can be further reduced. See below.]
 130 **
 131 ** Special care has to be taken to avoid narrowing across an operation
 132 ** which is potentially operating on non-integral operands. One obvious
 133 ** case is when an expression contains a non-integral constant, but ends
 134 ** up as an integer index at runtime (like t[x+1.5] with x=0.5).
 135 **
 136 ** Operations with two non-constant operands illustrate a similar problem
 137 ** (like t[a+b] with a=1.5 and b=2.5). Backpropagation has to stop there,
 138 ** unless it can be proven that either operand is integral (e.g. by CSEing
 139 ** a previous conversion). As a not-so-obvious corollary this logic also
 140 ** applies for a whole expression tree (e.g. t[(a+1)+(b+1)]).
 141 **
 142 ** Correctness of the transformation is guaranteed by avoiding to expand
 143 ** the tree by adding more conversions than the one we would need to emit
 144 ** if not backpropagating. TOBIT employs a more optimistic rule, because
 145 ** the conversion has special semantics, designed to make the life of the
 146 ** compiler writer easier. ;-)
 147 **
 148 ** Using on-the-fly backpropagation of an expression tree doesn't work
 149 ** because it's unknown whether the transform is correct until the end.
 150 ** This either requires IR rollback and cache invalidation for every
 151 ** subtree or a two-pass algorithm. The former didn't work out too well,
 152 ** so the code now combines a recursive collector with a stack-based
 153 ** emitter.
 154 **
 155 ** [A recursive backpropagation algorithm with backtracking, employing
 156 ** skip-list lookup and round-robin caching, emitting stack operations
 157 ** on-the-fly for a stack-based interpreter -- and all of that in a meager
 158 ** kilobyte? Yep, compilers are a great treasure chest. Throw away your
 159 ** textbooks and read the codebase of a compiler today!]
 160 **
 161 ** There's another optimization opportunity for array indexing: it's
 162 ** always accompanied by an array bounds-check. The outermost overflow
 163 ** check may be delegated to the ABC operation. This works because ABC is
 164 ** an unsigned comparison and wrap-around due to overflow creates negative
 165 ** numbers.
 166 **
 167 ** But this optimization is only valid for constants that cannot overflow
 168 ** an int32_t into the range of valid array indexes [0..2^27+1). A check
 169 ** for +-2^30 is safe since -2^31 - 2^30 wraps to 2^30 and 2^31-1 + 2^30
 170 ** wraps to -2^30-1.
 171 **
 172 ** It's also good enough in practice, since e.g. t[i+1] or t[i-10] are
 173 ** quite common. So the above example finally ends up as ADD(i, 1)!
 174 **
 175 ** Later on, the assembler is able to fuse the whole array reference and
 176 ** the ADD into the memory operands of loads and other instructions. This
 177 ** is why LuaJIT is able to generate very pretty (and fast) machine code
 178 ** for array indexing. And that, my dear, concludes another story about
 179 ** one of the hidden secrets of LuaJIT ...
 180 */
 181
 182 /* Maximum backpropagation depth and maximum stack size. */
 183 #define NARROW_MAX_BACKPROP     100
 184 #define NARROW_MAX_STACK        256
 185
 186 /* The stack machine has a 32 bit instruction format: [IROpT | IRRef1]
 187 ** The lower 16 bits hold a reference (or 0). The upper 16 bits hold
 188 ** the IR opcode + type or one of the following special opcodes:
 189 */
 190 enum {
 191   NARROW_REF,           /* Push ref. */
 192   NARROW_CONV,          /* Push conversion of ref. */
 193   NARROW_SEXT,          /* Push sign-extension of ref. */
 194   NARROW_INT            /* Push KINT ref. The next code holds an int32_t. */
 195 };
 196
 197 typedef uint32_t NarrowIns;
 198
 199 #define NARROWINS(op, ref)      (((op) << 16) + (ref))
 200 #define narrow_op(ins)          ((IROpT)((ins) >> 16))
 201 #define narrow_ref(ins)         ((IRRef1)(ins))
 202
 203 /* Context used for narrowing of type conversions. */
 204 typedef struct NarrowConv {
 205   jit_State *J;         /* JIT compiler state. */
 206   NarrowIns *sp;        /* Current stack pointer. */
 207   NarrowIns *maxsp;     /* Maximum stack pointer minus redzone. */
 208   int lim;              /* Limit on the number of emitted conversions. */
 209   IRRef mode;           /* Conversion mode (IRCONV_*). */
 210   IRType t;             /* Destination type: IRT_INT or IRT_I64. */
 211   NarrowIns stack[NARROW_MAX_STACK];  /* Stack holding stack-machine code. */
 212 } NarrowConv;
 213
 214 /* Lookup a reference in the backpropagation cache. */
 215 static BPropEntry *narrow_bpc_get(jit_State *J, IRRef1 key, IRRef mode)
 216 {
 217   ptrdiff_t i;
 218   for (i = 0; i < BPROP_SLOTS; i++) {
 219     BPropEntry *bp = &J->bpropcache[i];
 220     /* Stronger checks are ok, too. */
 221     if (bp->key == key && bp->mode >= mode &&
 222         ((bp->mode ^ mode) & IRCONV_MODEMASK) == 0)
 223       return bp;
 224   }
 225   return NULL;
 226 }
 227
 228 /* Add an entry to the backpropagation cache. */
 229 static void narrow_bpc_set(jit_State *J, IRRef1 key, IRRef1 val, IRRef mode)
 230 {
 231   uint32_t slot = J->bpropslot;
 232   BPropEntry *bp = &J->bpropcache[slot];
 233   J->bpropslot = (slot + 1) & (BPROP_SLOTS-1);
 234   bp->key = key;
 235   bp->val = val;
 236   bp->mode = mode;
 237 }
 238
 239 /* Backpropagate overflow stripping. */
 240 static void narrow_stripov_backprop(NarrowConv *nc, IRRef ref, int depth)
 241 {
 242   jit_State *J = nc->J;
 243   IRIns *ir = IR(ref);
 244   if (ir->o == IR_ADDOV || ir->o == IR_SUBOV ||
 245       (ir->o == IR_MULOV && (nc->mode & IRCONV_CONVMASK) == IRCONV_ANY)) {
 246     BPropEntry *bp = narrow_bpc_get(nc->J, ref, IRCONV_TOBIT);
 247     if (bp) {
 248       ref = bp->val;
 249     } else if (++depth < NARROW_MAX_BACKPROP && nc->sp < nc->maxsp) {
 250       narrow_stripov_backprop(nc, ir->op1, depth);
 251       narrow_stripov_backprop(nc, ir->op2, depth);
 252       *nc->sp++ = NARROWINS(IRT(ir->o - IR_ADDOV + IR_ADD, IRT_INT), ref);
 253       return;
 254     }
 255   }
 256   *nc->sp++ = NARROWINS(NARROW_REF, ref);
 257 }
 258
 259 /* Backpropagate narrowing conversion. Return number of needed conversions. */
 260 static int narrow_conv_backprop(NarrowConv *nc, IRRef ref, int depth)
 261 {
 262   jit_State *J = nc->J;
 263   IRIns *ir = IR(ref);
 264   IRRef cref;
 265
 266   /* Check the easy cases first. */
 267   if (ir->o == IR_CONV && (ir->op2 & IRCONV_SRCMASK) == IRT_INT) {
 268     if ((nc->mode & IRCONV_CONVMASK) <= IRCONV_ANY)
 269       narrow_stripov_backprop(nc, ir->op1, depth+1);
 270     else
 271       *nc->sp++ = NARROWINS(NARROW_REF, ir->op1);  /* Undo conversion. */
 272     if (nc->t == IRT_I64)
 273       *nc->sp++ = NARROWINS(NARROW_SEXT, 0);  /* Sign-extend integer. */
 274     return 0;
 275   } else if (ir->o == IR_KNUM) {  /* Narrow FP constant. */
 276     lua_Number n = ir_knum(ir)->n;
 277     if ((nc->mode & IRCONV_CONVMASK) == IRCONV_TOBIT) {
 278       /* Allows a wider range of constants. */
 279       int64_t k64 = (int64_t)n;
 280       if (n == (lua_Number)k64) {  /* Only if const doesn't lose precision. */
 281         *nc->sp++ = NARROWINS(NARROW_INT, 0);
 282         *nc->sp++ = (NarrowIns)k64;  /* But always truncate to 32 bits. */
 283         return 0;
 284       }
 285     } else {
 286       int32_t k = lj_num2int(n);
 287       /* Only if constant is a small integer. */
 288       if (checki16(k) && n == (lua_Number)k) {
 289         *nc->sp++ = NARROWINS(NARROW_INT, 0);
 290         *nc->sp++ = (NarrowIns)k;
 291         return 0;
 292       }
 293     }
 294     return 10;  /* Never narrow other FP constants (this is rare). */
 295   }
 296
 297   /* Try to CSE the conversion. Stronger checks are ok, too. */
 298   cref = J->chain[fins->o];
 299   while (cref > ref) {
 300     IRIns *cr = IR(cref);
 301     if (cr->op1 == ref &&
 302         (fins->o == IR_TOBIT ||
 303          ((cr->op2 & IRCONV_MODEMASK) == (nc->mode & IRCONV_MODEMASK) &&
 304           irt_isguard(cr->t) >= irt_isguard(fins->t)))) {
 305       *nc->sp++ = NARROWINS(NARROW_REF, cref);
 306       return 0;  /* Already there, no additional conversion needed. */
 307     }
 308     cref = cr->prev;
 309   }
 310
 311   /* Backpropagate across ADD/SUB. */
 312   if (ir->o == IR_ADD || ir->o == IR_SUB) {
 313     /* Try cache lookup first. */
 314     IRRef mode = nc->mode;
 315     BPropEntry *bp;
 316     /* Inner conversions need a stronger check. */
 317     if ((mode & IRCONV_CONVMASK) == IRCONV_INDEX && depth > 0)
 318       mode += IRCONV_CHECK-IRCONV_INDEX;
 319     bp = narrow_bpc_get(nc->J, (IRRef1)ref, mode);
 320     if (bp) {
 321       *nc->sp++ = NARROWINS(NARROW_REF, bp->val);
 322       return 0;
 323     } else if (nc->t == IRT_I64) {
 324       /* Try sign-extending from an existing (checked) conversion to int. */
 325       mode = (IRT_INT<<5)|IRT_NUM|IRCONV_INDEX;
 326       bp = narrow_bpc_get(nc->J, (IRRef1)ref, mode);
 327       if (bp) {
 328         *nc->sp++ = NARROWINS(NARROW_REF, bp->val);
 329         *nc->sp++ = NARROWINS(NARROW_SEXT, 0);
 330         return 0;
 331       }
 332     }
 333     if (++depth < NARROW_MAX_BACKPROP && nc->sp < nc->maxsp) {
 334       NarrowIns *savesp = nc->sp;
 335       int count = narrow_conv_backprop(nc, ir->op1, depth);
 336       count += narrow_conv_backprop(nc, ir->op2, depth);
 337       if (count <= nc->lim) {  /* Limit total number of conversions. */
 338         *nc->sp++ = NARROWINS(IRT(ir->o, nc->t), ref);
 339         return count;
 340       }
 341       nc->sp = savesp;  /* Too many conversions, need to backtrack. */
 342     }
 343   }
 344
 345   /* Otherwise add a conversion. */
 346   *nc->sp++ = NARROWINS(NARROW_CONV, ref);
 347   return 1;
 348 }
 349
 350 /* Emit the conversions collected during backpropagation. */
 351 static IRRef narrow_conv_emit(jit_State *J, NarrowConv *nc)
 352 {
 353   /* The fins fields must be saved now -- emitir() overwrites them. */
 354   IROpT guardot = irt_isguard(fins->t) ? IRTG(IR_ADDOV-IR_ADD, 0) : 0;
 355   IROpT convot = fins->ot;
 356   IRRef1 convop2 = fins->op2;
 357   NarrowIns *next = nc->stack;  /* List of instructions from backpropagation. */
 358   NarrowIns *last = nc->sp;
 359   NarrowIns *sp = nc->stack;  /* Recycle the stack to store operands. */
 360   while (next < last) {  /* Simple stack machine to process the ins. list. */
 361     NarrowIns ref = *next++;
 362     IROpT op = narrow_op(ref);
 363     if (op == NARROW_REF) {
 364       *sp++ = ref;
 365     } else if (op == NARROW_CONV) {
 366       *sp++ = emitir_raw(convot, ref, convop2);  /* Raw emit avoids a loop. */
 367     } else if (op == NARROW_SEXT) {
 368       lua_assert(sp >= nc->stack+1);
 369       sp[-1] = emitir(IRT(IR_CONV, IRT_I64), sp[-1],
 370                       (IRT_I64<<5)|IRT_INT|IRCONV_SEXT);
 371     } else if (op == NARROW_INT) {
 372       lua_assert(next < last);
 373       *sp++ = nc->t == IRT_I64 ?
 374               lj_ir_kint64(J, (int64_t)(int32_t)*next++) :
 375               lj_ir_kint(J, *next++);
 376     } else {  /* Regular IROpT. Pops two operands and pushes one result. */
 377       IRRef mode = nc->mode;
 378       lua_assert(sp >= nc->stack+2);
 379       sp--;
 380       /* Omit some overflow checks for array indexing. See comments above. */
 381       if ((mode & IRCONV_CONVMASK) == IRCONV_INDEX) {
 382         if (next == last && irref_isk(narrow_ref(sp[0])) &&
 383           (uint32_t)IR(narrow_ref(sp[0]))->i + 0x40000000u < 0x80000000u)
 384           guardot = 0;
 385         else  /* Otherwise cache a stronger check. */
 386           mode += IRCONV_CHECK-IRCONV_INDEX;
 387       }
 388       sp[-1] = emitir(op+guardot, sp[-1], sp[0]);
 389       /* Add to cache. */
 390       if (narrow_ref(ref))
 391         narrow_bpc_set(J, narrow_ref(ref), narrow_ref(sp[-1]), mode);
 392     }
 393   }
 394   lua_assert(sp == nc->stack+1);
 395   return nc->stack[0];
 396 }
 397
 398 /* Narrow a type conversion of an arithmetic operation. */
 399 TRef LJ_FASTCALL lj_opt_narrow_convert(jit_State *J)
 400 {
 401   if ((J->flags & JIT_F_OPT_NARROW)) {
 402     NarrowConv nc;
 403     nc.J = J;
 404     nc.sp = nc.stack;
 405     nc.maxsp = &nc.stack[NARROW_MAX_STACK-4];
 406     nc.t = irt_type(fins->t);
 407     if (fins->o == IR_TOBIT) {
 408       nc.mode = IRCONV_TOBIT;  /* Used only in the backpropagation cache. */
 409       nc.lim = 2;  /* TOBIT can use a more optimistic rule. */
 410     } else {
 411       nc.mode = fins->op2;
 412       nc.lim = 1;
 413     }
 414     if (narrow_conv_backprop(&nc, fins->op1, 0) <= nc.lim)
 415       return narrow_conv_emit(J, &nc);
 416   }
 417   return NEXTFOLD;
 418 }
 419
 420 /* -- Narrowing of implicit conversions ----------------------------------- */
 421
 422 /* Recursively strip overflow checks. */
 423 static TRef narrow_stripov(jit_State *J, TRef tr, int lastop, IRRef mode)
 424 {
 425   IRRef ref = tref_ref(tr);
 426   IRIns *ir = IR(ref);
 427   int op = ir->o;
 428   if (op >= IR_ADDOV && op <= lastop) {
 429     BPropEntry *bp = narrow_bpc_get(J, ref, mode);
 430     if (bp) {
 431       return TREF(bp->val, irt_t(IR(bp->val)->t));
 432     } else {
 433       IRRef op1 = ir->op1, op2 = ir->op2;  /* The IR may be reallocated. */
 434       op1 = narrow_stripov(J, op1, lastop, mode);
 435       op2 = narrow_stripov(J, op2, lastop, mode);
 436       tr = emitir(IRT(op - IR_ADDOV + IR_ADD,
 437                       ((mode & IRCONV_DSTMASK) >> IRCONV_DSH)), op1, op2);
 438       narrow_bpc_set(J, ref, tref_ref(tr), mode);
 439     }
 440   } else if (LJ_64 && (mode & IRCONV_SEXT) && !irt_is64(ir->t)) {
 441     tr = emitir(IRT(IR_CONV, IRT_INTP), tr, mode);
 442   }
 443   return tr;
 444 }
 445
 446 /* Narrow array index. */
 447 TRef LJ_FASTCALL lj_opt_narrow_index(jit_State *J, TRef tr)
 448 {
 449   IRIns *ir;
 450   lua_assert(tref_isnumber(tr));
 451   if (tref_isnum(tr))  /* Conversion may be narrowed, too. See above. */
 452     return emitir(IRTGI(IR_CONV), tr, IRCONV_INT_NUM|IRCONV_INDEX);
 453   /* Omit some overflow checks for array indexing. See comments above. */
 454   ir = IR(tref_ref(tr));
 455   if ((ir->o == IR_ADDOV || ir->o == IR_SUBOV) && irref_isk(ir->op2) &&
 456       (uint32_t)IR(ir->op2)->i + 0x40000000u < 0x80000000u)
 457     return emitir(IRTI(ir->o - IR_ADDOV + IR_ADD), ir->op1, ir->op2);
 458   return tr;
 459 }
 460
 461 /* Narrow conversion to integer operand (overflow undefined). */
 462 TRef LJ_FASTCALL lj_opt_narrow_toint(jit_State *J, TRef tr)
 463 {
 464   if (tref_isstr(tr))
 465     tr = emitir(IRTG(IR_STRTO, IRT_NUM), tr, 0);
 466   if (tref_isnum(tr))  /* Conversion may be narrowed, too. See above. */
 467     return emitir(IRTI(IR_CONV), tr, IRCONV_INT_NUM|IRCONV_ANY);
 468   if (!tref_isinteger(tr))
 469     lj_trace_err(J, LJ_TRERR_BADTYPE);
 470   /*
 471   ** Undefined overflow semantics allow stripping of ADDOV, SUBOV and MULOV.
 472   ** Use IRCONV_TOBIT for the cache entries, since the semantics are the same.
 473   */
 474   return narrow_stripov(J, tr, IR_MULOV, (IRT_INT<<5)|IRT_INT|IRCONV_TOBIT);
 475 }
 476
 477 /* Narrow conversion to bitop operand (overflow wrapped). */
 478 TRef LJ_FASTCALL lj_opt_narrow_tobit(jit_State *J, TRef tr)
 479 {
 480   if (tref_isstr(tr))
 481     tr = emitir(IRTG(IR_STRTO, IRT_NUM), tr, 0);
 482   if (tref_isnum(tr))  /* Conversion may be narrowed, too. See above. */
 483     return emitir(IRTI(IR_TOBIT), tr, lj_ir_knum_tobit(J));
 484   if (!tref_isinteger(tr))
 485     lj_trace_err(J, LJ_TRERR_BADTYPE);
 486   /*
 487   ** Wrapped overflow semantics allow stripping of ADDOV and SUBOV.
 488   ** MULOV cannot be stripped due to precision widening.
 489   */
 490   return narrow_stripov(J, tr, IR_SUBOV, (IRT_INT<<5)|IRT_INT|IRCONV_TOBIT);
 491 }
 492
 493 #if LJ_HASFFI
 494 /* Narrow C array index (overflow undefined). */
 495 TRef LJ_FASTCALL lj_opt_narrow_cindex(jit_State *J, TRef tr)
 496 {
 497   lua_assert(tref_isnumber(tr));
 498   if (tref_isnum(tr))
 499     return emitir(IRT(IR_CONV, IRT_INTP), tr, (IRT_INTP<<5)|IRT_NUM|IRCONV_ANY);
 500   /* Undefined overflow semantics allow stripping of ADDOV, SUBOV and MULOV. */
 501   return narrow_stripov(J, tr, IR_MULOV,
 502                         LJ_64 ? ((IRT_INTP<<5)|IRT_INT|IRCONV_SEXT) :
 503                                 ((IRT_INTP<<5)|IRT_INT|IRCONV_TOBIT));
 504 }
 505 #endif
 506
 507 /* -- Narrowing of arithmetic operators ----------------------------------- */
 508
 509 /* Check whether a number fits into an int32_t (-0 is ok, too). */
 510 static int numisint(lua_Number n)
 511 {
 512   return (n == (lua_Number)lj_num2int(n));
 513 }
 514
 515 /* Narrowing of arithmetic operations. */
 516 TRef lj_opt_narrow_arith(jit_State *J, TRef rb, TRef rc,
 517                          TValue *vb, TValue *vc, IROp op)
 518 {
 519   if (tref_isstr(rb)) {
 520     rb = emitir(IRTG(IR_STRTO, IRT_NUM), rb, 0);
 521     lj_strscan_num(strV(vb), vb);
 522   }
 523   if (tref_isstr(rc)) {
 524     rc = emitir(IRTG(IR_STRTO, IRT_NUM), rc, 0);
 525     lj_strscan_num(strV(vc), vc);
 526   }
 527   /* Must not narrow MUL in non-DUALNUM variant, because it loses -0. */
 528   if ((op >= IR_ADD && op <= (LJ_DUALNUM ? IR_MUL : IR_SUB)) &&
 529       tref_isinteger(rb) && tref_isinteger(rc) &&
 530       numisint(lj_vm_foldarith(numberVnum(vb), numberVnum(vc),
 531                                (int)op - (int)IR_ADD)))
 532     return emitir(IRTGI((int)op - (int)IR_ADD + (int)IR_ADDOV), rb, rc);
 533   if (!tref_isnum(rb)) rb = emitir(IRTN(IR_CONV), rb, IRCONV_NUM_INT);
 534   if (!tref_isnum(rc)) rc = emitir(IRTN(IR_CONV), rc, IRCONV_NUM_INT);
 535   return emitir(IRTN(op), rb, rc);
 536 }
 537
 538 /* Narrowing of unary minus operator. */
 539 TRef lj_opt_narrow_unm(jit_State *J, TRef rc, TValue *vc)
 540 {
 541   if (tref_isstr(rc)) {
 542     rc = emitir(IRTG(IR_STRTO, IRT_NUM), rc, 0);
 543     lj_strscan_num(strV(vc), vc);
 544   }
 545   if (tref_isinteger(rc)) {
 546     if ((uint32_t)numberVint(vc) != 0x80000000u)
 547       return emitir(IRTGI(IR_SUBOV), lj_ir_kint(J, 0), rc);
 548     rc = emitir(IRTN(IR_CONV), rc, IRCONV_NUM_INT);
 549   }
 550   return emitir(IRTN(IR_NEG), rc, lj_ir_knum_neg(J));
 551 }
 552
 553 /* Narrowing of modulo operator. */
 554 TRef lj_opt_narrow_mod(jit_State *J, TRef rb, TRef rc, TValue *vc)
 555 {
 556   TRef tmp;
 557   if (tvisstr(vc) && !lj_strscan_num(strV(vc), vc))
 558     lj_trace_err(J, LJ_TRERR_BADTYPE);
 559   if ((LJ_DUALNUM || (J->flags & JIT_F_OPT_NARROW)) &&
 560       tref_isinteger(rb) && tref_isinteger(rc) &&
 561       (tvisint(vc) ? intV(vc) != 0 : !tviszero(vc))) {
 562     emitir(IRTGI(IR_NE), rc, lj_ir_kint(J, 0));
 563     return emitir(IRTI(IR_MOD), rb, rc);
 564   }
 565   /* b % c ==> b - floor(b/c)*c */
 566   rb = lj_ir_tonum(J, rb);
 567   rc = lj_ir_tonum(J, rc);
 568   tmp = emitir(IRTN(IR_DIV), rb, rc);
 569   tmp = emitir(IRTN(IR_FPMATH), tmp, IRFPM_FLOOR);
 570   tmp = emitir(IRTN(IR_MUL), tmp, rc);
 571   return emitir(IRTN(IR_SUB), rb, tmp);
 572 }
 573
 574 /* Narrowing of power operator or math.pow. */
 575 TRef lj_opt_narrow_pow(jit_State *J, TRef rb, TRef rc, TValue *vc)
 576 {
 577   if (tvisstr(vc) && !lj_strscan_num(strV(vc), vc))
 578     lj_trace_err(J, LJ_TRERR_BADTYPE);
 579   /* Narrowing must be unconditional to preserve (-x)^i semantics. */
 580   if (tvisint(vc) || numisint(numV(vc))) {
 581     int checkrange = 0;
 582     /* Split pow is faster for bigger exponents. But do this only for (+k)^i. */
 583     if (tref_isk(rb) && (int32_t)ir_knum(IR(tref_ref(rb)))->u32.hi >= 0) {
 584       int32_t k = numberVint(vc);
 585       if (!(k >= -65536 && k <= 65536)) goto split_pow;
 586       checkrange = 1;
 587     }
 588     if (!tref_isinteger(rc)) {
 589       if (tref_isstr(rc))
 590         rc = emitir(IRTG(IR_STRTO, IRT_NUM), rc, 0);
 591       /* Guarded conversion to integer! */
 592       rc = emitir(IRTGI(IR_CONV), rc, IRCONV_INT_NUM|IRCONV_CHECK);
 593     }
 594     if (checkrange && !tref_isk(rc)) {  /* Range guard: -65536 <= i <= 65536 */
 595       TRef tmp = emitir(IRTI(IR_ADD), rc, lj_ir_kint(J, 65536));
 596       emitir(IRTGI(IR_ULE), tmp, lj_ir_kint(J, 2*65536));
 597     }
 598     return emitir(IRTN(IR_POW), rb, rc);
 599   }
 600 split_pow:
 601   /* FOLD covers most cases, but some are easier to do here. */
 602   if (tref_isk(rb) && tvispone(ir_knum(IR(tref_ref(rb)))))
 603     return rb;  /* 1 ^ x ==> 1 */
 604   rc = lj_ir_tonum(J, rc);
 605   if (tref_isk(rc) && ir_knum(IR(tref_ref(rc)))->n == 0.5)
 606     return emitir(IRTN(IR_FPMATH), rb, IRFPM_SQRT);  /* x ^ 0.5 ==> sqrt(x) */
 607   /* Split up b^c into exp2(c*log2(b)). Assembler may rejoin later. */
 608   rb = emitir(IRTN(IR_FPMATH), rb, IRFPM_LOG2);
 609   rc = emitir(IRTN(IR_MUL), rb, rc);
 610   return emitir(IRTN(IR_FPMATH), rc, IRFPM_EXP2);
 611 }
 612
 613 /* -- Predictive narrowing of induction variables ------------------------- */
 614
 615 /* Narrow a single runtime value. */
 616 static int narrow_forl(jit_State *J, cTValue *o)
 617 {
 618   if (tvisint(o)) return 1;
 619   if (LJ_DUALNUM || (J->flags & JIT_F_OPT_NARROW)) return numisint(numV(o));
 620   return 0;
 621 }
 622
 623 /* Narrow the FORL index type by looking at the runtime values. */
 624 IRType lj_opt_narrow_forl(jit_State *J, cTValue *tv)
 625 {
 626   lua_assert(tvisnumber(&tv[FORL_IDX]) &&
 627              tvisnumber(&tv[FORL_STOP]) &&
 628              tvisnumber(&tv[FORL_STEP]));
 629   /* Narrow only if the runtime values of start/stop/step are all integers. */
 630   if (narrow_forl(J, &tv[FORL_IDX]) &&
 631       narrow_forl(J, &tv[FORL_STOP]) &&
 632       narrow_forl(J, &tv[FORL_STEP])) {
 633     /* And if the loop index can't possibly overflow. */
 634     lua_Number step = numberVnum(&tv[FORL_STEP]);
 635     lua_Number sum = numberVnum(&tv[FORL_STOP]) + step;
 636     if (0 <= step ? (sum <= 2147483647.0) : (sum >= -2147483648.0))
 637       return IRT_INT;
 638   }
 639   return IRT_NUM;
 640 }
 641
 642 #undef IR
 643 #undef fins
 644 #undef emitir
 645 #undef emitir_raw
 646
 647 #endif