src/lj_jit.h

   1 /*
   2 ** Common definitions for the JIT compiler.
   3 ** Copyright (C) 2005-2011 Mike Pall. See Copyright Notice in luajit.h
   4 */
   5
   6 #ifndef _LJ_JIT_H
   7 #define _LJ_JIT_H
   8
   9 #include "lj_obj.h"
  10 #include "lj_ir.h"
  11
  12 /* JIT engine flags. */
  13 #define JIT_F_ON                0x00000001
  14
  15 /* CPU-specific JIT engine flags. */
  16 #if LJ_TARGET_X86ORX64
  17 #define JIT_F_CMOV              0x00000010
  18 #define JIT_F_SSE2              0x00000020
  19 #define JIT_F_SSE3              0x00000040
  20 #define JIT_F_SSE4_1            0x00000080
  21 #define JIT_F_P4                0x00000100
  22 #define JIT_F_PREFER_IMUL       0x00000200
  23 #define JIT_F_SPLIT_XMM         0x00000400
  24 #define JIT_F_LEA_AGU           0x00000800
  25
  26 /* Names for the CPU-specific flags. Must match the order above. */
  27 #define JIT_F_CPU_FIRST         JIT_F_CMOV
  28 #define JIT_F_CPUSTRING         "\4CMOV\4SSE2\4SSE3\6SSE4.1\2P4\3AMD\2K8\4ATOM"
  29 #elif LJ_TARGET_ARM
  30 #define JIT_F_ARMV6             0x00000010
  31 #define JIT_F_ARMV6T2           0x00000020
  32 #define JIT_F_ARMV7             0x00000040
  33
  34 /* Names for the CPU-specific flags. Must match the order above. */
  35 #define JIT_F_CPU_FIRST         JIT_F_ARMV6
  36 #define JIT_F_CPUSTRING         "\5ARMv6\7ARMv6T2\5ARMv7"
  37 #else
  38 #define JIT_F_CPU_FIRST         0
  39 #define JIT_F_CPUSTRING         ""
  40 #endif
  41
  42 /* Optimization flags. */
  43 #define JIT_F_OPT_MASK          0x0fff0000
  44
  45 #define JIT_F_OPT_FOLD          0x00010000
  46 #define JIT_F_OPT_CSE           0x00020000
  47 #define JIT_F_OPT_DCE           0x00040000
  48 #define JIT_F_OPT_FWD           0x00080000
  49 #define JIT_F_OPT_DSE           0x00100000
  50 #define JIT_F_OPT_NARROW        0x00200000
  51 #define JIT_F_OPT_LOOP          0x00400000
  52 #define JIT_F_OPT_ABC           0x00800000
  53 #define JIT_F_OPT_FUSE          0x01000000
  54
  55 /* Optimizations names for -O. Must match the order above. */
  56 #define JIT_F_OPT_FIRST         JIT_F_OPT_FOLD
  57 #define JIT_F_OPTSTRING \
  58   "\4fold\3cse\3dce\3fwd\3dse\6narrow\4loop\3abc\4fuse"
  59
  60 /* Optimization levels set a fixed combination of flags. */
  61 #define JIT_F_OPT_0     0
  62 #define JIT_F_OPT_1     (JIT_F_OPT_FOLD|JIT_F_OPT_CSE|JIT_F_OPT_DCE)
  63 #define JIT_F_OPT_2     (JIT_F_OPT_1|JIT_F_OPT_NARROW|JIT_F_OPT_LOOP)
  64 #define JIT_F_OPT_3 \
  65   (JIT_F_OPT_2|JIT_F_OPT_FWD|JIT_F_OPT_DSE|JIT_F_OPT_ABC|JIT_F_OPT_FUSE)
  66 #define JIT_F_OPT_DEFAULT       JIT_F_OPT_3
  67
  68 #if LJ_TARGET_WINDOWS || LJ_64
  69 /* See: http://blogs.msdn.com/oldnewthing/archive/2003/10/08/55239.aspx */
  70 #define JIT_P_sizemcode_DEFAULT         64
  71 #else
  72 /* Could go as low as 4K, but the mmap() overhead would be rather high. */
  73 #define JIT_P_sizemcode_DEFAULT         32
  74 #endif
  75
  76 /* Optimization parameters and their defaults. Length is a char in octal! */
  77 #define JIT_PARAMDEF(_) \
  78   _(\010, maxtrace,     1000)   /* Max. # of traces in cache. */ \
  79   _(\011, maxrecord,    4000)   /* Max. # of recorded IR instructions. */ \
  80   _(\012, maxirconst,   500)    /* Max. # of IR constants of a trace. */ \
  81   _(\007, maxside,      100)    /* Max. # of side traces of a root trace. */ \
  82   _(\007, maxsnap,      500)    /* Max. # of snapshots for a trace. */ \
  83   \
  84   _(\007, hotloop,      56)     /* # of iter. to detect a hot loop/call. */ \
  85   _(\007, hotexit,      10)     /* # of taken exits to start a side trace. */ \
  86   _(\007, tryside,      4)      /* # of attempts to compile a side trace. */ \
  87   \
  88   _(\012, instunroll,   4)      /* Max. unroll for instable loops. */ \
  89   _(\012, loopunroll,   15)     /* Max. unroll for loop ops in side traces. */ \
  90   _(\012, callunroll,   3)      /* Max. unroll for recursive calls. */ \
  91   _(\011, recunroll,    2)      /* Min. unroll for true recursion. */ \
  92   \
  93   /* Size of each machine code area (in KBytes). */ \
  94   _(\011, sizemcode,    JIT_P_sizemcode_DEFAULT) \
  95   /* Max. total size of all machine code areas (in KBytes). */ \
  96   _(\010, maxmcode,     512) \
  97   /* End of list. */
  98
  99 enum {
 100 #define JIT_PARAMENUM(len, name, value) JIT_P_##name,
 101 JIT_PARAMDEF(JIT_PARAMENUM)
 102 #undef JIT_PARAMENUM
 103   JIT_P__MAX
 104 };
 105
 106 #define JIT_PARAMSTR(len, name, value)  #len #name
 107 #define JIT_P_STRING    JIT_PARAMDEF(JIT_PARAMSTR)
 108
 109 /* Trace compiler state. */
 110 typedef enum {
 111   LJ_TRACE_IDLE,        /* Trace compiler idle. */
 112   LJ_TRACE_ACTIVE = 0x10,
 113   LJ_TRACE_RECORD,      /* Bytecode recording active. */
 114   LJ_TRACE_START,       /* New trace started. */
 115   LJ_TRACE_END,         /* End of trace. */
 116   LJ_TRACE_ASM,         /* Assemble trace. */
 117   LJ_TRACE_ERR          /* Trace aborted with error. */
 118 } TraceState;
 119
 120 /* Post-processing action. */
 121 typedef enum {
 122   LJ_POST_NONE,         /* No action. */
 123   LJ_POST_FIXCOMP,      /* Fixup comparison and emit pending guard. */
 124   LJ_POST_FIXGUARD,     /* Fixup and emit pending guard. */
 125   LJ_POST_FIXBOOL,      /* Fixup boolean result. */
 126   LJ_POST_FFRETRY       /* Suppress recording of retried fast functions. */
 127 } PostProc;
 128
 129 /* Machine code type. */
 130 #if LJ_TARGET_X86ORX64
 131 typedef uint8_t MCode;
 132 #else
 133 typedef uint32_t MCode;
 134 #endif
 135
 136 /* Stack snapshot header. */
 137 typedef struct SnapShot {
 138   uint16_t mapofs;      /* Offset into snapshot map. */
 139   IRRef1 ref;           /* First IR ref for this snapshot. */
 140   uint8_t nslots;       /* Number of valid slots. */
 141   uint8_t nent;         /* Number of compressed entries. */
 142   uint8_t depth;        /* Number of frame links. */
 143   uint8_t count;        /* Count of taken exits for this snapshot. */
 144 } SnapShot;
 145
 146 #define SNAPCOUNT_DONE  255     /* Already compiled and linked a side trace. */
 147
 148 /* Compressed snapshot entry. */
 149 typedef uint32_t SnapEntry;
 150
 151 #define SNAP_FRAME              0x010000        /* Frame slot. */
 152 #define SNAP_CONT               0x020000        /* Continuation slot. */
 153 #define SNAP_NORESTORE          0x040000        /* No need to restore slot. */
 154 #define SNAP_SOFTFPNUM          0x080000        /* Soft-float number. */
 155 LJ_STATIC_ASSERT(SNAP_FRAME == TREF_FRAME);
 156 LJ_STATIC_ASSERT(SNAP_CONT == TREF_CONT);
 157
 158 #define SNAP(slot, flags, ref)  (((SnapEntry)(slot) << 24) + (flags) + (ref))
 159 #define SNAP_TR(slot, tr) \
 160   (((SnapEntry)(slot) << 24) + ((tr) & (TREF_CONT|TREF_FRAME|TREF_REFMASK)))
 161 #define SNAP_MKPC(pc)           ((SnapEntry)u32ptr(pc))
 162 #define SNAP_MKFTSZ(ftsz)       ((SnapEntry)(ftsz))
 163 #define snap_ref(sn)            ((sn) & 0xffff)
 164 #define snap_slot(sn)           ((BCReg)((sn) >> 24))
 165 #define snap_isframe(sn)        ((sn) & SNAP_FRAME)
 166 #define snap_pc(sn)             ((const BCIns *)(uintptr_t)(sn))
 167 #define snap_setref(sn, ref)    (((sn) & (0xffff0000&~SNAP_NORESTORE)) | (ref))
 168
 169 /* Snapshot and exit numbers. */
 170 typedef uint32_t SnapNo;
 171 typedef uint32_t ExitNo;
 172
 173 /* Trace number. */
 174 typedef uint32_t TraceNo;       /* Used to pass around trace numbers. */
 175 typedef uint16_t TraceNo1;      /* Stored trace number. */
 176
 177 /* Type of link. ORDER LJ_TRLINK */
 178 typedef enum {
 179   LJ_TRLINK_NONE,               /* Incomplete trace. No link, yet. */
 180   LJ_TRLINK_ROOT,               /* Link to other root trace. */
 181   LJ_TRLINK_LOOP,               /* Loop to same trace. */
 182   LJ_TRLINK_TAILREC,            /* Tail-recursion. */
 183   LJ_TRLINK_UPREC,              /* Up-recursion. */
 184   LJ_TRLINK_DOWNREC,            /* Down-recursion. */
 185   LJ_TRLINK_INTERP,             /* Fallback to interpreter. */
 186   LJ_TRLINK_RETURN              /* Return to interpreter. */
 187 } TraceLink;
 188
 189 /* Trace object. */
 190 typedef struct GCtrace {
 191   GCHeader;
 192   uint8_t topslot;      /* Top stack slot already checked to be allocated. */
 193   uint8_t linktype;     /* Type of link. */
 194   IRRef nins;           /* Next IR instruction. Biased with REF_BIAS. */
 195   GCRef gclist;
 196   IRIns *ir;            /* IR instructions/constants. Biased with REF_BIAS. */
 197   IRRef nk;             /* Lowest IR constant. Biased with REF_BIAS. */
 198   uint16_t nsnap;       /* Number of snapshots. */
 199   uint16_t nsnapmap;    /* Number of snapshot map elements. */
 200   SnapShot *snap;       /* Snapshot array. */
 201   SnapEntry *snapmap;   /* Snapshot map. */
 202   GCRef startpt;        /* Starting prototype. */
 203   MRef startpc;         /* Bytecode PC of starting instruction. */
 204   BCIns startins;       /* Original bytecode of starting instruction. */
 205   MSize szmcode;        /* Size of machine code. */
 206   MCode *mcode;         /* Start of machine code. */
 207   MSize mcloop;         /* Offset of loop start in machine code. */
 208   uint16_t nchild;      /* Number of child traces (root trace only). */
 209   uint16_t spadjust;    /* Stack pointer adjustment (offset in bytes). */
 210   TraceNo1 traceno;     /* Trace number. */
 211   TraceNo1 link;        /* Linked trace (or self for loops). */
 212   TraceNo1 root;        /* Root trace of side trace (or 0 for root traces). */
 213   TraceNo1 nextroot;    /* Next root trace for same prototype. */
 214   TraceNo1 nextside;    /* Next side trace of same root trace. */
 215   uint16_t unused2;
 216 #ifdef LUAJIT_USE_GDBJIT
 217   void *gdbjit_entry;   /* GDB JIT entry. */
 218 #endif
 219 } GCtrace;
 220
 221 #define gco2trace(o)    check_exp((o)->gch.gct == ~LJ_TTRACE, (GCtrace *)(o))
 222 #define traceref(J, n) \
 223   check_exp((n)>0 && (MSize)(n)<J->sizetrace, (GCtrace *)gcref(J->trace[(n)]))
 224
 225 LJ_STATIC_ASSERT(offsetof(GChead, gclist) == offsetof(GCtrace, gclist));
 226
 227 /* Round-robin penalty cache for bytecodes leading to aborted traces. */
 228 typedef struct HotPenalty {
 229   MRef pc;              /* Starting bytecode PC. */
 230   uint16_t val;         /* Penalty value, i.e. hotcount start. */
 231   uint16_t reason;      /* Abort reason (really TraceErr). */
 232 } HotPenalty;
 233
 234 #define PENALTY_SLOTS   64      /* Penalty cache slot. Must be a power of 2. */
 235 #define PENALTY_MIN     36      /* Minimum penalty value. */
 236 #define PENALTY_MAX     60000   /* Maximum penalty value. */
 237 #define PENALTY_RNDBITS 4       /* # of random bits to add to penalty value. */
 238
 239 /* Round-robin backpropagation cache for narrowing conversions. */
 240 typedef struct BPropEntry {
 241   IRRef1 key;           /* Key: original reference. */
 242   IRRef1 val;           /* Value: reference after conversion. */
 243   IRRef mode;           /* Mode for this entry (currently IRCONV_*). */
 244 } BPropEntry;
 245
 246 /* Number of slots for the backpropagation cache. Must be a power of 2. */
 247 #define BPROP_SLOTS     16
 248
 249 /* Scalar evolution analysis cache. */
 250 typedef struct ScEvEntry {
 251   IRRef1 idx;           /* Index reference. */
 252   IRRef1 start;         /* Constant start reference. */
 253   IRRef1 stop;          /* Constant stop reference. */
 254   IRRef1 step;          /* Constant step reference. */
 255   IRType1 t;            /* Scalar type. */
 256   uint8_t dir;          /* Direction. 1: +, 0: -. */
 257 } ScEvEntry;
 258
 259 /* 128 bit SIMD constants. */
 260 enum {
 261   LJ_KSIMD_ABS,
 262   LJ_KSIMD_NEG,
 263   LJ_KSIMD__MAX
 264 };
 265
 266 /* Get 16 byte aligned pointer to SIMD constant. */
 267 #define LJ_KSIMD(J, n) \
 268   ((TValue *)(((intptr_t)&J->ksimd[2*(n)] + 15) & ~(intptr_t)15))
 269
 270 /* Set/reset flag to activate the SPLIT pass for the current trace. */
 271 #if LJ_SOFTFP || (LJ_32 && LJ_HASFFI)
 272 #define lj_needsplit(J)         (J->needsplit = 1)
 273 #define lj_resetsplit(J)        (J->needsplit = 0)
 274 #else
 275 #define lj_needsplit(J)         UNUSED(J)
 276 #define lj_resetsplit(J)        UNUSED(J)
 277 #endif
 278
 279 /* Fold state is used to fold instructions on-the-fly. */
 280 typedef struct FoldState {
 281   IRIns ins;            /* Currently emitted instruction. */
 282   IRIns left;           /* Instruction referenced by left operand. */
 283   IRIns right;          /* Instruction referenced by right operand. */
 284 } FoldState;
 285
 286 /* JIT compiler state. */
 287 typedef struct jit_State {
 288   GCtrace cur;          /* Current trace. */
 289
 290   lua_State *L;         /* Current Lua state. */
 291   const BCIns *pc;      /* Current PC. */
 292   GCfunc *fn;           /* Current function. */
 293   GCproto *pt;          /* Current prototype. */
 294   TRef *base;           /* Current frame base, points into J->slots. */
 295
 296   uint32_t flags;       /* JIT engine flags. */
 297   BCReg maxslot;        /* Relative to baseslot. */
 298   BCReg baseslot;       /* Current frame base, offset into J->slots. */
 299
 300   uint8_t mergesnap;    /* Allowed to merge with next snapshot. */
 301   uint8_t needsnap;     /* Need snapshot before recording next bytecode. */
 302   IRType1 guardemit;    /* Accumulated IRT_GUARD for emitted instructions. */
 303   uint8_t bcskip;       /* Number of bytecode instructions to skip. */
 304
 305   FoldState fold;       /* Fold state. */
 306
 307   const BCIns *bc_min;  /* Start of allowed bytecode range for root trace. */
 308   MSize bc_extent;      /* Extent of the range. */
 309
 310   TraceState state;     /* Trace compiler state. */
 311
 312   int32_t instunroll;   /* Unroll counter for instable loops. */
 313   int32_t loopunroll;   /* Unroll counter for loop ops in side traces. */
 314   int32_t tailcalled;   /* Number of successive tailcalls. */
 315   int32_t framedepth;   /* Current frame depth. */
 316   int32_t retdepth;     /* Return frame depth (count of RETF). */
 317
 318   MRef k64;             /* Pointer to chained array of 64 bit constants. */
 319   TValue ksimd[LJ_KSIMD__MAX*2+1];  /* 16 byte aligned SIMD constants. */
 320
 321   IRIns *irbuf;         /* Temp. IR instruction buffer. Biased with REF_BIAS. */
 322   IRRef irtoplim;       /* Upper limit of instuction buffer (biased). */
 323   IRRef irbotlim;       /* Lower limit of instuction buffer (biased). */
 324   IRRef loopref;        /* Last loop reference or ref of final LOOP (or 0). */
 325
 326   MSize sizesnap;       /* Size of temp. snapshot buffer. */
 327   SnapShot *snapbuf;    /* Temp. snapshot buffer. */
 328   SnapEntry *snapmapbuf;  /* Temp. snapshot map buffer. */
 329   MSize sizesnapmap;    /* Size of temp. snapshot map buffer. */
 330
 331   PostProc postproc;    /* Required post-processing after execution. */
 332 #if LJ_SOFTFP || (LJ_32 && LJ_HASFFI)
 333   int needsplit;        /* Need SPLIT pass. */
 334 #endif
 335
 336   GCRef *trace;         /* Array of traces. */
 337   TraceNo freetrace;    /* Start of scan for next free trace. */
 338   MSize sizetrace;      /* Size of trace array. */
 339
 340   IRRef1 chain[IR__MAX];  /* IR instruction skip-list chain anchors. */
 341   TRef slot[LJ_MAX_JSLOTS+LJ_STACK_EXTRA];  /* Stack slot map. */
 342
 343   int32_t param[JIT_P__MAX];  /* JIT engine parameters. */
 344
 345   MCode *exitstubgroup[LJ_MAX_EXITSTUBGR];  /* Exit stub group addresses. */
 346
 347   HotPenalty penalty[PENALTY_SLOTS];  /* Penalty slots. */
 348   uint32_t penaltyslot; /* Round-robin index into penalty slots. */
 349   uint32_t prngstate;   /* PRNG state. */
 350
 351   BPropEntry bpropcache[BPROP_SLOTS];  /* Backpropagation cache slots. */
 352   uint32_t bpropslot;   /* Round-robin index into bpropcache slots. */
 353
 354   ScEvEntry scev;       /* Scalar evolution analysis cache slots. */
 355
 356   const BCIns *startpc; /* Bytecode PC of starting instruction. */
 357   TraceNo parent;       /* Parent of current side trace (0 for root traces). */
 358   ExitNo exitno;        /* Exit number in parent of current side trace. */
 359
 360   BCIns *patchpc;       /* PC for pending re-patch. */
 361   BCIns patchins;       /* Instruction for pending re-patch. */
 362
 363   int mcprot;           /* Protection of current mcode area. */
 364   MCode *mcarea;        /* Base of current mcode area. */
 365   MCode *mctop;         /* Top of current mcode area. */
 366   MCode *mcbot;         /* Bottom of current mcode area. */
 367   size_t szmcarea;      /* Size of current mcode area. */
 368   size_t szallmcarea;   /* Total size of all allocated mcode areas. */
 369
 370   TValue errinfo;       /* Additional info element for trace errors. */
 371 } LJ_ALIGN(16) jit_State;
 372
 373 /* Trivial PRNG e.g. used for penalty randomization. */
 374 static LJ_AINLINE uint32_t LJ_PRNG_BITS(jit_State *J, int bits)
 375 {
 376   /* Yes, this LCG is very weak, but that doesn't matter for our use case. */
 377   J->prngstate = J->prngstate * 1103515245 + 12345;
 378   return J->prngstate >> (32-bits);
 379 }
 380
 381 #endif