src/gallium/drivers/nv50/codegen/nv50_ir.h

   1 /*
   2  * Copyright 2011 Christoph Bumiller
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice shall be included in
  12  * all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
  18  * WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF
  19  * OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  20  * SOFTWARE.
  21  */
  22
  23 #ifndef __NV50_IR_H__
  24 #define __NV50_IR_H__
  25
  26 #include <stdio.h>
  27 #include <stdlib.h>
  28 #include <stdint.h>
  29 #include <deque>
  30 #include <list>
  31 #include <vector>
  32
  33 #include "nv50_ir_util.h"
  34 #include "nv50_ir_graph.h"
  35
  36 #include "nv50_ir_driver.h"
  37
  38 namespace nv50_ir {
  39
  40 enum operation
  41 {
  42    OP_NOP = 0,
  43    OP_PHI,
  44    OP_UNION, // unify a new definition and several source values
  45    OP_SPLIT, // $r0d -> { $r0, $r1 } ($r0d and $r0/$r1 will be coalesced)
  46    OP_MERGE, // opposite of split, e.g. combine 2 32 bit into a 64 bit value
  47    OP_CONSTRAINT, // copy values into consecutive registers
  48    OP_MOV, // simple copy, no modifiers allowed
  49    OP_LOAD,
  50    OP_STORE,
  51    OP_ADD,
  52    OP_SUB,
  53    OP_MUL,
  54    OP_DIV,
  55    OP_MOD,
  56    OP_MAD,
  57    OP_FMA,
  58    OP_SAD, // abs(src0 - src1) + src2
  59    OP_ABS,
  60    OP_NEG,
  61    OP_NOT,
  62    OP_AND,
  63    OP_OR,
  64    OP_XOR,
  65    OP_SHL,
  66    OP_SHR,
  67    OP_MAX,
  68    OP_MIN,
  69    OP_SAT,  // CLAMP(f32, 0.0, 1.0)
  70    OP_CEIL,
  71    OP_FLOOR,
  72    OP_TRUNC,
  73    OP_CVT,
  74    OP_SET_AND, // dst = (src0 CMP src1) & src2
  75    OP_SET_OR,
  76    OP_SET_XOR,
  77    OP_SET,
  78    OP_SELP, // dst = src2 ? src0 : src1
  79    OP_SLCT, // dst = (src2 CMP 0) ? src0 : src1
  80    OP_RCP,
  81    OP_RSQ,
  82    OP_LG2,
  83    OP_SIN,
  84    OP_COS,
  85    OP_EX2,
  86    OP_EXP, // exponential (base M_E)
  87    OP_LOG, // natural logarithm
  88    OP_PRESIN,
  89    OP_PREEX2,
  90    OP_SQRT,
  91    OP_POW,
  92    OP_BRA,
  93    OP_CALL,
  94    OP_RET,
  95    OP_CONT,
  96    OP_BREAK,
  97    OP_PRERET,
  98    OP_PRECONT,
  99    OP_PREBREAK,
 100    OP_BRKPT,     // breakpoint (not related to loops)
 101    OP_JOINAT,    // push control flow convergence point
 102    OP_JOIN,      // converge
 103    OP_DISCARD,
 104    OP_EXIT,
 105    OP_MEMBAR,
 106    OP_VFETCH, // indirection 0 in attribute space, indirection 1 is vertex base
 107    OP_PFETCH, // fetch base address of vertex src0 (immediate) [+ src1]
 108    OP_EXPORT,
 109    OP_LINTERP,
 110    OP_PINTERP,
 111    OP_EMIT,    // emit vertex
 112    OP_RESTART, // restart primitive
 113    OP_TEX,
 114    OP_TXB, // texture bias
 115    OP_TXL, // texure lod
 116    OP_TXF, // texel fetch
 117    OP_TXQ, // texture size query
 118    OP_TXD, // texture derivatives
 119    OP_TXG, // texture gather
 120    OP_TEXCSAA,
 121    OP_SULD, // surface load
 122    OP_SUST, // surface store
 123    OP_DFDX,
 124    OP_DFDY,
 125    OP_RDSV, // read system value
 126    OP_WRSV, // write system value
 127    OP_PIXLD,
 128    OP_QUADOP,
 129    OP_QUADON,
 130    OP_QUADPOP,
 131    OP_POPCNT, // bitcount(src0 & src1)
 132    OP_INSBF,  // insert first src1[8:15] bits of src0 into src2 at src1[0:7]
 133    OP_EXTBF,
 134    OP_LAST
 135 };
 136
 137 #define NV50_IR_SUBOP_MUL_HIGH     1
 138 #define NV50_IR_SUBOP_EMIT_RESTART 1
 139 #define NV50_IR_SUBOP_LDC_IL       1
 140 #define NV50_IR_SUBOP_LDC_IS       2
 141 #define NV50_IR_SUBOP_LDC_ISL      3
 142 #define NV50_IR_SUBOP_SHIFT_WRAP   1
 143
 144 enum DataType
 145 {
 146    TYPE_NONE,
 147    TYPE_U8,
 148    TYPE_S8,
 149    TYPE_U16,
 150    TYPE_S16,
 151    TYPE_U32,
 152    TYPE_S32,
 153    TYPE_U64, // 64 bit operations are only lowered after register allocation
 154    TYPE_S64,
 155    TYPE_F16,
 156    TYPE_F32,
 157    TYPE_F64,
 158    TYPE_B96,
 159    TYPE_B128
 160 };
 161
 162 enum CondCode
 163 {
 164    CC_FL = 0,
 165    CC_NEVER = CC_FL, // when used with FILE_FLAGS
 166    CC_LT = 1,
 167    CC_EQ = 2,
 168    CC_NOT_P = CC_EQ, // when used with FILE_PREDICATE
 169    CC_LE = 3,
 170    CC_GT = 4,
 171    CC_NE = 5,
 172    CC_P  = CC_NE,
 173    CC_GE = 6,
 174    CC_TR = 7,
 175    CC_ALWAYS = CC_TR,
 176    CC_U  = 8,
 177    CC_LTU = 9,
 178    CC_EQU = 10,
 179    CC_LEU = 11,
 180    CC_GTU = 12,
 181    CC_NEU = 13,
 182    CC_GEU = 14,
 183    CC_NO = 0x10,
 184    CC_NC = 0x11,
 185    CC_NS = 0x12,
 186    CC_NA = 0x13,
 187    CC_A  = 0x14,
 188    CC_S  = 0x15,
 189    CC_C  = 0x16,
 190    CC_O  = 0x17
 191 };
 192
 193 enum RoundMode
 194 {
 195    ROUND_N, // nearest
 196    ROUND_M, // towards -inf
 197    ROUND_Z, // towards 0
 198    ROUND_P, // towards +inf
 199    ROUND_NI, // nearest integer
 200    ROUND_MI, // to integer towards -inf
 201    ROUND_ZI, // to integer towards 0
 202    ROUND_PI, // to integer towards +inf
 203 };
 204
 205 enum CacheMode
 206 {
 207    CACHE_CA,            // cache at all levels
 208    CACHE_WB = CACHE_CA, // cache write back
 209    CACHE_CG,            // cache at global level
 210    CACHE_CS,            // cache streaming
 211    CACHE_CV,            // cache as volatile
 212    CACHE_WT = CACHE_CV  // cache write-through
 213 };
 214
 215 enum DataFile
 216 {
 217    FILE_NULL = 0,
 218    FILE_GPR,
 219    FILE_PREDICATE,       // boolean predicate
 220    FILE_FLAGS,           // zero/sign/carry/overflow bits
 221    FILE_ADDRESS,
 222    LAST_REGISTER_FILE = FILE_ADDRESS,
 223    FILE_IMMEDIATE,
 224    FILE_MEMORY_CONST,
 225    FILE_SHADER_INPUT,
 226    FILE_SHADER_OUTPUT,
 227    FILE_MEMORY_GLOBAL,
 228    FILE_MEMORY_SHARED,
 229    FILE_MEMORY_LOCAL,
 230    FILE_SYSTEM_VALUE,
 231    DATA_FILE_COUNT
 232 };
 233
 234 enum TexTarget
 235 {
 236    TEX_TARGET_1D,
 237    TEX_TARGET_2D,
 238    TEX_TARGET_2D_MS,
 239    TEX_TARGET_3D,
 240    TEX_TARGET_CUBE,
 241    TEX_TARGET_1D_SHADOW,
 242    TEX_TARGET_2D_SHADOW,
 243    TEX_TARGET_CUBE_SHADOW,
 244    TEX_TARGET_1D_ARRAY,
 245    TEX_TARGET_2D_ARRAY,
 246    TEX_TARGET_2D_MS_ARRAY,
 247    TEX_TARGET_CUBE_ARRAY,
 248    TEX_TARGET_1D_ARRAY_SHADOW,
 249    TEX_TARGET_2D_ARRAY_SHADOW,
 250    TEX_TARGET_RECT,
 251    TEX_TARGET_RECT_SHADOW,
 252    TEX_TARGET_CUBE_ARRAY_SHADOW,
 253    TEX_TARGET_BUFFER,
 254    TEX_TARGET_COUNT
 255 };
 256
 257 enum SVSemantic
 258 {
 259    SV_POSITION, // WPOS
 260    SV_VERTEX_ID,
 261    SV_INSTANCE_ID,
 262    SV_INVOCATION_ID,
 263    SV_PRIMITIVE_ID,
 264    SV_VERTEX_COUNT, // gl_PatchVerticesIn
 265    SV_LAYER,
 266    SV_VIEWPORT_INDEX,
 267    SV_YDIR,
 268    SV_FACE,
 269    SV_POINT_SIZE,
 270    SV_POINT_COORD,
 271    SV_CLIP_DISTANCE,
 272    SV_SAMPLE_INDEX,
 273    SV_TESS_FACTOR,
 274    SV_TESS_COORD,
 275    SV_TID,
 276    SV_CTAID,
 277    SV_NTID,
 278    SV_GRIDID,
 279    SV_NCTAID,
 280    SV_LANEID,
 281    SV_PHYSID,
 282    SV_NPHYSID,
 283    SV_CLOCK,
 284    SV_LBASE,
 285    SV_SBASE,
 286    SV_UNDEFINED,
 287    SV_LAST
 288 };
 289
 290 class Program;
 291 class Function;
 292 class BasicBlock;
 293
 294 class Target;
 295
 296 class Instruction;
 297 class CmpInstruction;
 298 class TexInstruction;
 299 class FlowInstruction;
 300
 301 class Value;
 302 class LValue;
 303 class Symbol;
 304 class ImmediateValue;
 305
 306 struct Storage
 307 {
 308    DataFile file;
 309    int8_t fileIndex; // signed, may be indirect for CONST[]
 310    uint8_t size; // this should match the Instruction type's size
 311    DataType type; // mainly for pretty printing
 312    union {
 313       uint64_t u64;    // immediate values
 314       uint32_t u32;
 315       uint16_t u16;
 316       uint8_t u8;
 317       int64_t s64;
 318       int32_t s32;
 319       int16_t s16;
 320       int8_t s8;
 321       float f32;
 322       double f64;
 323       int32_t offset; // offset from 0 (base of address space)
 324       int32_t id;     // register id (< 0 if virtual/unassigned, in units <= 4)
 325       struct {
 326          SVSemantic sv;
 327          int index;
 328       } sv;
 329    } data;
 330 };
 331
 332 // precedence: NOT after SAT after NEG after ABS
 333 #define NV50_IR_MOD_ABS (1 << 0)
 334 #define NV50_IR_MOD_NEG (1 << 1)
 335 #define NV50_IR_MOD_SAT (1 << 2)
 336 #define NV50_IR_MOD_NOT (1 << 3)
 337 #define NV50_IR_MOD_NEG_ABS (NV50_IR_MOD_NEG | NV50_IR_MOD_ABS)
 338
 339 #define NV50_IR_INTERP_MODE_MASK   0x3
 340 #define NV50_IR_INTERP_LINEAR      (0 << 0)
 341 #define NV50_IR_INTERP_PERSPECTIVE (1 << 0)
 342 #define NV50_IR_INTERP_FLAT        (2 << 0)
 343 #define NV50_IR_INTERP_SC          (3 << 0) // what exactly is that ?
 344 #define NV50_IR_INTERP_SAMPLE_MASK 0xc
 345 #define NV50_IR_INTERP_DEFAULT     (0 << 2)
 346 #define NV50_IR_INTERP_CENTROID    (1 << 2)
 347 #define NV50_IR_INTERP_OFFSET      (2 << 2)
 348 #define NV50_IR_INTERP_SAMPLEID    (3 << 2)
 349
 350 // do we really want this to be a class ?
 351 class Modifier
 352 {
 353 public:
 354    Modifier() : bits(0) { }
 355    Modifier(unsigned int m) : bits(m) { }
 356    Modifier(operation op);
 357
 358    // @return new Modifier applying a after b (asserts if unrepresentable)
 359    Modifier operator*(const Modifier) const;
 360    Modifier operator*=(const Modifier m) { *this = *this * m; return *this; }
 361    Modifier operator==(const Modifier m) const { return m.bits == bits; }
 362    Modifier operator!=(const Modifier m) const { return m.bits != bits; }
 363
 364    inline Modifier operator&(const Modifier m) const { return bits & m.bits; }
 365    inline Modifier operator|(const Modifier m) const { return bits | m.bits; }
 366    inline Modifier operator^(const Modifier m) const { return bits ^ m.bits; }
 367
 368    operation getOp() const;
 369
 370    inline int neg() const { return (bits & NV50_IR_MOD_NEG) ? 1 : 0; }
 371    inline int abs() const { return (bits & NV50_IR_MOD_ABS) ? 1 : 0; }
 372
 373    inline operator bool() const { return bits ? true : false; }
 374
 375    void applyTo(ImmediateValue &imm) const;
 376
 377    int print(char *buf, size_t size) const;
 378
 379 private:
 380    uint8_t bits;
 381 };
 382
 383 class ValueRef
 384 {
 385 public:
 386    ValueRef(Value * = NULL);
 387    ValueRef(const ValueRef&);
 388    ~ValueRef();
 389
 390    inline bool exists() const { return value != NULL; }
 391
 392    void set(Value *);
 393    void set(const ValueRef&);
 394    inline Value *get() const { return value; }
 395    inline Value *rep() const;
 396
 397    inline Instruction *getInsn() const { return insn; }
 398    inline void setInsn(Instruction *inst) { insn = inst; }
 399
 400    inline bool isIndirect(int dim) const { return indirect[dim] >= 0; }
 401    inline const ValueRef *getIndirect(int dim) const;
 402
 403    inline DataFile getFile() const;
 404    inline unsigned getSize() const;
 405
 406    // SSA: return eventual (traverse MOVs) literal value, if it exists
 407    bool getImmediate(ImmediateValue&) const;
 408
 409 public:
 410    Modifier mod;
 411    int8_t indirect[2]; // >= 0 if relative to lvalue in insn->src(indirect[i])
 412    uint8_t swizzle;
 413
 414    bool usedAsPtr; // for printing
 415
 416 private:
 417    Value *value;
 418    Instruction *insn;
 419 };
 420
 421 class ValueDef
 422 {
 423 public:
 424    ValueDef(Value * = NULL);
 425    ValueDef(const ValueDef&);
 426    ~ValueDef();
 427
 428    inline bool exists() const { return value != NULL; }
 429
 430    inline Value *get() const { return value; }
 431    inline Value *rep() const;
 432    void set(Value *);
 433    bool mayReplace(const ValueRef &);
 434    void replace(const ValueRef &, bool doSet); // replace all uses of the old value
 435
 436    inline Instruction *getInsn() const { return insn; }
 437    inline void setInsn(Instruction *inst) { insn = inst; }
 438
 439    inline DataFile getFile() const;
 440    inline unsigned getSize() const;
 441
 442    inline void setSSA(LValue *);
 443    inline const LValue *preSSA() const;
 444
 445 private:
 446    Value *value;   // should make this LValue * ...
 447    LValue *origin; // pre SSA value
 448    Instruction *insn;
 449 };
 450
 451 class Value
 452 {
 453 public:
 454    Value();
 455    virtual ~Value() { }
 456
 457    virtual Value *clone(ClonePolicy<Function>&) const = 0;
 458
 459    virtual int print(char *, size_t, DataType ty = TYPE_NONE) const = 0;
 460
 461    virtual bool equals(const Value *, bool strict = false) const;
 462    virtual bool interfers(const Value *) const;
 463    virtual bool isUniform() const { return true; }
 464
 465    inline Value *rep() const { return join; }
 466
 467    inline Instruction *getUniqueInsn() const;
 468    inline Instruction *getInsn() const; // use when uniqueness is certain
 469
 470    inline int refCount() { return uses.size(); }
 471
 472    inline LValue *asLValue();
 473    inline Symbol *asSym();
 474    inline ImmediateValue *asImm();
 475    inline const Symbol *asSym() const;
 476    inline const ImmediateValue *asImm() const;
 477
 478    inline bool inFile(DataFile f) { return reg.file == f; }
 479
 480    static inline Value *get(Iterator&);
 481
 482    std::list<ValueRef *> uses;
 483    std::list<ValueDef *> defs;
 484    typedef std::list<ValueRef *>::iterator UseIterator;
 485    typedef std::list<ValueRef *>::const_iterator UseCIterator;
 486    typedef std::list<ValueDef *>::iterator DefIterator;
 487    typedef std::list<ValueDef *>::const_iterator DefCIterator;
 488
 489    int id;
 490    Storage reg;
 491
 492    // TODO: these should be in LValue:
 493    Interval livei;
 494    Value *join;
 495 };
 496
 497 class LValue : public Value
 498 {
 499 public:
 500    LValue(Function *, DataFile file);
 501    LValue(Function *, LValue *);
 502    ~LValue() { }
 503
 504    virtual bool isUniform() const;
 505
 506    virtual LValue *clone(ClonePolicy<Function>&) const;
 507
 508    virtual int print(char *, size_t, DataType ty = TYPE_NONE) const;
 509
 510 public:
 511    unsigned compMask : 8; // compound/component mask
 512    unsigned compound : 1; // used by RA, value involved in split/merge
 513    unsigned ssa      : 1;
 514    unsigned fixedReg : 1; // set & used by RA, earlier just use (id < 0)
 515    unsigned noSpill  : 1; // do not spill (e.g. if spill temporary already)
 516 };
 517
 518 class Symbol : public Value
 519 {
 520 public:
 521    Symbol(Program *, DataFile file = FILE_MEMORY_CONST, ubyte fileIdx = 0);
 522    ~Symbol() { }
 523
 524    virtual Symbol *clone(ClonePolicy<Function>&) const;
 525
 526    virtual bool equals(const Value *that, bool strict) const;
 527
 528    virtual bool isUniform() const;
 529
 530    virtual int print(char *, size_t, DataType ty = TYPE_NONE) const;
 531
 532    // print with indirect values
 533    int print(char *, size_t, Value *, Value *, DataType ty = TYPE_NONE) const;
 534
 535    inline void setFile(DataFile file, ubyte fileIndex = 0)
 536    {
 537       reg.file = file;
 538       reg.fileIndex = fileIndex;
 539    }
 540
 541    inline void setOffset(int32_t offset);
 542    inline void setAddress(Symbol *base, int32_t offset);
 543    inline void setSV(SVSemantic sv, uint32_t idx = 0);
 544
 545    inline const Symbol *getBase() const { return baseSym; }
 546
 547 private:
 548    Symbol *baseSym; // array base for Symbols representing array elements
 549 };
 550
 551 class ImmediateValue : public Value
 552 {
 553 public:
 554    ImmediateValue() { }
 555    ImmediateValue(Program *, uint32_t);
 556    ImmediateValue(Program *, float);
 557    ImmediateValue(Program *, double);
 558    // NOTE: not added to program with
 559    ImmediateValue(const ImmediateValue *, DataType ty);
 560    ~ImmediateValue() { };
 561
 562    virtual ImmediateValue *clone(ClonePolicy<Function>&) const;
 563
 564    virtual bool equals(const Value *that, bool strict) const;
 565
 566    // these only work if 'type' is valid (we mostly use untyped literals):
 567    bool isInteger(const int ival) const; // ival is cast to this' type
 568    bool isNegative() const;
 569    bool isPow2() const;
 570
 571    void applyLog2();
 572
 573    // for constant folding:
 574    ImmediateValue operator+(const ImmediateValue&) const;
 575    ImmediateValue operator-(const ImmediateValue&) const;
 576    ImmediateValue operator*(const ImmediateValue&) const;
 577    ImmediateValue operator/(const ImmediateValue&) const;
 578
 579    ImmediateValue& operator=(const ImmediateValue&); // only sets value !
 580
 581    bool compare(CondCode cc, float fval) const;
 582
 583    virtual int print(char *, size_t, DataType ty = TYPE_NONE) const;
 584 };
 585
 586 class Instruction
 587 {
 588 public:
 589    Instruction();
 590    Instruction(Function *, operation, DataType);
 591    virtual ~Instruction();
 592
 593    virtual Instruction *clone(ClonePolicy<Function>&,
 594                               Instruction * = NULL) const;
 595
 596    void setDef(int i, Value *);
 597    void setSrc(int s, Value *);
 598    void setSrc(int s, const ValueRef&);
 599    void swapSources(int a, int b);
 600    bool setIndirect(int s, int dim, Value *);
 601
 602    inline ValueRef& src(int s) { return srcs[s]; }
 603    inline ValueDef& def(int s) { return defs[s]; }
 604    inline const ValueRef& src(int s) const { return srcs[s]; }
 605    inline const ValueDef& def(int s) const { return defs[s]; }
 606
 607    inline Value *getDef(int d) const { return defs[d].get(); }
 608    inline Value *getSrc(int s) const { return srcs[s].get(); }
 609    inline Value *getIndirect(int s, int dim) const;
 610
 611    inline bool defExists(unsigned d) const
 612    {
 613       return d < defs.size() && defs[d].exists();
 614    }
 615    inline bool srcExists(unsigned s) const
 616    {
 617       return s < srcs.size() && srcs[s].exists();
 618    }
 619
 620    inline bool constrainedDefs() const;
 621
 622    bool setPredicate(CondCode ccode, Value *);
 623    inline Value *getPredicate() const;
 624    bool writesPredicate() const;
 625    inline bool isPredicated() const { return predSrc >= 0; }
 626
 627    inline void setFlagsSrc(int s, Value *);
 628    inline void setFlagsDef(int d, Value *);
 629
 630    unsigned int defCount() const { return defs.size(); };
 631    unsigned int defCount(unsigned int mask, bool singleFile = false) const;
 632    unsigned int srcCount() const { return srcs.size(); };
 633    unsigned int srcCount(unsigned int mask, bool singleFile = false) const;
 634
 635    // save & remove / set indirect[0,1] and predicate source
 636    void takeExtraSources(int s, Value *[3]);
 637    void putExtraSources(int s, Value *[3]);
 638
 639    inline void setType(DataType type) { dType = sType = type; }
 640
 641    inline void setType(DataType dtype, DataType stype)
 642    {
 643       dType = dtype;
 644       sType = stype;
 645    }
 646
 647    inline bool isPseudo() const { return op < OP_MOV; }
 648    bool isDead() const;
 649    bool isNop() const;
 650    bool isCommutationLegal(const Instruction *) const; // must be adjacent !
 651    bool isActionEqual(const Instruction *) const;
 652    bool isResultEqual(const Instruction *) const;
 653
 654    void print() const;
 655
 656    inline CmpInstruction *asCmp();
 657    inline TexInstruction *asTex();
 658    inline FlowInstruction *asFlow();
 659    inline const TexInstruction *asTex() const;
 660    inline const CmpInstruction *asCmp() const;
 661    inline const FlowInstruction *asFlow() const;
 662
 663 public:
 664    Instruction *next;
 665    Instruction *prev;
 666    int id;
 667    int serial; // CFG order
 668
 669    operation op;
 670    DataType dType; // destination or defining type
 671    DataType sType; // source or secondary type
 672    CondCode cc;
 673    RoundMode rnd;
 674    CacheMode cache;
 675
 676    uint8_t subOp; // quadop, 1 for mul-high, etc.
 677
 678    unsigned encSize    : 4; // encoding size in bytes
 679    unsigned saturate   : 1; // to [0.0f, 1.0f]
 680    unsigned join       : 1; // converge control flow (use OP_JOIN until end)
 681    unsigned fixed      : 1; // prevent dead code elimination
 682    unsigned terminator : 1; // end of basic block
 683    unsigned atomic     : 1;
 684    unsigned ftz        : 1; // flush denormal to zero
 685    unsigned dnz        : 1; // denormals, NaN are zero
 686    unsigned ipa        : 4; // interpolation mode
 687    unsigned lanes      : 4;
 688    unsigned perPatch   : 1;
 689    unsigned exit       : 1; // terminate program after insn
 690
 691    int8_t postFactor; // MUL/DIV(if < 0) by 1 << postFactor
 692
 693    int8_t predSrc;
 694    int8_t flagsDef;
 695    int8_t flagsSrc;
 696
 697    BasicBlock *bb;
 698
 699 protected:
 700    std::deque<ValueDef> defs; // no gaps !
 701    std::deque<ValueRef> srcs; // no gaps !
 702
 703    // instruction specific methods:
 704    // (don't want to subclass, would need more constructors and memory pools)
 705 public:
 706    inline void setInterpolate(unsigned int mode) { ipa = mode; }
 707
 708    unsigned int getInterpMode() const { return ipa & 0x3; }
 709    unsigned int getSampleMode() const { return ipa & 0xc; }
 710
 711 private:
 712    void init();
 713 };
 714
 715 enum TexQuery
 716 {
 717    TXQ_DIMS,
 718    TXQ_TYPE,
 719    TXQ_SAMPLE_POSITION,
 720    TXQ_FILTER,
 721    TXQ_LOD,
 722    TXQ_WRAP,
 723    TXQ_BORDER_COLOUR
 724 };
 725
 726 class TexInstruction : public Instruction
 727 {
 728 public:
 729    class Target
 730    {
 731    public:
 732       Target(TexTarget targ = TEX_TARGET_2D) : target(targ) { }
 733
 734       const char *getName() const { return descTable[target].name; }
 735       unsigned int getArgCount() const { return descTable[target].argc; }
 736       unsigned int getDim() const { return descTable[target].dim; }
 737       int isArray() const { return descTable[target].array ? 1 : 0; }
 738       int isCube() const { return descTable[target].cube ? 1 : 0; }
 739       int isShadow() const { return descTable[target].shadow ? 1 : 0; }
 740
 741       Target& operator=(TexTarget targ)
 742       {
 743          assert(targ < TEX_TARGET_COUNT);
 744          return *this;
 745       }
 746
 747       inline bool operator==(TexTarget targ) const { return target == targ; }
 748
 749    private:
 750       struct Desc
 751       {
 752          char name[19];
 753          uint8_t dim;
 754          uint8_t argc;
 755          bool array;
 756          bool cube;
 757          bool shadow;
 758       };
 759
 760       static const struct Desc descTable[TEX_TARGET_COUNT];
 761
 762    private:
 763       enum TexTarget target;
 764    };
 765
 766 public:
 767    TexInstruction(Function *, operation);
 768    virtual ~TexInstruction();
 769
 770    virtual TexInstruction *clone(ClonePolicy<Function>&,
 771                                  Instruction * = NULL) const;
 772
 773    inline void setTexture(Target targ, uint8_t r, uint8_t s)
 774    {
 775       tex.r = r;
 776       tex.s = s;
 777       tex.target = targ;
 778    }
 779
 780    inline Value *getIndirectR() const;
 781    inline Value *getIndirectS() const;
 782
 783 public:
 784    struct {
 785       Target target;
 786
 787       uint8_t r;
 788       int8_t rIndirectSrc;
 789       uint8_t s;
 790       int8_t sIndirectSrc;
 791
 792       uint8_t mask;
 793       uint8_t gatherComp;
 794
 795       bool liveOnly; // only execute on live pixels of a quad (optimization)
 796       bool levelZero;
 797       bool derivAll;
 798
 799       int8_t useOffsets; // 0, 1, or 4 for textureGatherOffsets
 800       int8_t offset[4][3];
 801
 802       enum TexQuery query;
 803    } tex;
 804
 805    ValueRef dPdx[3];
 806    ValueRef dPdy[3];
 807 };
 808
 809 class CmpInstruction : public Instruction
 810 {
 811 public:
 812    CmpInstruction(Function *, operation);
 813
 814    virtual CmpInstruction *clone(ClonePolicy<Function>&,
 815                                  Instruction * = NULL) const;
 816
 817    void setCondition(CondCode cond) { setCond = cond; }
 818    CondCode getCondition() const { return setCond; }
 819
 820 public:
 821    CondCode setCond;
 822 };
 823
 824 class FlowInstruction : public Instruction
 825 {
 826 public:
 827    FlowInstruction(Function *, operation, void *target);
 828
 829    virtual FlowInstruction *clone(ClonePolicy<Function>&,
 830                                   Instruction * = NULL) const;
 831
 832 public:
 833    unsigned allWarp  : 1;
 834    unsigned absolute : 1;
 835    unsigned limit    : 1;
 836    unsigned builtin  : 1; // true for calls to emulation code
 837
 838    union {
 839       BasicBlock *bb;
 840       int builtin;
 841       Function *fn;
 842    } target;
 843 };
 844
 845 class BasicBlock
 846 {
 847 public:
 848    BasicBlock(Function *);
 849    ~BasicBlock();
 850
 851    BasicBlock *clone(ClonePolicy<Function>&) const;
 852
 853    inline int getId() const { return id; }
 854    inline unsigned int getInsnCount() const { return numInsns; }
 855    inline bool isTerminated() const { return exit && exit->terminator; }
 856
 857    bool dominatedBy(BasicBlock *bb);
 858    inline bool reachableBy(BasicBlock *by, BasicBlock *term);
 859
 860    // returns mask of conditional out blocks
 861    // e.g. 3 for IF { .. } ELSE { .. } ENDIF, 1 for IF { .. } ENDIF
 862    unsigned int initiatesSimpleConditional() const;
 863
 864 public:
 865    Function *getFunction() const { return func; }
 866    Program *getProgram() const { return program; }
 867
 868    Instruction *getEntry() const { return entry; } // first non-phi instruction
 869    Instruction *getPhi() const { return phi; }
 870    Instruction *getFirst() const { return phi ? phi : entry; }
 871    Instruction *getExit() const { return exit; }
 872
 873    void insertHead(Instruction *);
 874    void insertTail(Instruction *);
 875    void insertBefore(Instruction *, Instruction *);
 876    void insertAfter(Instruction *, Instruction *);
 877    void remove(Instruction *);
 878    void permuteAdjacent(Instruction *, Instruction *);
 879
 880    BasicBlock *idom() const;
 881
 882    // NOTE: currently does not rebuild the dominator tree
 883    BasicBlock *splitBefore(Instruction *, bool attach = true);
 884    BasicBlock *splitAfter(Instruction *, bool attach = true);
 885
 886    DLList& getDF() { return df; }
 887    DLList::Iterator iterDF() { return df.iterator(); }
 888
 889    static inline BasicBlock *get(Iterator&);
 890    static inline BasicBlock *get(Graph::Node *);
 891
 892 public:
 893    Graph::Node cfg; // first edge is branch *taken* (the ELSE branch)
 894    Graph::Node dom;
 895
 896    BitSet liveSet;
 897    BitSet defSet;
 898
 899    uint32_t binPos;
 900    uint32_t binSize;
 901
 902    Instruction *joinAt; // for quick reference
 903
 904    bool explicitCont; // loop headers: true if loop contains continue stmts
 905
 906 private:
 907    int id;
 908    DLList df;
 909
 910    Instruction *phi;
 911    Instruction *entry;
 912    Instruction *exit;
 913
 914    unsigned int numInsns;
 915
 916 private:
 917    Function *func;
 918    Program *program;
 919
 920    void splitCommon(Instruction *, BasicBlock *, bool attach);
 921 };
 922
 923 class Function
 924 {
 925 public:
 926    Function(Program *, const char *name, uint32_t label);
 927    ~Function();
 928
 929    static inline Function *get(Graph::Node *node);
 930
 931    inline Program *getProgram() const { return prog; }
 932    inline const char *getName() const { return name; }
 933    inline int getId() const { return id; }
 934    inline uint32_t getLabel() const { return label; }
 935
 936    void print();
 937    void printLiveIntervals() const;
 938    void printCFGraph(const char *filePath);
 939
 940    bool setEntry(BasicBlock *);
 941    bool setExit(BasicBlock *);
 942
 943    unsigned int orderInstructions(ArrayList&);
 944
 945    inline void add(BasicBlock *bb, int& id) { allBBlocks.insert(bb, id); }
 946    inline void add(Instruction *insn, int& id) { allInsns.insert(insn, id); }
 947    inline void add(LValue *lval, int& id) { allLValues.insert(lval, id); }
 948
 949    inline LValue *getLValue(int id);
 950
 951    void buildLiveSets();
 952    void buildDefSets();
 953    bool convertToSSA();
 954
 955 public:
 956    std::deque<ValueDef> ins;
 957    std::deque<ValueRef> outs;
 958    std::deque<Value *> clobbers;
 959
 960    Graph cfg;
 961    Graph::Node *cfgExit;
 962    Graph *domTree;
 963    Graph::Node call; // node in the call graph
 964
 965    BasicBlock **bbArray; // BBs in emission order
 966    int bbCount;
 967
 968    unsigned int loopNestingBound;
 969    int regClobberMax;
 970
 971    uint32_t binPos;
 972    uint32_t binSize;
 973
 974    Value *stackPtr;
 975
 976    uint32_t tlsBase; // base address for l[] space (if no stack pointer is used)
 977    uint32_t tlsSize;
 978
 979    ArrayList allBBlocks;
 980    ArrayList allInsns;
 981    ArrayList allLValues;
 982
 983 private:
 984    void buildLiveSetsPreSSA(BasicBlock *, const int sequence);
 985    void buildDefSetsPreSSA(BasicBlock *bb, const int seq);
 986
 987 private:
 988    uint32_t label;
 989    int id;
 990    const char *const name;
 991    Program *prog;
 992 };
 993
 994 enum CGStage
 995 {
 996    CG_STAGE_PRE_SSA,
 997    CG_STAGE_SSA, // expected directly before register allocation
 998    CG_STAGE_POST_RA
 999 };
1000
1001 class Program
1002 {
1003 public:
1004    enum Type
1005    {
1006       TYPE_VERTEX,
1007       TYPE_TESSELLATION_CONTROL,
1008       TYPE_TESSELLATION_EVAL,
1009       TYPE_GEOMETRY,
1010       TYPE_FRAGMENT,
1011       TYPE_COMPUTE
1012    };
1013
1014    Program(Type type, Target *targ);
1015    ~Program();
1016
1017    void print();
1018
1019    Type getType() const { return progType; }
1020
1021    inline void add(Function *fn, int& id) { allFuncs.insert(fn, id); }
1022    inline void del(Function *fn, int& id) { allFuncs.remove(id); }
1023    inline void add(Value *rval, int& id) { allRValues.insert(rval, id); }
1024
1025    bool makeFromTGSI(struct nv50_ir_prog_info *);
1026    bool makeFromSM4(struct nv50_ir_prog_info *);
1027    bool convertToSSA();
1028    bool optimizeSSA(int level);
1029    bool optimizePostRA(int level);
1030    bool registerAllocation();
1031    bool emitBinary(struct nv50_ir_prog_info *);
1032
1033    const Target *getTarget() const { return target; }
1034
1035 private:
1036    void emitSymbolTable(struct nv50_ir_prog_info *);
1037
1038    Type progType;
1039    Target *target;
1040
1041 public:
1042    Function *main;
1043    Graph calls;
1044
1045    ArrayList allFuncs;
1046    ArrayList allRValues;
1047
1048    uint32_t *code;
1049    uint32_t binSize;
1050    uint32_t tlsSize; // size required for FILE_MEMORY_LOCAL
1051
1052    int maxGPR;
1053
1054    MemoryPool mem_Instruction;
1055    MemoryPool mem_CmpInstruction;
1056    MemoryPool mem_TexInstruction;
1057    MemoryPool mem_FlowInstruction;
1058    MemoryPool mem_LValue;
1059    MemoryPool mem_Symbol;
1060    MemoryPool mem_ImmediateValue;
1061
1062    uint32_t dbgFlags;
1063
1064    void releaseInstruction(Instruction *);
1065    void releaseValue(Value *);
1066 };
1067
1068 // TODO: add const version
1069 class Pass
1070 {
1071 public:
1072    bool run(Program *, bool ordered = false, bool skipPhi = false);
1073    bool run(Function *, bool ordered = false, bool skipPhi = false);
1074
1075 private:
1076    // return false to continue with next entity on next higher level
1077    virtual bool visit(Function *) { return true; }
1078    virtual bool visit(BasicBlock *) { return true; }
1079    virtual bool visit(Instruction *) { return false; }
1080
1081    bool doRun(Program *, bool ordered, bool skipPhi);
1082    bool doRun(Function *, bool ordered, bool skipPhi);
1083
1084 protected:
1085    bool err;
1086    Function *func;
1087    Program *prog;
1088 };
1089
1090 // =============================================================================
1091
1092 #include "nv50_ir_inlines.h"
1093
1094 } // namespace nv50_ir
1095
1096 #endif // __NV50_IR_H__