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,
  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    FILE_IMMEDIATE,
 223    FILE_MEMORY_CONST,
 224    FILE_SHADER_INPUT,
 225    FILE_SHADER_OUTPUT,
 226    FILE_MEMORY_GLOBAL,
 227    FILE_MEMORY_SHARED,
 228    FILE_MEMORY_LOCAL,
 229    FILE_SYSTEM_VALUE,
 230    DATA_FILE_COUNT
 231 };
 232
 233 enum TexTarget
 234 {
 235    TEX_TARGET_1D,
 236    TEX_TARGET_2D,
 237    TEX_TARGET_2D_MS,
 238    TEX_TARGET_3D,
 239    TEX_TARGET_CUBE,
 240    TEX_TARGET_1D_SHADOW,
 241    TEX_TARGET_2D_SHADOW,
 242    TEX_TARGET_CUBE_SHADOW,
 243    TEX_TARGET_1D_ARRAY,
 244    TEX_TARGET_2D_ARRAY,
 245    TEX_TARGET_2D_MS_ARRAY,
 246    TEX_TARGET_CUBE_ARRAY,
 247    TEX_TARGET_1D_ARRAY_SHADOW,
 248    TEX_TARGET_2D_ARRAY_SHADOW,
 249    TEX_TARGET_RECT,
 250    TEX_TARGET_RECT_SHADOW,
 251    TEX_TARGET_CUBE_ARRAY_SHADOW,
 252    TEX_TARGET_BUFFER,
 253    TEX_TARGET_COUNT
 254 };
 255
 256 enum SVSemantic
 257 {
 258    SV_POSITION, // WPOS
 259    SV_VERTEX_ID,
 260    SV_INSTANCE_ID,
 261    SV_INVOCATION_ID,
 262    SV_PRIMITIVE_ID,
 263    SV_VERTEX_COUNT, // gl_PatchVerticesIn
 264    SV_LAYER,
 265    SV_VIEWPORT_INDEX,
 266    SV_YDIR,
 267    SV_FACE,
 268    SV_POINT_SIZE,
 269    SV_POINT_COORD,
 270    SV_CLIP_DISTANCE,
 271    SV_SAMPLE_INDEX,
 272    SV_TESS_FACTOR,
 273    SV_TESS_COORD,
 274    SV_TID,
 275    SV_CTAID,
 276    SV_NTID,
 277    SV_GRIDID,
 278    SV_NCTAID,
 279    SV_LANEID,
 280    SV_PHYSID,
 281    SV_NPHYSID,
 282    SV_CLOCK,
 283    SV_LBASE,
 284    SV_SBASE,
 285    SV_UNDEFINED,
 286    SV_LAST
 287 };
 288
 289 class Program;
 290 class Function;
 291 class BasicBlock;
 292
 293 class Target;
 294
 295 class Instruction;
 296 class CmpInstruction;
 297 class TexInstruction;
 298 class FlowInstruction;
 299
 300 class Value;
 301 class LValue;
 302 class Symbol;
 303 class ImmediateValue;
 304
 305 struct Storage
 306 {
 307    DataFile file;
 308    int8_t fileIndex; // signed, may be indirect for CONST[]
 309    uint8_t size; // this should match the Instruction type's size
 310    DataType type; // mainly for pretty printing
 311    union {
 312       uint64_t u64;    // immediate values
 313       uint32_t u32;
 314       uint16_t u16;
 315       uint8_t u8;
 316       int64_t s64;
 317       int32_t s32;
 318       int16_t s16;
 319       int8_t s8;
 320       float f32;
 321       double f64;
 322       int32_t offset; // offset from 0 (base of address space)
 323       int32_t id;     // register id (< 0 if virtual/unassigned)
 324       struct {
 325          SVSemantic sv;
 326          int index;
 327       } sv;
 328    } data;
 329 };
 330
 331 // precedence: NOT after SAT after NEG after ABS
 332 #define NV50_IR_MOD_ABS (1 << 0)
 333 #define NV50_IR_MOD_NEG (1 << 1)
 334 #define NV50_IR_MOD_SAT (1 << 2)
 335 #define NV50_IR_MOD_NOT (1 << 3)
 336 #define NV50_IR_MOD_NEG_ABS (NV50_IR_MOD_NEG | NV50_IR_MOD_ABS)
 337
 338 #define NV50_IR_INTERP_MODE_MASK   0x3
 339 #define NV50_IR_INTERP_LINEAR      (0 << 0)
 340 #define NV50_IR_INTERP_PERSPECTIVE (1 << 0)
 341 #define NV50_IR_INTERP_FLAT        (2 << 0)
 342 #define NV50_IR_INTERP_SC          (3 << 0) // what exactly is that ?
 343 #define NV50_IR_INTERP_SAMPLE_MASK 0xc
 344 #define NV50_IR_INTERP_DEFAULT     (0 << 2)
 345 #define NV50_IR_INTERP_CENTROID    (1 << 2)
 346 #define NV50_IR_INTERP_OFFSET      (2 << 2)
 347 #define NV50_IR_INTERP_SAMPLEID    (3 << 2)
 348
 349 // do we really want this to be a class ?
 350 class Modifier
 351 {
 352 public:
 353    Modifier() : bits(0) { }
 354    Modifier(unsigned int m) : bits(m) { }
 355    Modifier(operation op);
 356
 357    // @return new Modifier applying a after b (asserts if unrepresentable)
 358    Modifier operator*(const Modifier) const;
 359    Modifier operator==(const Modifier m) const { return m.bits == bits; }
 360    Modifier operator!=(const Modifier m) const { return m.bits != bits; }
 361
 362    inline Modifier operator&(const Modifier m) const { return bits & m.bits; }
 363    inline Modifier operator|(const Modifier m) const { return bits | m.bits; }
 364    inline Modifier operator^(const Modifier m) const { return bits ^ m.bits; }
 365
 366    operation getOp() const;
 367
 368    inline int neg() const { return (bits & NV50_IR_MOD_NEG) ? 1 : 0; }
 369    inline int abs() const { return (bits & NV50_IR_MOD_ABS) ? 1 : 0; }
 370
 371    inline operator bool() { return bits ? true : false; }
 372
 373    void applyTo(ImmediateValue &imm) const;
 374
 375    int print(char *buf, size_t size) const;
 376
 377 private:
 378    uint8_t bits;
 379 };
 380
 381 class ValueRef
 382 {
 383 public:
 384    ValueRef();
 385    ValueRef(const ValueRef&);
 386    ~ValueRef();
 387
 388    inline ValueRef& operator=(Value *val) { this->set(val); return *this; }
 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    ImmediateValue *getImmediate() 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();
 425    ValueDef(const ValueDef&);
 426    ~ValueDef();
 427
 428    inline ValueDef& operator=(Value *val) { this->set(val); return *this; }
 429
 430    inline bool exists() const { return value != NULL; }
 431
 432    inline Value *get() const { return value; }
 433    inline Value *rep() const;
 434    void set(Value *);
 435    void replace(Value *, bool doSet); // replace all uses of the old value
 436
 437    inline Instruction *getInsn() const { return insn; }
 438    inline void setInsn(Instruction *inst) { insn = inst; }
 439
 440    inline DataFile getFile() const;
 441    inline unsigned getSize() const;
 442
 443    inline void setSSA(LValue *);
 444    inline const LValue *preSSA() const;
 445
 446 private:
 447    Value *value;   // should make this LValue * ...
 448    LValue *origin; // pre SSA value
 449    Instruction *insn;
 450 };
 451
 452 class Value
 453 {
 454 public:
 455    Value();
 456
 457    virtual Value *clone(Function *) const { return NULL; }
 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
 464    inline Instruction *getUniqueInsn() const;
 465    inline Instruction *getInsn() const; // use when uniqueness is certain
 466
 467    inline int refCount() { return uses.size(); }
 468
 469    inline LValue *asLValue();
 470    inline Symbol *asSym();
 471    inline ImmediateValue *asImm();
 472    inline const Symbol *asSym() const;
 473    inline const ImmediateValue *asImm() const;
 474
 475    bool coalesce(Value *, bool force = false);
 476
 477    inline bool inFile(DataFile f) { return reg.file == f; }
 478
 479    static inline Value *get(Iterator&);
 480
 481    std::list<ValueRef *> uses;
 482    std::list<ValueDef *> defs;
 483    typedef std::list<ValueRef *>::iterator UseIterator;
 484    typedef std::list<ValueRef *>::const_iterator UseCIterator;
 485    typedef std::list<ValueDef *>::iterator DefIterator;
 486    typedef std::list<ValueDef *>::const_iterator DefCIterator;
 487
 488    int id;
 489    Storage reg;
 490
 491    // TODO: these should be in LValue:
 492    Interval livei;
 493    Value *join;
 494 };
 495
 496 class LValue : public Value
 497 {
 498 public:
 499    LValue(Function *, DataFile file);
 500    LValue(Function *, LValue *);
 501
 502    virtual Value *clone(Function *) const;
 503
 504    virtual int print(char *, size_t, DataType ty = TYPE_NONE) const;
 505
 506 public:
 507    unsigned ssa : 1;
 508
 509    int affinity;
 510 };
 511
 512 class Symbol : public Value
 513 {
 514 public:
 515    Symbol(Program *, DataFile file = FILE_MEMORY_CONST, ubyte fileIdx = 0);
 516
 517    virtual Value *clone(Function *) const;
 518
 519    virtual bool equals(const Value *that, bool strict) const;
 520
 521    virtual int print(char *, size_t, DataType ty = TYPE_NONE) const;
 522
 523    // print with indirect values
 524    int print(char *, size_t, Value *, Value *, DataType ty = TYPE_NONE) const;
 525
 526    inline void setFile(DataFile file, ubyte fileIndex = 0)
 527    {
 528       reg.file = file;
 529       reg.fileIndex = fileIndex;
 530    }
 531
 532    inline void setOffset(int32_t offset);
 533    inline void setAddress(Symbol *base, int32_t offset);
 534    inline void setSV(SVSemantic sv, uint32_t idx = 0);
 535
 536    inline const Symbol *getBase() const { return baseSym; }
 537
 538 private:
 539    Symbol *baseSym; // array base for Symbols representing array elements
 540 };
 541
 542 class ImmediateValue : public Value
 543 {
 544 public:
 545    ImmediateValue(Program *, uint32_t);
 546    ImmediateValue(Program *, float);
 547    ImmediateValue(Program *, double);
 548
 549    // NOTE: not added to program with
 550    ImmediateValue(const ImmediateValue *, DataType ty);
 551
 552    virtual bool equals(const Value *that, bool strict) const;
 553
 554    // these only work if 'type' is valid (we mostly use untyped literals):
 555    bool isInteger(const int ival) const; // ival is cast to this' type
 556    bool isNegative() const;
 557    bool isPow2() const;
 558
 559    void applyLog2();
 560
 561    // for constant folding:
 562    ImmediateValue operator+(const ImmediateValue&) const;
 563    ImmediateValue operator-(const ImmediateValue&) const;
 564    ImmediateValue operator*(const ImmediateValue&) const;
 565    ImmediateValue operator/(const ImmediateValue&) const;
 566
 567    bool compare(CondCode cc, float fval) const;
 568
 569    virtual int print(char *, size_t, DataType ty = TYPE_NONE) const;
 570 };
 571
 572 class Instruction
 573 {
 574 public:
 575    Instruction();
 576    Instruction(Function *, operation, DataType);
 577    virtual ~Instruction();
 578
 579    virtual Instruction *clone(bool deep) const;
 580
 581    void setDef(int i, Value *);
 582    void setSrc(int s, Value *);
 583    void setSrc(int s, const ValueRef&);
 584    void swapSources(int a, int b);
 585    bool setIndirect(int s, int dim, Value *);
 586
 587    inline Value *getDef(int d) const { return def[d].get(); }
 588    inline Value *getSrc(int s) const { return src[s].get(); }
 589    inline Value *getIndirect(int s, int dim) const;
 590
 591    inline bool defExists(unsigned d) const
 592    {
 593       return d < def.size() && def[d].exists();
 594    }
 595    inline bool srcExists(unsigned s) const
 596    {
 597       return s < src.size() && src[s].exists();
 598    }
 599
 600    inline bool constrainedDefs() const { return defExists(1); }
 601
 602    bool setPredicate(CondCode ccode, Value *);
 603    inline Value *getPredicate() const;
 604    bool writesPredicate() const;
 605
 606    inline void setFlagsSrc(int s, Value *);
 607    inline void setFlagsDef(int d, Value *);
 608
 609    unsigned int defCount(unsigned int mask) const;
 610    unsigned int srcCount(unsigned int mask) const;
 611
 612    // save & remove / set indirect[0,1] and predicate source
 613    void takeExtraSources(int s, Value *[3]);
 614    void putExtraSources(int s, Value *[3]);
 615
 616    inline void setType(DataType type) { dType = sType = type; }
 617
 618    inline void setType(DataType dtype, DataType stype)
 619    {
 620       dType = dtype;
 621       sType = stype;
 622    }
 623
 624    inline bool isPseudo() const { return op < OP_MOV; }
 625    bool isDead() const;
 626    bool isNop() const;
 627    bool isCommutationLegal(const Instruction *) const; // must be adjacent !
 628    bool isActionEqual(const Instruction *) const;
 629    bool isResultEqual(const Instruction *) const;
 630
 631    void print() const;
 632
 633    inline CmpInstruction *asCmp();
 634    inline TexInstruction *asTex();
 635    inline FlowInstruction *asFlow();
 636    inline const TexInstruction *asTex() const;
 637    inline const CmpInstruction *asCmp() const;
 638    inline const FlowInstruction *asFlow() const;
 639
 640 public:
 641    Instruction *next;
 642    Instruction *prev;
 643    int id;
 644    int serial; // CFG order
 645
 646    operation op;
 647    DataType dType; // destination or defining type
 648    DataType sType; // source or secondary type
 649    CondCode cc;
 650    RoundMode rnd;
 651    CacheMode cache;
 652
 653    uint8_t subOp; // quadop, 1 for mul-high, etc.
 654
 655    unsigned encSize    : 4; // encoding size in bytes
 656    unsigned saturate   : 1; // to [0.0f, 1.0f]
 657    unsigned join       : 1; // converge control flow (use OP_JOIN until end)
 658    unsigned fixed      : 1; // prevent dead code elimination
 659    unsigned terminator : 1; // end of basic block
 660    unsigned atomic     : 1;
 661    unsigned ftz        : 1; // flush denormal to zero
 662    unsigned dnz        : 1; // denormals, NaN are zero
 663    unsigned ipa        : 4; // interpolation mode
 664    unsigned lanes      : 4;
 665    unsigned perPatch   : 1;
 666    unsigned exit       : 1; // terminate program after insn
 667
 668    int8_t postFactor; // MUL/DIV(if < 0) by 1 << postFactor
 669
 670    int8_t predSrc;
 671    int8_t flagsDef;
 672    int8_t flagsSrc;
 673
 674    std::deque<ValueDef> def; // no gaps !
 675    std::deque<ValueRef> src; // no gaps !
 676
 677    BasicBlock *bb;
 678
 679    // instruction specific methods:
 680    // (don't want to subclass, would need more constructors and memory pools)
 681 public:
 682    inline void setInterpolate(unsigned int mode) { ipa = mode; }
 683
 684    unsigned int getInterpMode() const { return ipa & 0x3; }
 685    unsigned int getSampleMode() const { return ipa & 0xc; }
 686
 687 private:
 688    void init();
 689 protected:
 690    void cloneBase(Instruction *clone, bool deep) const;
 691 };
 692
 693 enum TexQuery
 694 {
 695    TXQ_DIMS,
 696    TXQ_TYPE,
 697    TXQ_SAMPLE_POSITION,
 698    TXQ_FILTER,
 699    TXQ_LOD,
 700    TXQ_WRAP,
 701    TXQ_BORDER_COLOUR
 702 };
 703
 704 class TexInstruction : public Instruction
 705 {
 706 public:
 707    class Target
 708    {
 709    public:
 710       Target(TexTarget targ = TEX_TARGET_2D) : target(targ) { }
 711
 712       const char *getName() const { return descTable[target].name; }
 713       unsigned int getArgCount() const { return descTable[target].argc; }
 714       unsigned int getDim() const { return descTable[target].dim; }
 715       int isArray() const { return descTable[target].array ? 1 : 0; }
 716       int isCube() const { return descTable[target].cube ? 1 : 0; }
 717       int isShadow() const { return descTable[target].shadow ? 1 : 0; }
 718
 719       Target& operator=(TexTarget targ)
 720       {
 721          assert(targ < TEX_TARGET_COUNT);
 722          return *this;
 723       }
 724
 725       inline bool operator==(TexTarget targ) const { return target == targ; }
 726
 727    private:
 728       struct Desc
 729       {
 730          char name[19];
 731          uint8_t dim;
 732          uint8_t argc;
 733          bool array;
 734          bool cube;
 735          bool shadow;
 736       };
 737
 738       static const struct Desc descTable[TEX_TARGET_COUNT];
 739
 740    private:
 741       enum TexTarget target;
 742    };
 743
 744 public:
 745    TexInstruction(Function *, operation);
 746    virtual ~TexInstruction();
 747
 748    virtual Instruction *clone(bool deep) const;
 749
 750    inline void setTexture(Target targ, uint8_t r, uint8_t s)
 751    {
 752       tex.r = r;
 753       tex.s = s;
 754       tex.target = targ;
 755    }
 756
 757    inline Value *getIndirectR() const;
 758    inline Value *getIndirectS() const;
 759
 760 public:
 761    struct {
 762       Target target;
 763
 764       uint8_t r;
 765       int8_t rIndirectSrc;
 766       uint8_t s;
 767       int8_t sIndirectSrc;
 768
 769       uint8_t mask;
 770       uint8_t gatherComp;
 771
 772       bool liveOnly; // only execute on live pixels of a quad (optimization)
 773       bool levelZero;
 774       bool derivAll;
 775
 776       int8_t useOffsets; // 0, 1, or 4 for textureGatherOffsets
 777       int8_t offset[4][3];
 778
 779       enum TexQuery query;
 780    } tex;
 781
 782    ValueRef dPdx[3];
 783    ValueRef dPdy[3];
 784 };
 785
 786 class CmpInstruction : public Instruction
 787 {
 788 public:
 789    CmpInstruction(Function *, operation);
 790
 791    virtual Instruction *clone(bool deep) const;
 792
 793    void setCondition(CondCode cond) { setCond = cond; }
 794    CondCode getCondition() const { return setCond; }
 795
 796 public:
 797    CondCode setCond;
 798 };
 799
 800 class FlowInstruction : public Instruction
 801 {
 802 public:
 803    FlowInstruction(Function *, operation, BasicBlock *target);
 804
 805 public:
 806    unsigned allWarp  : 1;
 807    unsigned absolute : 1;
 808    unsigned limit    : 1;
 809    unsigned builtin  : 1; // true for calls to emulation code
 810
 811    union {
 812       BasicBlock *bb;
 813       int builtin;
 814       Function *fn;
 815    } target;
 816 };
 817
 818 class BasicBlock
 819 {
 820 public:
 821    BasicBlock(Function *);
 822    ~BasicBlock();
 823
 824    inline int getId() const { return id; }
 825    inline unsigned int getInsnCount() const { return numInsns; }
 826    inline bool isTerminated() const { return exit && exit->terminator; }
 827
 828    bool dominatedBy(BasicBlock *bb);
 829    inline bool reachableBy(BasicBlock *by, BasicBlock *term);
 830
 831    // returns mask of conditional out blocks
 832    // e.g. 3 for IF { .. } ELSE { .. } ENDIF, 1 for IF { .. } ENDIF
 833    unsigned int initiatesSimpleConditional() const;
 834
 835 public:
 836    Function *getFunction() const { return func; }
 837    Program *getProgram() const { return program; }
 838
 839    Instruction *getEntry() const { return entry; } // first non-phi instruction
 840    Instruction *getPhi() const { return phi; }
 841    Instruction *getFirst() const { return phi ? phi : entry; }
 842    Instruction *getExit() const { return exit; }
 843
 844    void insertHead(Instruction *);
 845    void insertTail(Instruction *);
 846    void insertBefore(Instruction *, Instruction *);
 847    void insertAfter(Instruction *, Instruction *);
 848    void remove(Instruction *);
 849    void permuteAdjacent(Instruction *, Instruction *);
 850
 851    BasicBlock *idom() const;
 852
 853    // NOTE: currently does not rebuild the dominator tree
 854    BasicBlock *splitBefore(Instruction *, bool attach = true);
 855    BasicBlock *splitAfter(Instruction *, bool attach = true);
 856
 857    DLList& getDF() { return df; }
 858    DLList::Iterator iterDF() { return df.iterator(); }
 859
 860    static inline BasicBlock *get(Iterator&);
 861    static inline BasicBlock *get(Graph::Node *);
 862
 863 public:
 864    Graph::Node cfg; // first edge is branch *taken* (the ELSE branch)
 865    Graph::Node dom;
 866
 867    BitSet liveSet;
 868
 869    uint32_t binPos;
 870    uint32_t binSize;
 871
 872    Instruction *joinAt; // for quick reference
 873
 874    bool explicitCont; // loop headers: true if loop contains continue stmts
 875
 876 private:
 877    int id;
 878    DLList df;
 879
 880    Instruction *phi;
 881    Instruction *entry;
 882    Instruction *exit;
 883
 884    unsigned int numInsns;
 885
 886 private:
 887    Function *func;
 888    Program *program;
 889
 890    void splitCommon(Instruction *, BasicBlock *, bool attach);
 891 };
 892
 893 class Function
 894 {
 895 public:
 896    Function(Program *, const char *name);
 897    ~Function();
 898
 899    inline Program *getProgram() const { return prog; }
 900    inline const char *getName() const { return name; }
 901    inline int getId() const { return id; }
 902
 903    void print();
 904    void printLiveIntervals() const;
 905    void printCFGraph(const char *filePath);
 906
 907    bool setEntry(BasicBlock *);
 908    bool setExit(BasicBlock *);
 909
 910    unsigned int orderInstructions(ArrayList&);
 911
 912    inline void add(BasicBlock *bb, int& id) { allBBlocks.insert(bb, id); }
 913    inline void add(Instruction *insn, int& id) { allInsns.insert(insn, id); }
 914    inline void add(LValue *lval, int& id) { allLValues.insert(lval, id); }
 915
 916    inline LValue *getLValue(int id);
 917
 918    bool convertToSSA();
 919
 920 public:
 921    Graph cfg;
 922    Graph::Node *cfgExit;
 923    Graph *domTree;
 924    Graph::Node call; // node in the call graph
 925
 926    BasicBlock **bbArray; // BBs in emission order
 927    int bbCount;
 928
 929    unsigned int loopNestingBound;
 930    int regClobberMax;
 931
 932    uint32_t binPos;
 933    uint32_t binSize;
 934
 935    ArrayList allBBlocks;
 936    ArrayList allInsns;
 937    ArrayList allLValues;
 938
 939 private:
 940    void buildLiveSetsPreSSA(BasicBlock *, const int sequence);
 941
 942 private:
 943    int id;
 944    const char *const name;
 945    Program *prog;
 946 };
 947
 948 enum CGStage
 949 {
 950    CG_STAGE_PRE_SSA,
 951    CG_STAGE_SSA, // expected directly before register allocation
 952    CG_STAGE_POST_RA
 953 };
 954
 955 class Program
 956 {
 957 public:
 958    enum Type
 959    {
 960       TYPE_VERTEX,
 961       TYPE_TESSELLATION_CONTROL,
 962       TYPE_TESSELLATION_EVAL,
 963       TYPE_GEOMETRY,
 964       TYPE_FRAGMENT,
 965       TYPE_COMPUTE
 966    };
 967
 968    Program(Type type, Target *targ);
 969    ~Program();
 970
 971    void print();
 972
 973    Type getType() const { return progType; }
 974
 975    inline void add(Function *fn, int& id) { allFuncs.insert(fn, id); }
 976    inline void add(Value *rval, int& id) { allRValues.insert(rval, id); }
 977
 978    bool makeFromTGSI(struct nv50_ir_prog_info *);
 979    bool makeFromSM4(struct nv50_ir_prog_info *);
 980    bool convertToSSA();
 981    bool optimizeSSA(int level);
 982    bool optimizePostRA(int level);
 983    bool registerAllocation();
 984    bool emitBinary(struct nv50_ir_prog_info *);
 985
 986    const Target *getTarget() const { return target; }
 987
 988 private:
 989    Type progType;
 990    Target *target;
 991
 992 public:
 993    Function *main;
 994    Graph calls;
 995
 996    ArrayList allFuncs;
 997    ArrayList allRValues;
 998
 999    uint32_t *code;
1000    uint32_t binSize;
1001
1002    int maxGPR;
1003
1004    MemoryPool mem_Instruction;
1005    MemoryPool mem_CmpInstruction;
1006    MemoryPool mem_TexInstruction;
1007    MemoryPool mem_FlowInstruction;
1008    MemoryPool mem_LValue;
1009    MemoryPool mem_Symbol;
1010    MemoryPool mem_ImmediateValue;
1011
1012    uint32_t dbgFlags;
1013
1014    void releaseInstruction(Instruction *);
1015    void releaseValue(Value *);
1016 };
1017
1018 // TODO: add const version
1019 class Pass
1020 {
1021 public:
1022    bool run(Program *, bool ordered = false, bool skipPhi = false);
1023    bool run(Function *, bool ordered = false, bool skipPhi = false);
1024
1025 private:
1026    // return false to continue with next entity on next higher level
1027    virtual bool visit(Function *) { return true; }
1028    virtual bool visit(BasicBlock *) { return true; }
1029    virtual bool visit(Instruction *) { return false; }
1030
1031    bool doRun(Program *, bool ordered, bool skipPhi);
1032    bool doRun(Function *, bool ordered, bool skipPhi);
1033
1034 protected:
1035    bool err;
1036    Function *func;
1037    Program *prog;
1038 };
1039
1040 // =============================================================================
1041
1042 #include "nv50_ir_inlines.h"
1043
1044 } // namespace nv50_ir
1045
1046 #endif // __NV50_IR_H__