src/gallium/drivers/nouveau/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 OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
  18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  20  * OTHER DEALINGS IN THE 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 "codegen/nv50_ir_util.h"
  34 #include "codegen/nv50_ir_graph.h"
  35
  36 #include "codegen/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, // NOTE: add u64 + u32 is legal for targets w/o 64-bit integer adds
  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, // memory barrier (mfence, lfence, sfence)
 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, // texture op for coverage sampling
 121    OP_TEXPREP, // turn cube map array into 2d array coordinates
 122    OP_SULDB, // surface load (raw)
 123    OP_SULDP, // surface load (formatted)
 124    OP_SUSTB, // surface store (raw)
 125    OP_SUSTP, // surface store (formatted)
 126    OP_SUREDB,
 127    OP_SUREDP, // surface reduction (atomic op)
 128    OP_SULEA,   // surface load effective address
 129    OP_SUBFM,   // surface bitfield manipulation
 130    OP_SUCLAMP, // clamp surface coordinates
 131    OP_SUEAU,   // surface effective address
 132    OP_MADSP,   // special integer multiply-add
 133    OP_TEXBAR, // texture dependency barrier
 134    OP_DFDX,
 135    OP_DFDY,
 136    OP_RDSV, // read system value
 137    OP_WRSV, // write system value
 138    OP_QUADOP,
 139    OP_QUADON,
 140    OP_QUADPOP,
 141    OP_POPCNT, // bitcount(src0 & src1)
 142    OP_INSBF,  // insert first src1[8:15] bits of src0 into src2 at src1[0:7]
 143    OP_EXTBF,  // place bits [K,K+N) of src0 into dst, src1 = 0xNNKK
 144    OP_PERMT,  // dst = bytes from src2,src0 selected by src1 (nvc0's src order)
 145    OP_ATOM,
 146    OP_BAR,    // execution barrier, sources = { id, thread count, predicate }
 147    OP_VADD,   // byte/word vector operations
 148    OP_VAVG,
 149    OP_VMIN,
 150    OP_VMAX,
 151    OP_VSAD,
 152    OP_VSET,
 153    OP_VSHR,
 154    OP_VSHL,
 155    OP_VSEL,
 156    OP_CCTL, // cache control
 157    OP_LAST
 158 };
 159
 160 // various instruction-specific modifier definitions Instruction::subOp
 161 // MOV_FINAL marks a MOV originating from an EXPORT (used for placing TEXBARs)
 162 #define NV50_IR_SUBOP_MUL_HIGH     1
 163 #define NV50_IR_SUBOP_EMIT_RESTART 1
 164 #define NV50_IR_SUBOP_LDC_IL       1
 165 #define NV50_IR_SUBOP_LDC_IS       2
 166 #define NV50_IR_SUBOP_LDC_ISL      3
 167 #define NV50_IR_SUBOP_SHIFT_WRAP   1
 168 #define NV50_IR_SUBOP_EMU_PRERET   1
 169 #define NV50_IR_SUBOP_TEXBAR(n)    n
 170 #define NV50_IR_SUBOP_MOV_FINAL    1
 171 #define NV50_IR_SUBOP_EXTBF_REV    1
 172 #define NV50_IR_SUBOP_PERMT_F4E    1
 173 #define NV50_IR_SUBOP_PERMT_B4E    2
 174 #define NV50_IR_SUBOP_PERMT_RC8    3
 175 #define NV50_IR_SUBOP_PERMT_ECL    4
 176 #define NV50_IR_SUBOP_PERMT_ECR    5
 177 #define NV50_IR_SUBOP_PERMT_RC16   6
 178 #define NV50_IR_SUBOP_BAR_SYNC     0
 179 #define NV50_IR_SUBOP_BAR_ARRIVE   1
 180 #define NV50_IR_SUBOP_BAR_RED_AND  2
 181 #define NV50_IR_SUBOP_BAR_RED_OR   3
 182 #define NV50_IR_SUBOP_BAR_RED_POPC 4
 183 #define NV50_IR_SUBOP_MEMBAR_L     1
 184 #define NV50_IR_SUBOP_MEMBAR_S     2
 185 #define NV50_IR_SUBOP_MEMBAR_M     3
 186 #define NV50_IR_SUBOP_MEMBAR_CTA  (0 << 2)
 187 #define NV50_IR_SUBOP_MEMBAR_GL   (1 << 2)
 188 #define NV50_IR_SUBOP_MEMBAR_SYS  (2 << 2)
 189 #define NV50_IR_SUBOP_MEMBAR_DIR(m)   ((m) & 0x3)
 190 #define NV50_IR_SUBOP_MEMBAR_SCOPE(m) ((m) & ~0x3)
 191 #define NV50_IR_SUBOP_MEMBAR(d,s) \
 192    (NV50_IR_SUBOP_MEMBAR_##d | NV50_IR_SUBOP_MEMBAR_##s)
 193 #define NV50_IR_SUBOP_ATOM_ADD      0
 194 #define NV50_IR_SUBOP_ATOM_MIN      1
 195 #define NV50_IR_SUBOP_ATOM_MAX      2
 196 #define NV50_IR_SUBOP_ATOM_INC      3
 197 #define NV50_IR_SUBOP_ATOM_DEC      4
 198 #define NV50_IR_SUBOP_ATOM_AND      5
 199 #define NV50_IR_SUBOP_ATOM_OR       6
 200 #define NV50_IR_SUBOP_ATOM_XOR      7
 201 #define NV50_IR_SUBOP_ATOM_CAS      8
 202 #define NV50_IR_SUBOP_ATOM_EXCH     9
 203 #define NV50_IR_SUBOP_CCTL_IV      5
 204 #define NV50_IR_SUBOP_CCTL_IVALL   6
 205 #define NV50_IR_SUBOP_SUST_IGN     0
 206 #define NV50_IR_SUBOP_SUST_TRAP    1
 207 #define NV50_IR_SUBOP_SUST_SDCL    3
 208 #define NV50_IR_SUBOP_SULD_ZERO    0
 209 #define NV50_IR_SUBOP_SULD_TRAP    1
 210 #define NV50_IR_SUBOP_SULD_SDCL    3
 211 #define NV50_IR_SUBOP_SUBFM_3D     1
 212 #define NV50_IR_SUBOP_SUCLAMP_2D   0x10
 213 #define NV50_IR_SUBOP_SUCLAMP_SD(r, d) (( 0 + (r)) | ((d == 2) ? 0x10 : 0))
 214 #define NV50_IR_SUBOP_SUCLAMP_PL(r, d) (( 5 + (r)) | ((d == 2) ? 0x10 : 0))
 215 #define NV50_IR_SUBOP_SUCLAMP_BL(r, d) ((10 + (r)) | ((d == 2) ? 0x10 : 0))
 216 #define NV50_IR_SUBOP_MADSP_SD     0xffff
 217 // Yes, we could represent those with DataType.
 218 // Or put the type into operation and have a couple 1000 values in that enum.
 219 // This will have to do for now.
 220 // The bitfields are supposed to correspond to nve4 ISA.
 221 #define NV50_IR_SUBOP_MADSP(a,b,c) (((c) << 8) | ((b) << 4) | (a))
 222 #define NV50_IR_SUBOP_V1(d,a,b)    (((d) << 10) | ((b) << 5) | (a) | 0x0000)
 223 #define NV50_IR_SUBOP_V2(d,a,b)    (((d) << 10) | ((b) << 5) | (a) | 0x4000)
 224 #define NV50_IR_SUBOP_V4(d,a,b)    (((d) << 10) | ((b) << 5) | (a) | 0x8000)
 225 #define NV50_IR_SUBOP_Vn(n)        ((n) >> 14)
 226
 227 enum DataType
 228 {
 229    TYPE_NONE,
 230    TYPE_U8,
 231    TYPE_S8,
 232    TYPE_U16,
 233    TYPE_S16,
 234    TYPE_U32,
 235    TYPE_S32,
 236    TYPE_U64, // 64 bit operations are only lowered after register allocation
 237    TYPE_S64,
 238    TYPE_F16,
 239    TYPE_F32,
 240    TYPE_F64,
 241    TYPE_B96,
 242    TYPE_B128
 243 };
 244
 245 enum CondCode
 246 {
 247    CC_FL = 0,
 248    CC_NEVER = CC_FL, // when used with FILE_FLAGS
 249    CC_LT = 1,
 250    CC_EQ = 2,
 251    CC_NOT_P = CC_EQ, // when used with FILE_PREDICATE
 252    CC_LE = 3,
 253    CC_GT = 4,
 254    CC_NE = 5,
 255    CC_P  = CC_NE,
 256    CC_GE = 6,
 257    CC_TR = 7,
 258    CC_ALWAYS = CC_TR,
 259    CC_U  = 8,
 260    CC_LTU = 9,
 261    CC_EQU = 10,
 262    CC_LEU = 11,
 263    CC_GTU = 12,
 264    CC_NEU = 13,
 265    CC_GEU = 14,
 266    CC_NO = 0x10,
 267    CC_NC = 0x11,
 268    CC_NS = 0x12,
 269    CC_NA = 0x13,
 270    CC_A  = 0x14,
 271    CC_S  = 0x15,
 272    CC_C  = 0x16,
 273    CC_O  = 0x17
 274 };
 275
 276 enum RoundMode
 277 {
 278    ROUND_N, // nearest
 279    ROUND_M, // towards -inf
 280    ROUND_Z, // towards 0
 281    ROUND_P, // towards +inf
 282    ROUND_NI, // nearest integer
 283    ROUND_MI, // to integer towards -inf
 284    ROUND_ZI, // to integer towards 0
 285    ROUND_PI, // to integer towards +inf
 286 };
 287
 288 enum CacheMode
 289 {
 290    CACHE_CA,            // cache at all levels
 291    CACHE_WB = CACHE_CA, // cache write back
 292    CACHE_CG,            // cache at global level
 293    CACHE_CS,            // cache streaming
 294    CACHE_CV,            // cache as volatile
 295    CACHE_WT = CACHE_CV  // cache write-through
 296 };
 297
 298 enum DataFile
 299 {
 300    FILE_NULL = 0,
 301    FILE_GPR,
 302    FILE_PREDICATE,       // boolean predicate
 303    FILE_FLAGS,           // zero/sign/carry/overflow bits
 304    FILE_ADDRESS,
 305    LAST_REGISTER_FILE = FILE_ADDRESS,
 306    FILE_IMMEDIATE,
 307    FILE_MEMORY_CONST,
 308    FILE_SHADER_INPUT,
 309    FILE_SHADER_OUTPUT,
 310    FILE_MEMORY_GLOBAL,
 311    FILE_MEMORY_SHARED,
 312    FILE_MEMORY_LOCAL,
 313    FILE_SYSTEM_VALUE,
 314    DATA_FILE_COUNT
 315 };
 316
 317 enum TexTarget
 318 {
 319    TEX_TARGET_1D,
 320    TEX_TARGET_2D,
 321    TEX_TARGET_2D_MS,
 322    TEX_TARGET_3D,
 323    TEX_TARGET_CUBE,
 324    TEX_TARGET_1D_SHADOW,
 325    TEX_TARGET_2D_SHADOW,
 326    TEX_TARGET_CUBE_SHADOW,
 327    TEX_TARGET_1D_ARRAY,
 328    TEX_TARGET_2D_ARRAY,
 329    TEX_TARGET_2D_MS_ARRAY,
 330    TEX_TARGET_CUBE_ARRAY,
 331    TEX_TARGET_1D_ARRAY_SHADOW,
 332    TEX_TARGET_2D_ARRAY_SHADOW,
 333    TEX_TARGET_RECT,
 334    TEX_TARGET_RECT_SHADOW,
 335    TEX_TARGET_CUBE_ARRAY_SHADOW,
 336    TEX_TARGET_BUFFER,
 337    TEX_TARGET_COUNT
 338 };
 339
 340 enum SVSemantic
 341 {
 342    SV_POSITION, // WPOS
 343    SV_VERTEX_ID,
 344    SV_INSTANCE_ID,
 345    SV_INVOCATION_ID,
 346    SV_PRIMITIVE_ID,
 347    SV_VERTEX_COUNT, // gl_PatchVerticesIn
 348    SV_LAYER,
 349    SV_VIEWPORT_INDEX,
 350    SV_YDIR,
 351    SV_FACE,
 352    SV_POINT_SIZE,
 353    SV_POINT_COORD,
 354    SV_CLIP_DISTANCE,
 355    SV_SAMPLE_INDEX,
 356    SV_TESS_FACTOR,
 357    SV_TESS_COORD,
 358    SV_TID,
 359    SV_CTAID,
 360    SV_NTID,
 361    SV_GRIDID,
 362    SV_NCTAID,
 363    SV_LANEID,
 364    SV_PHYSID,
 365    SV_NPHYSID,
 366    SV_CLOCK,
 367    SV_LBASE,
 368    SV_SBASE,
 369    SV_VERTEX_STRIDE,
 370    SV_UNDEFINED,
 371    SV_LAST
 372 };
 373
 374 class Program;
 375 class Function;
 376 class BasicBlock;
 377
 378 class Target;
 379
 380 class Instruction;
 381 class CmpInstruction;
 382 class TexInstruction;
 383 class FlowInstruction;
 384
 385 class Value;
 386 class LValue;
 387 class Symbol;
 388 class ImmediateValue;
 389
 390 struct Storage
 391 {
 392    DataFile file;
 393    int8_t fileIndex; // signed, may be indirect for CONST[]
 394    uint8_t size; // this should match the Instruction type's size
 395    DataType type; // mainly for pretty printing
 396    union {
 397       uint64_t u64;    // immediate values
 398       uint32_t u32;
 399       uint16_t u16;
 400       uint8_t u8;
 401       int64_t s64;
 402       int32_t s32;
 403       int16_t s16;
 404       int8_t s8;
 405       float f32;
 406       double f64;
 407       int32_t offset; // offset from 0 (base of address space)
 408       int32_t id;     // register id (< 0 if virtual/unassigned, in units <= 4)
 409       struct {
 410          SVSemantic sv;
 411          int index;
 412       } sv;
 413    } data;
 414 };
 415
 416 // precedence: NOT after SAT after NEG after ABS
 417 #define NV50_IR_MOD_ABS (1 << 0)
 418 #define NV50_IR_MOD_NEG (1 << 1)
 419 #define NV50_IR_MOD_SAT (1 << 2)
 420 #define NV50_IR_MOD_NOT (1 << 3)
 421 #define NV50_IR_MOD_NEG_ABS (NV50_IR_MOD_NEG | NV50_IR_MOD_ABS)
 422
 423 #define NV50_IR_INTERP_MODE_MASK   0x3
 424 #define NV50_IR_INTERP_LINEAR      (0 << 0)
 425 #define NV50_IR_INTERP_PERSPECTIVE (1 << 0)
 426 #define NV50_IR_INTERP_FLAT        (2 << 0)
 427 #define NV50_IR_INTERP_SC          (3 << 0) // what exactly is that ?
 428 #define NV50_IR_INTERP_SAMPLE_MASK 0xc
 429 #define NV50_IR_INTERP_DEFAULT     (0 << 2)
 430 #define NV50_IR_INTERP_CENTROID    (1 << 2)
 431 #define NV50_IR_INTERP_OFFSET      (2 << 2)
 432 #define NV50_IR_INTERP_SAMPLEID    (3 << 2)
 433
 434 // do we really want this to be a class ?
 435 class Modifier
 436 {
 437 public:
 438    Modifier() : bits(0) { }
 439    Modifier(unsigned int m) : bits(m) { }
 440    Modifier(operation op);
 441
 442    // @return new Modifier applying a after b (asserts if unrepresentable)
 443    Modifier operator*(const Modifier) const;
 444    Modifier operator*=(const Modifier m) { *this = *this * m; return *this; }
 445    Modifier operator==(const Modifier m) const { return m.bits == bits; }
 446    Modifier operator!=(const Modifier m) const { return m.bits != bits; }
 447
 448    inline Modifier operator&(const Modifier m) const { return bits & m.bits; }
 449    inline Modifier operator|(const Modifier m) const { return bits | m.bits; }
 450    inline Modifier operator^(const Modifier m) const { return bits ^ m.bits; }
 451
 452    operation getOp() const;
 453
 454    inline int neg() const { return (bits & NV50_IR_MOD_NEG) ? 1 : 0; }
 455    inline int abs() const { return (bits & NV50_IR_MOD_ABS) ? 1 : 0; }
 456
 457    inline operator bool() const { return bits ? true : false; }
 458
 459    void applyTo(ImmediateValue &imm) const;
 460
 461    int print(char *buf, size_t size) const;
 462
 463 private:
 464    uint8_t bits;
 465 };
 466
 467 class ValueRef
 468 {
 469 public:
 470    ValueRef(Value * = NULL);
 471    ValueRef(const ValueRef&);
 472    ~ValueRef();
 473
 474    inline bool exists() const { return value != NULL; }
 475
 476    void set(Value *);
 477    void set(const ValueRef&);
 478    inline Value *get() const { return value; }
 479    inline Value *rep() const;
 480
 481    inline Instruction *getInsn() const { return insn; }
 482    inline void setInsn(Instruction *inst) { insn = inst; }
 483
 484    inline bool isIndirect(int dim) const { return indirect[dim] >= 0; }
 485    inline const ValueRef *getIndirect(int dim) const;
 486
 487    inline DataFile getFile() const;
 488    inline unsigned getSize() const;
 489
 490    // SSA: return eventual (traverse MOVs) literal value, if it exists
 491    bool getImmediate(ImmediateValue&) const;
 492
 493 public:
 494    Modifier mod;
 495    int8_t indirect[2]; // >= 0 if relative to lvalue in insn->src(indirect[i])
 496    uint8_t swizzle;
 497
 498    bool usedAsPtr; // for printing
 499
 500 private:
 501    Value *value;
 502    Instruction *insn;
 503 };
 504
 505 class ValueDef
 506 {
 507 public:
 508    ValueDef(Value * = NULL);
 509    ValueDef(const ValueDef&);
 510    ~ValueDef();
 511
 512    inline bool exists() const { return value != NULL; }
 513
 514    inline Value *get() const { return value; }
 515    inline Value *rep() const;
 516    void set(Value *);
 517    bool mayReplace(const ValueRef &);
 518    void replace(const ValueRef &, bool doSet); // replace all uses of the old value
 519
 520    inline Instruction *getInsn() const { return insn; }
 521    inline void setInsn(Instruction *inst) { insn = inst; }
 522
 523    inline DataFile getFile() const;
 524    inline unsigned getSize() const;
 525
 526    inline void setSSA(LValue *);
 527    inline const LValue *preSSA() const;
 528
 529 private:
 530    Value *value;   // should make this LValue * ...
 531    LValue *origin; // pre SSA value
 532    Instruction *insn;
 533 };
 534
 535 class Value
 536 {
 537 public:
 538    Value();
 539    virtual ~Value() { }
 540
 541    virtual Value *clone(ClonePolicy<Function>&) const = 0;
 542
 543    virtual int print(char *, size_t, DataType ty = TYPE_NONE) const = 0;
 544
 545    virtual bool equals(const Value *, bool strict = false) const;
 546    virtual bool interfers(const Value *) const;
 547    virtual bool isUniform() const { return true; }
 548
 549    inline Value *rep() const { return join; }
 550
 551    inline Instruction *getUniqueInsn() const;
 552    inline Instruction *getInsn() const; // use when uniqueness is certain
 553
 554    inline int refCount() { return uses.size(); }
 555
 556    inline LValue *asLValue();
 557    inline Symbol *asSym();
 558    inline ImmediateValue *asImm();
 559    inline const Symbol *asSym() const;
 560    inline const ImmediateValue *asImm() const;
 561
 562    inline bool inFile(DataFile f) { return reg.file == f; }
 563
 564    static inline Value *get(Iterator&);
 565
 566    std::list<ValueRef *> uses;
 567    std::list<ValueDef *> defs;
 568    typedef std::list<ValueRef *>::iterator UseIterator;
 569    typedef std::list<ValueRef *>::const_iterator UseCIterator;
 570    typedef std::list<ValueDef *>::iterator DefIterator;
 571    typedef std::list<ValueDef *>::const_iterator DefCIterator;
 572
 573    int id;
 574    Storage reg;
 575
 576    // TODO: these should be in LValue:
 577    Interval livei;
 578    Value *join;
 579 };
 580
 581 class LValue : public Value
 582 {
 583 public:
 584    LValue(Function *, DataFile file);
 585    LValue(Function *, LValue *);
 586    ~LValue() { }
 587
 588    virtual bool isUniform() const;
 589
 590    virtual LValue *clone(ClonePolicy<Function>&) const;
 591
 592    virtual int print(char *, size_t, DataType ty = TYPE_NONE) const;
 593
 594 public:
 595    unsigned compMask : 8; // compound/component mask
 596    unsigned compound : 1; // used by RA, value involved in split/merge
 597    unsigned ssa      : 1;
 598    unsigned fixedReg : 1; // set & used by RA, earlier just use (id < 0)
 599    unsigned noSpill  : 1; // do not spill (e.g. if spill temporary already)
 600 };
 601
 602 class Symbol : public Value
 603 {
 604 public:
 605    Symbol(Program *, DataFile file = FILE_MEMORY_CONST, ubyte fileIdx = 0);
 606    ~Symbol() { }
 607
 608    virtual Symbol *clone(ClonePolicy<Function>&) const;
 609
 610    virtual bool equals(const Value *that, bool strict) const;
 611
 612    virtual bool isUniform() const;
 613
 614    virtual int print(char *, size_t, DataType ty = TYPE_NONE) const;
 615
 616    // print with indirect values
 617    int print(char *, size_t, Value *, Value *, DataType ty = TYPE_NONE) const;
 618
 619    inline void setFile(DataFile file, ubyte fileIndex = 0)
 620    {
 621       reg.file = file;
 622       reg.fileIndex = fileIndex;
 623    }
 624
 625    inline void setOffset(int32_t offset);
 626    inline void setAddress(Symbol *base, int32_t offset);
 627    inline void setSV(SVSemantic sv, uint32_t idx = 0);
 628
 629    inline const Symbol *getBase() const { return baseSym; }
 630
 631 private:
 632    Symbol *baseSym; // array base for Symbols representing array elements
 633 };
 634
 635 class ImmediateValue : public Value
 636 {
 637 public:
 638    ImmediateValue() { }
 639    ImmediateValue(Program *, uint32_t);
 640    ImmediateValue(Program *, float);
 641    ImmediateValue(Program *, double);
 642    // NOTE: not added to program with
 643    ImmediateValue(const ImmediateValue *, DataType ty);
 644    ~ImmediateValue() { };
 645
 646    virtual ImmediateValue *clone(ClonePolicy<Function>&) const;
 647
 648    virtual bool equals(const Value *that, bool strict) const;
 649
 650    // these only work if 'type' is valid (we mostly use untyped literals):
 651    bool isInteger(const int ival) const; // ival is cast to this' type
 652    bool isNegative() const;
 653    bool isPow2() const;
 654
 655    void applyLog2();
 656
 657    // for constant folding:
 658    ImmediateValue operator+(const ImmediateValue&) const;
 659    ImmediateValue operator-(const ImmediateValue&) const;
 660    ImmediateValue operator*(const ImmediateValue&) const;
 661    ImmediateValue operator/(const ImmediateValue&) const;
 662
 663    ImmediateValue& operator=(const ImmediateValue&); // only sets value !
 664
 665    bool compare(CondCode cc, float fval) const;
 666
 667    virtual int print(char *, size_t, DataType ty = TYPE_NONE) const;
 668 };
 669
 670 class Instruction
 671 {
 672 public:
 673    Instruction();
 674    Instruction(Function *, operation, DataType);
 675    virtual ~Instruction();
 676
 677    virtual Instruction *clone(ClonePolicy<Function>&,
 678                               Instruction * = NULL) const;
 679
 680    void setDef(int i, Value *);
 681    void setSrc(int s, Value *);
 682    void setSrc(int s, const ValueRef&);
 683    void swapSources(int a, int b);
 684    void moveSources(int s, int delta);
 685    bool setIndirect(int s, int dim, Value *);
 686
 687    inline ValueRef& src(int s) { return srcs[s]; }
 688    inline ValueDef& def(int s) { return defs[s]; }
 689    inline const ValueRef& src(int s) const { return srcs[s]; }
 690    inline const ValueDef& def(int s) const { return defs[s]; }
 691
 692    inline Value *getDef(int d) const { return defs[d].get(); }
 693    inline Value *getSrc(int s) const { return srcs[s].get(); }
 694    inline Value *getIndirect(int s, int dim) const;
 695
 696    inline bool defExists(unsigned d) const
 697    {
 698       return d < defs.size() && defs[d].exists();
 699    }
 700    inline bool srcExists(unsigned s) const
 701    {
 702       return s < srcs.size() && srcs[s].exists();
 703    }
 704
 705    inline bool constrainedDefs() const;
 706
 707    bool setPredicate(CondCode ccode, Value *);
 708    inline Value *getPredicate() const;
 709    bool writesPredicate() const;
 710    inline bool isPredicated() const { return predSrc >= 0; }
 711
 712    inline void setFlagsSrc(int s, Value *);
 713    inline void setFlagsDef(int d, Value *);
 714    inline bool usesFlags() const { return flagsSrc >= 0; }
 715
 716    unsigned int defCount() const { return defs.size(); };
 717    unsigned int defCount(unsigned int mask, bool singleFile = false) const;
 718    unsigned int srcCount() const { return srcs.size(); };
 719    unsigned int srcCount(unsigned int mask, bool singleFile = false) const;
 720
 721    // save & remove / set indirect[0,1] and predicate source
 722    void takeExtraSources(int s, Value *[3]);
 723    void putExtraSources(int s, Value *[3]);
 724
 725    inline void setType(DataType type) { dType = sType = type; }
 726
 727    inline void setType(DataType dtype, DataType stype)
 728    {
 729       dType = dtype;
 730       sType = stype;
 731    }
 732
 733    inline bool isPseudo() const { return op < OP_MOV; }
 734    bool isDead() const;
 735    bool isNop() const;
 736    bool isCommutationLegal(const Instruction *) const; // must be adjacent !
 737    bool isActionEqual(const Instruction *) const;
 738    bool isResultEqual(const Instruction *) const;
 739
 740    void print() const;
 741
 742    inline CmpInstruction *asCmp();
 743    inline TexInstruction *asTex();
 744    inline FlowInstruction *asFlow();
 745    inline const TexInstruction *asTex() const;
 746    inline const CmpInstruction *asCmp() const;
 747    inline const FlowInstruction *asFlow() const;
 748
 749 public:
 750    Instruction *next;
 751    Instruction *prev;
 752    int id;
 753    int serial; // CFG order
 754
 755    operation op;
 756    DataType dType; // destination or defining type
 757    DataType sType; // source or secondary type
 758    CondCode cc;
 759    RoundMode rnd;
 760    CacheMode cache;
 761
 762    uint16_t subOp; // quadop, 1 for mul-high, etc.
 763
 764    unsigned encSize    : 4; // encoding size in bytes
 765    unsigned saturate   : 1; // to [0.0f, 1.0f]
 766    unsigned join       : 1; // converge control flow (use OP_JOIN until end)
 767    unsigned fixed      : 1; // prevent dead code elimination
 768    unsigned terminator : 1; // end of basic block
 769    unsigned ftz        : 1; // flush denormal to zero
 770    unsigned dnz        : 1; // denormals, NaN are zero
 771    unsigned ipa        : 4; // interpolation mode
 772    unsigned lanes      : 4;
 773    unsigned perPatch   : 1;
 774    unsigned exit       : 1; // terminate program after insn
 775    unsigned mask       : 4; // for vector ops
 776
 777    int8_t postFactor; // MUL/DIV(if < 0) by 1 << postFactor
 778
 779    int8_t predSrc;
 780    int8_t flagsDef;
 781    int8_t flagsSrc;
 782
 783    uint8_t sched; // scheduling data (NOTE: maybe move to separate storage)
 784
 785    BasicBlock *bb;
 786
 787 protected:
 788    std::deque<ValueDef> defs; // no gaps !
 789    std::deque<ValueRef> srcs; // no gaps !
 790
 791    // instruction specific methods:
 792    // (don't want to subclass, would need more constructors and memory pools)
 793 public:
 794    inline void setInterpolate(unsigned int mode) { ipa = mode; }
 795
 796    unsigned int getInterpMode() const { return ipa & 0x3; }
 797    unsigned int getSampleMode() const { return ipa & 0xc; }
 798
 799 private:
 800    void init();
 801 };
 802
 803 enum TexQuery
 804 {
 805    TXQ_DIMS,
 806    TXQ_TYPE,
 807    TXQ_SAMPLE_POSITION,
 808    TXQ_FILTER,
 809    TXQ_LOD,
 810    TXQ_WRAP,
 811    TXQ_BORDER_COLOUR
 812 };
 813
 814 class TexInstruction : public Instruction
 815 {
 816 public:
 817    class Target
 818    {
 819    public:
 820       Target(TexTarget targ = TEX_TARGET_2D) : target(targ) { }
 821
 822       const char *getName() const { return descTable[target].name; }
 823       unsigned int getArgCount() const { return descTable[target].argc; }
 824       unsigned int getDim() const { return descTable[target].dim; }
 825       int isArray() const { return descTable[target].array ? 1 : 0; }
 826       int isCube() const { return descTable[target].cube ? 1 : 0; }
 827       int isShadow() const { return descTable[target].shadow ? 1 : 0; }
 828       int isMS() const {
 829         return target == TEX_TARGET_2D_MS || target == TEX_TARGET_2D_MS_ARRAY; }
 830       void clearMS() {
 831          if (isMS()) {
 832             if (isArray())
 833                target = TEX_TARGET_2D_ARRAY;
 834             else
 835                target = TEX_TARGET_2D;
 836          }
 837       }
 838
 839       Target& operator=(TexTarget targ)
 840       {
 841          assert(targ < TEX_TARGET_COUNT);
 842          target = targ;
 843          return *this;
 844       }
 845
 846       inline bool operator==(TexTarget targ) const { return target == targ; }
 847       inline bool operator!=(TexTarget targ) const { return target != targ; }
 848
 849       enum TexTarget getEnum() const { return target; }
 850
 851    private:
 852       struct Desc
 853       {
 854          char name[19];
 855          uint8_t dim;
 856          uint8_t argc;
 857          bool array;
 858          bool cube;
 859          bool shadow;
 860       };
 861
 862       static const struct Desc descTable[TEX_TARGET_COUNT];
 863
 864    private:
 865       enum TexTarget target;
 866    };
 867
 868 public:
 869    TexInstruction(Function *, operation);
 870    virtual ~TexInstruction();
 871
 872    virtual TexInstruction *clone(ClonePolicy<Function>&,
 873                                  Instruction * = NULL) const;
 874
 875    inline void setTexture(Target targ, uint8_t r, uint8_t s)
 876    {
 877       tex.r = r;
 878       tex.s = s;
 879       tex.target = targ;
 880    }
 881
 882    void setIndirectR(Value *);
 883    void setIndirectS(Value *);
 884    inline Value *getIndirectR() const;
 885    inline Value *getIndirectS() const;
 886
 887 public:
 888    struct {
 889       Target target;
 890
 891       uint16_t r;
 892       uint16_t s;
 893       int8_t rIndirectSrc;
 894       int8_t sIndirectSrc;
 895
 896       uint8_t mask;
 897       uint8_t gatherComp;
 898
 899       bool liveOnly; // only execute on live pixels of a quad (optimization)
 900       bool levelZero;
 901       bool derivAll;
 902
 903       int8_t useOffsets; // 0, 1, or 4 for textureGatherOffsets
 904       int8_t offset[4][3];
 905
 906       enum TexQuery query;
 907    } tex;
 908
 909    ValueRef dPdx[3];
 910    ValueRef dPdy[3];
 911 };
 912
 913 class CmpInstruction : public Instruction
 914 {
 915 public:
 916    CmpInstruction(Function *, operation);
 917
 918    virtual CmpInstruction *clone(ClonePolicy<Function>&,
 919                                  Instruction * = NULL) const;
 920
 921    void setCondition(CondCode cond) { setCond = cond; }
 922    CondCode getCondition() const { return setCond; }
 923
 924 public:
 925    CondCode setCond;
 926 };
 927
 928 class FlowInstruction : public Instruction
 929 {
 930 public:
 931    FlowInstruction(Function *, operation, void *target);
 932
 933    virtual FlowInstruction *clone(ClonePolicy<Function>&,
 934                                   Instruction * = NULL) const;
 935
 936 public:
 937    unsigned allWarp  : 1;
 938    unsigned absolute : 1;
 939    unsigned limit    : 1;
 940    unsigned builtin  : 1; // true for calls to emulation code
 941    unsigned indirect : 1; // target in src(0)
 942
 943    union {
 944       BasicBlock *bb;
 945       int builtin;
 946       Function *fn;
 947    } target;
 948 };
 949
 950 class BasicBlock
 951 {
 952 public:
 953    BasicBlock(Function *);
 954    ~BasicBlock();
 955
 956    BasicBlock *clone(ClonePolicy<Function>&) const;
 957
 958    inline int getId() const { return id; }
 959    inline unsigned int getInsnCount() const { return numInsns; }
 960    inline bool isTerminated() const { return exit && exit->terminator; }
 961
 962    bool dominatedBy(BasicBlock *bb);
 963    inline bool reachableBy(const BasicBlock *by, const BasicBlock *term);
 964
 965    // returns mask of conditional out blocks
 966    // e.g. 3 for IF { .. } ELSE { .. } ENDIF, 1 for IF { .. } ENDIF
 967    unsigned int initiatesSimpleConditional() const;
 968
 969 public:
 970    Function *getFunction() const { return func; }
 971    Program *getProgram() const { return program; }
 972
 973    Instruction *getEntry() const { return entry; } // first non-phi instruction
 974    Instruction *getPhi() const { return phi; }
 975    Instruction *getFirst() const { return phi ? phi : entry; }
 976    Instruction *getExit() const { return exit; }
 977
 978    void insertHead(Instruction *);
 979    void insertTail(Instruction *);
 980    void insertBefore(Instruction *, Instruction *);
 981    void insertAfter(Instruction *, Instruction *);
 982    void remove(Instruction *);
 983    void permuteAdjacent(Instruction *, Instruction *);
 984
 985    BasicBlock *idom() const;
 986
 987    // NOTE: currently does not rebuild the dominator tree
 988    BasicBlock *splitBefore(Instruction *, bool attach = true);
 989    BasicBlock *splitAfter(Instruction *, bool attach = true);
 990
 991    DLList& getDF() { return df; }
 992    DLList::Iterator iterDF() { return df.iterator(); }
 993
 994    static inline BasicBlock *get(Iterator&);
 995    static inline BasicBlock *get(Graph::Node *);
 996
 997 public:
 998    Graph::Node cfg; // first edge is branch *taken* (the ELSE branch)
 999    Graph::Node dom;
1000
1001    BitSet liveSet;
1002    BitSet defSet;
1003
1004    uint32_t binPos;
1005    uint32_t binSize;
1006
1007    Instruction *joinAt; // for quick reference
1008
1009    bool explicitCont; // loop headers: true if loop contains continue stmts
1010
1011 private:
1012    int id;
1013    DLList df;
1014
1015    Instruction *phi;
1016    Instruction *entry;
1017    Instruction *exit;
1018
1019    unsigned int numInsns;
1020
1021 private:
1022    Function *func;
1023    Program *program;
1024
1025    void splitCommon(Instruction *, BasicBlock *, bool attach);
1026 };
1027
1028 class Function
1029 {
1030 public:
1031    Function(Program *, const char *name, uint32_t label);
1032    ~Function();
1033
1034    static inline Function *get(Graph::Node *node);
1035
1036    inline Program *getProgram() const { return prog; }
1037    inline const char *getName() const { return name; }
1038    inline int getId() const { return id; }
1039    inline uint32_t getLabel() const { return label; }
1040
1041    void print();
1042    void printLiveIntervals() const;
1043    void printCFGraph(const char *filePath);
1044
1045    bool setEntry(BasicBlock *);
1046    bool setExit(BasicBlock *);
1047
1048    unsigned int orderInstructions(ArrayList&);
1049
1050    inline void add(BasicBlock *bb, int& id) { allBBlocks.insert(bb, id); }
1051    inline void add(Instruction *insn, int& id) { allInsns.insert(insn, id); }
1052    inline void add(LValue *lval, int& id) { allLValues.insert(lval, id); }
1053
1054    inline LValue *getLValue(int id);
1055
1056    void buildLiveSets();
1057    void buildDefSets();
1058    bool convertToSSA();
1059
1060 public:
1061    std::deque<ValueDef> ins;
1062    std::deque<ValueRef> outs;
1063    std::deque<Value *> clobbers;
1064
1065    Graph cfg;
1066    Graph::Node *cfgExit;
1067    Graph *domTree;
1068    Graph::Node call; // node in the call graph
1069
1070    BasicBlock **bbArray; // BBs in emission order
1071    int bbCount;
1072
1073    unsigned int loopNestingBound;
1074    int regClobberMax;
1075
1076    uint32_t binPos;
1077    uint32_t binSize;
1078
1079    Value *stackPtr;
1080
1081    uint32_t tlsBase; // base address for l[] space (if no stack pointer is used)
1082    uint32_t tlsSize;
1083
1084    ArrayList allBBlocks;
1085    ArrayList allInsns;
1086    ArrayList allLValues;
1087
1088 private:
1089    void buildLiveSetsPreSSA(BasicBlock *, const int sequence);
1090    void buildDefSetsPreSSA(BasicBlock *bb, const int seq);
1091
1092 private:
1093    uint32_t label;
1094    int id;
1095    const char *const name;
1096    Program *prog;
1097 };
1098
1099 enum CGStage
1100 {
1101    CG_STAGE_PRE_SSA,
1102    CG_STAGE_SSA, // expected directly before register allocation
1103    CG_STAGE_POST_RA
1104 };
1105
1106 class Program
1107 {
1108 public:
1109    enum Type
1110    {
1111       TYPE_VERTEX,
1112       TYPE_TESSELLATION_CONTROL,
1113       TYPE_TESSELLATION_EVAL,
1114       TYPE_GEOMETRY,
1115       TYPE_FRAGMENT,
1116       TYPE_COMPUTE
1117    };
1118
1119    Program(Type type, Target *targ);
1120    ~Program();
1121
1122    void print();
1123
1124    Type getType() const { return progType; }
1125
1126    inline void add(Function *fn, int& id) { allFuncs.insert(fn, id); }
1127    inline void del(Function *fn, int& id) { allFuncs.remove(id); }
1128    inline void add(Value *rval, int& id) { allRValues.insert(rval, id); }
1129
1130    bool makeFromTGSI(struct nv50_ir_prog_info *);
1131    bool makeFromSM4(struct nv50_ir_prog_info *);
1132    bool convertToSSA();
1133    bool optimizeSSA(int level);
1134    bool optimizePostRA(int level);
1135    bool registerAllocation();
1136    bool emitBinary(struct nv50_ir_prog_info *);
1137
1138    const Target *getTarget() const { return target; }
1139
1140 private:
1141    void emitSymbolTable(struct nv50_ir_prog_info *);
1142
1143    Type progType;
1144    Target *target;
1145
1146 public:
1147    Function *main;
1148    Graph calls;
1149
1150    ArrayList allFuncs;
1151    ArrayList allRValues;
1152
1153    uint32_t *code;
1154    uint32_t binSize;
1155    uint32_t tlsSize; // size required for FILE_MEMORY_LOCAL
1156
1157    int maxGPR;
1158
1159    MemoryPool mem_Instruction;
1160    MemoryPool mem_CmpInstruction;
1161    MemoryPool mem_TexInstruction;
1162    MemoryPool mem_FlowInstruction;
1163    MemoryPool mem_LValue;
1164    MemoryPool mem_Symbol;
1165    MemoryPool mem_ImmediateValue;
1166
1167    uint32_t dbgFlags;
1168    uint8_t  optLevel;
1169
1170    void *targetPriv; // e.g. to carry information between passes
1171
1172    const struct nv50_ir_prog_info *driver; // for driver configuration
1173
1174    void releaseInstruction(Instruction *);
1175    void releaseValue(Value *);
1176 };
1177
1178 // TODO: add const version
1179 class Pass
1180 {
1181 public:
1182    bool run(Program *, bool ordered = false, bool skipPhi = false);
1183    bool run(Function *, bool ordered = false, bool skipPhi = false);
1184
1185 private:
1186    // return false to continue with next entity on next higher level
1187    virtual bool visit(Function *) { return true; }
1188    virtual bool visit(BasicBlock *) { return true; }
1189    virtual bool visit(Instruction *) { return false; }
1190
1191    bool doRun(Program *, bool ordered, bool skipPhi);
1192    bool doRun(Function *, bool ordered, bool skipPhi);
1193
1194 protected:
1195    bool err;
1196    Function *func;
1197    Program *prog;
1198 };
1199
1200 // =============================================================================
1201
1202 #include "codegen/nv50_ir_inlines.h"
1203
1204 } // namespace nv50_ir
1205
1206 #endif // __NV50_IR_H__