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_TXLQ, // texture query lod
 121    OP_TEXCSAA, // texture op for coverage sampling
 122    OP_TEXPREP, // turn cube map array into 2d array coordinates
 123    OP_SULDB, // surface load (raw)
 124    OP_SULDP, // surface load (formatted)
 125    OP_SUSTB, // surface store (raw)
 126    OP_SUSTP, // surface store (formatted)
 127    OP_SUREDB,
 128    OP_SUREDP, // surface reduction (atomic op)
 129    OP_SULEA,   // surface load effective address
 130    OP_SUBFM,   // surface bitfield manipulation
 131    OP_SUCLAMP, // clamp surface coordinates
 132    OP_SUEAU,   // surface effective address
 133    OP_MADSP,   // special integer multiply-add
 134    OP_TEXBAR, // texture dependency barrier
 135    OP_DFDX,
 136    OP_DFDY,
 137    OP_RDSV, // read system value
 138    OP_WRSV, // write system value
 139    OP_QUADOP,
 140    OP_QUADON,
 141    OP_QUADPOP,
 142    OP_POPCNT, // bitcount(src0 & src1)
 143    OP_INSBF,  // insert first src1[8:15] bits of src0 into src2 at src1[0:7]
 144    OP_EXTBF,  // place bits [K,K+N) of src0 into dst, src1 = 0xNNKK
 145    OP_PERMT,  // dst = bytes from src2,src0 selected by src1 (nvc0's src order)
 146    OP_ATOM,
 147    OP_BAR,    // execution barrier, sources = { id, thread count, predicate }
 148    OP_VADD,   // byte/word vector operations
 149    OP_VAVG,
 150    OP_VMIN,
 151    OP_VMAX,
 152    OP_VSAD,
 153    OP_VSET,
 154    OP_VSHR,
 155    OP_VSHL,
 156    OP_VSEL,
 157    OP_CCTL, // cache control
 158    OP_LAST
 159 };
 160
 161 // various instruction-specific modifier definitions Instruction::subOp
 162 // MOV_FINAL marks a MOV originating from an EXPORT (used for placing TEXBARs)
 163 #define NV50_IR_SUBOP_MUL_HIGH     1
 164 #define NV50_IR_SUBOP_EMIT_RESTART 1
 165 #define NV50_IR_SUBOP_LDC_IL       1
 166 #define NV50_IR_SUBOP_LDC_IS       2
 167 #define NV50_IR_SUBOP_LDC_ISL      3
 168 #define NV50_IR_SUBOP_SHIFT_WRAP   1
 169 #define NV50_IR_SUBOP_EMU_PRERET   1
 170 #define NV50_IR_SUBOP_TEXBAR(n)    n
 171 #define NV50_IR_SUBOP_MOV_FINAL    1
 172 #define NV50_IR_SUBOP_EXTBF_REV    1
 173 #define NV50_IR_SUBOP_PERMT_F4E    1
 174 #define NV50_IR_SUBOP_PERMT_B4E    2
 175 #define NV50_IR_SUBOP_PERMT_RC8    3
 176 #define NV50_IR_SUBOP_PERMT_ECL    4
 177 #define NV50_IR_SUBOP_PERMT_ECR    5
 178 #define NV50_IR_SUBOP_PERMT_RC16   6
 179 #define NV50_IR_SUBOP_BAR_SYNC     0
 180 #define NV50_IR_SUBOP_BAR_ARRIVE   1
 181 #define NV50_IR_SUBOP_BAR_RED_AND  2
 182 #define NV50_IR_SUBOP_BAR_RED_OR   3
 183 #define NV50_IR_SUBOP_BAR_RED_POPC 4
 184 #define NV50_IR_SUBOP_MEMBAR_L     1
 185 #define NV50_IR_SUBOP_MEMBAR_S     2
 186 #define NV50_IR_SUBOP_MEMBAR_M     3
 187 #define NV50_IR_SUBOP_MEMBAR_CTA  (0 << 2)
 188 #define NV50_IR_SUBOP_MEMBAR_GL   (1 << 2)
 189 #define NV50_IR_SUBOP_MEMBAR_SYS  (2 << 2)
 190 #define NV50_IR_SUBOP_MEMBAR_DIR(m)   ((m) & 0x3)
 191 #define NV50_IR_SUBOP_MEMBAR_SCOPE(m) ((m) & ~0x3)
 192 #define NV50_IR_SUBOP_MEMBAR(d,s) \
 193    (NV50_IR_SUBOP_MEMBAR_##d | NV50_IR_SUBOP_MEMBAR_##s)
 194 #define NV50_IR_SUBOP_ATOM_ADD      0
 195 #define NV50_IR_SUBOP_ATOM_MIN      1
 196 #define NV50_IR_SUBOP_ATOM_MAX      2
 197 #define NV50_IR_SUBOP_ATOM_INC      3
 198 #define NV50_IR_SUBOP_ATOM_DEC      4
 199 #define NV50_IR_SUBOP_ATOM_AND      5
 200 #define NV50_IR_SUBOP_ATOM_OR       6
 201 #define NV50_IR_SUBOP_ATOM_XOR      7
 202 #define NV50_IR_SUBOP_ATOM_CAS      8
 203 #define NV50_IR_SUBOP_ATOM_EXCH     9
 204 #define NV50_IR_SUBOP_CCTL_IV      5
 205 #define NV50_IR_SUBOP_CCTL_IVALL   6
 206 #define NV50_IR_SUBOP_SUST_IGN     0
 207 #define NV50_IR_SUBOP_SUST_TRAP    1
 208 #define NV50_IR_SUBOP_SUST_SDCL    3
 209 #define NV50_IR_SUBOP_SULD_ZERO    0
 210 #define NV50_IR_SUBOP_SULD_TRAP    1
 211 #define NV50_IR_SUBOP_SULD_SDCL    3
 212 #define NV50_IR_SUBOP_SUBFM_3D     1
 213 #define NV50_IR_SUBOP_SUCLAMP_2D   0x10
 214 #define NV50_IR_SUBOP_SUCLAMP_SD(r, d) (( 0 + (r)) | ((d == 2) ? 0x10 : 0))
 215 #define NV50_IR_SUBOP_SUCLAMP_PL(r, d) (( 5 + (r)) | ((d == 2) ? 0x10 : 0))
 216 #define NV50_IR_SUBOP_SUCLAMP_BL(r, d) ((10 + (r)) | ((d == 2) ? 0x10 : 0))
 217 #define NV50_IR_SUBOP_MADSP_SD     0xffff
 218 // Yes, we could represent those with DataType.
 219 // Or put the type into operation and have a couple 1000 values in that enum.
 220 // This will have to do for now.
 221 // The bitfields are supposed to correspond to nve4 ISA.
 222 #define NV50_IR_SUBOP_MADSP(a,b,c) (((c) << 8) | ((b) << 4) | (a))
 223 #define NV50_IR_SUBOP_V1(d,a,b)    (((d) << 10) | ((b) << 5) | (a) | 0x0000)
 224 #define NV50_IR_SUBOP_V2(d,a,b)    (((d) << 10) | ((b) << 5) | (a) | 0x4000)
 225 #define NV50_IR_SUBOP_V4(d,a,b)    (((d) << 10) | ((b) << 5) | (a) | 0x8000)
 226 #define NV50_IR_SUBOP_Vn(n)        ((n) >> 14)
 227
 228 enum DataType
 229 {
 230    TYPE_NONE,
 231    TYPE_U8,
 232    TYPE_S8,
 233    TYPE_U16,
 234    TYPE_S16,
 235    TYPE_U32,
 236    TYPE_S32,
 237    TYPE_U64, // 64 bit operations are only lowered after register allocation
 238    TYPE_S64,
 239    TYPE_F16,
 240    TYPE_F32,
 241    TYPE_F64,
 242    TYPE_B96,
 243    TYPE_B128
 244 };
 245
 246 enum CondCode
 247 {
 248    CC_FL = 0,
 249    CC_NEVER = CC_FL, // when used with FILE_FLAGS
 250    CC_LT = 1,
 251    CC_EQ = 2,
 252    CC_NOT_P = CC_EQ, // when used with FILE_PREDICATE
 253    CC_LE = 3,
 254    CC_GT = 4,
 255    CC_NE = 5,
 256    CC_P  = CC_NE,
 257    CC_GE = 6,
 258    CC_TR = 7,
 259    CC_ALWAYS = CC_TR,
 260    CC_U  = 8,
 261    CC_LTU = 9,
 262    CC_EQU = 10,
 263    CC_LEU = 11,
 264    CC_GTU = 12,
 265    CC_NEU = 13,
 266    CC_GEU = 14,
 267    CC_NO = 0x10,
 268    CC_NC = 0x11,
 269    CC_NS = 0x12,
 270    CC_NA = 0x13,
 271    CC_A  = 0x14,
 272    CC_S  = 0x15,
 273    CC_C  = 0x16,
 274    CC_O  = 0x17
 275 };
 276
 277 enum RoundMode
 278 {
 279    ROUND_N, // nearest
 280    ROUND_M, // towards -inf
 281    ROUND_Z, // towards 0
 282    ROUND_P, // towards +inf
 283    ROUND_NI, // nearest integer
 284    ROUND_MI, // to integer towards -inf
 285    ROUND_ZI, // to integer towards 0
 286    ROUND_PI, // to integer towards +inf
 287 };
 288
 289 enum CacheMode
 290 {
 291    CACHE_CA,            // cache at all levels
 292    CACHE_WB = CACHE_CA, // cache write back
 293    CACHE_CG,            // cache at global level
 294    CACHE_CS,            // cache streaming
 295    CACHE_CV,            // cache as volatile
 296    CACHE_WT = CACHE_CV  // cache write-through
 297 };
 298
 299 enum DataFile
 300 {
 301    FILE_NULL = 0,
 302    FILE_GPR,
 303    FILE_PREDICATE,       // boolean predicate
 304    FILE_FLAGS,           // zero/sign/carry/overflow bits
 305    FILE_ADDRESS,
 306    LAST_REGISTER_FILE = FILE_ADDRESS,
 307    FILE_IMMEDIATE,
 308    FILE_MEMORY_CONST,
 309    FILE_SHADER_INPUT,
 310    FILE_SHADER_OUTPUT,
 311    FILE_MEMORY_GLOBAL,
 312    FILE_MEMORY_SHARED,
 313    FILE_MEMORY_LOCAL,
 314    FILE_SYSTEM_VALUE,
 315    DATA_FILE_COUNT
 316 };
 317
 318 enum TexTarget
 319 {
 320    TEX_TARGET_1D,
 321    TEX_TARGET_2D,
 322    TEX_TARGET_2D_MS,
 323    TEX_TARGET_3D,
 324    TEX_TARGET_CUBE,
 325    TEX_TARGET_1D_SHADOW,
 326    TEX_TARGET_2D_SHADOW,
 327    TEX_TARGET_CUBE_SHADOW,
 328    TEX_TARGET_1D_ARRAY,
 329    TEX_TARGET_2D_ARRAY,
 330    TEX_TARGET_2D_MS_ARRAY,
 331    TEX_TARGET_CUBE_ARRAY,
 332    TEX_TARGET_1D_ARRAY_SHADOW,
 333    TEX_TARGET_2D_ARRAY_SHADOW,
 334    TEX_TARGET_RECT,
 335    TEX_TARGET_RECT_SHADOW,
 336    TEX_TARGET_CUBE_ARRAY_SHADOW,
 337    TEX_TARGET_BUFFER,
 338    TEX_TARGET_COUNT
 339 };
 340
 341 enum SVSemantic
 342 {
 343    SV_POSITION, // WPOS
 344    SV_VERTEX_ID,
 345    SV_INSTANCE_ID,
 346    SV_INVOCATION_ID,
 347    SV_PRIMITIVE_ID,
 348    SV_VERTEX_COUNT, // gl_PatchVerticesIn
 349    SV_LAYER,
 350    SV_VIEWPORT_INDEX,
 351    SV_YDIR,
 352    SV_FACE,
 353    SV_POINT_SIZE,
 354    SV_POINT_COORD,
 355    SV_CLIP_DISTANCE,
 356    SV_SAMPLE_INDEX,
 357    SV_SAMPLE_POS,
 358    SV_TESS_FACTOR,
 359    SV_TESS_COORD,
 360    SV_TID,
 361    SV_CTAID,
 362    SV_NTID,
 363    SV_GRIDID,
 364    SV_NCTAID,
 365    SV_LANEID,
 366    SV_PHYSID,
 367    SV_NPHYSID,
 368    SV_CLOCK,
 369    SV_LBASE,
 370    SV_SBASE,
 371    SV_VERTEX_STRIDE,
 372    SV_UNDEFINED,
 373    SV_LAST
 374 };
 375
 376 class Program;
 377 class Function;
 378 class BasicBlock;
 379
 380 class Target;
 381
 382 class Instruction;
 383 class CmpInstruction;
 384 class TexInstruction;
 385 class FlowInstruction;
 386
 387 class Value;
 388 class LValue;
 389 class Symbol;
 390 class ImmediateValue;
 391
 392 struct Storage
 393 {
 394    DataFile file;
 395    int8_t fileIndex; // signed, may be indirect for CONST[]
 396    uint8_t size; // this should match the Instruction type's size
 397    DataType type; // mainly for pretty printing
 398    union {
 399       uint64_t u64;    // immediate values
 400       uint32_t u32;
 401       uint16_t u16;
 402       uint8_t u8;
 403       int64_t s64;
 404       int32_t s32;
 405       int16_t s16;
 406       int8_t s8;
 407       float f32;
 408       double f64;
 409       int32_t offset; // offset from 0 (base of address space)
 410       int32_t id;     // register id (< 0 if virtual/unassigned, in units <= 4)
 411       struct {
 412          SVSemantic sv;
 413          int index;
 414       } sv;
 415    } data;
 416 };
 417
 418 // precedence: NOT after SAT after NEG after ABS
 419 #define NV50_IR_MOD_ABS (1 << 0)
 420 #define NV50_IR_MOD_NEG (1 << 1)
 421 #define NV50_IR_MOD_SAT (1 << 2)
 422 #define NV50_IR_MOD_NOT (1 << 3)
 423 #define NV50_IR_MOD_NEG_ABS (NV50_IR_MOD_NEG | NV50_IR_MOD_ABS)
 424
 425 #define NV50_IR_INTERP_MODE_MASK   0x3
 426 #define NV50_IR_INTERP_LINEAR      (0 << 0)
 427 #define NV50_IR_INTERP_PERSPECTIVE (1 << 0)
 428 #define NV50_IR_INTERP_FLAT        (2 << 0)
 429 #define NV50_IR_INTERP_SC          (3 << 0) // what exactly is that ?
 430 #define NV50_IR_INTERP_SAMPLE_MASK 0xc
 431 #define NV50_IR_INTERP_DEFAULT     (0 << 2)
 432 #define NV50_IR_INTERP_CENTROID    (1 << 2)
 433 #define NV50_IR_INTERP_OFFSET      (2 << 2)
 434 #define NV50_IR_INTERP_SAMPLEID    (3 << 2)
 435
 436 // do we really want this to be a class ?
 437 class Modifier
 438 {
 439 public:
 440    Modifier() : bits(0) { }
 441    Modifier(unsigned int m) : bits(m) { }
 442    Modifier(operation op);
 443
 444    // @return new Modifier applying a after b (asserts if unrepresentable)
 445    Modifier operator*(const Modifier) const;
 446    Modifier operator*=(const Modifier m) { *this = *this * m; return *this; }
 447    Modifier operator==(const Modifier m) const { return m.bits == bits; }
 448    Modifier operator!=(const Modifier m) const { return m.bits != bits; }
 449
 450    inline Modifier operator&(const Modifier m) const { return bits & m.bits; }
 451    inline Modifier operator|(const Modifier m) const { return bits | m.bits; }
 452    inline Modifier operator^(const Modifier m) const { return bits ^ m.bits; }
 453
 454    operation getOp() const;
 455
 456    inline int neg() const { return (bits & NV50_IR_MOD_NEG) ? 1 : 0; }
 457    inline int abs() const { return (bits & NV50_IR_MOD_ABS) ? 1 : 0; }
 458
 459    inline operator bool() const { return bits ? true : false; }
 460
 461    void applyTo(ImmediateValue &imm) const;
 462
 463    int print(char *buf, size_t size) const;
 464
 465 private:
 466    uint8_t bits;
 467 };
 468
 469 class ValueRef
 470 {
 471 public:
 472    ValueRef(Value * = NULL);
 473    ValueRef(const ValueRef&);
 474    ~ValueRef();
 475
 476    inline bool exists() const { return value != NULL; }
 477
 478    void set(Value *);
 479    void set(const ValueRef&);
 480    inline Value *get() const { return value; }
 481    inline Value *rep() const;
 482
 483    inline Instruction *getInsn() const { return insn; }
 484    inline void setInsn(Instruction *inst) { insn = inst; }
 485
 486    inline bool isIndirect(int dim) const { return indirect[dim] >= 0; }
 487    inline const ValueRef *getIndirect(int dim) const;
 488
 489    inline DataFile getFile() const;
 490    inline unsigned getSize() const;
 491
 492    // SSA: return eventual (traverse MOVs) literal value, if it exists
 493    bool getImmediate(ImmediateValue&) const;
 494
 495 public:
 496    Modifier mod;
 497    int8_t indirect[2]; // >= 0 if relative to lvalue in insn->src(indirect[i])
 498    uint8_t swizzle;
 499
 500    bool usedAsPtr; // for printing
 501
 502 private:
 503    Value *value;
 504    Instruction *insn;
 505 };
 506
 507 class ValueDef
 508 {
 509 public:
 510    ValueDef(Value * = NULL);
 511    ValueDef(const ValueDef&);
 512    ~ValueDef();
 513
 514    inline bool exists() const { return value != NULL; }
 515
 516    inline Value *get() const { return value; }
 517    inline Value *rep() const;
 518    void set(Value *);
 519    bool mayReplace(const ValueRef &);
 520    void replace(const ValueRef &, bool doSet); // replace all uses of the old value
 521
 522    inline Instruction *getInsn() const { return insn; }
 523    inline void setInsn(Instruction *inst) { insn = inst; }
 524
 525    inline DataFile getFile() const;
 526    inline unsigned getSize() const;
 527
 528    inline void setSSA(LValue *);
 529    inline const LValue *preSSA() const;
 530
 531 private:
 532    Value *value;   // should make this LValue * ...
 533    LValue *origin; // pre SSA value
 534    Instruction *insn;
 535 };
 536
 537 class Value
 538 {
 539 public:
 540    Value();
 541    virtual ~Value() { }
 542
 543    virtual Value *clone(ClonePolicy<Function>&) const = 0;
 544
 545    virtual int print(char *, size_t, DataType ty = TYPE_NONE) const = 0;
 546
 547    virtual bool equals(const Value *, bool strict = false) const;
 548    virtual bool interfers(const Value *) const;
 549    virtual bool isUniform() const { return true; }
 550
 551    inline Value *rep() const { return join; }
 552
 553    inline Instruction *getUniqueInsn() const;
 554    inline Instruction *getInsn() const; // use when uniqueness is certain
 555
 556    inline int refCount() { return uses.size(); }
 557
 558    inline LValue *asLValue();
 559    inline Symbol *asSym();
 560    inline ImmediateValue *asImm();
 561    inline const Symbol *asSym() const;
 562    inline const ImmediateValue *asImm() const;
 563
 564    inline bool inFile(DataFile f) { return reg.file == f; }
 565
 566    static inline Value *get(Iterator&);
 567
 568    std::list<ValueRef *> uses;
 569    std::list<ValueDef *> defs;
 570    typedef std::list<ValueRef *>::iterator UseIterator;
 571    typedef std::list<ValueRef *>::const_iterator UseCIterator;
 572    typedef std::list<ValueDef *>::iterator DefIterator;
 573    typedef std::list<ValueDef *>::const_iterator DefCIterator;
 574
 575    int id;
 576    Storage reg;
 577
 578    // TODO: these should be in LValue:
 579    Interval livei;
 580    Value *join;
 581 };
 582
 583 class LValue : public Value
 584 {
 585 public:
 586    LValue(Function *, DataFile file);
 587    LValue(Function *, LValue *);
 588    ~LValue() { }
 589
 590    virtual bool isUniform() const;
 591
 592    virtual LValue *clone(ClonePolicy<Function>&) const;
 593
 594    virtual int print(char *, size_t, DataType ty = TYPE_NONE) const;
 595
 596 public:
 597    unsigned compMask : 8; // compound/component mask
 598    unsigned compound : 1; // used by RA, value involved in split/merge
 599    unsigned ssa      : 1;
 600    unsigned fixedReg : 1; // set & used by RA, earlier just use (id < 0)
 601    unsigned noSpill  : 1; // do not spill (e.g. if spill temporary already)
 602 };
 603
 604 class Symbol : public Value
 605 {
 606 public:
 607    Symbol(Program *, DataFile file = FILE_MEMORY_CONST, ubyte fileIdx = 0);
 608    ~Symbol() { }
 609
 610    virtual Symbol *clone(ClonePolicy<Function>&) const;
 611
 612    virtual bool equals(const Value *that, bool strict) const;
 613
 614    virtual bool isUniform() const;
 615
 616    virtual int print(char *, size_t, DataType ty = TYPE_NONE) const;
 617
 618    // print with indirect values
 619    int print(char *, size_t, Value *, Value *, DataType ty = TYPE_NONE) const;
 620
 621    inline void setFile(DataFile file, ubyte fileIndex = 0)
 622    {
 623       reg.file = file;
 624       reg.fileIndex = fileIndex;
 625    }
 626
 627    inline void setOffset(int32_t offset);
 628    inline void setAddress(Symbol *base, int32_t offset);
 629    inline void setSV(SVSemantic sv, uint32_t idx = 0);
 630
 631    inline const Symbol *getBase() const { return baseSym; }
 632
 633 private:
 634    Symbol *baseSym; // array base for Symbols representing array elements
 635 };
 636
 637 class ImmediateValue : public Value
 638 {
 639 public:
 640    ImmediateValue() { }
 641    ImmediateValue(Program *, uint32_t);
 642    ImmediateValue(Program *, float);
 643    ImmediateValue(Program *, double);
 644    // NOTE: not added to program with
 645    ImmediateValue(const ImmediateValue *, DataType ty);
 646    ~ImmediateValue() { };
 647
 648    virtual ImmediateValue *clone(ClonePolicy<Function>&) const;
 649
 650    virtual bool equals(const Value *that, bool strict) const;
 651
 652    // these only work if 'type' is valid (we mostly use untyped literals):
 653    bool isInteger(const int ival) const; // ival is cast to this' type
 654    bool isNegative() const;
 655    bool isPow2() const;
 656
 657    void applyLog2();
 658
 659    // for constant folding:
 660    ImmediateValue operator+(const ImmediateValue&) const;
 661    ImmediateValue operator-(const ImmediateValue&) const;
 662    ImmediateValue operator*(const ImmediateValue&) const;
 663    ImmediateValue operator/(const ImmediateValue&) const;
 664
 665    ImmediateValue& operator=(const ImmediateValue&); // only sets value !
 666
 667    bool compare(CondCode cc, float fval) const;
 668
 669    virtual int print(char *, size_t, DataType ty = TYPE_NONE) const;
 670 };
 671
 672 class Instruction
 673 {
 674 public:
 675    Instruction();
 676    Instruction(Function *, operation, DataType);
 677    virtual ~Instruction();
 678
 679    virtual Instruction *clone(ClonePolicy<Function>&,
 680                               Instruction * = NULL) const;
 681
 682    void setDef(int i, Value *);
 683    void setSrc(int s, Value *);
 684    void setSrc(int s, const ValueRef&);
 685    void swapSources(int a, int b);
 686    void moveSources(int s, int delta);
 687    bool setIndirect(int s, int dim, Value *);
 688
 689    inline ValueRef& src(int s) { return srcs[s]; }
 690    inline ValueDef& def(int s) { return defs[s]; }
 691    inline const ValueRef& src(int s) const { return srcs[s]; }
 692    inline const ValueDef& def(int s) const { return defs[s]; }
 693
 694    inline Value *getDef(int d) const { return defs[d].get(); }
 695    inline Value *getSrc(int s) const { return srcs[s].get(); }
 696    inline Value *getIndirect(int s, int dim) const;
 697
 698    inline bool defExists(unsigned d) const
 699    {
 700       return d < defs.size() && defs[d].exists();
 701    }
 702    inline bool srcExists(unsigned s) const
 703    {
 704       return s < srcs.size() && srcs[s].exists();
 705    }
 706
 707    inline bool constrainedDefs() const;
 708
 709    bool setPredicate(CondCode ccode, Value *);
 710    inline Value *getPredicate() const;
 711    bool writesPredicate() const;
 712    inline bool isPredicated() const { return predSrc >= 0; }
 713
 714    inline void setFlagsSrc(int s, Value *);
 715    inline void setFlagsDef(int d, Value *);
 716    inline bool usesFlags() const { return flagsSrc >= 0; }
 717
 718    unsigned int defCount() const { return defs.size(); };
 719    unsigned int defCount(unsigned int mask, bool singleFile = false) const;
 720    unsigned int srcCount() const { return srcs.size(); };
 721    unsigned int srcCount(unsigned int mask, bool singleFile = false) const;
 722
 723    // save & remove / set indirect[0,1] and predicate source
 724    void takeExtraSources(int s, Value *[3]);
 725    void putExtraSources(int s, Value *[3]);
 726
 727    inline void setType(DataType type) { dType = sType = type; }
 728
 729    inline void setType(DataType dtype, DataType stype)
 730    {
 731       dType = dtype;
 732       sType = stype;
 733    }
 734
 735    inline bool isPseudo() const { return op < OP_MOV; }
 736    bool isDead() const;
 737    bool isNop() const;
 738    bool isCommutationLegal(const Instruction *) const; // must be adjacent !
 739    bool isActionEqual(const Instruction *) const;
 740    bool isResultEqual(const Instruction *) const;
 741
 742    void print() const;
 743
 744    inline CmpInstruction *asCmp();
 745    inline TexInstruction *asTex();
 746    inline FlowInstruction *asFlow();
 747    inline const TexInstruction *asTex() const;
 748    inline const CmpInstruction *asCmp() const;
 749    inline const FlowInstruction *asFlow() const;
 750
 751 public:
 752    Instruction *next;
 753    Instruction *prev;
 754    int id;
 755    int serial; // CFG order
 756
 757    operation op;
 758    DataType dType; // destination or defining type
 759    DataType sType; // source or secondary type
 760    CondCode cc;
 761    RoundMode rnd;
 762    CacheMode cache;
 763
 764    uint16_t subOp; // quadop, 1 for mul-high, etc.
 765
 766    unsigned encSize    : 4; // encoding size in bytes
 767    unsigned saturate   : 1; // to [0.0f, 1.0f]
 768    unsigned join       : 1; // converge control flow (use OP_JOIN until end)
 769    unsigned fixed      : 1; // prevent dead code elimination
 770    unsigned terminator : 1; // end of basic block
 771    unsigned ftz        : 1; // flush denormal to zero
 772    unsigned dnz        : 1; // denormals, NaN are zero
 773    unsigned ipa        : 4; // interpolation mode
 774    unsigned lanes      : 4;
 775    unsigned perPatch   : 1;
 776    unsigned exit       : 1; // terminate program after insn
 777    unsigned mask       : 4; // for vector ops
 778
 779    int8_t postFactor; // MUL/DIV(if < 0) by 1 << postFactor
 780
 781    int8_t predSrc;
 782    int8_t flagsDef;
 783    int8_t flagsSrc;
 784
 785    uint8_t sched; // scheduling data (NOTE: maybe move to separate storage)
 786
 787    BasicBlock *bb;
 788
 789 protected:
 790    std::deque<ValueDef> defs; // no gaps !
 791    std::deque<ValueRef> srcs; // no gaps !
 792
 793    // instruction specific methods:
 794    // (don't want to subclass, would need more constructors and memory pools)
 795 public:
 796    inline void setInterpolate(unsigned int mode) { ipa = mode; }
 797
 798    unsigned int getInterpMode() const { return ipa & 0x3; }
 799    unsigned int getSampleMode() const { return ipa & 0xc; }
 800
 801 private:
 802    void init();
 803 };
 804
 805 enum TexQuery
 806 {
 807    TXQ_DIMS,
 808    TXQ_TYPE,
 809    TXQ_SAMPLE_POSITION,
 810    TXQ_FILTER,
 811    TXQ_LOD,
 812    TXQ_WRAP,
 813    TXQ_BORDER_COLOUR
 814 };
 815
 816 class TexInstruction : public Instruction
 817 {
 818 public:
 819    class Target
 820    {
 821    public:
 822       Target(TexTarget targ = TEX_TARGET_2D) : target(targ) { }
 823
 824       const char *getName() const { return descTable[target].name; }
 825       unsigned int getArgCount() const { return descTable[target].argc; }
 826       unsigned int getDim() const { return descTable[target].dim; }
 827       int isArray() const { return descTable[target].array ? 1 : 0; }
 828       int isCube() const { return descTable[target].cube ? 1 : 0; }
 829       int isShadow() const { return descTable[target].shadow ? 1 : 0; }
 830       int isMS() const {
 831         return target == TEX_TARGET_2D_MS || target == TEX_TARGET_2D_MS_ARRAY; }
 832       void clearMS() {
 833          if (isMS()) {
 834             if (isArray())
 835                target = TEX_TARGET_2D_ARRAY;
 836             else
 837                target = TEX_TARGET_2D;
 838          }
 839       }
 840
 841       Target& operator=(TexTarget targ)
 842       {
 843          assert(targ < TEX_TARGET_COUNT);
 844          target = targ;
 845          return *this;
 846       }
 847
 848       inline bool operator==(TexTarget targ) const { return target == targ; }
 849       inline bool operator!=(TexTarget targ) const { return target != targ; }
 850
 851       enum TexTarget getEnum() const { return target; }
 852
 853    private:
 854       struct Desc
 855       {
 856          char name[19];
 857          uint8_t dim;
 858          uint8_t argc;
 859          bool array;
 860          bool cube;
 861          bool shadow;
 862       };
 863
 864       static const struct Desc descTable[TEX_TARGET_COUNT];
 865
 866    private:
 867       enum TexTarget target;
 868    };
 869
 870 public:
 871    TexInstruction(Function *, operation);
 872    virtual ~TexInstruction();
 873
 874    virtual TexInstruction *clone(ClonePolicy<Function>&,
 875                                  Instruction * = NULL) const;
 876
 877    inline void setTexture(Target targ, uint8_t r, uint8_t s)
 878    {
 879       tex.r = r;
 880       tex.s = s;
 881       tex.target = targ;
 882    }
 883
 884    void setIndirectR(Value *);
 885    void setIndirectS(Value *);
 886    inline Value *getIndirectR() const;
 887    inline Value *getIndirectS() const;
 888
 889 public:
 890    struct {
 891       Target target;
 892
 893       uint16_t r;
 894       uint16_t s;
 895       int8_t rIndirectSrc;
 896       int8_t sIndirectSrc;
 897
 898       uint8_t mask;
 899       uint8_t gatherComp;
 900
 901       bool liveOnly; // only execute on live pixels of a quad (optimization)
 902       bool levelZero;
 903       bool derivAll;
 904
 905       int8_t useOffsets; // 0, 1, or 4 for textureGatherOffsets
 906       int8_t offset[4][3];
 907
 908       enum TexQuery query;
 909    } tex;
 910
 911    ValueRef dPdx[3];
 912    ValueRef dPdy[3];
 913 };
 914
 915 class CmpInstruction : public Instruction
 916 {
 917 public:
 918    CmpInstruction(Function *, operation);
 919
 920    virtual CmpInstruction *clone(ClonePolicy<Function>&,
 921                                  Instruction * = NULL) const;
 922
 923    void setCondition(CondCode cond) { setCond = cond; }
 924    CondCode getCondition() const { return setCond; }
 925
 926 public:
 927    CondCode setCond;
 928 };
 929
 930 class FlowInstruction : public Instruction
 931 {
 932 public:
 933    FlowInstruction(Function *, operation, void *target);
 934
 935    virtual FlowInstruction *clone(ClonePolicy<Function>&,
 936                                   Instruction * = NULL) const;
 937
 938 public:
 939    unsigned allWarp  : 1;
 940    unsigned absolute : 1;
 941    unsigned limit    : 1;
 942    unsigned builtin  : 1; // true for calls to emulation code
 943    unsigned indirect : 1; // target in src(0)
 944
 945    union {
 946       BasicBlock *bb;
 947       int builtin;
 948       Function *fn;
 949    } target;
 950 };
 951
 952 class BasicBlock
 953 {
 954 public:
 955    BasicBlock(Function *);
 956    ~BasicBlock();
 957
 958    BasicBlock *clone(ClonePolicy<Function>&) const;
 959
 960    inline int getId() const { return id; }
 961    inline unsigned int getInsnCount() const { return numInsns; }
 962    inline bool isTerminated() const { return exit && exit->terminator; }
 963
 964    bool dominatedBy(BasicBlock *bb);
 965    inline bool reachableBy(const BasicBlock *by, const BasicBlock *term);
 966
 967    // returns mask of conditional out blocks
 968    // e.g. 3 for IF { .. } ELSE { .. } ENDIF, 1 for IF { .. } ENDIF
 969    unsigned int initiatesSimpleConditional() const;
 970
 971 public:
 972    Function *getFunction() const { return func; }
 973    Program *getProgram() const { return program; }
 974
 975    Instruction *getEntry() const { return entry; } // first non-phi instruction
 976    Instruction *getPhi() const { return phi; }
 977    Instruction *getFirst() const { return phi ? phi : entry; }
 978    Instruction *getExit() const { return exit; }
 979
 980    void insertHead(Instruction *);
 981    void insertTail(Instruction *);
 982    void insertBefore(Instruction *, Instruction *);
 983    void insertAfter(Instruction *, Instruction *);
 984    void remove(Instruction *);
 985    void permuteAdjacent(Instruction *, Instruction *);
 986
 987    BasicBlock *idom() const;
 988
 989    // NOTE: currently does not rebuild the dominator tree
 990    BasicBlock *splitBefore(Instruction *, bool attach = true);
 991    BasicBlock *splitAfter(Instruction *, bool attach = true);
 992
 993    DLList& getDF() { return df; }
 994    DLList::Iterator iterDF() { return df.iterator(); }
 995
 996    static inline BasicBlock *get(Iterator&);
 997    static inline BasicBlock *get(Graph::Node *);
 998
 999 public:
1000    Graph::Node cfg; // first edge is branch *taken* (the ELSE branch)
1001    Graph::Node dom;
1002
1003    BitSet liveSet;
1004    BitSet defSet;
1005
1006    uint32_t binPos;
1007    uint32_t binSize;
1008
1009    Instruction *joinAt; // for quick reference
1010
1011    bool explicitCont; // loop headers: true if loop contains continue stmts
1012
1013 private:
1014    int id;
1015    DLList df;
1016
1017    Instruction *phi;
1018    Instruction *entry;
1019    Instruction *exit;
1020
1021    unsigned int numInsns;
1022
1023 private:
1024    Function *func;
1025    Program *program;
1026
1027    void splitCommon(Instruction *, BasicBlock *, bool attach);
1028 };
1029
1030 class Function
1031 {
1032 public:
1033    Function(Program *, const char *name, uint32_t label);
1034    ~Function();
1035
1036    static inline Function *get(Graph::Node *node);
1037
1038    inline Program *getProgram() const { return prog; }
1039    inline const char *getName() const { return name; }
1040    inline int getId() const { return id; }
1041    inline uint32_t getLabel() const { return label; }
1042
1043    void print();
1044    void printLiveIntervals() const;
1045    void printCFGraph(const char *filePath);
1046
1047    bool setEntry(BasicBlock *);
1048    bool setExit(BasicBlock *);
1049
1050    unsigned int orderInstructions(ArrayList&);
1051
1052    inline void add(BasicBlock *bb, int& id) { allBBlocks.insert(bb, id); }
1053    inline void add(Instruction *insn, int& id) { allInsns.insert(insn, id); }
1054    inline void add(LValue *lval, int& id) { allLValues.insert(lval, id); }
1055
1056    inline LValue *getLValue(int id);
1057
1058    void buildLiveSets();
1059    void buildDefSets();
1060    bool convertToSSA();
1061
1062 public:
1063    std::deque<ValueDef> ins;
1064    std::deque<ValueRef> outs;
1065    std::deque<Value *> clobbers;
1066
1067    Graph cfg;
1068    Graph::Node *cfgExit;
1069    Graph *domTree;
1070    Graph::Node call; // node in the call graph
1071
1072    BasicBlock **bbArray; // BBs in emission order
1073    int bbCount;
1074
1075    unsigned int loopNestingBound;
1076    int regClobberMax;
1077
1078    uint32_t binPos;
1079    uint32_t binSize;
1080
1081    Value *stackPtr;
1082
1083    uint32_t tlsBase; // base address for l[] space (if no stack pointer is used)
1084    uint32_t tlsSize;
1085
1086    ArrayList allBBlocks;
1087    ArrayList allInsns;
1088    ArrayList allLValues;
1089
1090 private:
1091    void buildLiveSetsPreSSA(BasicBlock *, const int sequence);
1092    void buildDefSetsPreSSA(BasicBlock *bb, const int seq);
1093
1094 private:
1095    uint32_t label;
1096    int id;
1097    const char *const name;
1098    Program *prog;
1099 };
1100
1101 enum CGStage
1102 {
1103    CG_STAGE_PRE_SSA,
1104    CG_STAGE_SSA, // expected directly before register allocation
1105    CG_STAGE_POST_RA
1106 };
1107
1108 class Program
1109 {
1110 public:
1111    enum Type
1112    {
1113       TYPE_VERTEX,
1114       TYPE_TESSELLATION_CONTROL,
1115       TYPE_TESSELLATION_EVAL,
1116       TYPE_GEOMETRY,
1117       TYPE_FRAGMENT,
1118       TYPE_COMPUTE
1119    };
1120
1121    Program(Type type, Target *targ);
1122    ~Program();
1123
1124    void print();
1125
1126    Type getType() const { return progType; }
1127
1128    inline void add(Function *fn, int& id) { allFuncs.insert(fn, id); }
1129    inline void del(Function *fn, int& id) { allFuncs.remove(id); }
1130    inline void add(Value *rval, int& id) { allRValues.insert(rval, id); }
1131
1132    bool makeFromTGSI(struct nv50_ir_prog_info *);
1133    bool makeFromSM4(struct nv50_ir_prog_info *);
1134    bool convertToSSA();
1135    bool optimizeSSA(int level);
1136    bool optimizePostRA(int level);
1137    bool registerAllocation();
1138    bool emitBinary(struct nv50_ir_prog_info *);
1139
1140    const Target *getTarget() const { return target; }
1141
1142 private:
1143    void emitSymbolTable(struct nv50_ir_prog_info *);
1144
1145    Type progType;
1146    Target *target;
1147
1148 public:
1149    Function *main;
1150    Graph calls;
1151
1152    ArrayList allFuncs;
1153    ArrayList allRValues;
1154
1155    uint32_t *code;
1156    uint32_t binSize;
1157    uint32_t tlsSize; // size required for FILE_MEMORY_LOCAL
1158
1159    int maxGPR;
1160
1161    MemoryPool mem_Instruction;
1162    MemoryPool mem_CmpInstruction;
1163    MemoryPool mem_TexInstruction;
1164    MemoryPool mem_FlowInstruction;
1165    MemoryPool mem_LValue;
1166    MemoryPool mem_Symbol;
1167    MemoryPool mem_ImmediateValue;
1168
1169    uint32_t dbgFlags;
1170    uint8_t  optLevel;
1171
1172    void *targetPriv; // e.g. to carry information between passes
1173
1174    const struct nv50_ir_prog_info *driver; // for driver configuration
1175
1176    void releaseInstruction(Instruction *);
1177    void releaseValue(Value *);
1178 };
1179
1180 // TODO: add const version
1181 class Pass
1182 {
1183 public:
1184    bool run(Program *, bool ordered = false, bool skipPhi = false);
1185    bool run(Function *, bool ordered = false, bool skipPhi = false);
1186
1187 private:
1188    // return false to continue with next entity on next higher level
1189    virtual bool visit(Function *) { return true; }
1190    virtual bool visit(BasicBlock *) { return true; }
1191    virtual bool visit(Instruction *) { return false; }
1192
1193    bool doRun(Program *, bool ordered, bool skipPhi);
1194    bool doRun(Function *, bool ordered, bool skipPhi);
1195
1196 protected:
1197    bool err;
1198    Function *func;
1199    Program *prog;
1200 };
1201
1202 // =============================================================================
1203
1204 #include "codegen/nv50_ir_inlines.h"
1205
1206 } // namespace nv50_ir
1207
1208 #endif // __NV50_IR_H__