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_PIXLD, // get info about raster object or surfaces
 140    OP_QUADOP,
 141    OP_QUADON,
 142    OP_QUADPOP,
 143    OP_POPCNT, // bitcount(src0 & src1)
 144    OP_INSBF,  // insert first src1[8:15] bits of src0 into src2 at src1[0:7]
 145    OP_EXTBF,  // place bits [K,K+N) of src0 into dst, src1 = 0xNNKK
 146    OP_PERMT,  // dst = bytes from src2,src0 selected by src1 (nvc0's src order)
 147    OP_ATOM,
 148    OP_BAR,    // execution barrier, sources = { id, thread count, predicate }
 149    OP_VADD,   // byte/word vector operations
 150    OP_VAVG,
 151    OP_VMIN,
 152    OP_VMAX,
 153    OP_VSAD,
 154    OP_VSET,
 155    OP_VSHR,
 156    OP_VSHL,
 157    OP_VSEL,
 158    OP_CCTL, // cache control
 159    OP_LAST
 160 };
 161
 162 // various instruction-specific modifier definitions Instruction::subOp
 163 // MOV_FINAL marks a MOV originating from an EXPORT (used for placing TEXBARs)
 164 #define NV50_IR_SUBOP_MUL_HIGH     1
 165 #define NV50_IR_SUBOP_EMIT_RESTART 1
 166 #define NV50_IR_SUBOP_LDC_IL       1
 167 #define NV50_IR_SUBOP_LDC_IS       2
 168 #define NV50_IR_SUBOP_LDC_ISL      3
 169 #define NV50_IR_SUBOP_SHIFT_WRAP   1
 170 #define NV50_IR_SUBOP_EMU_PRERET   1
 171 #define NV50_IR_SUBOP_TEXBAR(n)    n
 172 #define NV50_IR_SUBOP_MOV_FINAL    1
 173 #define NV50_IR_SUBOP_EXTBF_REV    1
 174 #define NV50_IR_SUBOP_PERMT_F4E    1
 175 #define NV50_IR_SUBOP_PERMT_B4E    2
 176 #define NV50_IR_SUBOP_PERMT_RC8    3
 177 #define NV50_IR_SUBOP_PERMT_ECL    4
 178 #define NV50_IR_SUBOP_PERMT_ECR    5
 179 #define NV50_IR_SUBOP_PERMT_RC16   6
 180 #define NV50_IR_SUBOP_BAR_SYNC     0
 181 #define NV50_IR_SUBOP_BAR_ARRIVE   1
 182 #define NV50_IR_SUBOP_BAR_RED_AND  2
 183 #define NV50_IR_SUBOP_BAR_RED_OR   3
 184 #define NV50_IR_SUBOP_BAR_RED_POPC 4
 185 #define NV50_IR_SUBOP_MEMBAR_L     1
 186 #define NV50_IR_SUBOP_MEMBAR_S     2
 187 #define NV50_IR_SUBOP_MEMBAR_M     3
 188 #define NV50_IR_SUBOP_MEMBAR_CTA  (0 << 2)
 189 #define NV50_IR_SUBOP_MEMBAR_GL   (1 << 2)
 190 #define NV50_IR_SUBOP_MEMBAR_SYS  (2 << 2)
 191 #define NV50_IR_SUBOP_MEMBAR_DIR(m)   ((m) & 0x3)
 192 #define NV50_IR_SUBOP_MEMBAR_SCOPE(m) ((m) & ~0x3)
 193 #define NV50_IR_SUBOP_MEMBAR(d,s) \
 194    (NV50_IR_SUBOP_MEMBAR_##d | NV50_IR_SUBOP_MEMBAR_##s)
 195 #define NV50_IR_SUBOP_ATOM_ADD      0
 196 #define NV50_IR_SUBOP_ATOM_MIN      1
 197 #define NV50_IR_SUBOP_ATOM_MAX      2
 198 #define NV50_IR_SUBOP_ATOM_INC      3
 199 #define NV50_IR_SUBOP_ATOM_DEC      4
 200 #define NV50_IR_SUBOP_ATOM_AND      5
 201 #define NV50_IR_SUBOP_ATOM_OR       6
 202 #define NV50_IR_SUBOP_ATOM_XOR      7
 203 #define NV50_IR_SUBOP_ATOM_CAS      8
 204 #define NV50_IR_SUBOP_ATOM_EXCH     9
 205 #define NV50_IR_SUBOP_CCTL_IV      5
 206 #define NV50_IR_SUBOP_CCTL_IVALL   6
 207 #define NV50_IR_SUBOP_SUST_IGN     0
 208 #define NV50_IR_SUBOP_SUST_TRAP    1
 209 #define NV50_IR_SUBOP_SUST_SDCL    3
 210 #define NV50_IR_SUBOP_SULD_ZERO    0
 211 #define NV50_IR_SUBOP_SULD_TRAP    1
 212 #define NV50_IR_SUBOP_SULD_SDCL    3
 213 #define NV50_IR_SUBOP_SUBFM_3D     1
 214 #define NV50_IR_SUBOP_SUCLAMP_2D   0x10
 215 #define NV50_IR_SUBOP_SUCLAMP_SD(r, d) (( 0 + (r)) | ((d == 2) ? 0x10 : 0))
 216 #define NV50_IR_SUBOP_SUCLAMP_PL(r, d) (( 5 + (r)) | ((d == 2) ? 0x10 : 0))
 217 #define NV50_IR_SUBOP_SUCLAMP_BL(r, d) ((10 + (r)) | ((d == 2) ? 0x10 : 0))
 218 #define NV50_IR_SUBOP_PIXLD_COUNT       0
 219 #define NV50_IR_SUBOP_PIXLD_COVMASK     1
 220 #define NV50_IR_SUBOP_PIXLD_COVERED     2
 221 #define NV50_IR_SUBOP_PIXLD_OFFSET      3
 222 #define NV50_IR_SUBOP_PIXLD_CENT_OFFSET 4
 223 #define NV50_IR_SUBOP_PIXLD_SAMPLEID    5
 224 #define NV50_IR_SUBOP_MADSP_SD     0xffff
 225 // Yes, we could represent those with DataType.
 226 // Or put the type into operation and have a couple 1000 values in that enum.
 227 // This will have to do for now.
 228 // The bitfields are supposed to correspond to nve4 ISA.
 229 #define NV50_IR_SUBOP_MADSP(a,b,c) (((c) << 8) | ((b) << 4) | (a))
 230 #define NV50_IR_SUBOP_V1(d,a,b)    (((d) << 10) | ((b) << 5) | (a) | 0x0000)
 231 #define NV50_IR_SUBOP_V2(d,a,b)    (((d) << 10) | ((b) << 5) | (a) | 0x4000)
 232 #define NV50_IR_SUBOP_V4(d,a,b)    (((d) << 10) | ((b) << 5) | (a) | 0x8000)
 233 #define NV50_IR_SUBOP_Vn(n)        ((n) >> 14)
 234
 235 enum DataType
 236 {
 237    TYPE_NONE,
 238    TYPE_U8,
 239    TYPE_S8,
 240    TYPE_U16,
 241    TYPE_S16,
 242    TYPE_U32,
 243    TYPE_S32,
 244    TYPE_U64, // 64 bit operations are only lowered after register allocation
 245    TYPE_S64,
 246    TYPE_F16,
 247    TYPE_F32,
 248    TYPE_F64,
 249    TYPE_B96,
 250    TYPE_B128
 251 };
 252
 253 enum CondCode
 254 {
 255    CC_FL = 0,
 256    CC_NEVER = CC_FL, // when used with FILE_FLAGS
 257    CC_LT = 1,
 258    CC_EQ = 2,
 259    CC_NOT_P = CC_EQ, // when used with FILE_PREDICATE
 260    CC_LE = 3,
 261    CC_GT = 4,
 262    CC_NE = 5,
 263    CC_P  = CC_NE,
 264    CC_GE = 6,
 265    CC_TR = 7,
 266    CC_ALWAYS = CC_TR,
 267    CC_U  = 8,
 268    CC_LTU = 9,
 269    CC_EQU = 10,
 270    CC_LEU = 11,
 271    CC_GTU = 12,
 272    CC_NEU = 13,
 273    CC_GEU = 14,
 274    CC_NO = 0x10,
 275    CC_NC = 0x11,
 276    CC_NS = 0x12,
 277    CC_NA = 0x13,
 278    CC_A  = 0x14,
 279    CC_S  = 0x15,
 280    CC_C  = 0x16,
 281    CC_O  = 0x17
 282 };
 283
 284 enum RoundMode
 285 {
 286    ROUND_N, // nearest
 287    ROUND_M, // towards -inf
 288    ROUND_Z, // towards 0
 289    ROUND_P, // towards +inf
 290    ROUND_NI, // nearest integer
 291    ROUND_MI, // to integer towards -inf
 292    ROUND_ZI, // to integer towards 0
 293    ROUND_PI, // to integer towards +inf
 294 };
 295
 296 enum CacheMode
 297 {
 298    CACHE_CA,            // cache at all levels
 299    CACHE_WB = CACHE_CA, // cache write back
 300    CACHE_CG,            // cache at global level
 301    CACHE_CS,            // cache streaming
 302    CACHE_CV,            // cache as volatile
 303    CACHE_WT = CACHE_CV  // cache write-through
 304 };
 305
 306 enum DataFile
 307 {
 308    FILE_NULL = 0,
 309    FILE_GPR,
 310    FILE_PREDICATE,       // boolean predicate
 311    FILE_FLAGS,           // zero/sign/carry/overflow bits
 312    FILE_ADDRESS,
 313    LAST_REGISTER_FILE = FILE_ADDRESS,
 314    FILE_IMMEDIATE,
 315    FILE_MEMORY_CONST,
 316    FILE_SHADER_INPUT,
 317    FILE_SHADER_OUTPUT,
 318    FILE_MEMORY_GLOBAL,
 319    FILE_MEMORY_SHARED,
 320    FILE_MEMORY_LOCAL,
 321    FILE_SYSTEM_VALUE,
 322    DATA_FILE_COUNT
 323 };
 324
 325 enum TexTarget
 326 {
 327    TEX_TARGET_1D,
 328    TEX_TARGET_2D,
 329    TEX_TARGET_2D_MS,
 330    TEX_TARGET_3D,
 331    TEX_TARGET_CUBE,
 332    TEX_TARGET_1D_SHADOW,
 333    TEX_TARGET_2D_SHADOW,
 334    TEX_TARGET_CUBE_SHADOW,
 335    TEX_TARGET_1D_ARRAY,
 336    TEX_TARGET_2D_ARRAY,
 337    TEX_TARGET_2D_MS_ARRAY,
 338    TEX_TARGET_CUBE_ARRAY,
 339    TEX_TARGET_1D_ARRAY_SHADOW,
 340    TEX_TARGET_2D_ARRAY_SHADOW,
 341    TEX_TARGET_RECT,
 342    TEX_TARGET_RECT_SHADOW,
 343    TEX_TARGET_CUBE_ARRAY_SHADOW,
 344    TEX_TARGET_BUFFER,
 345    TEX_TARGET_COUNT
 346 };
 347
 348 enum SVSemantic
 349 {
 350    SV_POSITION, // WPOS
 351    SV_VERTEX_ID,
 352    SV_INSTANCE_ID,
 353    SV_INVOCATION_ID,
 354    SV_PRIMITIVE_ID,
 355    SV_VERTEX_COUNT, // gl_PatchVerticesIn
 356    SV_LAYER,
 357    SV_VIEWPORT_INDEX,
 358    SV_YDIR,
 359    SV_FACE,
 360    SV_POINT_SIZE,
 361    SV_POINT_COORD,
 362    SV_CLIP_DISTANCE,
 363    SV_SAMPLE_INDEX,
 364    SV_SAMPLE_POS,
 365    SV_TESS_FACTOR,
 366    SV_TESS_COORD,
 367    SV_TID,
 368    SV_CTAID,
 369    SV_NTID,
 370    SV_GRIDID,
 371    SV_NCTAID,
 372    SV_LANEID,
 373    SV_PHYSID,
 374    SV_NPHYSID,
 375    SV_CLOCK,
 376    SV_LBASE,
 377    SV_SBASE,
 378    SV_VERTEX_STRIDE,
 379    SV_UNDEFINED,
 380    SV_LAST
 381 };
 382
 383 class Program;
 384 class Function;
 385 class BasicBlock;
 386
 387 class Target;
 388
 389 class Instruction;
 390 class CmpInstruction;
 391 class TexInstruction;
 392 class FlowInstruction;
 393
 394 class Value;
 395 class LValue;
 396 class Symbol;
 397 class ImmediateValue;
 398
 399 struct Storage
 400 {
 401    DataFile file;
 402    int8_t fileIndex; // signed, may be indirect for CONST[]
 403    uint8_t size; // this should match the Instruction type's size
 404    DataType type; // mainly for pretty printing
 405    union {
 406       uint64_t u64;    // immediate values
 407       uint32_t u32;
 408       uint16_t u16;
 409       uint8_t u8;
 410       int64_t s64;
 411       int32_t s32;
 412       int16_t s16;
 413       int8_t s8;
 414       float f32;
 415       double f64;
 416       int32_t offset; // offset from 0 (base of address space)
 417       int32_t id;     // register id (< 0 if virtual/unassigned, in units <= 4)
 418       struct {
 419          SVSemantic sv;
 420          int index;
 421       } sv;
 422    } data;
 423 };
 424
 425 // precedence: NOT after SAT after NEG after ABS
 426 #define NV50_IR_MOD_ABS (1 << 0)
 427 #define NV50_IR_MOD_NEG (1 << 1)
 428 #define NV50_IR_MOD_SAT (1 << 2)
 429 #define NV50_IR_MOD_NOT (1 << 3)
 430 #define NV50_IR_MOD_NEG_ABS (NV50_IR_MOD_NEG | NV50_IR_MOD_ABS)
 431
 432 #define NV50_IR_INTERP_MODE_MASK   0x3
 433 #define NV50_IR_INTERP_LINEAR      (0 << 0)
 434 #define NV50_IR_INTERP_PERSPECTIVE (1 << 0)
 435 #define NV50_IR_INTERP_FLAT        (2 << 0)
 436 #define NV50_IR_INTERP_SC          (3 << 0) // what exactly is that ?
 437 #define NV50_IR_INTERP_SAMPLE_MASK 0xc
 438 #define NV50_IR_INTERP_DEFAULT     (0 << 2)
 439 #define NV50_IR_INTERP_CENTROID    (1 << 2)
 440 #define NV50_IR_INTERP_OFFSET      (2 << 2)
 441 #define NV50_IR_INTERP_SAMPLEID    (3 << 2)
 442
 443 // do we really want this to be a class ?
 444 class Modifier
 445 {
 446 public:
 447    Modifier() : bits(0) { }
 448    Modifier(unsigned int m) : bits(m) { }
 449    Modifier(operation op);
 450
 451    // @return new Modifier applying a after b (asserts if unrepresentable)
 452    Modifier operator*(const Modifier) const;
 453    Modifier operator*=(const Modifier m) { *this = *this * m; return *this; }
 454    Modifier operator==(const Modifier m) const { return m.bits == bits; }
 455    Modifier operator!=(const Modifier m) const { return m.bits != bits; }
 456
 457    inline Modifier operator&(const Modifier m) const { return bits & m.bits; }
 458    inline Modifier operator|(const Modifier m) const { return bits | m.bits; }
 459    inline Modifier operator^(const Modifier m) const { return bits ^ m.bits; }
 460
 461    operation getOp() const;
 462
 463    inline int neg() const { return (bits & NV50_IR_MOD_NEG) ? 1 : 0; }
 464    inline int abs() const { return (bits & NV50_IR_MOD_ABS) ? 1 : 0; }
 465
 466    inline operator bool() const { return bits ? true : false; }
 467
 468    void applyTo(ImmediateValue &imm) const;
 469
 470    int print(char *buf, size_t size) const;
 471
 472 private:
 473    uint8_t bits;
 474 };
 475
 476 class ValueRef
 477 {
 478 public:
 479    ValueRef(Value * = NULL);
 480    ValueRef(const ValueRef&);
 481    ~ValueRef();
 482
 483    inline bool exists() const { return value != NULL; }
 484
 485    void set(Value *);
 486    void set(const ValueRef&);
 487    inline Value *get() const { return value; }
 488    inline Value *rep() const;
 489
 490    inline Instruction *getInsn() const { return insn; }
 491    inline void setInsn(Instruction *inst) { insn = inst; }
 492
 493    inline bool isIndirect(int dim) const { return indirect[dim] >= 0; }
 494    inline const ValueRef *getIndirect(int dim) const;
 495
 496    inline DataFile getFile() const;
 497    inline unsigned getSize() const;
 498
 499    // SSA: return eventual (traverse MOVs) literal value, if it exists
 500    bool getImmediate(ImmediateValue&) const;
 501
 502 public:
 503    Modifier mod;
 504    int8_t indirect[2]; // >= 0 if relative to lvalue in insn->src(indirect[i])
 505    uint8_t swizzle;
 506
 507    bool usedAsPtr; // for printing
 508
 509 private:
 510    Value *value;
 511    Instruction *insn;
 512 };
 513
 514 class ValueDef
 515 {
 516 public:
 517    ValueDef(Value * = NULL);
 518    ValueDef(const ValueDef&);
 519    ~ValueDef();
 520
 521    inline bool exists() const { return value != NULL; }
 522
 523    inline Value *get() const { return value; }
 524    inline Value *rep() const;
 525    void set(Value *);
 526    bool mayReplace(const ValueRef &);
 527    void replace(const ValueRef &, bool doSet); // replace all uses of the old value
 528
 529    inline Instruction *getInsn() const { return insn; }
 530    inline void setInsn(Instruction *inst) { insn = inst; }
 531
 532    inline DataFile getFile() const;
 533    inline unsigned getSize() const;
 534
 535    inline void setSSA(LValue *);
 536    inline const LValue *preSSA() const;
 537
 538 private:
 539    Value *value;   // should make this LValue * ...
 540    LValue *origin; // pre SSA value
 541    Instruction *insn;
 542 };
 543
 544 class Value
 545 {
 546 public:
 547    Value();
 548    virtual ~Value() { }
 549
 550    virtual Value *clone(ClonePolicy<Function>&) const = 0;
 551
 552    virtual int print(char *, size_t, DataType ty = TYPE_NONE) const = 0;
 553
 554    virtual bool equals(const Value *, bool strict = false) const;
 555    virtual bool interfers(const Value *) const;
 556    virtual bool isUniform() const { return true; }
 557
 558    inline Value *rep() const { return join; }
 559
 560    inline Instruction *getUniqueInsn() const;
 561    inline Instruction *getInsn() const; // use when uniqueness is certain
 562
 563    inline int refCount() { return uses.size(); }
 564
 565    inline LValue *asLValue();
 566    inline Symbol *asSym();
 567    inline ImmediateValue *asImm();
 568    inline const Symbol *asSym() const;
 569    inline const ImmediateValue *asImm() const;
 570
 571    inline bool inFile(DataFile f) { return reg.file == f; }
 572
 573    static inline Value *get(Iterator&);
 574
 575    std::list<ValueRef *> uses;
 576    std::list<ValueDef *> defs;
 577    typedef std::list<ValueRef *>::iterator UseIterator;
 578    typedef std::list<ValueRef *>::const_iterator UseCIterator;
 579    typedef std::list<ValueDef *>::iterator DefIterator;
 580    typedef std::list<ValueDef *>::const_iterator DefCIterator;
 581
 582    int id;
 583    Storage reg;
 584
 585    // TODO: these should be in LValue:
 586    Interval livei;
 587    Value *join;
 588 };
 589
 590 class LValue : public Value
 591 {
 592 public:
 593    LValue(Function *, DataFile file);
 594    LValue(Function *, LValue *);
 595    ~LValue() { }
 596
 597    virtual bool isUniform() const;
 598
 599    virtual LValue *clone(ClonePolicy<Function>&) const;
 600
 601    virtual int print(char *, size_t, DataType ty = TYPE_NONE) const;
 602
 603 public:
 604    unsigned compMask : 8; // compound/component mask
 605    unsigned compound : 1; // used by RA, value involved in split/merge
 606    unsigned ssa      : 1;
 607    unsigned fixedReg : 1; // set & used by RA, earlier just use (id < 0)
 608    unsigned noSpill  : 1; // do not spill (e.g. if spill temporary already)
 609 };
 610
 611 class Symbol : public Value
 612 {
 613 public:
 614    Symbol(Program *, DataFile file = FILE_MEMORY_CONST, ubyte fileIdx = 0);
 615    ~Symbol() { }
 616
 617    virtual Symbol *clone(ClonePolicy<Function>&) const;
 618
 619    virtual bool equals(const Value *that, bool strict) const;
 620
 621    virtual bool isUniform() const;
 622
 623    virtual int print(char *, size_t, DataType ty = TYPE_NONE) const;
 624
 625    // print with indirect values
 626    int print(char *, size_t, Value *, Value *, DataType ty = TYPE_NONE) const;
 627
 628    inline void setFile(DataFile file, ubyte fileIndex = 0)
 629    {
 630       reg.file = file;
 631       reg.fileIndex = fileIndex;
 632    }
 633
 634    inline void setOffset(int32_t offset);
 635    inline void setAddress(Symbol *base, int32_t offset);
 636    inline void setSV(SVSemantic sv, uint32_t idx = 0);
 637
 638    inline const Symbol *getBase() const { return baseSym; }
 639
 640 private:
 641    Symbol *baseSym; // array base for Symbols representing array elements
 642 };
 643
 644 class ImmediateValue : public Value
 645 {
 646 public:
 647    ImmediateValue() { }
 648    ImmediateValue(Program *, uint32_t);
 649    ImmediateValue(Program *, float);
 650    ImmediateValue(Program *, double);
 651    // NOTE: not added to program with
 652    ImmediateValue(const ImmediateValue *, DataType ty);
 653    ~ImmediateValue() { };
 654
 655    virtual ImmediateValue *clone(ClonePolicy<Function>&) const;
 656
 657    virtual bool equals(const Value *that, bool strict) const;
 658
 659    // these only work if 'type' is valid (we mostly use untyped literals):
 660    bool isInteger(const int ival) const; // ival is cast to this' type
 661    bool isNegative() const;
 662    bool isPow2() const;
 663
 664    void applyLog2();
 665
 666    // for constant folding:
 667    ImmediateValue operator+(const ImmediateValue&) const;
 668    ImmediateValue operator-(const ImmediateValue&) const;
 669    ImmediateValue operator*(const ImmediateValue&) const;
 670    ImmediateValue operator/(const ImmediateValue&) const;
 671
 672    ImmediateValue& operator=(const ImmediateValue&); // only sets value !
 673
 674    bool compare(CondCode cc, float fval) const;
 675
 676    virtual int print(char *, size_t, DataType ty = TYPE_NONE) const;
 677 };
 678
 679 class Instruction
 680 {
 681 public:
 682    Instruction();
 683    Instruction(Function *, operation, DataType);
 684    virtual ~Instruction();
 685
 686    virtual Instruction *clone(ClonePolicy<Function>&,
 687                               Instruction * = NULL) const;
 688
 689    void setDef(int i, Value *);
 690    void setSrc(int s, Value *);
 691    void setSrc(int s, const ValueRef&);
 692    void swapSources(int a, int b);
 693    void moveSources(int s, int delta);
 694    bool setIndirect(int s, int dim, Value *);
 695
 696    inline ValueRef& src(int s) { return srcs[s]; }
 697    inline ValueDef& def(int s) { return defs[s]; }
 698    inline const ValueRef& src(int s) const { return srcs[s]; }
 699    inline const ValueDef& def(int s) const { return defs[s]; }
 700
 701    inline Value *getDef(int d) const { return defs[d].get(); }
 702    inline Value *getSrc(int s) const { return srcs[s].get(); }
 703    inline Value *getIndirect(int s, int dim) const;
 704
 705    inline bool defExists(unsigned d) const
 706    {
 707       return d < defs.size() && defs[d].exists();
 708    }
 709    inline bool srcExists(unsigned s) const
 710    {
 711       return s < srcs.size() && srcs[s].exists();
 712    }
 713
 714    inline bool constrainedDefs() const;
 715
 716    bool setPredicate(CondCode ccode, Value *);
 717    inline Value *getPredicate() const;
 718    bool writesPredicate() const;
 719    inline bool isPredicated() const { return predSrc >= 0; }
 720
 721    inline void setFlagsSrc(int s, Value *);
 722    inline void setFlagsDef(int d, Value *);
 723    inline bool usesFlags() const { return flagsSrc >= 0; }
 724
 725    unsigned int defCount() const { return defs.size(); };
 726    unsigned int defCount(unsigned int mask, bool singleFile = false) const;
 727    unsigned int srcCount() const { return srcs.size(); };
 728    unsigned int srcCount(unsigned int mask, bool singleFile = false) const;
 729
 730    // save & remove / set indirect[0,1] and predicate source
 731    void takeExtraSources(int s, Value *[3]);
 732    void putExtraSources(int s, Value *[3]);
 733
 734    inline void setType(DataType type) { dType = sType = type; }
 735
 736    inline void setType(DataType dtype, DataType stype)
 737    {
 738       dType = dtype;
 739       sType = stype;
 740    }
 741
 742    inline bool isPseudo() const { return op < OP_MOV; }
 743    bool isDead() const;
 744    bool isNop() const;
 745    bool isCommutationLegal(const Instruction *) const; // must be adjacent !
 746    bool isActionEqual(const Instruction *) const;
 747    bool isResultEqual(const Instruction *) const;
 748
 749    void print() const;
 750
 751    inline CmpInstruction *asCmp();
 752    inline TexInstruction *asTex();
 753    inline FlowInstruction *asFlow();
 754    inline const TexInstruction *asTex() const;
 755    inline const CmpInstruction *asCmp() const;
 756    inline const FlowInstruction *asFlow() const;
 757
 758 public:
 759    Instruction *next;
 760    Instruction *prev;
 761    int id;
 762    int serial; // CFG order
 763
 764    operation op;
 765    DataType dType; // destination or defining type
 766    DataType sType; // source or secondary type
 767    CondCode cc;
 768    RoundMode rnd;
 769    CacheMode cache;
 770
 771    uint16_t subOp; // quadop, 1 for mul-high, etc.
 772
 773    unsigned encSize    : 4; // encoding size in bytes
 774    unsigned saturate   : 1; // to [0.0f, 1.0f]
 775    unsigned join       : 1; // converge control flow (use OP_JOIN until end)
 776    unsigned fixed      : 1; // prevent dead code elimination
 777    unsigned terminator : 1; // end of basic block
 778    unsigned ftz        : 1; // flush denormal to zero
 779    unsigned dnz        : 1; // denormals, NaN are zero
 780    unsigned ipa        : 4; // interpolation mode
 781    unsigned lanes      : 4;
 782    unsigned perPatch   : 1;
 783    unsigned exit       : 1; // terminate program after insn
 784    unsigned mask       : 4; // for vector ops
 785
 786    int8_t postFactor; // MUL/DIV(if < 0) by 1 << postFactor
 787
 788    int8_t predSrc;
 789    int8_t flagsDef;
 790    int8_t flagsSrc;
 791
 792    uint8_t sched; // scheduling data (NOTE: maybe move to separate storage)
 793
 794    BasicBlock *bb;
 795
 796 protected:
 797    std::deque<ValueDef> defs; // no gaps !
 798    std::deque<ValueRef> srcs; // no gaps !
 799
 800    // instruction specific methods:
 801    // (don't want to subclass, would need more constructors and memory pools)
 802 public:
 803    inline void setInterpolate(unsigned int mode) { ipa = mode; }
 804
 805    unsigned int getInterpMode() const { return ipa & 0x3; }
 806    unsigned int getSampleMode() const { return ipa & 0xc; }
 807
 808 private:
 809    void init();
 810 };
 811
 812 enum TexQuery
 813 {
 814    TXQ_DIMS,
 815    TXQ_TYPE,
 816    TXQ_SAMPLE_POSITION,
 817    TXQ_FILTER,
 818    TXQ_LOD,
 819    TXQ_WRAP,
 820    TXQ_BORDER_COLOUR
 821 };
 822
 823 class TexInstruction : public Instruction
 824 {
 825 public:
 826    class Target
 827    {
 828    public:
 829       Target(TexTarget targ = TEX_TARGET_2D) : target(targ) { }
 830
 831       const char *getName() const { return descTable[target].name; }
 832       unsigned int getArgCount() const { return descTable[target].argc; }
 833       unsigned int getDim() const { return descTable[target].dim; }
 834       int isArray() const { return descTable[target].array ? 1 : 0; }
 835       int isCube() const { return descTable[target].cube ? 1 : 0; }
 836       int isShadow() const { return descTable[target].shadow ? 1 : 0; }
 837       int isMS() const {
 838         return target == TEX_TARGET_2D_MS || target == TEX_TARGET_2D_MS_ARRAY; }
 839       void clearMS() {
 840          if (isMS()) {
 841             if (isArray())
 842                target = TEX_TARGET_2D_ARRAY;
 843             else
 844                target = TEX_TARGET_2D;
 845          }
 846       }
 847
 848       Target& operator=(TexTarget targ)
 849       {
 850          assert(targ < TEX_TARGET_COUNT);
 851          target = targ;
 852          return *this;
 853       }
 854
 855       inline bool operator==(TexTarget targ) const { return target == targ; }
 856       inline bool operator!=(TexTarget targ) const { return target != targ; }
 857
 858       enum TexTarget getEnum() const { return target; }
 859
 860    private:
 861       struct Desc
 862       {
 863          char name[19];
 864          uint8_t dim;
 865          uint8_t argc;
 866          bool array;
 867          bool cube;
 868          bool shadow;
 869       };
 870
 871       static const struct Desc descTable[TEX_TARGET_COUNT];
 872
 873    private:
 874       enum TexTarget target;
 875    };
 876
 877 public:
 878    TexInstruction(Function *, operation);
 879    virtual ~TexInstruction();
 880
 881    virtual TexInstruction *clone(ClonePolicy<Function>&,
 882                                  Instruction * = NULL) const;
 883
 884    inline void setTexture(Target targ, uint8_t r, uint8_t s)
 885    {
 886       tex.r = r;
 887       tex.s = s;
 888       tex.target = targ;
 889    }
 890
 891    void setIndirectR(Value *);
 892    void setIndirectS(Value *);
 893    inline Value *getIndirectR() const;
 894    inline Value *getIndirectS() const;
 895
 896 public:
 897    struct {
 898       Target target;
 899
 900       uint16_t r;
 901       uint16_t s;
 902       int8_t rIndirectSrc;
 903       int8_t sIndirectSrc;
 904
 905       uint8_t mask;
 906       uint8_t gatherComp;
 907
 908       bool liveOnly; // only execute on live pixels of a quad (optimization)
 909       bool levelZero;
 910       bool derivAll;
 911
 912       int8_t useOffsets; // 0, 1, or 4 for textureGatherOffsets
 913       int8_t offset[4][3];
 914
 915       enum TexQuery query;
 916    } tex;
 917
 918    ValueRef dPdx[3];
 919    ValueRef dPdy[3];
 920 };
 921
 922 class CmpInstruction : public Instruction
 923 {
 924 public:
 925    CmpInstruction(Function *, operation);
 926
 927    virtual CmpInstruction *clone(ClonePolicy<Function>&,
 928                                  Instruction * = NULL) const;
 929
 930    void setCondition(CondCode cond) { setCond = cond; }
 931    CondCode getCondition() const { return setCond; }
 932
 933 public:
 934    CondCode setCond;
 935 };
 936
 937 class FlowInstruction : public Instruction
 938 {
 939 public:
 940    FlowInstruction(Function *, operation, void *target);
 941
 942    virtual FlowInstruction *clone(ClonePolicy<Function>&,
 943                                   Instruction * = NULL) const;
 944
 945 public:
 946    unsigned allWarp  : 1;
 947    unsigned absolute : 1;
 948    unsigned limit    : 1;
 949    unsigned builtin  : 1; // true for calls to emulation code
 950    unsigned indirect : 1; // target in src(0)
 951
 952    union {
 953       BasicBlock *bb;
 954       int builtin;
 955       Function *fn;
 956    } target;
 957 };
 958
 959 class BasicBlock
 960 {
 961 public:
 962    BasicBlock(Function *);
 963    ~BasicBlock();
 964
 965    BasicBlock *clone(ClonePolicy<Function>&) const;
 966
 967    inline int getId() const { return id; }
 968    inline unsigned int getInsnCount() const { return numInsns; }
 969    inline bool isTerminated() const { return exit && exit->terminator; }
 970
 971    bool dominatedBy(BasicBlock *bb);
 972    inline bool reachableBy(const BasicBlock *by, const BasicBlock *term);
 973
 974    // returns mask of conditional out blocks
 975    // e.g. 3 for IF { .. } ELSE { .. } ENDIF, 1 for IF { .. } ENDIF
 976    unsigned int initiatesSimpleConditional() const;
 977
 978 public:
 979    Function *getFunction() const { return func; }
 980    Program *getProgram() const { return program; }
 981
 982    Instruction *getEntry() const { return entry; } // first non-phi instruction
 983    Instruction *getPhi() const { return phi; }
 984    Instruction *getFirst() const { return phi ? phi : entry; }
 985    Instruction *getExit() const { return exit; }
 986
 987    void insertHead(Instruction *);
 988    void insertTail(Instruction *);
 989    void insertBefore(Instruction *, Instruction *);
 990    void insertAfter(Instruction *, Instruction *);
 991    void remove(Instruction *);
 992    void permuteAdjacent(Instruction *, Instruction *);
 993
 994    BasicBlock *idom() const;
 995
 996    // NOTE: currently does not rebuild the dominator tree
 997    BasicBlock *splitBefore(Instruction *, bool attach = true);
 998    BasicBlock *splitAfter(Instruction *, bool attach = true);
 999
1000    DLList& getDF() { return df; }
1001    DLList::Iterator iterDF() { return df.iterator(); }
1002
1003    static inline BasicBlock *get(Iterator&);
1004    static inline BasicBlock *get(Graph::Node *);
1005
1006 public:
1007    Graph::Node cfg; // first edge is branch *taken* (the ELSE branch)
1008    Graph::Node dom;
1009
1010    BitSet liveSet;
1011    BitSet defSet;
1012
1013    uint32_t binPos;
1014    uint32_t binSize;
1015
1016    Instruction *joinAt; // for quick reference
1017
1018    bool explicitCont; // loop headers: true if loop contains continue stmts
1019
1020 private:
1021    int id;
1022    DLList df;
1023
1024    Instruction *phi;
1025    Instruction *entry;
1026    Instruction *exit;
1027
1028    unsigned int numInsns;
1029
1030 private:
1031    Function *func;
1032    Program *program;
1033
1034    void splitCommon(Instruction *, BasicBlock *, bool attach);
1035 };
1036
1037 class Function
1038 {
1039 public:
1040    Function(Program *, const char *name, uint32_t label);
1041    ~Function();
1042
1043    static inline Function *get(Graph::Node *node);
1044
1045    inline Program *getProgram() const { return prog; }
1046    inline const char *getName() const { return name; }
1047    inline int getId() const { return id; }
1048    inline uint32_t getLabel() const { return label; }
1049
1050    void print();
1051    void printLiveIntervals() const;
1052    void printCFGraph(const char *filePath);
1053
1054    bool setEntry(BasicBlock *);
1055    bool setExit(BasicBlock *);
1056
1057    unsigned int orderInstructions(ArrayList&);
1058
1059    inline void add(BasicBlock *bb, int& id) { allBBlocks.insert(bb, id); }
1060    inline void add(Instruction *insn, int& id) { allInsns.insert(insn, id); }
1061    inline void add(LValue *lval, int& id) { allLValues.insert(lval, id); }
1062
1063    inline LValue *getLValue(int id);
1064
1065    void buildLiveSets();
1066    void buildDefSets();
1067    bool convertToSSA();
1068
1069 public:
1070    std::deque<ValueDef> ins;
1071    std::deque<ValueRef> outs;
1072    std::deque<Value *> clobbers;
1073
1074    Graph cfg;
1075    Graph::Node *cfgExit;
1076    Graph *domTree;
1077    Graph::Node call; // node in the call graph
1078
1079    BasicBlock **bbArray; // BBs in emission order
1080    int bbCount;
1081
1082    unsigned int loopNestingBound;
1083    int regClobberMax;
1084
1085    uint32_t binPos;
1086    uint32_t binSize;
1087
1088    Value *stackPtr;
1089
1090    uint32_t tlsBase; // base address for l[] space (if no stack pointer is used)
1091    uint32_t tlsSize;
1092
1093    ArrayList allBBlocks;
1094    ArrayList allInsns;
1095    ArrayList allLValues;
1096
1097 private:
1098    void buildLiveSetsPreSSA(BasicBlock *, const int sequence);
1099    void buildDefSetsPreSSA(BasicBlock *bb, const int seq);
1100
1101 private:
1102    uint32_t label;
1103    int id;
1104    const char *const name;
1105    Program *prog;
1106 };
1107
1108 enum CGStage
1109 {
1110    CG_STAGE_PRE_SSA,
1111    CG_STAGE_SSA, // expected directly before register allocation
1112    CG_STAGE_POST_RA
1113 };
1114
1115 class Program
1116 {
1117 public:
1118    enum Type
1119    {
1120       TYPE_VERTEX,
1121       TYPE_TESSELLATION_CONTROL,
1122       TYPE_TESSELLATION_EVAL,
1123       TYPE_GEOMETRY,
1124       TYPE_FRAGMENT,
1125       TYPE_COMPUTE
1126    };
1127
1128    Program(Type type, Target *targ);
1129    ~Program();
1130
1131    void print();
1132
1133    Type getType() const { return progType; }
1134
1135    inline void add(Function *fn, int& id) { allFuncs.insert(fn, id); }
1136    inline void del(Function *fn, int& id) { allFuncs.remove(id); }
1137    inline void add(Value *rval, int& id) { allRValues.insert(rval, id); }
1138
1139    bool makeFromTGSI(struct nv50_ir_prog_info *);
1140    bool makeFromSM4(struct nv50_ir_prog_info *);
1141    bool convertToSSA();
1142    bool optimizeSSA(int level);
1143    bool optimizePostRA(int level);
1144    bool registerAllocation();
1145    bool emitBinary(struct nv50_ir_prog_info *);
1146
1147    const Target *getTarget() const { return target; }
1148
1149 private:
1150    void emitSymbolTable(struct nv50_ir_prog_info *);
1151
1152    Type progType;
1153    Target *target;
1154
1155 public:
1156    Function *main;
1157    Graph calls;
1158
1159    ArrayList allFuncs;
1160    ArrayList allRValues;
1161
1162    uint32_t *code;
1163    uint32_t binSize;
1164    uint32_t tlsSize; // size required for FILE_MEMORY_LOCAL
1165
1166    int maxGPR;
1167
1168    MemoryPool mem_Instruction;
1169    MemoryPool mem_CmpInstruction;
1170    MemoryPool mem_TexInstruction;
1171    MemoryPool mem_FlowInstruction;
1172    MemoryPool mem_LValue;
1173    MemoryPool mem_Symbol;
1174    MemoryPool mem_ImmediateValue;
1175
1176    uint32_t dbgFlags;
1177    uint8_t  optLevel;
1178
1179    void *targetPriv; // e.g. to carry information between passes
1180
1181    const struct nv50_ir_prog_info *driver; // for driver configuration
1182
1183    void releaseInstruction(Instruction *);
1184    void releaseValue(Value *);
1185 };
1186
1187 // TODO: add const version
1188 class Pass
1189 {
1190 public:
1191    bool run(Program *, bool ordered = false, bool skipPhi = false);
1192    bool run(Function *, bool ordered = false, bool skipPhi = false);
1193
1194 private:
1195    // return false to continue with next entity on next higher level
1196    virtual bool visit(Function *) { return true; }
1197    virtual bool visit(BasicBlock *) { return true; }
1198    virtual bool visit(Instruction *) { return false; }
1199
1200    bool doRun(Program *, bool ordered, bool skipPhi);
1201    bool doRun(Function *, bool ordered, bool skipPhi);
1202
1203 protected:
1204    bool err;
1205    Function *func;
1206    Program *prog;
1207 };
1208
1209 // =============================================================================
1210
1211 #include "codegen/nv50_ir_inlines.h"
1212
1213 } // namespace nv50_ir
1214
1215 #endif // __NV50_IR_H__