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/unordered_set.h"
  34 #include "codegen/nv50_ir_util.h"
  35 #include "codegen/nv50_ir_graph.h"
  36
  37 #include "codegen/nv50_ir_driver.h"
  38
  39 namespace nv50_ir {
  40
  41 enum operation
  42 {
  43    OP_NOP = 0,
  44    OP_PHI,
  45    OP_UNION, // unify a new definition and several source values
  46    OP_SPLIT, // $r0d -> { $r0, $r1 } ($r0d and $r0/$r1 will be coalesced)
  47    OP_MERGE, // opposite of split, e.g. combine 2 32 bit into a 64 bit value
  48    OP_CONSTRAINT, // copy values into consecutive registers
  49    OP_MOV, // simple copy, no modifiers allowed
  50    OP_LOAD,
  51    OP_STORE,
  52    OP_ADD, // NOTE: add u64 + u32 is legal for targets w/o 64-bit integer adds
  53    OP_SUB,
  54    OP_MUL,
  55    OP_DIV,
  56    OP_MOD,
  57    OP_MAD,
  58    OP_FMA,
  59    OP_SAD, // abs(src0 - src1) + src2
  60    OP_ABS,
  61    OP_NEG,
  62    OP_NOT,
  63    OP_AND,
  64    OP_OR,
  65    OP_XOR,
  66    OP_SHL,
  67    OP_SHR,
  68    OP_MAX,
  69    OP_MIN,
  70    OP_SAT, // CLAMP(f32, 0.0, 1.0)
  71    OP_CEIL,
  72    OP_FLOOR,
  73    OP_TRUNC,
  74    OP_CVT,
  75    OP_SET_AND, // dst = (src0 CMP src1) & src2
  76    OP_SET_OR,
  77    OP_SET_XOR,
  78    OP_SET,
  79    OP_SELP, // dst = src2 ? src0 : src1
  80    OP_SLCT, // dst = (src2 CMP 0) ? src0 : src1
  81    OP_RCP,
  82    OP_RSQ,
  83    OP_LG2,
  84    OP_SIN,
  85    OP_COS,
  86    OP_EX2,
  87    OP_EXP, // exponential (base M_E)
  88    OP_LOG, // natural logarithm
  89    OP_PRESIN,
  90    OP_PREEX2,
  91    OP_SQRT,
  92    OP_POW,
  93    OP_BRA,
  94    OP_CALL,
  95    OP_RET,
  96    OP_CONT,
  97    OP_BREAK,
  98    OP_PRERET,
  99    OP_PRECONT,
 100    OP_PREBREAK,
 101    OP_BRKPT,     // breakpoint (not related to loops)
 102    OP_JOINAT,    // push control flow convergence point
 103    OP_JOIN,      // converge
 104    OP_DISCARD,
 105    OP_EXIT,
 106    OP_MEMBAR, // memory barrier (mfence, lfence, sfence)
 107    OP_VFETCH, // indirection 0 in attribute space, indirection 1 is vertex base
 108    OP_PFETCH, // fetch base address of vertex src0 (immediate) [+ src1]
 109    OP_AFETCH, // fetch base address of shader input (a[%r1+0x10])
 110    OP_EXPORT,
 111    OP_LINTERP,
 112    OP_PINTERP,
 113    OP_EMIT,    // emit vertex
 114    OP_RESTART, // restart primitive
 115    OP_TEX,
 116    OP_TXB, // texture bias
 117    OP_TXL, // texure lod
 118    OP_TXF, // texel fetch
 119    OP_TXQ, // texture size query
 120    OP_TXD, // texture derivatives
 121    OP_TXG, // texture gather
 122    OP_TXLQ, // texture query lod
 123    OP_TEXCSAA, // texture op for coverage sampling
 124    OP_TEXPREP, // turn cube map array into 2d array coordinates
 125    OP_SULDB, // surface load (raw)
 126    OP_SULDP, // surface load (formatted)
 127    OP_SUSTB, // surface store (raw)
 128    OP_SUSTP, // surface store (formatted)
 129    OP_SUREDB,
 130    OP_SUREDP, // surface reduction (atomic op)
 131    OP_SULEA,   // surface load effective address
 132    OP_SUBFM,   // surface bitfield manipulation
 133    OP_SUCLAMP, // clamp surface coordinates
 134    OP_SUEAU,   // surface effective address
 135    OP_MADSP,   // special integer multiply-add
 136    OP_TEXBAR, // texture dependency barrier
 137    OP_DFDX,
 138    OP_DFDY,
 139    OP_RDSV, // read system value
 140    OP_WRSV, // write system value
 141    OP_PIXLD, // get info about raster object or surfaces
 142    OP_QUADOP,
 143    OP_QUADON,
 144    OP_QUADPOP,
 145    OP_POPCNT, // bitcount(src0 & src1)
 146    OP_INSBF,  // insert first src1[8:15] bits of src0 into src2 at src1[0:7]
 147    OP_EXTBF,  // place bits [K,K+N) of src0 into dst, src1 = 0xNNKK
 148    OP_BFIND,  // find highest/lowest set bit
 149    OP_PERMT,  // dst = bytes from src2,src0 selected by src1 (nvc0's src order)
 150    OP_ATOM,
 151    OP_BAR,    // execution barrier, sources = { id, thread count, predicate }
 152    OP_VADD,   // byte/word vector operations
 153    OP_VAVG,
 154    OP_VMIN,
 155    OP_VMAX,
 156    OP_VSAD,
 157    OP_VSET,
 158    OP_VSHR,
 159    OP_VSHL,
 160    OP_VSEL,
 161    OP_CCTL, // cache control
 162    OP_SHFL, // warp shuffle
 163    OP_LAST
 164 };
 165
 166 // various instruction-specific modifier definitions Instruction::subOp
 167 // MOV_FINAL marks a MOV originating from an EXPORT (used for placing TEXBARs)
 168 #define NV50_IR_SUBOP_MUL_HIGH     1
 169 #define NV50_IR_SUBOP_EMIT_RESTART 1
 170 #define NV50_IR_SUBOP_LDC_IL       1
 171 #define NV50_IR_SUBOP_LDC_IS       2
 172 #define NV50_IR_SUBOP_LDC_ISL      3
 173 #define NV50_IR_SUBOP_SHIFT_WRAP   1
 174 #define NV50_IR_SUBOP_EMU_PRERET   1
 175 #define NV50_IR_SUBOP_TEXBAR(n)    n
 176 #define NV50_IR_SUBOP_MOV_FINAL    1
 177 #define NV50_IR_SUBOP_EXTBF_REV    1
 178 #define NV50_IR_SUBOP_BFIND_SAMT   1
 179 #define NV50_IR_SUBOP_RCPRSQ_64H   1
 180 #define NV50_IR_SUBOP_PERMT_F4E    1
 181 #define NV50_IR_SUBOP_PERMT_B4E    2
 182 #define NV50_IR_SUBOP_PERMT_RC8    3
 183 #define NV50_IR_SUBOP_PERMT_ECL    4
 184 #define NV50_IR_SUBOP_PERMT_ECR    5
 185 #define NV50_IR_SUBOP_PERMT_RC16   6
 186 #define NV50_IR_SUBOP_BAR_SYNC     0
 187 #define NV50_IR_SUBOP_BAR_ARRIVE   1
 188 #define NV50_IR_SUBOP_BAR_RED_AND  2
 189 #define NV50_IR_SUBOP_BAR_RED_OR   3
 190 #define NV50_IR_SUBOP_BAR_RED_POPC 4
 191 #define NV50_IR_SUBOP_MEMBAR_L     1
 192 #define NV50_IR_SUBOP_MEMBAR_S     2
 193 #define NV50_IR_SUBOP_MEMBAR_M     3
 194 #define NV50_IR_SUBOP_MEMBAR_CTA  (0 << 2)
 195 #define NV50_IR_SUBOP_MEMBAR_GL   (1 << 2)
 196 #define NV50_IR_SUBOP_MEMBAR_SYS  (2 << 2)
 197 #define NV50_IR_SUBOP_MEMBAR_DIR(m)   ((m) & 0x3)
 198 #define NV50_IR_SUBOP_MEMBAR_SCOPE(m) ((m) & ~0x3)
 199 #define NV50_IR_SUBOP_MEMBAR(d,s) \
 200    (NV50_IR_SUBOP_MEMBAR_##d | NV50_IR_SUBOP_MEMBAR_##s)
 201 #define NV50_IR_SUBOP_ATOM_ADD      0
 202 #define NV50_IR_SUBOP_ATOM_MIN      1
 203 #define NV50_IR_SUBOP_ATOM_MAX      2
 204 #define NV50_IR_SUBOP_ATOM_INC      3
 205 #define NV50_IR_SUBOP_ATOM_DEC      4
 206 #define NV50_IR_SUBOP_ATOM_AND      5
 207 #define NV50_IR_SUBOP_ATOM_OR       6
 208 #define NV50_IR_SUBOP_ATOM_XOR      7
 209 #define NV50_IR_SUBOP_ATOM_CAS      8
 210 #define NV50_IR_SUBOP_ATOM_EXCH     9
 211 #define NV50_IR_SUBOP_CCTL_IV      5
 212 #define NV50_IR_SUBOP_CCTL_IVALL   6
 213 #define NV50_IR_SUBOP_SUST_IGN     0
 214 #define NV50_IR_SUBOP_SUST_TRAP    1
 215 #define NV50_IR_SUBOP_SUST_SDCL    3
 216 #define NV50_IR_SUBOP_SULD_ZERO    0
 217 #define NV50_IR_SUBOP_SULD_TRAP    1
 218 #define NV50_IR_SUBOP_SULD_SDCL    3
 219 #define NV50_IR_SUBOP_SUBFM_3D     1
 220 #define NV50_IR_SUBOP_SUCLAMP_2D   0x10
 221 #define NV50_IR_SUBOP_SUCLAMP_SD(r, d) (( 0 + (r)) | ((d == 2) ? 0x10 : 0))
 222 #define NV50_IR_SUBOP_SUCLAMP_PL(r, d) (( 5 + (r)) | ((d == 2) ? 0x10 : 0))
 223 #define NV50_IR_SUBOP_SUCLAMP_BL(r, d) ((10 + (r)) | ((d == 2) ? 0x10 : 0))
 224 #define NV50_IR_SUBOP_PIXLD_COUNT       0
 225 #define NV50_IR_SUBOP_PIXLD_COVMASK     1
 226 #define NV50_IR_SUBOP_PIXLD_COVERED     2
 227 #define NV50_IR_SUBOP_PIXLD_OFFSET      3
 228 #define NV50_IR_SUBOP_PIXLD_CENT_OFFSET 4
 229 #define NV50_IR_SUBOP_PIXLD_SAMPLEID    5
 230 #define NV50_IR_SUBOP_SHFL_IDX  0
 231 #define NV50_IR_SUBOP_SHFL_UP   1
 232 #define NV50_IR_SUBOP_SHFL_DOWN 2
 233 #define NV50_IR_SUBOP_SHFL_BFLY 3
 234 #define NV50_IR_SUBOP_MADSP_SD     0xffff
 235 // Yes, we could represent those with DataType.
 236 // Or put the type into operation and have a couple 1000 values in that enum.
 237 // This will have to do for now.
 238 // The bitfields are supposed to correspond to nve4 ISA.
 239 #define NV50_IR_SUBOP_MADSP(a,b,c) (((c) << 8) | ((b) << 4) | (a))
 240 #define NV50_IR_SUBOP_V1(d,a,b)    (((d) << 10) | ((b) << 5) | (a) | 0x0000)
 241 #define NV50_IR_SUBOP_V2(d,a,b)    (((d) << 10) | ((b) << 5) | (a) | 0x4000)
 242 #define NV50_IR_SUBOP_V4(d,a,b)    (((d) << 10) | ((b) << 5) | (a) | 0x8000)
 243 #define NV50_IR_SUBOP_Vn(n)        ((n) >> 14)
 244
 245 enum DataType
 246 {
 247    TYPE_NONE,
 248    TYPE_U8,
 249    TYPE_S8,
 250    TYPE_U16,
 251    TYPE_S16,
 252    TYPE_U32,
 253    TYPE_S32,
 254    TYPE_U64, // 64 bit operations are only lowered after register allocation
 255    TYPE_S64,
 256    TYPE_F16,
 257    TYPE_F32,
 258    TYPE_F64,
 259    TYPE_B96,
 260    TYPE_B128
 261 };
 262
 263 enum CondCode
 264 {
 265    CC_FL = 0,
 266    CC_NEVER = CC_FL, // when used with FILE_FLAGS
 267    CC_LT = 1,
 268    CC_EQ = 2,
 269    CC_NOT_P = CC_EQ, // when used with FILE_PREDICATE
 270    CC_LE = 3,
 271    CC_GT = 4,
 272    CC_NE = 5,
 273    CC_P  = CC_NE,
 274    CC_GE = 6,
 275    CC_TR = 7,
 276    CC_ALWAYS = CC_TR,
 277    CC_U  = 8,
 278    CC_LTU = 9,
 279    CC_EQU = 10,
 280    CC_LEU = 11,
 281    CC_GTU = 12,
 282    CC_NEU = 13,
 283    CC_GEU = 14,
 284    CC_NO = 0x10,
 285    CC_NC = 0x11,
 286    CC_NS = 0x12,
 287    CC_NA = 0x13,
 288    CC_A  = 0x14,
 289    CC_S  = 0x15,
 290    CC_C  = 0x16,
 291    CC_O  = 0x17
 292 };
 293
 294 enum RoundMode
 295 {
 296    ROUND_N, // nearest
 297    ROUND_M, // towards -inf
 298    ROUND_Z, // towards 0
 299    ROUND_P, // towards +inf
 300    ROUND_NI, // nearest integer
 301    ROUND_MI, // to integer towards -inf
 302    ROUND_ZI, // to integer towards 0
 303    ROUND_PI, // to integer towards +inf
 304 };
 305
 306 enum CacheMode
 307 {
 308    CACHE_CA,            // cache at all levels
 309    CACHE_WB = CACHE_CA, // cache write back
 310    CACHE_CG,            // cache at global level
 311    CACHE_CS,            // cache streaming
 312    CACHE_CV,            // cache as volatile
 313    CACHE_WT = CACHE_CV  // cache write-through
 314 };
 315
 316 enum DataFile
 317 {
 318    FILE_NULL = 0,
 319    FILE_GPR,
 320    FILE_PREDICATE,       // boolean predicate
 321    FILE_FLAGS,           // zero/sign/carry/overflow bits
 322    FILE_ADDRESS,
 323    LAST_REGISTER_FILE = FILE_ADDRESS,
 324    FILE_IMMEDIATE,
 325    FILE_MEMORY_CONST,
 326    FILE_SHADER_INPUT,
 327    FILE_SHADER_OUTPUT,
 328    FILE_MEMORY_GLOBAL,
 329    FILE_MEMORY_SHARED,
 330    FILE_MEMORY_LOCAL,
 331    FILE_SYSTEM_VALUE,
 332    DATA_FILE_COUNT
 333 };
 334
 335 enum TexTarget
 336 {
 337    TEX_TARGET_1D,
 338    TEX_TARGET_2D,
 339    TEX_TARGET_2D_MS,
 340    TEX_TARGET_3D,
 341    TEX_TARGET_CUBE,
 342    TEX_TARGET_1D_SHADOW,
 343    TEX_TARGET_2D_SHADOW,
 344    TEX_TARGET_CUBE_SHADOW,
 345    TEX_TARGET_1D_ARRAY,
 346    TEX_TARGET_2D_ARRAY,
 347    TEX_TARGET_2D_MS_ARRAY,
 348    TEX_TARGET_CUBE_ARRAY,
 349    TEX_TARGET_1D_ARRAY_SHADOW,
 350    TEX_TARGET_2D_ARRAY_SHADOW,
 351    TEX_TARGET_RECT,
 352    TEX_TARGET_RECT_SHADOW,
 353    TEX_TARGET_CUBE_ARRAY_SHADOW,
 354    TEX_TARGET_BUFFER,
 355    TEX_TARGET_COUNT
 356 };
 357
 358 enum SVSemantic
 359 {
 360    SV_POSITION, // WPOS
 361    SV_VERTEX_ID,
 362    SV_INSTANCE_ID,
 363    SV_INVOCATION_ID,
 364    SV_PRIMITIVE_ID,
 365    SV_VERTEX_COUNT, // gl_PatchVerticesIn
 366    SV_LAYER,
 367    SV_VIEWPORT_INDEX,
 368    SV_YDIR,
 369    SV_FACE,
 370    SV_POINT_SIZE,
 371    SV_POINT_COORD,
 372    SV_CLIP_DISTANCE,
 373    SV_SAMPLE_INDEX,
 374    SV_SAMPLE_POS,
 375    SV_SAMPLE_MASK,
 376    SV_TESS_OUTER,
 377    SV_TESS_INNER,
 378    SV_TESS_COORD,
 379    SV_TID,
 380    SV_CTAID,
 381    SV_NTID,
 382    SV_GRIDID,
 383    SV_NCTAID,
 384    SV_LANEID,
 385    SV_PHYSID,
 386    SV_NPHYSID,
 387    SV_CLOCK,
 388    SV_LBASE,
 389    SV_SBASE,
 390    SV_VERTEX_STRIDE,
 391    SV_INVOCATION_INFO,
 392    SV_THREAD_KILL,
 393    SV_BASEVERTEX,
 394    SV_BASEINSTANCE,
 395    SV_DRAWID,
 396    SV_UNDEFINED,
 397    SV_LAST
 398 };
 399
 400 class Program;
 401 class Function;
 402 class BasicBlock;
 403
 404 class Target;
 405
 406 class Instruction;
 407 class CmpInstruction;
 408 class TexInstruction;
 409 class FlowInstruction;
 410
 411 class Value;
 412 class LValue;
 413 class Symbol;
 414 class ImmediateValue;
 415
 416 struct Storage
 417 {
 418    DataFile file;
 419    int8_t fileIndex; // signed, may be indirect for CONST[]
 420    uint8_t size; // this should match the Instruction type's size
 421    DataType type; // mainly for pretty printing
 422    union {
 423       uint64_t u64;    // immediate values
 424       uint32_t u32;
 425       uint16_t u16;
 426       uint8_t u8;
 427       int64_t s64;
 428       int32_t s32;
 429       int16_t s16;
 430       int8_t s8;
 431       float f32;
 432       double f64;
 433       int32_t offset; // offset from 0 (base of address space)
 434       int32_t id;     // register id (< 0 if virtual/unassigned, in units <= 4)
 435       struct {
 436          SVSemantic sv;
 437          int index;
 438       } sv;
 439    } data;
 440 };
 441
 442 // precedence: NOT after SAT after NEG after ABS
 443 #define NV50_IR_MOD_ABS (1 << 0)
 444 #define NV50_IR_MOD_NEG (1 << 1)
 445 #define NV50_IR_MOD_SAT (1 << 2)
 446 #define NV50_IR_MOD_NOT (1 << 3)
 447 #define NV50_IR_MOD_NEG_ABS (NV50_IR_MOD_NEG | NV50_IR_MOD_ABS)
 448
 449 #define NV50_IR_INTERP_MODE_MASK   0x3
 450 #define NV50_IR_INTERP_LINEAR      (0 << 0)
 451 #define NV50_IR_INTERP_PERSPECTIVE (1 << 0)
 452 #define NV50_IR_INTERP_FLAT        (2 << 0)
 453 #define NV50_IR_INTERP_SC          (3 << 0) // what exactly is that ?
 454 #define NV50_IR_INTERP_SAMPLE_MASK 0xc
 455 #define NV50_IR_INTERP_DEFAULT     (0 << 2)
 456 #define NV50_IR_INTERP_CENTROID    (1 << 2)
 457 #define NV50_IR_INTERP_OFFSET      (2 << 2)
 458 #define NV50_IR_INTERP_SAMPLEID    (3 << 2)
 459
 460 // do we really want this to be a class ?
 461 class Modifier
 462 {
 463 public:
 464    Modifier() : bits(0) { }
 465    Modifier(unsigned int m) : bits(m) { }
 466    Modifier(operation op);
 467
 468    // @return new Modifier applying a after b (asserts if unrepresentable)
 469    Modifier operator*(const Modifier) const;
 470    Modifier operator*=(const Modifier m) { *this = *this * m; return *this; }
 471    Modifier operator==(const Modifier m) const { return m.bits == bits; }
 472    Modifier operator!=(const Modifier m) const { return m.bits != bits; }
 473
 474    inline Modifier operator&(const Modifier m) const { return bits & m.bits; }
 475    inline Modifier operator|(const Modifier m) const { return bits | m.bits; }
 476    inline Modifier operator^(const Modifier m) const { return bits ^ m.bits; }
 477
 478    operation getOp() const;
 479
 480    inline int neg() const { return (bits & NV50_IR_MOD_NEG) ? 1 : 0; }
 481    inline int abs() const { return (bits & NV50_IR_MOD_ABS) ? 1 : 0; }
 482
 483    inline operator bool() const { return bits ? true : false; }
 484
 485    void applyTo(ImmediateValue &imm) const;
 486
 487    int print(char *buf, size_t size) const;
 488
 489 private:
 490    uint8_t bits;
 491 };
 492
 493 class ValueRef
 494 {
 495 public:
 496    ValueRef(Value * = NULL);
 497    ValueRef(const ValueRef&);
 498    ~ValueRef();
 499
 500    inline bool exists() const { return value != NULL; }
 501
 502    void set(Value *);
 503    void set(const ValueRef&);
 504    inline Value *get() const { return value; }
 505    inline Value *rep() const;
 506
 507    inline Instruction *getInsn() const { return insn; }
 508    inline void setInsn(Instruction *inst) { insn = inst; }
 509
 510    inline bool isIndirect(int dim) const { return indirect[dim] >= 0; }
 511    inline const ValueRef *getIndirect(int dim) const;
 512
 513    inline DataFile getFile() const;
 514    inline unsigned getSize() const;
 515
 516    // SSA: return eventual (traverse MOVs) literal value, if it exists
 517    bool getImmediate(ImmediateValue&) const;
 518
 519 public:
 520    Modifier mod;
 521    int8_t indirect[2]; // >= 0 if relative to lvalue in insn->src(indirect[i])
 522    uint8_t swizzle;
 523
 524    bool usedAsPtr; // for printing
 525
 526 private:
 527    Value *value;
 528    Instruction *insn;
 529 };
 530
 531 class ValueDef
 532 {
 533 public:
 534    ValueDef(Value * = NULL);
 535    ValueDef(const ValueDef&);
 536    ~ValueDef();
 537
 538    inline bool exists() const { return value != NULL; }
 539
 540    inline Value *get() const { return value; }
 541    inline Value *rep() const;
 542    void set(Value *);
 543    bool mayReplace(const ValueRef &);
 544    void replace(const ValueRef &, bool doSet); // replace all uses of the old value
 545
 546    inline Instruction *getInsn() const { return insn; }
 547    inline void setInsn(Instruction *inst) { insn = inst; }
 548
 549    inline DataFile getFile() const;
 550    inline unsigned getSize() const;
 551
 552    inline void setSSA(LValue *);
 553    inline const LValue *preSSA() const;
 554
 555 private:
 556    Value *value;   // should make this LValue * ...
 557    LValue *origin; // pre SSA value
 558    Instruction *insn;
 559 };
 560
 561 class Value
 562 {
 563 public:
 564    Value();
 565    virtual ~Value() { }
 566
 567    virtual Value *clone(ClonePolicy<Function>&) const = 0;
 568
 569    virtual int print(char *, size_t, DataType ty = TYPE_NONE) const = 0;
 570
 571    virtual bool equals(const Value *, bool strict = false) const;
 572    virtual bool interfers(const Value *) const;
 573    virtual bool isUniform() const { return true; }
 574
 575    inline Value *rep() const { return join; }
 576
 577    inline Instruction *getUniqueInsn() const;
 578    inline Instruction *getInsn() const; // use when uniqueness is certain
 579
 580    inline int refCount() { return uses.size(); }
 581
 582    inline LValue *asLValue();
 583    inline Symbol *asSym();
 584    inline ImmediateValue *asImm();
 585    inline const Symbol *asSym() const;
 586    inline const ImmediateValue *asImm() const;
 587
 588    inline bool inFile(DataFile f) { return reg.file == f; }
 589
 590    static inline Value *get(Iterator&);
 591
 592    unordered_set<ValueRef *> uses;
 593    std::list<ValueDef *> defs;
 594    typedef unordered_set<ValueRef *>::iterator UseIterator;
 595    typedef unordered_set<ValueRef *>::const_iterator UseCIterator;
 596    typedef std::list<ValueDef *>::iterator DefIterator;
 597    typedef std::list<ValueDef *>::const_iterator DefCIterator;
 598
 599    int id;
 600    Storage reg;
 601
 602    // TODO: these should be in LValue:
 603    Interval livei;
 604    Value *join;
 605 };
 606
 607 class LValue : public Value
 608 {
 609 public:
 610    LValue(Function *, DataFile file);
 611    LValue(Function *, LValue *);
 612    ~LValue() { }
 613
 614    virtual bool isUniform() const;
 615
 616    virtual LValue *clone(ClonePolicy<Function>&) const;
 617
 618    virtual int print(char *, size_t, DataType ty = TYPE_NONE) const;
 619
 620 public:
 621    unsigned compMask : 8; // compound/component mask
 622    unsigned compound : 1; // used by RA, value involved in split/merge
 623    unsigned ssa      : 1;
 624    unsigned fixedReg : 1; // set & used by RA, earlier just use (id < 0)
 625    unsigned noSpill  : 1; // do not spill (e.g. if spill temporary already)
 626 };
 627
 628 class Symbol : public Value
 629 {
 630 public:
 631    Symbol(Program *, DataFile file = FILE_MEMORY_CONST, ubyte fileIdx = 0);
 632    ~Symbol() { }
 633
 634    virtual Symbol *clone(ClonePolicy<Function>&) const;
 635
 636    virtual bool equals(const Value *that, bool strict) const;
 637
 638    virtual bool isUniform() const;
 639
 640    virtual int print(char *, size_t, DataType ty = TYPE_NONE) const;
 641
 642    // print with indirect values
 643    int print(char *, size_t, Value *, Value *, DataType ty = TYPE_NONE) const;
 644
 645    inline void setFile(DataFile file, ubyte fileIndex = 0)
 646    {
 647       reg.file = file;
 648       reg.fileIndex = fileIndex;
 649    }
 650
 651    inline void setOffset(int32_t offset);
 652    inline void setAddress(Symbol *base, int32_t offset);
 653    inline void setSV(SVSemantic sv, uint32_t idx = 0);
 654
 655    inline const Symbol *getBase() const { return baseSym; }
 656
 657 private:
 658    Symbol *baseSym; // array base for Symbols representing array elements
 659 };
 660
 661 class ImmediateValue : public Value
 662 {
 663 public:
 664    ImmediateValue() { }
 665    ImmediateValue(Program *, uint32_t);
 666    ImmediateValue(Program *, float);
 667    ImmediateValue(Program *, double);
 668    // NOTE: not added to program with
 669    ImmediateValue(const ImmediateValue *, DataType ty);
 670    ~ImmediateValue() { };
 671
 672    virtual ImmediateValue *clone(ClonePolicy<Function>&) const;
 673
 674    virtual bool equals(const Value *that, bool strict) const;
 675
 676    // these only work if 'type' is valid (we mostly use untyped literals):
 677    bool isInteger(const int ival) const; // ival is cast to this' type
 678    bool isNegative() const;
 679    bool isPow2() const;
 680
 681    void applyLog2();
 682
 683    // for constant folding:
 684    ImmediateValue operator+(const ImmediateValue&) const;
 685    ImmediateValue operator-(const ImmediateValue&) const;
 686    ImmediateValue operator*(const ImmediateValue&) const;
 687    ImmediateValue operator/(const ImmediateValue&) const;
 688
 689    ImmediateValue& operator=(const ImmediateValue&); // only sets value !
 690
 691    bool compare(CondCode cc, float fval) const;
 692
 693    virtual int print(char *, size_t, DataType ty = TYPE_NONE) const;
 694 };
 695
 696 class Instruction
 697 {
 698 public:
 699    Instruction();
 700    Instruction(Function *, operation, DataType);
 701    virtual ~Instruction();
 702
 703    virtual Instruction *clone(ClonePolicy<Function>&,
 704                               Instruction * = NULL) const;
 705
 706    void setDef(int i, Value *);
 707    void setSrc(int s, Value *);
 708    void setSrc(int s, const ValueRef&);
 709    void swapSources(int a, int b);
 710    void moveSources(int s, int delta);
 711    bool setIndirect(int s, int dim, Value *);
 712
 713    inline ValueRef& src(int s) { return srcs[s]; }
 714    inline ValueDef& def(int s) { return defs[s]; }
 715    inline const ValueRef& src(int s) const { return srcs[s]; }
 716    inline const ValueDef& def(int s) const { return defs[s]; }
 717
 718    inline Value *getDef(int d) const { return defs[d].get(); }
 719    inline Value *getSrc(int s) const { return srcs[s].get(); }
 720    inline Value *getIndirect(int s, int dim) const;
 721
 722    inline bool defExists(unsigned d) const
 723    {
 724       return d < defs.size() && defs[d].exists();
 725    }
 726    inline bool srcExists(unsigned s) const
 727    {
 728       return s < srcs.size() && srcs[s].exists();
 729    }
 730
 731    inline bool constrainedDefs() const;
 732
 733    bool setPredicate(CondCode ccode, Value *);
 734    inline Value *getPredicate() const;
 735    bool writesPredicate() const;
 736    inline bool isPredicated() const { return predSrc >= 0; }
 737
 738    inline void setFlagsSrc(int s, Value *);
 739    inline void setFlagsDef(int d, Value *);
 740    inline bool usesFlags() const { return flagsSrc >= 0; }
 741
 742    unsigned int defCount() const { return defs.size(); };
 743    unsigned int defCount(unsigned int mask, bool singleFile = false) const;
 744    unsigned int srcCount() const { return srcs.size(); };
 745    unsigned int srcCount(unsigned int mask, bool singleFile = false) const;
 746
 747    // save & remove / set indirect[0,1] and predicate source
 748    void takeExtraSources(int s, Value *[3]);
 749    void putExtraSources(int s, Value *[3]);
 750
 751    inline void setType(DataType type) { dType = sType = type; }
 752
 753    inline void setType(DataType dtype, DataType stype)
 754    {
 755       dType = dtype;
 756       sType = stype;
 757    }
 758
 759    inline bool isPseudo() const { return op < OP_MOV; }
 760    bool isDead() const;
 761    bool isNop() const;
 762    bool isCommutationLegal(const Instruction *) const; // must be adjacent !
 763    bool isActionEqual(const Instruction *) const;
 764    bool isResultEqual(const Instruction *) const;
 765
 766    void print() const;
 767
 768    inline CmpInstruction *asCmp();
 769    inline TexInstruction *asTex();
 770    inline FlowInstruction *asFlow();
 771    inline const TexInstruction *asTex() const;
 772    inline const CmpInstruction *asCmp() const;
 773    inline const FlowInstruction *asFlow() const;
 774
 775 public:
 776    Instruction *next;
 777    Instruction *prev;
 778    int id;
 779    int serial; // CFG order
 780
 781    operation op;
 782    DataType dType; // destination or defining type
 783    DataType sType; // source or secondary type
 784    CondCode cc;
 785    RoundMode rnd;
 786    CacheMode cache;
 787
 788    uint16_t subOp; // quadop, 1 for mul-high, etc.
 789
 790    unsigned encSize    : 4; // encoding size in bytes
 791    unsigned saturate   : 1; // to [0.0f, 1.0f]
 792    unsigned join       : 1; // converge control flow (use OP_JOIN until end)
 793    unsigned fixed      : 1; // prevent dead code elimination
 794    unsigned terminator : 1; // end of basic block
 795    unsigned ftz        : 1; // flush denormal to zero
 796    unsigned dnz        : 1; // denormals, NaN are zero
 797    unsigned ipa        : 4; // interpolation mode
 798    unsigned lanes      : 4;
 799    unsigned perPatch   : 1;
 800    unsigned exit       : 1; // terminate program after insn
 801    unsigned mask       : 4; // for vector ops
 802
 803    int8_t postFactor; // MUL/DIV(if < 0) by 1 << postFactor
 804
 805    int8_t predSrc;
 806    int8_t flagsDef;
 807    int8_t flagsSrc;
 808
 809    uint32_t sched; // scheduling data (NOTE: maybe move to separate storage)
 810
 811    BasicBlock *bb;
 812
 813 protected:
 814    std::deque<ValueDef> defs; // no gaps !
 815    std::deque<ValueRef> srcs; // no gaps !
 816
 817    // instruction specific methods:
 818    // (don't want to subclass, would need more constructors and memory pools)
 819 public:
 820    inline void setInterpolate(unsigned int mode) { ipa = mode; }
 821
 822    unsigned int getInterpMode() const { return ipa & 0x3; }
 823    unsigned int getSampleMode() const { return ipa & 0xc; }
 824
 825 private:
 826    void init();
 827 };
 828
 829 enum TexQuery
 830 {
 831    TXQ_DIMS, /* x, y, z, levels */
 832    TXQ_TYPE, /* ?, ?, samples, ? */
 833    TXQ_SAMPLE_POSITION,
 834    TXQ_FILTER,
 835    TXQ_LOD,
 836    TXQ_WRAP,
 837    TXQ_BORDER_COLOUR
 838 };
 839
 840 class TexInstruction : public Instruction
 841 {
 842 public:
 843    class Target
 844    {
 845    public:
 846       Target(TexTarget targ = TEX_TARGET_2D) : target(targ) { }
 847
 848       const char *getName() const { return descTable[target].name; }
 849       unsigned int getArgCount() const { return descTable[target].argc; }
 850       unsigned int getDim() const { return descTable[target].dim; }
 851       int isArray() const { return descTable[target].array ? 1 : 0; }
 852       int isCube() const { return descTable[target].cube ? 1 : 0; }
 853       int isShadow() const { return descTable[target].shadow ? 1 : 0; }
 854       int isMS() const {
 855         return target == TEX_TARGET_2D_MS || target == TEX_TARGET_2D_MS_ARRAY; }
 856       void clearMS() {
 857          if (isMS()) {
 858             if (isArray())
 859                target = TEX_TARGET_2D_ARRAY;
 860             else
 861                target = TEX_TARGET_2D;
 862          }
 863       }
 864
 865       Target& operator=(TexTarget targ)
 866       {
 867          assert(targ < TEX_TARGET_COUNT);
 868          target = targ;
 869          return *this;
 870       }
 871
 872       inline bool operator==(TexTarget targ) const { return target == targ; }
 873       inline bool operator!=(TexTarget targ) const { return target != targ; }
 874
 875       enum TexTarget getEnum() const { return target; }
 876
 877    private:
 878       struct Desc
 879       {
 880          char name[19];
 881          uint8_t dim;
 882          uint8_t argc;
 883          bool array;
 884          bool cube;
 885          bool shadow;
 886       };
 887
 888       static const struct Desc descTable[TEX_TARGET_COUNT];
 889
 890    private:
 891       enum TexTarget target;
 892    };
 893
 894 public:
 895    TexInstruction(Function *, operation);
 896    virtual ~TexInstruction();
 897
 898    virtual TexInstruction *clone(ClonePolicy<Function>&,
 899                                  Instruction * = NULL) const;
 900
 901    inline void setTexture(Target targ, uint8_t r, uint8_t s)
 902    {
 903       tex.r = r;
 904       tex.s = s;
 905       tex.target = targ;
 906    }
 907
 908    void setIndirectR(Value *);
 909    void setIndirectS(Value *);
 910    inline Value *getIndirectR() const;
 911    inline Value *getIndirectS() const;
 912
 913 public:
 914    struct {
 915       Target target;
 916
 917       uint16_t r;
 918       uint16_t s;
 919       int8_t rIndirectSrc;
 920       int8_t sIndirectSrc;
 921
 922       uint8_t mask;
 923       uint8_t gatherComp;
 924
 925       bool liveOnly; // only execute on live pixels of a quad (optimization)
 926       bool levelZero;
 927       bool derivAll;
 928
 929       int8_t useOffsets; // 0, 1, or 4 for textureGatherOffsets
 930       int8_t offset[3]; // only used on nv50
 931
 932       enum TexQuery query;
 933    } tex;
 934
 935    ValueRef dPdx[3];
 936    ValueRef dPdy[3];
 937    ValueRef offset[4][3];
 938 };
 939
 940 class CmpInstruction : public Instruction
 941 {
 942 public:
 943    CmpInstruction(Function *, operation);
 944
 945    virtual CmpInstruction *clone(ClonePolicy<Function>&,
 946                                  Instruction * = NULL) const;
 947
 948    void setCondition(CondCode cond) { setCond = cond; }
 949    CondCode getCondition() const { return setCond; }
 950
 951 public:
 952    CondCode setCond;
 953 };
 954
 955 class FlowInstruction : public Instruction
 956 {
 957 public:
 958    FlowInstruction(Function *, operation, void *target);
 959
 960    virtual FlowInstruction *clone(ClonePolicy<Function>&,
 961                                   Instruction * = NULL) const;
 962
 963 public:
 964    unsigned allWarp  : 1;
 965    unsigned absolute : 1;
 966    unsigned limit    : 1;
 967    unsigned builtin  : 1; // true for calls to emulation code
 968    unsigned indirect : 1; // target in src(0)
 969
 970    union {
 971       BasicBlock *bb;
 972       int builtin;
 973       Function *fn;
 974    } target;
 975 };
 976
 977 class BasicBlock
 978 {
 979 public:
 980    BasicBlock(Function *);
 981    ~BasicBlock();
 982
 983    BasicBlock *clone(ClonePolicy<Function>&) const;
 984
 985    inline int getId() const { return id; }
 986    inline unsigned int getInsnCount() const { return numInsns; }
 987    inline bool isTerminated() const { return exit && exit->terminator; }
 988
 989    bool dominatedBy(BasicBlock *bb);
 990    inline bool reachableBy(const BasicBlock *by, const BasicBlock *term);
 991
 992    // returns mask of conditional out blocks
 993    // e.g. 3 for IF { .. } ELSE { .. } ENDIF, 1 for IF { .. } ENDIF
 994    unsigned int initiatesSimpleConditional() const;
 995
 996 public:
 997    Function *getFunction() const { return func; }
 998    Program *getProgram() const { return program; }
 999
1000    Instruction *getEntry() const { return entry; } // first non-phi instruction
1001    Instruction *getPhi() const { return phi; }
1002    Instruction *getFirst() const { return phi ? phi : entry; }
1003    Instruction *getExit() const { return exit; }
1004
1005    void insertHead(Instruction *);
1006    void insertTail(Instruction *);
1007    void insertBefore(Instruction *, Instruction *);
1008    void insertAfter(Instruction *, Instruction *);
1009    void remove(Instruction *);
1010    void permuteAdjacent(Instruction *, Instruction *);
1011
1012    BasicBlock *idom() const;
1013
1014    // NOTE: currently does not rebuild the dominator tree
1015    BasicBlock *splitBefore(Instruction *, bool attach = true);
1016    BasicBlock *splitAfter(Instruction *, bool attach = true);
1017
1018    DLList& getDF() { return df; }
1019    DLList::Iterator iterDF() { return df.iterator(); }
1020
1021    static inline BasicBlock *get(Iterator&);
1022    static inline BasicBlock *get(Graph::Node *);
1023
1024 public:
1025    Graph::Node cfg; // first edge is branch *taken* (the ELSE branch)
1026    Graph::Node dom;
1027
1028    BitSet liveSet;
1029    BitSet defSet;
1030
1031    uint32_t binPos;
1032    uint32_t binSize;
1033
1034    Instruction *joinAt; // for quick reference
1035
1036    bool explicitCont; // loop headers: true if loop contains continue stmts
1037
1038 private:
1039    int id;
1040    DLList df;
1041
1042    Instruction *phi;
1043    Instruction *entry;
1044    Instruction *exit;
1045
1046    unsigned int numInsns;
1047
1048 private:
1049    Function *func;
1050    Program *program;
1051
1052    void splitCommon(Instruction *, BasicBlock *, bool attach);
1053 };
1054
1055 class Function
1056 {
1057 public:
1058    Function(Program *, const char *name, uint32_t label);
1059    ~Function();
1060
1061    static inline Function *get(Graph::Node *node);
1062
1063    inline Program *getProgram() const { return prog; }
1064    inline const char *getName() const { return name; }
1065    inline int getId() const { return id; }
1066    inline uint32_t getLabel() const { return label; }
1067
1068    void print();
1069    void printLiveIntervals() const;
1070    void printCFGraph(const char *filePath);
1071
1072    bool setEntry(BasicBlock *);
1073    bool setExit(BasicBlock *);
1074
1075    unsigned int orderInstructions(ArrayList&);
1076
1077    inline void add(BasicBlock *bb, int& id) { allBBlocks.insert(bb, id); }
1078    inline void add(Instruction *insn, int& id) { allInsns.insert(insn, id); }
1079    inline void add(LValue *lval, int& id) { allLValues.insert(lval, id); }
1080
1081    inline LValue *getLValue(int id);
1082
1083    void buildLiveSets();
1084    void buildDefSets();
1085    bool convertToSSA();
1086
1087 public:
1088    std::deque<ValueDef> ins;
1089    std::deque<ValueRef> outs;
1090    std::deque<Value *> clobbers;
1091
1092    Graph cfg;
1093    Graph::Node *cfgExit;
1094    Graph *domTree;
1095    Graph::Node call; // node in the call graph
1096
1097    BasicBlock **bbArray; // BBs in emission order
1098    int bbCount;
1099
1100    unsigned int loopNestingBound;
1101    int regClobberMax;
1102
1103    uint32_t binPos;
1104    uint32_t binSize;
1105
1106    Value *stackPtr;
1107
1108    uint32_t tlsBase; // base address for l[] space (if no stack pointer is used)
1109    uint32_t tlsSize;
1110
1111    ArrayList allBBlocks;
1112    ArrayList allInsns;
1113    ArrayList allLValues;
1114
1115 private:
1116    void buildLiveSetsPreSSA(BasicBlock *, const int sequence);
1117    void buildDefSetsPreSSA(BasicBlock *bb, const int seq);
1118
1119 private:
1120    uint32_t label;
1121    int id;
1122    const char *const name;
1123    Program *prog;
1124 };
1125
1126 enum CGStage
1127 {
1128    CG_STAGE_PRE_SSA,
1129    CG_STAGE_SSA, // expected directly before register allocation
1130    CG_STAGE_POST_RA
1131 };
1132
1133 class Program
1134 {
1135 public:
1136    enum Type
1137    {
1138       TYPE_VERTEX,
1139       TYPE_TESSELLATION_CONTROL,
1140       TYPE_TESSELLATION_EVAL,
1141       TYPE_GEOMETRY,
1142       TYPE_FRAGMENT,
1143       TYPE_COMPUTE
1144    };
1145
1146    Program(Type type, Target *targ);
1147    ~Program();
1148
1149    void print();
1150
1151    Type getType() const { return progType; }
1152
1153    inline void add(Function *fn, int& id) { allFuncs.insert(fn, id); }
1154    inline void del(Function *fn, int& id) { allFuncs.remove(id); }
1155    inline void add(Value *rval, int& id) { allRValues.insert(rval, id); }
1156
1157    bool makeFromTGSI(struct nv50_ir_prog_info *);
1158    bool makeFromSM4(struct nv50_ir_prog_info *);
1159    bool convertToSSA();
1160    bool optimizeSSA(int level);
1161    bool optimizePostRA(int level);
1162    bool registerAllocation();
1163    bool emitBinary(struct nv50_ir_prog_info *);
1164
1165    const Target *getTarget() const { return target; }
1166
1167 private:
1168    void emitSymbolTable(struct nv50_ir_prog_info *);
1169
1170    Type progType;
1171    Target *target;
1172
1173 public:
1174    Function *main;
1175    Graph calls;
1176
1177    ArrayList allFuncs;
1178    ArrayList allRValues;
1179
1180    uint32_t *code;
1181    uint32_t binSize;
1182    uint32_t tlsSize; // size required for FILE_MEMORY_LOCAL
1183
1184    int maxGPR;
1185
1186    MemoryPool mem_Instruction;
1187    MemoryPool mem_CmpInstruction;
1188    MemoryPool mem_TexInstruction;
1189    MemoryPool mem_FlowInstruction;
1190    MemoryPool mem_LValue;
1191    MemoryPool mem_Symbol;
1192    MemoryPool mem_ImmediateValue;
1193
1194    uint32_t dbgFlags;
1195    uint8_t  optLevel;
1196
1197    void *targetPriv; // e.g. to carry information between passes
1198
1199    const struct nv50_ir_prog_info *driver; // for driver configuration
1200
1201    void releaseInstruction(Instruction *);
1202    void releaseValue(Value *);
1203 };
1204
1205 // TODO: add const version
1206 class Pass
1207 {
1208 public:
1209    bool run(Program *, bool ordered = false, bool skipPhi = false);
1210    bool run(Function *, bool ordered = false, bool skipPhi = false);
1211
1212 private:
1213    // return false to continue with next entity on next higher level
1214    virtual bool visit(Function *) { return true; }
1215    virtual bool visit(BasicBlock *) { return true; }
1216    virtual bool visit(Instruction *) { return false; }
1217
1218    bool doRun(Program *, bool ordered, bool skipPhi);
1219    bool doRun(Function *, bool ordered, bool skipPhi);
1220
1221 protected:
1222    bool err;
1223    Function *func;
1224    Program *prog;
1225 };
1226
1227 // =============================================================================
1228
1229 #include "codegen/nv50_ir_inlines.h"
1230
1231 } // namespace nv50_ir
1232
1233 #endif // __NV50_IR_H__