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_SUQ,     // surface query
 136    OP_MADSP,   // special integer multiply-add
 137    OP_TEXBAR, // texture dependency barrier
 138    OP_DFDX,
 139    OP_DFDY,
 140    OP_RDSV, // read system value
 141    OP_WRSV, // write system value
 142    OP_PIXLD, // get info about raster object or surfaces
 143    OP_QUADOP,
 144    OP_QUADON,
 145    OP_QUADPOP,
 146    OP_POPCNT, // bitcount(src0 & src1)
 147    OP_INSBF,  // insert first src1[8:15] bits of src0 into src2 at src1[0:7]
 148    OP_EXTBF,  // place bits [K,K+N) of src0 into dst, src1 = 0xNNKK
 149    OP_BFIND,  // find highest/lowest set bit
 150    OP_PERMT,  // dst = bytes from src2,src0 selected by src1 (nvc0's src order)
 151    OP_ATOM,
 152    OP_BAR,    // execution barrier, sources = { id, thread count, predicate }
 153    OP_VADD,   // byte/word vector operations
 154    OP_VAVG,
 155    OP_VMIN,
 156    OP_VMAX,
 157    OP_VSAD,
 158    OP_VSET,
 159    OP_VSHR,
 160    OP_VSHL,
 161    OP_VSEL,
 162    OP_CCTL, // cache control
 163    OP_SHFL, // warp shuffle
 164    OP_VOTE,
 165    OP_LAST
 166 };
 167
 168 // various instruction-specific modifier definitions Instruction::subOp
 169 // MOV_FINAL marks a MOV originating from an EXPORT (used for placing TEXBARs)
 170 #define NV50_IR_SUBOP_MUL_HIGH     1
 171 #define NV50_IR_SUBOP_EMIT_RESTART 1
 172 #define NV50_IR_SUBOP_LDC_IL       1
 173 #define NV50_IR_SUBOP_LDC_IS       2
 174 #define NV50_IR_SUBOP_LDC_ISL      3
 175 #define NV50_IR_SUBOP_SHIFT_WRAP   1
 176 #define NV50_IR_SUBOP_EMU_PRERET   1
 177 #define NV50_IR_SUBOP_TEXBAR(n)    n
 178 #define NV50_IR_SUBOP_MOV_FINAL    1
 179 #define NV50_IR_SUBOP_EXTBF_REV    1
 180 #define NV50_IR_SUBOP_BFIND_SAMT   1
 181 #define NV50_IR_SUBOP_RCPRSQ_64H   1
 182 #define NV50_IR_SUBOP_PERMT_F4E    1
 183 #define NV50_IR_SUBOP_PERMT_B4E    2
 184 #define NV50_IR_SUBOP_PERMT_RC8    3
 185 #define NV50_IR_SUBOP_PERMT_ECL    4
 186 #define NV50_IR_SUBOP_PERMT_ECR    5
 187 #define NV50_IR_SUBOP_PERMT_RC16   6
 188 #define NV50_IR_SUBOP_BAR_SYNC     0
 189 #define NV50_IR_SUBOP_BAR_ARRIVE   1
 190 #define NV50_IR_SUBOP_BAR_RED_AND  2
 191 #define NV50_IR_SUBOP_BAR_RED_OR   3
 192 #define NV50_IR_SUBOP_BAR_RED_POPC 4
 193 #define NV50_IR_SUBOP_MEMBAR_L     1
 194 #define NV50_IR_SUBOP_MEMBAR_S     2
 195 #define NV50_IR_SUBOP_MEMBAR_M     3
 196 #define NV50_IR_SUBOP_MEMBAR_CTA  (0 << 2)
 197 #define NV50_IR_SUBOP_MEMBAR_GL   (1 << 2)
 198 #define NV50_IR_SUBOP_MEMBAR_SYS  (2 << 2)
 199 #define NV50_IR_SUBOP_MEMBAR_DIR(m)   ((m) & 0x3)
 200 #define NV50_IR_SUBOP_MEMBAR_SCOPE(m) ((m) & ~0x3)
 201 #define NV50_IR_SUBOP_MEMBAR(d,s) \
 202    (NV50_IR_SUBOP_MEMBAR_##d | NV50_IR_SUBOP_MEMBAR_##s)
 203 #define NV50_IR_SUBOP_ATOM_ADD      0
 204 #define NV50_IR_SUBOP_ATOM_MIN      1
 205 #define NV50_IR_SUBOP_ATOM_MAX      2
 206 #define NV50_IR_SUBOP_ATOM_INC      3
 207 #define NV50_IR_SUBOP_ATOM_DEC      4
 208 #define NV50_IR_SUBOP_ATOM_AND      5
 209 #define NV50_IR_SUBOP_ATOM_OR       6
 210 #define NV50_IR_SUBOP_ATOM_XOR      7
 211 #define NV50_IR_SUBOP_ATOM_CAS      8
 212 #define NV50_IR_SUBOP_ATOM_EXCH     9
 213 #define NV50_IR_SUBOP_CCTL_IV      5
 214 #define NV50_IR_SUBOP_CCTL_IVALL   6
 215 #define NV50_IR_SUBOP_SUST_IGN     0
 216 #define NV50_IR_SUBOP_SUST_TRAP    1
 217 #define NV50_IR_SUBOP_SUST_SDCL    3
 218 #define NV50_IR_SUBOP_SULD_ZERO    0
 219 #define NV50_IR_SUBOP_SULD_TRAP    1
 220 #define NV50_IR_SUBOP_SULD_SDCL    3
 221 #define NV50_IR_SUBOP_SUBFM_3D     1
 222 #define NV50_IR_SUBOP_SUCLAMP_2D   0x10
 223 #define NV50_IR_SUBOP_SUCLAMP_SD(r, d) (( 0 + (r)) | ((d == 2) ? 0x10 : 0))
 224 #define NV50_IR_SUBOP_SUCLAMP_PL(r, d) (( 5 + (r)) | ((d == 2) ? 0x10 : 0))
 225 #define NV50_IR_SUBOP_SUCLAMP_BL(r, d) ((10 + (r)) | ((d == 2) ? 0x10 : 0))
 226 #define NV50_IR_SUBOP_PIXLD_COUNT       0
 227 #define NV50_IR_SUBOP_PIXLD_COVMASK     1
 228 #define NV50_IR_SUBOP_PIXLD_COVERED     2
 229 #define NV50_IR_SUBOP_PIXLD_OFFSET      3
 230 #define NV50_IR_SUBOP_PIXLD_CENT_OFFSET 4
 231 #define NV50_IR_SUBOP_PIXLD_SAMPLEID    5
 232 #define NV50_IR_SUBOP_SHFL_IDX  0
 233 #define NV50_IR_SUBOP_SHFL_UP   1
 234 #define NV50_IR_SUBOP_SHFL_DOWN 2
 235 #define NV50_IR_SUBOP_SHFL_BFLY 3
 236 #define NV50_IR_SUBOP_LOAD_LOCKED    1
 237 #define NV50_IR_SUBOP_STORE_UNLOCKED 2
 238 #define NV50_IR_SUBOP_MADSP_SD     0xffff
 239 // Yes, we could represent those with DataType.
 240 // Or put the type into operation and have a couple 1000 values in that enum.
 241 // This will have to do for now.
 242 // The bitfields are supposed to correspond to nve4 ISA.
 243 #define NV50_IR_SUBOP_MADSP(a,b,c) (((c) << 8) | ((b) << 4) | (a))
 244 #define NV50_IR_SUBOP_V1(d,a,b)    (((d) << 10) | ((b) << 5) | (a) | 0x0000)
 245 #define NV50_IR_SUBOP_V2(d,a,b)    (((d) << 10) | ((b) << 5) | (a) | 0x4000)
 246 #define NV50_IR_SUBOP_V4(d,a,b)    (((d) << 10) | ((b) << 5) | (a) | 0x8000)
 247 #define NV50_IR_SUBOP_Vn(n)        ((n) >> 14)
 248 #define NV50_IR_SUBOP_VOTE_ALL 0
 249 #define NV50_IR_SUBOP_VOTE_ANY 1
 250 #define NV50_IR_SUBOP_VOTE_UNI 2
 251
 252 enum DataType
 253 {
 254    TYPE_NONE,
 255    TYPE_U8,
 256    TYPE_S8,
 257    TYPE_U16,
 258    TYPE_S16,
 259    TYPE_U32,
 260    TYPE_S32,
 261    TYPE_U64, // 64 bit operations are only lowered after register allocation
 262    TYPE_S64,
 263    TYPE_F16,
 264    TYPE_F32,
 265    TYPE_F64,
 266    TYPE_B96,
 267    TYPE_B128
 268 };
 269
 270 enum CondCode
 271 {
 272    CC_FL = 0,
 273    CC_NEVER = CC_FL, // when used with FILE_FLAGS
 274    CC_LT = 1,
 275    CC_EQ = 2,
 276    CC_NOT_P = CC_EQ, // when used with FILE_PREDICATE
 277    CC_LE = 3,
 278    CC_GT = 4,
 279    CC_NE = 5,
 280    CC_P  = CC_NE,
 281    CC_GE = 6,
 282    CC_TR = 7,
 283    CC_ALWAYS = CC_TR,
 284    CC_U  = 8,
 285    CC_LTU = 9,
 286    CC_EQU = 10,
 287    CC_LEU = 11,
 288    CC_GTU = 12,
 289    CC_NEU = 13,
 290    CC_GEU = 14,
 291    CC_NO = 0x10,
 292    CC_NC = 0x11,
 293    CC_NS = 0x12,
 294    CC_NA = 0x13,
 295    CC_A  = 0x14,
 296    CC_S  = 0x15,
 297    CC_C  = 0x16,
 298    CC_O  = 0x17
 299 };
 300
 301 enum RoundMode
 302 {
 303    ROUND_N, // nearest
 304    ROUND_M, // towards -inf
 305    ROUND_Z, // towards 0
 306    ROUND_P, // towards +inf
 307    ROUND_NI, // nearest integer
 308    ROUND_MI, // to integer towards -inf
 309    ROUND_ZI, // to integer towards 0
 310    ROUND_PI, // to integer towards +inf
 311 };
 312
 313 enum CacheMode
 314 {
 315    CACHE_CA,            // cache at all levels
 316    CACHE_WB = CACHE_CA, // cache write back
 317    CACHE_CG,            // cache at global level
 318    CACHE_CS,            // cache streaming
 319    CACHE_CV,            // cache as volatile
 320    CACHE_WT = CACHE_CV  // cache write-through
 321 };
 322
 323 enum DataFile
 324 {
 325    FILE_NULL = 0,
 326    FILE_GPR,
 327    FILE_PREDICATE,       // boolean predicate
 328    FILE_FLAGS,           // zero/sign/carry/overflow bits
 329    FILE_ADDRESS,
 330    LAST_REGISTER_FILE = FILE_ADDRESS,
 331    FILE_IMMEDIATE,
 332    FILE_MEMORY_CONST,
 333    FILE_SHADER_INPUT,
 334    FILE_SHADER_OUTPUT,
 335    FILE_MEMORY_GLOBAL,
 336    FILE_MEMORY_SHARED,
 337    FILE_MEMORY_LOCAL,
 338    FILE_SYSTEM_VALUE,
 339    DATA_FILE_COUNT
 340 };
 341
 342 enum TexTarget
 343 {
 344    TEX_TARGET_1D,
 345    TEX_TARGET_2D,
 346    TEX_TARGET_2D_MS,
 347    TEX_TARGET_3D,
 348    TEX_TARGET_CUBE,
 349    TEX_TARGET_1D_SHADOW,
 350    TEX_TARGET_2D_SHADOW,
 351    TEX_TARGET_CUBE_SHADOW,
 352    TEX_TARGET_1D_ARRAY,
 353    TEX_TARGET_2D_ARRAY,
 354    TEX_TARGET_2D_MS_ARRAY,
 355    TEX_TARGET_CUBE_ARRAY,
 356    TEX_TARGET_1D_ARRAY_SHADOW,
 357    TEX_TARGET_2D_ARRAY_SHADOW,
 358    TEX_TARGET_RECT,
 359    TEX_TARGET_RECT_SHADOW,
 360    TEX_TARGET_CUBE_ARRAY_SHADOW,
 361    TEX_TARGET_BUFFER,
 362    TEX_TARGET_COUNT
 363 };
 364
 365 enum SVSemantic
 366 {
 367    SV_POSITION, // WPOS
 368    SV_VERTEX_ID,
 369    SV_INSTANCE_ID,
 370    SV_INVOCATION_ID,
 371    SV_PRIMITIVE_ID,
 372    SV_VERTEX_COUNT, // gl_PatchVerticesIn
 373    SV_LAYER,
 374    SV_VIEWPORT_INDEX,
 375    SV_YDIR,
 376    SV_FACE,
 377    SV_POINT_SIZE,
 378    SV_POINT_COORD,
 379    SV_CLIP_DISTANCE,
 380    SV_SAMPLE_INDEX,
 381    SV_SAMPLE_POS,
 382    SV_SAMPLE_MASK,
 383    SV_TESS_OUTER,
 384    SV_TESS_INNER,
 385    SV_TESS_COORD,
 386    SV_TID,
 387    SV_CTAID,
 388    SV_NTID,
 389    SV_GRIDID,
 390    SV_NCTAID,
 391    SV_LANEID,
 392    SV_PHYSID,
 393    SV_NPHYSID,
 394    SV_CLOCK,
 395    SV_LBASE,
 396    SV_SBASE,
 397    SV_VERTEX_STRIDE,
 398    SV_INVOCATION_INFO,
 399    SV_THREAD_KILL,
 400    SV_BASEVERTEX,
 401    SV_BASEINSTANCE,
 402    SV_DRAWID,
 403    SV_UNDEFINED,
 404    SV_LAST
 405 };
 406
 407 class Program;
 408 class Function;
 409 class BasicBlock;
 410
 411 class Target;
 412
 413 class Instruction;
 414 class CmpInstruction;
 415 class TexInstruction;
 416 class FlowInstruction;
 417
 418 class Value;
 419 class LValue;
 420 class Symbol;
 421 class ImmediateValue;
 422
 423 struct Storage
 424 {
 425    DataFile file;
 426    int8_t fileIndex; // signed, may be indirect for CONST[]
 427    uint8_t size; // this should match the Instruction type's size
 428    DataType type; // mainly for pretty printing
 429    union {
 430       uint64_t u64;    // immediate values
 431       uint32_t u32;
 432       uint16_t u16;
 433       uint8_t u8;
 434       int64_t s64;
 435       int32_t s32;
 436       int16_t s16;
 437       int8_t s8;
 438       float f32;
 439       double f64;
 440       int32_t offset; // offset from 0 (base of address space)
 441       int32_t id;     // register id (< 0 if virtual/unassigned, in units <= 4)
 442       struct {
 443          SVSemantic sv;
 444          int index;
 445       } sv;
 446    } data;
 447 };
 448
 449 // precedence: NOT after SAT after NEG after ABS
 450 #define NV50_IR_MOD_ABS (1 << 0)
 451 #define NV50_IR_MOD_NEG (1 << 1)
 452 #define NV50_IR_MOD_SAT (1 << 2)
 453 #define NV50_IR_MOD_NOT (1 << 3)
 454 #define NV50_IR_MOD_NEG_ABS (NV50_IR_MOD_NEG | NV50_IR_MOD_ABS)
 455
 456 #define NV50_IR_INTERP_MODE_MASK   0x3
 457 #define NV50_IR_INTERP_LINEAR      (0 << 0)
 458 #define NV50_IR_INTERP_PERSPECTIVE (1 << 0)
 459 #define NV50_IR_INTERP_FLAT        (2 << 0)
 460 #define NV50_IR_INTERP_SC          (3 << 0) // what exactly is that ?
 461 #define NV50_IR_INTERP_SAMPLE_MASK 0xc
 462 #define NV50_IR_INTERP_DEFAULT     (0 << 2)
 463 #define NV50_IR_INTERP_CENTROID    (1 << 2)
 464 #define NV50_IR_INTERP_OFFSET      (2 << 2)
 465 #define NV50_IR_INTERP_SAMPLEID    (3 << 2)
 466
 467 // do we really want this to be a class ?
 468 class Modifier
 469 {
 470 public:
 471    Modifier() : bits(0) { }
 472    Modifier(unsigned int m) : bits(m) { }
 473    Modifier(operation op);
 474
 475    // @return new Modifier applying a after b (asserts if unrepresentable)
 476    Modifier operator*(const Modifier) const;
 477    Modifier operator*=(const Modifier m) { *this = *this * m; return *this; }
 478    Modifier operator==(const Modifier m) const { return m.bits == bits; }
 479    Modifier operator!=(const Modifier m) const { return m.bits != bits; }
 480
 481    inline Modifier operator&(const Modifier m) const { return bits & m.bits; }
 482    inline Modifier operator|(const Modifier m) const { return bits | m.bits; }
 483    inline Modifier operator^(const Modifier m) const { return bits ^ m.bits; }
 484
 485    operation getOp() const;
 486
 487    inline int neg() const { return (bits & NV50_IR_MOD_NEG) ? 1 : 0; }
 488    inline int abs() const { return (bits & NV50_IR_MOD_ABS) ? 1 : 0; }
 489
 490    inline operator bool() const { return bits ? true : false; }
 491
 492    void applyTo(ImmediateValue &imm) const;
 493
 494    int print(char *buf, size_t size) const;
 495
 496 private:
 497    uint8_t bits;
 498 };
 499
 500 class ValueRef
 501 {
 502 public:
 503    ValueRef(Value * = NULL);
 504    ValueRef(const ValueRef&);
 505    ~ValueRef();
 506
 507    inline bool exists() const { return value != NULL; }
 508
 509    void set(Value *);
 510    void set(const ValueRef&);
 511    inline Value *get() const { return value; }
 512    inline Value *rep() const;
 513
 514    inline Instruction *getInsn() const { return insn; }
 515    inline void setInsn(Instruction *inst) { insn = inst; }
 516
 517    inline bool isIndirect(int dim) const { return indirect[dim] >= 0; }
 518    inline const ValueRef *getIndirect(int dim) const;
 519
 520    inline DataFile getFile() const;
 521    inline unsigned getSize() const;
 522
 523    // SSA: return eventual (traverse MOVs) literal value, if it exists
 524    bool getImmediate(ImmediateValue&) const;
 525
 526 public:
 527    Modifier mod;
 528    int8_t indirect[2]; // >= 0 if relative to lvalue in insn->src(indirect[i])
 529    uint8_t swizzle;
 530
 531    bool usedAsPtr; // for printing
 532
 533 private:
 534    Value *value;
 535    Instruction *insn;
 536 };
 537
 538 class ValueDef
 539 {
 540 public:
 541    ValueDef(Value * = NULL);
 542    ValueDef(const ValueDef&);
 543    ~ValueDef();
 544
 545    inline bool exists() const { return value != NULL; }
 546
 547    inline Value *get() const { return value; }
 548    inline Value *rep() const;
 549    void set(Value *);
 550    bool mayReplace(const ValueRef &);
 551    void replace(const ValueRef &, bool doSet); // replace all uses of the old value
 552
 553    inline Instruction *getInsn() const { return insn; }
 554    inline void setInsn(Instruction *inst) { insn = inst; }
 555
 556    inline DataFile getFile() const;
 557    inline unsigned getSize() const;
 558
 559    inline void setSSA(LValue *);
 560    inline const LValue *preSSA() const;
 561
 562 private:
 563    Value *value;   // should make this LValue * ...
 564    LValue *origin; // pre SSA value
 565    Instruction *insn;
 566 };
 567
 568 class Value
 569 {
 570 public:
 571    Value();
 572    virtual ~Value() { }
 573
 574    virtual Value *clone(ClonePolicy<Function>&) const = 0;
 575
 576    virtual int print(char *, size_t, DataType ty = TYPE_NONE) const = 0;
 577
 578    virtual bool equals(const Value *, bool strict = false) const;
 579    virtual bool interfers(const Value *) const;
 580    virtual bool isUniform() const { return true; }
 581
 582    inline Value *rep() const { return join; }
 583
 584    inline Instruction *getUniqueInsn() const;
 585    inline Instruction *getInsn() const; // use when uniqueness is certain
 586
 587    inline int refCount() { return uses.size(); }
 588
 589    inline LValue *asLValue();
 590    inline Symbol *asSym();
 591    inline ImmediateValue *asImm();
 592    inline const Symbol *asSym() const;
 593    inline const ImmediateValue *asImm() const;
 594
 595    inline bool inFile(DataFile f) { return reg.file == f; }
 596
 597    static inline Value *get(Iterator&);
 598
 599    unordered_set<ValueRef *> uses;
 600    std::list<ValueDef *> defs;
 601    typedef unordered_set<ValueRef *>::iterator UseIterator;
 602    typedef unordered_set<ValueRef *>::const_iterator UseCIterator;
 603    typedef std::list<ValueDef *>::iterator DefIterator;
 604    typedef std::list<ValueDef *>::const_iterator DefCIterator;
 605
 606    int id;
 607    Storage reg;
 608
 609    // TODO: these should be in LValue:
 610    Interval livei;
 611    Value *join;
 612 };
 613
 614 class LValue : public Value
 615 {
 616 public:
 617    LValue(Function *, DataFile file);
 618    LValue(Function *, LValue *);
 619    ~LValue() { }
 620
 621    virtual bool isUniform() const;
 622
 623    virtual LValue *clone(ClonePolicy<Function>&) const;
 624
 625    virtual int print(char *, size_t, DataType ty = TYPE_NONE) const;
 626
 627 public:
 628    unsigned compMask : 8; // compound/component mask
 629    unsigned compound : 1; // used by RA, value involved in split/merge
 630    unsigned ssa      : 1;
 631    unsigned fixedReg : 1; // set & used by RA, earlier just use (id < 0)
 632    unsigned noSpill  : 1; // do not spill (e.g. if spill temporary already)
 633 };
 634
 635 class Symbol : public Value
 636 {
 637 public:
 638    Symbol(Program *, DataFile file = FILE_MEMORY_CONST, ubyte fileIdx = 0);
 639    ~Symbol() { }
 640
 641    virtual Symbol *clone(ClonePolicy<Function>&) const;
 642
 643    virtual bool equals(const Value *that, bool strict) const;
 644
 645    virtual bool isUniform() const;
 646
 647    virtual int print(char *, size_t, DataType ty = TYPE_NONE) const;
 648
 649    // print with indirect values
 650    int print(char *, size_t, Value *, Value *, DataType ty = TYPE_NONE) const;
 651
 652    inline void setFile(DataFile file, ubyte fileIndex = 0)
 653    {
 654       reg.file = file;
 655       reg.fileIndex = fileIndex;
 656    }
 657
 658    inline void setOffset(int32_t offset);
 659    inline void setAddress(Symbol *base, int32_t offset);
 660    inline void setSV(SVSemantic sv, uint32_t idx = 0);
 661
 662    inline const Symbol *getBase() const { return baseSym; }
 663
 664 private:
 665    Symbol *baseSym; // array base for Symbols representing array elements
 666 };
 667
 668 class ImmediateValue : public Value
 669 {
 670 public:
 671    ImmediateValue() { }
 672    ImmediateValue(Program *, uint32_t);
 673    ImmediateValue(Program *, float);
 674    ImmediateValue(Program *, double);
 675    // NOTE: not added to program with
 676    ImmediateValue(const ImmediateValue *, DataType ty);
 677    ~ImmediateValue() { };
 678
 679    virtual ImmediateValue *clone(ClonePolicy<Function>&) const;
 680
 681    virtual bool equals(const Value *that, bool strict) const;
 682
 683    // these only work if 'type' is valid (we mostly use untyped literals):
 684    bool isInteger(const int ival) const; // ival is cast to this' type
 685    bool isNegative() const;
 686    bool isPow2() const;
 687
 688    void applyLog2();
 689
 690    // for constant folding:
 691    ImmediateValue operator+(const ImmediateValue&) const;
 692    ImmediateValue operator-(const ImmediateValue&) const;
 693    ImmediateValue operator*(const ImmediateValue&) const;
 694    ImmediateValue operator/(const ImmediateValue&) const;
 695
 696    ImmediateValue& operator=(const ImmediateValue&); // only sets value !
 697
 698    bool compare(CondCode cc, float fval) const;
 699
 700    virtual int print(char *, size_t, DataType ty = TYPE_NONE) const;
 701 };
 702
 703 class Instruction
 704 {
 705 public:
 706    Instruction();
 707    Instruction(Function *, operation, DataType);
 708    virtual ~Instruction();
 709
 710    virtual Instruction *clone(ClonePolicy<Function>&,
 711                               Instruction * = NULL) const;
 712
 713    void setDef(int i, Value *);
 714    void setSrc(int s, Value *);
 715    void setSrc(int s, const ValueRef&);
 716    void swapSources(int a, int b);
 717    void moveSources(int s, int delta);
 718    bool setIndirect(int s, int dim, Value *);
 719
 720    inline ValueRef& src(int s) { return srcs[s]; }
 721    inline ValueDef& def(int s) { return defs[s]; }
 722    inline const ValueRef& src(int s) const { return srcs[s]; }
 723    inline const ValueDef& def(int s) const { return defs[s]; }
 724
 725    inline Value *getDef(int d) const { return defs[d].get(); }
 726    inline Value *getSrc(int s) const { return srcs[s].get(); }
 727    inline Value *getIndirect(int s, int dim) const;
 728
 729    inline bool defExists(unsigned d) const
 730    {
 731       return d < defs.size() && defs[d].exists();
 732    }
 733    inline bool srcExists(unsigned s) const
 734    {
 735       return s < srcs.size() && srcs[s].exists();
 736    }
 737
 738    inline bool constrainedDefs() const;
 739
 740    bool setPredicate(CondCode ccode, Value *);
 741    inline Value *getPredicate() const;
 742    bool writesPredicate() const;
 743    inline bool isPredicated() const { return predSrc >= 0; }
 744
 745    inline void setFlagsSrc(int s, Value *);
 746    inline void setFlagsDef(int d, Value *);
 747    inline bool usesFlags() const { return flagsSrc >= 0; }
 748
 749    unsigned int defCount() const { return defs.size(); };
 750    unsigned int defCount(unsigned int mask, bool singleFile = false) const;
 751    unsigned int srcCount() const { return srcs.size(); };
 752    unsigned int srcCount(unsigned int mask, bool singleFile = false) const;
 753
 754    // save & remove / set indirect[0,1] and predicate source
 755    void takeExtraSources(int s, Value *[3]);
 756    void putExtraSources(int s, Value *[3]);
 757
 758    inline void setType(DataType type) { dType = sType = type; }
 759
 760    inline void setType(DataType dtype, DataType stype)
 761    {
 762       dType = dtype;
 763       sType = stype;
 764    }
 765
 766    inline bool isPseudo() const { return op < OP_MOV; }
 767    bool isDead() const;
 768    bool isNop() const;
 769    bool isCommutationLegal(const Instruction *) const; // must be adjacent !
 770    bool isActionEqual(const Instruction *) const;
 771    bool isResultEqual(const Instruction *) const;
 772
 773    void print() const;
 774
 775    inline CmpInstruction *asCmp();
 776    inline TexInstruction *asTex();
 777    inline FlowInstruction *asFlow();
 778    inline const TexInstruction *asTex() const;
 779    inline const CmpInstruction *asCmp() const;
 780    inline const FlowInstruction *asFlow() const;
 781
 782 public:
 783    Instruction *next;
 784    Instruction *prev;
 785    int id;
 786    int serial; // CFG order
 787
 788    operation op;
 789    DataType dType; // destination or defining type
 790    DataType sType; // source or secondary type
 791    CondCode cc;
 792    RoundMode rnd;
 793    CacheMode cache;
 794
 795    uint16_t subOp; // quadop, 1 for mul-high, etc.
 796
 797    unsigned encSize    : 4; // encoding size in bytes
 798    unsigned saturate   : 1; // to [0.0f, 1.0f]
 799    unsigned join       : 1; // converge control flow (use OP_JOIN until end)
 800    unsigned fixed      : 1; // prevent dead code elimination
 801    unsigned terminator : 1; // end of basic block
 802    unsigned ftz        : 1; // flush denormal to zero
 803    unsigned dnz        : 1; // denormals, NaN are zero
 804    unsigned ipa        : 4; // interpolation mode
 805    unsigned lanes      : 4;
 806    unsigned perPatch   : 1;
 807    unsigned exit       : 1; // terminate program after insn
 808    unsigned mask       : 4; // for vector ops
 809
 810    int8_t postFactor; // MUL/DIV(if < 0) by 1 << postFactor
 811
 812    int8_t predSrc;
 813    int8_t flagsDef;
 814    int8_t flagsSrc;
 815
 816    uint32_t sched; // scheduling data (NOTE: maybe move to separate storage)
 817
 818    BasicBlock *bb;
 819
 820 protected:
 821    std::deque<ValueDef> defs; // no gaps !
 822    std::deque<ValueRef> srcs; // no gaps !
 823
 824    // instruction specific methods:
 825    // (don't want to subclass, would need more constructors and memory pools)
 826 public:
 827    inline void setInterpolate(unsigned int mode) { ipa = mode; }
 828
 829    unsigned int getInterpMode() const { return ipa & 0x3; }
 830    unsigned int getSampleMode() const { return ipa & 0xc; }
 831
 832 private:
 833    void init();
 834 };
 835
 836 enum TexQuery
 837 {
 838    TXQ_DIMS, /* x, y, z, levels */
 839    TXQ_TYPE, /* ?, ?, samples, ? */
 840    TXQ_SAMPLE_POSITION,
 841    TXQ_FILTER,
 842    TXQ_LOD,
 843    TXQ_WRAP,
 844    TXQ_BORDER_COLOUR
 845 };
 846
 847 class TexInstruction : public Instruction
 848 {
 849 public:
 850    class Target
 851    {
 852    public:
 853       Target(TexTarget targ = TEX_TARGET_2D) : target(targ) { }
 854
 855       const char *getName() const { return descTable[target].name; }
 856       unsigned int getArgCount() const { return descTable[target].argc; }
 857       unsigned int getDim() const { return descTable[target].dim; }
 858       int isArray() const { return descTable[target].array ? 1 : 0; }
 859       int isCube() const { return descTable[target].cube ? 1 : 0; }
 860       int isShadow() const { return descTable[target].shadow ? 1 : 0; }
 861       int isMS() const {
 862         return target == TEX_TARGET_2D_MS || target == TEX_TARGET_2D_MS_ARRAY; }
 863       void clearMS() {
 864          if (isMS()) {
 865             if (isArray())
 866                target = TEX_TARGET_2D_ARRAY;
 867             else
 868                target = TEX_TARGET_2D;
 869          }
 870       }
 871
 872       Target& operator=(TexTarget targ)
 873       {
 874          assert(targ < TEX_TARGET_COUNT);
 875          target = targ;
 876          return *this;
 877       }
 878
 879       inline bool operator==(TexTarget targ) const { return target == targ; }
 880       inline bool operator!=(TexTarget targ) const { return target != targ; }
 881
 882       enum TexTarget getEnum() const { return target; }
 883
 884    private:
 885       struct Desc
 886       {
 887          char name[19];
 888          uint8_t dim;
 889          uint8_t argc;
 890          bool array;
 891          bool cube;
 892          bool shadow;
 893       };
 894
 895       static const struct Desc descTable[TEX_TARGET_COUNT];
 896
 897    private:
 898       enum TexTarget target;
 899    };
 900
 901 public:
 902    TexInstruction(Function *, operation);
 903    virtual ~TexInstruction();
 904
 905    virtual TexInstruction *clone(ClonePolicy<Function>&,
 906                                  Instruction * = NULL) const;
 907
 908    inline void setTexture(Target targ, uint8_t r, uint8_t s)
 909    {
 910       tex.r = r;
 911       tex.s = s;
 912       tex.target = targ;
 913    }
 914
 915    void setIndirectR(Value *);
 916    void setIndirectS(Value *);
 917    inline Value *getIndirectR() const;
 918    inline Value *getIndirectS() const;
 919
 920 public:
 921    struct {
 922       Target target;
 923
 924       uint16_t r;
 925       uint16_t s;
 926       int8_t rIndirectSrc;
 927       int8_t sIndirectSrc;
 928
 929       uint8_t mask;
 930       uint8_t gatherComp;
 931
 932       bool liveOnly; // only execute on live pixels of a quad (optimization)
 933       bool levelZero;
 934       bool derivAll;
 935
 936       int8_t useOffsets; // 0, 1, or 4 for textureGatherOffsets
 937       int8_t offset[3]; // only used on nv50
 938
 939       enum TexQuery query;
 940    } tex;
 941
 942    ValueRef dPdx[3];
 943    ValueRef dPdy[3];
 944    ValueRef offset[4][3];
 945 };
 946
 947 class CmpInstruction : public Instruction
 948 {
 949 public:
 950    CmpInstruction(Function *, operation);
 951
 952    virtual CmpInstruction *clone(ClonePolicy<Function>&,
 953                                  Instruction * = NULL) const;
 954
 955    void setCondition(CondCode cond) { setCond = cond; }
 956    CondCode getCondition() const { return setCond; }
 957
 958 public:
 959    CondCode setCond;
 960 };
 961
 962 class FlowInstruction : public Instruction
 963 {
 964 public:
 965    FlowInstruction(Function *, operation, void *target);
 966
 967    virtual FlowInstruction *clone(ClonePolicy<Function>&,
 968                                   Instruction * = NULL) const;
 969
 970 public:
 971    unsigned allWarp  : 1;
 972    unsigned absolute : 1;
 973    unsigned limit    : 1;
 974    unsigned builtin  : 1; // true for calls to emulation code
 975    unsigned indirect : 1; // target in src(0)
 976
 977    union {
 978       BasicBlock *bb;
 979       int builtin;
 980       Function *fn;
 981    } target;
 982 };
 983
 984 class BasicBlock
 985 {
 986 public:
 987    BasicBlock(Function *);
 988    ~BasicBlock();
 989
 990    BasicBlock *clone(ClonePolicy<Function>&) const;
 991
 992    inline int getId() const { return id; }
 993    inline unsigned int getInsnCount() const { return numInsns; }
 994    inline bool isTerminated() const { return exit && exit->terminator; }
 995
 996    bool dominatedBy(BasicBlock *bb);
 997    inline bool reachableBy(const BasicBlock *by, const BasicBlock *term);
 998
 999    // returns mask of conditional out blocks
1000    // e.g. 3 for IF { .. } ELSE { .. } ENDIF, 1 for IF { .. } ENDIF
1001    unsigned int initiatesSimpleConditional() const;
1002
1003 public:
1004    Function *getFunction() const { return func; }
1005    Program *getProgram() const { return program; }
1006
1007    Instruction *getEntry() const { return entry; } // first non-phi instruction
1008    Instruction *getPhi() const { return phi; }
1009    Instruction *getFirst() const { return phi ? phi : entry; }
1010    Instruction *getExit() const { return exit; }
1011
1012    void insertHead(Instruction *);
1013    void insertTail(Instruction *);
1014    void insertBefore(Instruction *, Instruction *);
1015    void insertAfter(Instruction *, Instruction *);
1016    void remove(Instruction *);
1017    void permuteAdjacent(Instruction *, Instruction *);
1018
1019    BasicBlock *idom() const;
1020
1021    // NOTE: currently does not rebuild the dominator tree
1022    BasicBlock *splitBefore(Instruction *, bool attach = true);
1023    BasicBlock *splitAfter(Instruction *, bool attach = true);
1024
1025    DLList& getDF() { return df; }
1026    DLList::Iterator iterDF() { return df.iterator(); }
1027
1028    static inline BasicBlock *get(Iterator&);
1029    static inline BasicBlock *get(Graph::Node *);
1030
1031 public:
1032    Graph::Node cfg; // first edge is branch *taken* (the ELSE branch)
1033    Graph::Node dom;
1034
1035    BitSet liveSet;
1036    BitSet defSet;
1037
1038    uint32_t binPos;
1039    uint32_t binSize;
1040
1041    Instruction *joinAt; // for quick reference
1042
1043    bool explicitCont; // loop headers: true if loop contains continue stmts
1044
1045 private:
1046    int id;
1047    DLList df;
1048
1049    Instruction *phi;
1050    Instruction *entry;
1051    Instruction *exit;
1052
1053    unsigned int numInsns;
1054
1055 private:
1056    Function *func;
1057    Program *program;
1058
1059    void splitCommon(Instruction *, BasicBlock *, bool attach);
1060 };
1061
1062 class Function
1063 {
1064 public:
1065    Function(Program *, const char *name, uint32_t label);
1066    ~Function();
1067
1068    static inline Function *get(Graph::Node *node);
1069
1070    inline Program *getProgram() const { return prog; }
1071    inline const char *getName() const { return name; }
1072    inline int getId() const { return id; }
1073    inline uint32_t getLabel() const { return label; }
1074
1075    void print();
1076    void printLiveIntervals() const;
1077    void printCFGraph(const char *filePath);
1078
1079    bool setEntry(BasicBlock *);
1080    bool setExit(BasicBlock *);
1081
1082    unsigned int orderInstructions(ArrayList&);
1083
1084    inline void add(BasicBlock *bb, int& id) { allBBlocks.insert(bb, id); }
1085    inline void add(Instruction *insn, int& id) { allInsns.insert(insn, id); }
1086    inline void add(LValue *lval, int& id) { allLValues.insert(lval, id); }
1087
1088    inline LValue *getLValue(int id);
1089
1090    void buildLiveSets();
1091    void buildDefSets();
1092    bool convertToSSA();
1093
1094 public:
1095    std::deque<ValueDef> ins;
1096    std::deque<ValueRef> outs;
1097    std::deque<Value *> clobbers;
1098
1099    Graph cfg;
1100    Graph::Node *cfgExit;
1101    Graph *domTree;
1102    Graph::Node call; // node in the call graph
1103
1104    BasicBlock **bbArray; // BBs in emission order
1105    int bbCount;
1106
1107    unsigned int loopNestingBound;
1108    int regClobberMax;
1109
1110    uint32_t binPos;
1111    uint32_t binSize;
1112
1113    Value *stackPtr;
1114
1115    uint32_t tlsBase; // base address for l[] space (if no stack pointer is used)
1116    uint32_t tlsSize;
1117
1118    ArrayList allBBlocks;
1119    ArrayList allInsns;
1120    ArrayList allLValues;
1121
1122 private:
1123    void buildLiveSetsPreSSA(BasicBlock *, const int sequence);
1124    void buildDefSetsPreSSA(BasicBlock *bb, const int seq);
1125
1126 private:
1127    uint32_t label;
1128    int id;
1129    const char *const name;
1130    Program *prog;
1131 };
1132
1133 enum CGStage
1134 {
1135    CG_STAGE_PRE_SSA,
1136    CG_STAGE_SSA, // expected directly before register allocation
1137    CG_STAGE_POST_RA
1138 };
1139
1140 class Program
1141 {
1142 public:
1143    enum Type
1144    {
1145       TYPE_VERTEX,
1146       TYPE_TESSELLATION_CONTROL,
1147       TYPE_TESSELLATION_EVAL,
1148       TYPE_GEOMETRY,
1149       TYPE_FRAGMENT,
1150       TYPE_COMPUTE
1151    };
1152
1153    Program(Type type, Target *targ);
1154    ~Program();
1155
1156    void print();
1157
1158    Type getType() const { return progType; }
1159
1160    inline void add(Function *fn, int& id) { allFuncs.insert(fn, id); }
1161    inline void del(Function *fn, int& id) { allFuncs.remove(id); }
1162    inline void add(Value *rval, int& id) { allRValues.insert(rval, id); }
1163
1164    bool makeFromTGSI(struct nv50_ir_prog_info *);
1165    bool makeFromSM4(struct nv50_ir_prog_info *);
1166    bool convertToSSA();
1167    bool optimizeSSA(int level);
1168    bool optimizePostRA(int level);
1169    bool registerAllocation();
1170    bool emitBinary(struct nv50_ir_prog_info *);
1171
1172    const Target *getTarget() const { return target; }
1173
1174 private:
1175    void emitSymbolTable(struct nv50_ir_prog_info *);
1176
1177    Type progType;
1178    Target *target;
1179
1180 public:
1181    Function *main;
1182    Graph calls;
1183
1184    ArrayList allFuncs;
1185    ArrayList allRValues;
1186
1187    uint32_t *code;
1188    uint32_t binSize;
1189    uint32_t tlsSize; // size required for FILE_MEMORY_LOCAL
1190
1191    int maxGPR;
1192
1193    MemoryPool mem_Instruction;
1194    MemoryPool mem_CmpInstruction;
1195    MemoryPool mem_TexInstruction;
1196    MemoryPool mem_FlowInstruction;
1197    MemoryPool mem_LValue;
1198    MemoryPool mem_Symbol;
1199    MemoryPool mem_ImmediateValue;
1200
1201    uint32_t dbgFlags;
1202    uint8_t  optLevel;
1203
1204    void *targetPriv; // e.g. to carry information between passes
1205
1206    const struct nv50_ir_prog_info *driver; // for driver configuration
1207
1208    void releaseInstruction(Instruction *);
1209    void releaseValue(Value *);
1210 };
1211
1212 // TODO: add const version
1213 class Pass
1214 {
1215 public:
1216    bool run(Program *, bool ordered = false, bool skipPhi = false);
1217    bool run(Function *, bool ordered = false, bool skipPhi = false);
1218
1219 private:
1220    // return false to continue with next entity on next higher level
1221    virtual bool visit(Function *) { return true; }
1222    virtual bool visit(BasicBlock *) { return true; }
1223    virtual bool visit(Instruction *) { return false; }
1224
1225    bool doRun(Program *, bool ordered, bool skipPhi);
1226    bool doRun(Function *, bool ordered, bool skipPhi);
1227
1228 protected:
1229    bool err;
1230    Function *func;
1231    Program *prog;
1232 };
1233
1234 // =============================================================================
1235
1236 #include "codegen/nv50_ir_inlines.h"
1237
1238 } // namespace nv50_ir
1239
1240 #endif // __NV50_IR_H__