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_SHLADD,
  61    OP_ABS,
  62    OP_NEG,
  63    OP_NOT,
  64    OP_AND,
  65    OP_OR,
  66    OP_XOR,
  67    OP_SHL,
  68    OP_SHR,
  69    OP_MAX,
  70    OP_MIN,
  71    OP_SAT, // CLAMP(f32, 0.0, 1.0)
  72    OP_CEIL,
  73    OP_FLOOR,
  74    OP_TRUNC,
  75    OP_CVT,
  76    OP_SET_AND, // dst = (src0 CMP src1) & src2
  77    OP_SET_OR,
  78    OP_SET_XOR,
  79    OP_SET,
  80    OP_SELP, // dst = src2 ? src0 : src1
  81    OP_SLCT, // dst = (src2 CMP 0) ? src0 : src1
  82    OP_RCP,
  83    OP_RSQ,
  84    OP_LG2,
  85    OP_SIN,
  86    OP_COS,
  87    OP_EX2,
  88    OP_EXP, // exponential (base M_E)
  89    OP_LOG, // natural logarithm
  90    OP_PRESIN,
  91    OP_PREEX2,
  92    OP_SQRT,
  93    OP_POW,
  94    OP_BRA,
  95    OP_CALL,
  96    OP_RET,
  97    OP_CONT,
  98    OP_BREAK,
  99    OP_PRERET,
 100    OP_PRECONT,
 101    OP_PREBREAK,
 102    OP_BRKPT,     // breakpoint (not related to loops)
 103    OP_JOINAT,    // push control flow convergence point
 104    OP_JOIN,      // converge
 105    OP_DISCARD,
 106    OP_EXIT,
 107    OP_MEMBAR, // memory barrier (mfence, lfence, sfence)
 108    OP_VFETCH, // indirection 0 in attribute space, indirection 1 is vertex base
 109    OP_PFETCH, // fetch base address of vertex src0 (immediate) [+ src1]
 110    OP_AFETCH, // fetch base address of shader input (a[%r1+0x10])
 111    OP_EXPORT,
 112    OP_LINTERP,
 113    OP_PINTERP,
 114    OP_EMIT,    // emit vertex
 115    OP_RESTART, // restart primitive
 116    OP_TEX,
 117    OP_TXB, // texture bias
 118    OP_TXL, // texure lod
 119    OP_TXF, // texel fetch
 120    OP_TXQ, // texture size query
 121    OP_TXD, // texture derivatives
 122    OP_TXG, // texture gather
 123    OP_TXLQ, // texture query lod
 124    OP_TEXCSAA, // texture op for coverage sampling
 125    OP_TEXPREP, // turn cube map array into 2d array coordinates
 126    OP_SULDB, // surface load (raw)
 127    OP_SULDP, // surface load (formatted)
 128    OP_SUSTB, // surface store (raw)
 129    OP_SUSTP, // surface store (formatted)
 130    OP_SUREDB,
 131    OP_SUREDP, // surface reduction (atomic op)
 132    OP_SULEA,   // surface load effective address
 133    OP_SUBFM,   // surface bitfield manipulation
 134    OP_SUCLAMP, // clamp surface coordinates
 135    OP_SUEAU,   // surface effective address
 136    OP_SUQ,     // surface query
 137    OP_MADSP,   // special integer multiply-add
 138    OP_TEXBAR, // texture dependency barrier
 139    OP_DFDX,
 140    OP_DFDY,
 141    OP_RDSV, // read system value
 142    OP_WRSV, // write system value
 143    OP_PIXLD, // get info about raster object or surfaces
 144    OP_QUADOP,
 145    OP_QUADON,
 146    OP_QUADPOP,
 147    OP_POPCNT, // bitcount(src0 & src1)
 148    OP_INSBF,  // insert first src1[8:15] bits of src0 into src2 at src1[0:7]
 149    OP_EXTBF,  // place bits [K,K+N) of src0 into dst, src1 = 0xNNKK
 150    OP_BFIND,  // find highest/lowest set bit
 151    OP_PERMT,  // dst = bytes from src2,src0 selected by src1 (nvc0's src order)
 152    OP_ATOM,
 153    OP_BAR,    // execution barrier, sources = { id, thread count, predicate }
 154    OP_VADD,   // byte/word vector operations
 155    OP_VAVG,
 156    OP_VMIN,
 157    OP_VMAX,
 158    OP_VSAD,
 159    OP_VSET,
 160    OP_VSHR,
 161    OP_VSHL,
 162    OP_VSEL,
 163    OP_CCTL, // cache control
 164    OP_SHFL, // warp shuffle
 165    OP_VOTE,
 166    OP_BUFQ, // buffer query
 167    OP_LAST
 168 };
 169
 170 // various instruction-specific modifier definitions Instruction::subOp
 171 // MOV_FINAL marks a MOV originating from an EXPORT (used for placing TEXBARs)
 172 #define NV50_IR_SUBOP_MUL_HIGH     1
 173 #define NV50_IR_SUBOP_EMIT_RESTART 1
 174 #define NV50_IR_SUBOP_LDC_IL       1
 175 #define NV50_IR_SUBOP_LDC_IS       2
 176 #define NV50_IR_SUBOP_LDC_ISL      3
 177 #define NV50_IR_SUBOP_SHIFT_WRAP   1
 178 #define NV50_IR_SUBOP_EMU_PRERET   1
 179 #define NV50_IR_SUBOP_TEXBAR(n)    n
 180 #define NV50_IR_SUBOP_MOV_FINAL    1
 181 #define NV50_IR_SUBOP_EXTBF_REV    1
 182 #define NV50_IR_SUBOP_BFIND_SAMT   1
 183 #define NV50_IR_SUBOP_RCPRSQ_64H   1
 184 #define NV50_IR_SUBOP_PERMT_F4E    1
 185 #define NV50_IR_SUBOP_PERMT_B4E    2
 186 #define NV50_IR_SUBOP_PERMT_RC8    3
 187 #define NV50_IR_SUBOP_PERMT_ECL    4
 188 #define NV50_IR_SUBOP_PERMT_ECR    5
 189 #define NV50_IR_SUBOP_PERMT_RC16   6
 190 #define NV50_IR_SUBOP_BAR_SYNC     0
 191 #define NV50_IR_SUBOP_BAR_ARRIVE   1
 192 #define NV50_IR_SUBOP_BAR_RED_AND  2
 193 #define NV50_IR_SUBOP_BAR_RED_OR   3
 194 #define NV50_IR_SUBOP_BAR_RED_POPC 4
 195 #define NV50_IR_SUBOP_MEMBAR_L     1
 196 #define NV50_IR_SUBOP_MEMBAR_S     2
 197 #define NV50_IR_SUBOP_MEMBAR_M     3
 198 #define NV50_IR_SUBOP_MEMBAR_CTA  (0 << 2)
 199 #define NV50_IR_SUBOP_MEMBAR_GL   (1 << 2)
 200 #define NV50_IR_SUBOP_MEMBAR_SYS  (2 << 2)
 201 #define NV50_IR_SUBOP_MEMBAR_DIR(m)   ((m) & 0x3)
 202 #define NV50_IR_SUBOP_MEMBAR_SCOPE(m) ((m) & ~0x3)
 203 #define NV50_IR_SUBOP_MEMBAR(d,s) \
 204    (NV50_IR_SUBOP_MEMBAR_##d | NV50_IR_SUBOP_MEMBAR_##s)
 205 #define NV50_IR_SUBOP_ATOM_ADD      0
 206 #define NV50_IR_SUBOP_ATOM_MIN      1
 207 #define NV50_IR_SUBOP_ATOM_MAX      2
 208 #define NV50_IR_SUBOP_ATOM_INC      3
 209 #define NV50_IR_SUBOP_ATOM_DEC      4
 210 #define NV50_IR_SUBOP_ATOM_AND      5
 211 #define NV50_IR_SUBOP_ATOM_OR       6
 212 #define NV50_IR_SUBOP_ATOM_XOR      7
 213 #define NV50_IR_SUBOP_ATOM_CAS      8
 214 #define NV50_IR_SUBOP_ATOM_EXCH     9
 215 #define NV50_IR_SUBOP_CCTL_IV      5
 216 #define NV50_IR_SUBOP_CCTL_IVALL   6
 217 #define NV50_IR_SUBOP_SUST_IGN     0
 218 #define NV50_IR_SUBOP_SUST_TRAP    1
 219 #define NV50_IR_SUBOP_SUST_SDCL    3
 220 #define NV50_IR_SUBOP_SULD_ZERO    0
 221 #define NV50_IR_SUBOP_SULD_TRAP    1
 222 #define NV50_IR_SUBOP_SULD_SDCL    3
 223 #define NV50_IR_SUBOP_SUBFM_3D     1
 224 #define NV50_IR_SUBOP_SUCLAMP_2D   0x10
 225 #define NV50_IR_SUBOP_SUCLAMP_SD(r, d) (( 0 + (r)) | ((d == 2) ? 0x10 : 0))
 226 #define NV50_IR_SUBOP_SUCLAMP_PL(r, d) (( 5 + (r)) | ((d == 2) ? 0x10 : 0))
 227 #define NV50_IR_SUBOP_SUCLAMP_BL(r, d) ((10 + (r)) | ((d == 2) ? 0x10 : 0))
 228 #define NV50_IR_SUBOP_PIXLD_COUNT       0
 229 #define NV50_IR_SUBOP_PIXLD_COVMASK     1
 230 #define NV50_IR_SUBOP_PIXLD_COVERED     2
 231 #define NV50_IR_SUBOP_PIXLD_OFFSET      3
 232 #define NV50_IR_SUBOP_PIXLD_CENT_OFFSET 4
 233 #define NV50_IR_SUBOP_PIXLD_SAMPLEID    5
 234 #define NV50_IR_SUBOP_SHFL_IDX  0
 235 #define NV50_IR_SUBOP_SHFL_UP   1
 236 #define NV50_IR_SUBOP_SHFL_DOWN 2
 237 #define NV50_IR_SUBOP_SHFL_BFLY 3
 238 #define NV50_IR_SUBOP_LOAD_LOCKED    1
 239 #define NV50_IR_SUBOP_STORE_UNLOCKED 2
 240 #define NV50_IR_SUBOP_MADSP_SD     0xffff
 241 // Yes, we could represent those with DataType.
 242 // Or put the type into operation and have a couple 1000 values in that enum.
 243 // This will have to do for now.
 244 // The bitfields are supposed to correspond to nve4 ISA.
 245 #define NV50_IR_SUBOP_MADSP(a,b,c) (((c) << 8) | ((b) << 4) | (a))
 246 #define NV50_IR_SUBOP_V1(d,a,b)    (((d) << 10) | ((b) << 5) | (a) | 0x0000)
 247 #define NV50_IR_SUBOP_V2(d,a,b)    (((d) << 10) | ((b) << 5) | (a) | 0x4000)
 248 #define NV50_IR_SUBOP_V4(d,a,b)    (((d) << 10) | ((b) << 5) | (a) | 0x8000)
 249 #define NV50_IR_SUBOP_Vn(n)        ((n) >> 14)
 250 #define NV50_IR_SUBOP_VOTE_ALL 0
 251 #define NV50_IR_SUBOP_VOTE_ANY 1
 252 #define NV50_IR_SUBOP_VOTE_UNI 2
 253
 254 #define NV50_IR_SUBOP_MINMAX_LOW  1
 255 #define NV50_IR_SUBOP_MINMAX_MED  2
 256 #define NV50_IR_SUBOP_MINMAX_HIGH 3
 257
 258 enum DataType
 259 {
 260    TYPE_NONE,
 261    TYPE_U8,
 262    TYPE_S8,
 263    TYPE_U16,
 264    TYPE_S16,
 265    TYPE_U32,
 266    TYPE_S32,
 267    TYPE_U64, // 64 bit operations are only lowered after register allocation
 268    TYPE_S64,
 269    TYPE_F16,
 270    TYPE_F32,
 271    TYPE_F64,
 272    TYPE_B96,
 273    TYPE_B128
 274 };
 275
 276 enum CondCode
 277 {
 278    CC_FL = 0,
 279    CC_NEVER = CC_FL, // when used with FILE_FLAGS
 280    CC_LT = 1,
 281    CC_EQ = 2,
 282    CC_NOT_P = CC_EQ, // when used with FILE_PREDICATE
 283    CC_LE = 3,
 284    CC_GT = 4,
 285    CC_NE = 5,
 286    CC_P  = CC_NE,
 287    CC_GE = 6,
 288    CC_TR = 7,
 289    CC_ALWAYS = CC_TR,
 290    CC_U  = 8,
 291    CC_LTU = 9,
 292    CC_EQU = 10,
 293    CC_LEU = 11,
 294    CC_GTU = 12,
 295    CC_NEU = 13,
 296    CC_GEU = 14,
 297    CC_NO = 0x10,
 298    CC_NC = 0x11,
 299    CC_NS = 0x12,
 300    CC_NA = 0x13,
 301    CC_A  = 0x14,
 302    CC_S  = 0x15,
 303    CC_C  = 0x16,
 304    CC_O  = 0x17
 305 };
 306
 307 enum RoundMode
 308 {
 309    ROUND_N, // nearest
 310    ROUND_M, // towards -inf
 311    ROUND_Z, // towards 0
 312    ROUND_P, // towards +inf
 313    ROUND_NI, // nearest integer
 314    ROUND_MI, // to integer towards -inf
 315    ROUND_ZI, // to integer towards 0
 316    ROUND_PI, // to integer towards +inf
 317 };
 318
 319 enum CacheMode
 320 {
 321    CACHE_CA,            // cache at all levels
 322    CACHE_WB = CACHE_CA, // cache write back
 323    CACHE_CG,            // cache at global level
 324    CACHE_CS,            // cache streaming
 325    CACHE_CV,            // cache as volatile
 326    CACHE_WT = CACHE_CV  // cache write-through
 327 };
 328
 329 enum DataFile
 330 {
 331    FILE_NULL = 0,
 332    FILE_GPR,
 333    FILE_PREDICATE,       // boolean predicate
 334    FILE_FLAGS,           // zero/sign/carry/overflow bits
 335    FILE_ADDRESS,
 336    LAST_REGISTER_FILE = FILE_ADDRESS,
 337    FILE_IMMEDIATE,
 338    FILE_MEMORY_CONST,
 339    FILE_SHADER_INPUT,
 340    FILE_SHADER_OUTPUT,
 341    FILE_MEMORY_BUFFER,
 342    FILE_MEMORY_GLOBAL,
 343    FILE_MEMORY_SHARED,
 344    FILE_MEMORY_LOCAL,
 345    FILE_SYSTEM_VALUE,
 346    DATA_FILE_COUNT
 347 };
 348
 349 enum TexTarget
 350 {
 351    TEX_TARGET_1D,
 352    TEX_TARGET_2D,
 353    TEX_TARGET_2D_MS,
 354    TEX_TARGET_3D,
 355    TEX_TARGET_CUBE,
 356    TEX_TARGET_1D_SHADOW,
 357    TEX_TARGET_2D_SHADOW,
 358    TEX_TARGET_CUBE_SHADOW,
 359    TEX_TARGET_1D_ARRAY,
 360    TEX_TARGET_2D_ARRAY,
 361    TEX_TARGET_2D_MS_ARRAY,
 362    TEX_TARGET_CUBE_ARRAY,
 363    TEX_TARGET_1D_ARRAY_SHADOW,
 364    TEX_TARGET_2D_ARRAY_SHADOW,
 365    TEX_TARGET_RECT,
 366    TEX_TARGET_RECT_SHADOW,
 367    TEX_TARGET_CUBE_ARRAY_SHADOW,
 368    TEX_TARGET_BUFFER,
 369    TEX_TARGET_COUNT
 370 };
 371
 372 enum ImgFormat
 373 {
 374    FMT_NONE,
 375
 376    FMT_RGBA32F,
 377    FMT_RGBA16F,
 378    FMT_RG32F,
 379    FMT_RG16F,
 380    FMT_R11G11B10F,
 381    FMT_R32F,
 382    FMT_R16F,
 383
 384    FMT_RGBA32UI,
 385    FMT_RGBA16UI,
 386    FMT_RGB10A2UI,
 387    FMT_RGBA8UI,
 388    FMT_RG32UI,
 389    FMT_RG16UI,
 390    FMT_RG8UI,
 391    FMT_R32UI,
 392    FMT_R16UI,
 393    FMT_R8UI,
 394
 395    FMT_RGBA32I,
 396    FMT_RGBA16I,
 397    FMT_RGBA8I,
 398    FMT_RG32I,
 399    FMT_RG16I,
 400    FMT_RG8I,
 401    FMT_R32I,
 402    FMT_R16I,
 403    FMT_R8I,
 404
 405    FMT_RGBA16,
 406    FMT_RGB10A2,
 407    FMT_RGBA8,
 408    FMT_RG16,
 409    FMT_RG8,
 410    FMT_R16,
 411    FMT_R8,
 412
 413    FMT_RGBA16_SNORM,
 414    FMT_RGBA8_SNORM,
 415    FMT_RG16_SNORM,
 416    FMT_RG8_SNORM,
 417    FMT_R16_SNORM,
 418    FMT_R8_SNORM,
 419
 420    FMT_BGRA8,
 421
 422    IMG_FORMAT_COUNT,
 423 };
 424
 425 enum ImgType {
 426    UINT,
 427    SINT,
 428    UNORM,
 429    SNORM,
 430    FLOAT,
 431 };
 432
 433 enum SVSemantic
 434 {
 435    SV_POSITION, // WPOS
 436    SV_VERTEX_ID,
 437    SV_INSTANCE_ID,
 438    SV_INVOCATION_ID,
 439    SV_PRIMITIVE_ID,
 440    SV_VERTEX_COUNT, // gl_PatchVerticesIn
 441    SV_LAYER,
 442    SV_VIEWPORT_INDEX,
 443    SV_YDIR,
 444    SV_FACE,
 445    SV_POINT_SIZE,
 446    SV_POINT_COORD,
 447    SV_CLIP_DISTANCE,
 448    SV_SAMPLE_INDEX,
 449    SV_SAMPLE_POS,
 450    SV_SAMPLE_MASK,
 451    SV_TESS_OUTER,
 452    SV_TESS_INNER,
 453    SV_TESS_COORD,
 454    SV_TID,
 455    SV_CTAID,
 456    SV_NTID,
 457    SV_GRIDID,
 458    SV_NCTAID,
 459    SV_LANEID,
 460    SV_PHYSID,
 461    SV_NPHYSID,
 462    SV_CLOCK,
 463    SV_LBASE,
 464    SV_SBASE,
 465    SV_VERTEX_STRIDE,
 466    SV_INVOCATION_INFO,
 467    SV_THREAD_KILL,
 468    SV_BASEVERTEX,
 469    SV_BASEINSTANCE,
 470    SV_DRAWID,
 471    SV_WORK_DIM,
 472    SV_UNDEFINED,
 473    SV_LAST
 474 };
 475
 476 class Program;
 477 class Function;
 478 class BasicBlock;
 479
 480 class Target;
 481
 482 class Instruction;
 483 class CmpInstruction;
 484 class TexInstruction;
 485 class FlowInstruction;
 486
 487 class Value;
 488 class LValue;
 489 class Symbol;
 490 class ImmediateValue;
 491
 492 struct Storage
 493 {
 494    DataFile file;
 495    int8_t fileIndex; // signed, may be indirect for CONST[]
 496    uint8_t size; // this should match the Instruction type's size
 497    DataType type; // mainly for pretty printing
 498    union {
 499       uint64_t u64;    // immediate values
 500       uint32_t u32;
 501       uint16_t u16;
 502       uint8_t u8;
 503       int64_t s64;
 504       int32_t s32;
 505       int16_t s16;
 506       int8_t s8;
 507       float f32;
 508       double f64;
 509       int32_t offset; // offset from 0 (base of address space)
 510       int32_t id;     // register id (< 0 if virtual/unassigned, in units <= 4)
 511       struct {
 512          SVSemantic sv;
 513          int index;
 514       } sv;
 515    } data;
 516 };
 517
 518 // precedence: NOT after SAT after NEG after ABS
 519 #define NV50_IR_MOD_ABS (1 << 0)
 520 #define NV50_IR_MOD_NEG (1 << 1)
 521 #define NV50_IR_MOD_SAT (1 << 2)
 522 #define NV50_IR_MOD_NOT (1 << 3)
 523 #define NV50_IR_MOD_NEG_ABS (NV50_IR_MOD_NEG | NV50_IR_MOD_ABS)
 524
 525 #define NV50_IR_INTERP_MODE_MASK   0x3
 526 #define NV50_IR_INTERP_LINEAR      (0 << 0)
 527 #define NV50_IR_INTERP_PERSPECTIVE (1 << 0)
 528 #define NV50_IR_INTERP_FLAT        (2 << 0)
 529 #define NV50_IR_INTERP_SC          (3 << 0) // what exactly is that ?
 530 #define NV50_IR_INTERP_SAMPLE_MASK 0xc
 531 #define NV50_IR_INTERP_DEFAULT     (0 << 2)
 532 #define NV50_IR_INTERP_CENTROID    (1 << 2)
 533 #define NV50_IR_INTERP_OFFSET      (2 << 2)
 534 #define NV50_IR_INTERP_SAMPLEID    (3 << 2)
 535
 536 // do we really want this to be a class ?
 537 class Modifier
 538 {
 539 public:
 540    Modifier() : bits(0) { }
 541    Modifier(unsigned int m) : bits(m) { }
 542    Modifier(operation op);
 543
 544    // @return new Modifier applying a after b (asserts if unrepresentable)
 545    Modifier operator*(const Modifier) const;
 546    Modifier operator*=(const Modifier m) { *this = *this * m; return *this; }
 547    Modifier operator==(const Modifier m) const { return m.bits == bits; }
 548    Modifier operator!=(const Modifier m) const { return m.bits != bits; }
 549
 550    inline Modifier operator&(const Modifier m) const { return bits & m.bits; }
 551    inline Modifier operator|(const Modifier m) const { return bits | m.bits; }
 552    inline Modifier operator^(const Modifier m) const { return bits ^ m.bits; }
 553
 554    operation getOp() const;
 555
 556    inline int neg() const { return (bits & NV50_IR_MOD_NEG) ? 1 : 0; }
 557    inline int abs() const { return (bits & NV50_IR_MOD_ABS) ? 1 : 0; }
 558
 559    inline operator bool() const { return bits ? true : false; }
 560
 561    void applyTo(ImmediateValue &imm) const;
 562
 563    int print(char *buf, size_t size) const;
 564
 565 private:
 566    uint8_t bits;
 567 };
 568
 569 class ValueRef
 570 {
 571 public:
 572    ValueRef(Value * = NULL);
 573    ValueRef(const ValueRef&);
 574    ~ValueRef();
 575
 576    inline bool exists() const { return value != NULL; }
 577
 578    void set(Value *);
 579    void set(const ValueRef&);
 580    inline Value *get() const { return value; }
 581    inline Value *rep() const;
 582
 583    inline Instruction *getInsn() const { return insn; }
 584    inline void setInsn(Instruction *inst) { insn = inst; }
 585
 586    inline bool isIndirect(int dim) const { return indirect[dim] >= 0; }
 587    inline const ValueRef *getIndirect(int dim) const;
 588
 589    inline DataFile getFile() const;
 590    inline unsigned getSize() const;
 591
 592    // SSA: return eventual (traverse MOVs) literal value, if it exists
 593    bool getImmediate(ImmediateValue&) const;
 594
 595 public:
 596    Modifier mod;
 597    int8_t indirect[2]; // >= 0 if relative to lvalue in insn->src(indirect[i])
 598    uint8_t swizzle;
 599
 600    bool usedAsPtr; // for printing
 601
 602 private:
 603    Value *value;
 604    Instruction *insn;
 605 };
 606
 607 class ValueDef
 608 {
 609 public:
 610    ValueDef(Value * = NULL);
 611    ValueDef(const ValueDef&);
 612    ~ValueDef();
 613
 614    inline bool exists() const { return value != NULL; }
 615
 616    inline Value *get() const { return value; }
 617    inline Value *rep() const;
 618    void set(Value *);
 619    bool mayReplace(const ValueRef &);
 620    void replace(const ValueRef &, bool doSet); // replace all uses of the old value
 621
 622    inline Instruction *getInsn() const { return insn; }
 623    inline void setInsn(Instruction *inst) { insn = inst; }
 624
 625    inline DataFile getFile() const;
 626    inline unsigned getSize() const;
 627
 628    inline void setSSA(LValue *);
 629    inline const LValue *preSSA() const;
 630
 631 private:
 632    Value *value;   // should make this LValue * ...
 633    LValue *origin; // pre SSA value
 634    Instruction *insn;
 635 };
 636
 637 class Value
 638 {
 639 public:
 640    Value();
 641    virtual ~Value() { }
 642
 643    virtual Value *clone(ClonePolicy<Function>&) const = 0;
 644
 645    virtual int print(char *, size_t, DataType ty = TYPE_NONE) const = 0;
 646
 647    virtual bool equals(const Value *, bool strict = false) const;
 648    virtual bool interfers(const Value *) const;
 649    virtual bool isUniform() const { return true; }
 650
 651    inline Value *rep() const { return join; }
 652
 653    inline Instruction *getUniqueInsn() const;
 654    inline Instruction *getInsn() const; // use when uniqueness is certain
 655
 656    inline int refCount() { return uses.size(); }
 657
 658    inline LValue *asLValue();
 659    inline Symbol *asSym();
 660    inline ImmediateValue *asImm();
 661    inline const Symbol *asSym() const;
 662    inline const ImmediateValue *asImm() const;
 663
 664    inline bool inFile(DataFile f) { return reg.file == f; }
 665
 666    static inline Value *get(Iterator&);
 667
 668    unordered_set<ValueRef *> uses;
 669    std::list<ValueDef *> defs;
 670    typedef unordered_set<ValueRef *>::iterator UseIterator;
 671    typedef unordered_set<ValueRef *>::const_iterator UseCIterator;
 672    typedef std::list<ValueDef *>::iterator DefIterator;
 673    typedef std::list<ValueDef *>::const_iterator DefCIterator;
 674
 675    int id;
 676    Storage reg;
 677
 678    // TODO: these should be in LValue:
 679    Interval livei;
 680    Value *join;
 681 };
 682
 683 class LValue : public Value
 684 {
 685 public:
 686    LValue(Function *, DataFile file);
 687    LValue(Function *, LValue *);
 688    ~LValue() { }
 689
 690    virtual bool isUniform() const;
 691
 692    virtual LValue *clone(ClonePolicy<Function>&) const;
 693
 694    virtual int print(char *, size_t, DataType ty = TYPE_NONE) const;
 695
 696 public:
 697    unsigned compMask : 8; // compound/component mask
 698    unsigned compound : 1; // used by RA, value involved in split/merge
 699    unsigned ssa      : 1;
 700    unsigned fixedReg : 1; // set & used by RA, earlier just use (id < 0)
 701    unsigned noSpill  : 1; // do not spill (e.g. if spill temporary already)
 702 };
 703
 704 class Symbol : public Value
 705 {
 706 public:
 707    Symbol(Program *, DataFile file = FILE_MEMORY_CONST, ubyte fileIdx = 0);
 708    ~Symbol() { }
 709
 710    virtual Symbol *clone(ClonePolicy<Function>&) const;
 711
 712    virtual bool equals(const Value *that, bool strict) const;
 713
 714    virtual bool isUniform() const;
 715
 716    virtual int print(char *, size_t, DataType ty = TYPE_NONE) const;
 717
 718    // print with indirect values
 719    int print(char *, size_t, Value *, Value *, DataType ty = TYPE_NONE) const;
 720
 721    inline void setFile(DataFile file, ubyte fileIndex = 0)
 722    {
 723       reg.file = file;
 724       reg.fileIndex = fileIndex;
 725    }
 726
 727    inline void setOffset(int32_t offset);
 728    inline void setAddress(Symbol *base, int32_t offset);
 729    inline void setSV(SVSemantic sv, uint32_t idx = 0);
 730
 731    inline const Symbol *getBase() const { return baseSym; }
 732
 733 private:
 734    Symbol *baseSym; // array base for Symbols representing array elements
 735 };
 736
 737 class ImmediateValue : public Value
 738 {
 739 public:
 740    ImmediateValue() { }
 741    ImmediateValue(Program *, uint32_t);
 742    ImmediateValue(Program *, float);
 743    ImmediateValue(Program *, double);
 744    // NOTE: not added to program with
 745    ImmediateValue(const ImmediateValue *, DataType ty);
 746    ~ImmediateValue() { };
 747
 748    virtual ImmediateValue *clone(ClonePolicy<Function>&) const;
 749
 750    virtual bool equals(const Value *that, bool strict) const;
 751
 752    // these only work if 'type' is valid (we mostly use untyped literals):
 753    bool isInteger(const int ival) const; // ival is cast to this' type
 754    bool isNegative() const;
 755    bool isPow2() const;
 756
 757    void applyLog2();
 758
 759    // for constant folding:
 760    ImmediateValue operator+(const ImmediateValue&) const;
 761    ImmediateValue operator-(const ImmediateValue&) const;
 762    ImmediateValue operator*(const ImmediateValue&) const;
 763    ImmediateValue operator/(const ImmediateValue&) const;
 764
 765    ImmediateValue& operator=(const ImmediateValue&); // only sets value !
 766
 767    bool compare(CondCode cc, float fval) const;
 768
 769    virtual int print(char *, size_t, DataType ty = TYPE_NONE) const;
 770 };
 771
 772 class Instruction
 773 {
 774 public:
 775    Instruction();
 776    Instruction(Function *, operation, DataType);
 777    virtual ~Instruction();
 778
 779    virtual Instruction *clone(ClonePolicy<Function>&,
 780                               Instruction * = NULL) const;
 781
 782    void setDef(int i, Value *);
 783    void setSrc(int s, Value *);
 784    void setSrc(int s, const ValueRef&);
 785    void swapSources(int a, int b);
 786    void moveSources(int s, int delta);
 787    bool setIndirect(int s, int dim, Value *);
 788
 789    inline ValueRef& src(int s) { return srcs[s]; }
 790    inline ValueDef& def(int s) { return defs[s]; }
 791    inline const ValueRef& src(int s) const { return srcs[s]; }
 792    inline const ValueDef& def(int s) const { return defs[s]; }
 793
 794    inline Value *getDef(int d) const { return defs[d].get(); }
 795    inline Value *getSrc(int s) const { return srcs[s].get(); }
 796    inline Value *getIndirect(int s, int dim) const;
 797
 798    inline bool defExists(unsigned d) const
 799    {
 800       return d < defs.size() && defs[d].exists();
 801    }
 802    inline bool srcExists(unsigned s) const
 803    {
 804       return s < srcs.size() && srcs[s].exists();
 805    }
 806
 807    inline bool constrainedDefs() const;
 808
 809    bool setPredicate(CondCode ccode, Value *);
 810    inline Value *getPredicate() const;
 811    bool writesPredicate() const;
 812    inline bool isPredicated() const { return predSrc >= 0; }
 813
 814    inline void setFlagsSrc(int s, Value *);
 815    inline void setFlagsDef(int d, Value *);
 816    inline bool usesFlags() const { return flagsSrc >= 0; }
 817
 818    unsigned int defCount() const { return defs.size(); };
 819    unsigned int defCount(unsigned int mask, bool singleFile = false) const;
 820    unsigned int srcCount() const { return srcs.size(); };
 821    unsigned int srcCount(unsigned int mask, bool singleFile = false) const;
 822
 823    // save & remove / set indirect[0,1] and predicate source
 824    void takeExtraSources(int s, Value *[3]);
 825    void putExtraSources(int s, Value *[3]);
 826
 827    inline void setType(DataType type) { dType = sType = type; }
 828
 829    inline void setType(DataType dtype, DataType stype)
 830    {
 831       dType = dtype;
 832       sType = stype;
 833    }
 834
 835    inline bool isPseudo() const { return op < OP_MOV; }
 836    bool isDead() const;
 837    bool isNop() const;
 838    bool isCommutationLegal(const Instruction *) const; // must be adjacent !
 839    bool isActionEqual(const Instruction *) const;
 840    bool isResultEqual(const Instruction *) const;
 841
 842    // check whether the defs interfere with srcs and defs of another instruction
 843    bool canCommuteDefDef(const Instruction *) const;
 844    bool canCommuteDefSrc(const Instruction *) const;
 845
 846    void print() const;
 847
 848    inline CmpInstruction *asCmp();
 849    inline TexInstruction *asTex();
 850    inline FlowInstruction *asFlow();
 851    inline const TexInstruction *asTex() const;
 852    inline const CmpInstruction *asCmp() const;
 853    inline const FlowInstruction *asFlow() const;
 854
 855 public:
 856    Instruction *next;
 857    Instruction *prev;
 858    int id;
 859    int serial; // CFG order
 860
 861    operation op;
 862    DataType dType; // destination or defining type
 863    DataType sType; // source or secondary type
 864    CondCode cc;
 865    RoundMode rnd;
 866    CacheMode cache;
 867
 868    uint16_t subOp; // quadop, 1 for mul-high, etc.
 869
 870    unsigned encSize    : 4; // encoding size in bytes
 871    unsigned saturate   : 1; // to [0.0f, 1.0f]
 872    unsigned join       : 1; // converge control flow (use OP_JOIN until end)
 873    unsigned fixed      : 1; // prevent dead code elimination
 874    unsigned terminator : 1; // end of basic block
 875    unsigned ftz        : 1; // flush denormal to zero
 876    unsigned dnz        : 1; // denormals, NaN are zero
 877    unsigned ipa        : 4; // interpolation mode
 878    unsigned lanes      : 4;
 879    unsigned perPatch   : 1;
 880    unsigned exit       : 1; // terminate program after insn
 881    unsigned mask       : 4; // for vector ops
 882
 883    int8_t postFactor; // MUL/DIV(if < 0) by 1 << postFactor
 884
 885    int8_t predSrc;
 886    int8_t flagsDef;
 887    int8_t flagsSrc;
 888
 889    uint32_t sched; // scheduling data (NOTE: maybe move to separate storage)
 890
 891    BasicBlock *bb;
 892
 893 protected:
 894    std::deque<ValueDef> defs; // no gaps !
 895    std::deque<ValueRef> srcs; // no gaps !
 896
 897    // instruction specific methods:
 898    // (don't want to subclass, would need more constructors and memory pools)
 899 public:
 900    inline void setInterpolate(unsigned int mode) { ipa = mode; }
 901
 902    unsigned int getInterpMode() const { return ipa & 0x3; }
 903    unsigned int getSampleMode() const { return ipa & 0xc; }
 904
 905 private:
 906    void init();
 907 };
 908
 909 enum TexQuery
 910 {
 911    TXQ_DIMS, /* x, y, z, levels */
 912    TXQ_TYPE, /* ?, ?, samples, ? */
 913    TXQ_SAMPLE_POSITION,
 914    TXQ_FILTER,
 915    TXQ_LOD,
 916    TXQ_WRAP,
 917    TXQ_BORDER_COLOUR
 918 };
 919
 920 class TexInstruction : public Instruction
 921 {
 922 public:
 923    class Target
 924    {
 925    public:
 926       Target(TexTarget targ = TEX_TARGET_2D) : target(targ) { }
 927
 928       const char *getName() const { return descTable[target].name; }
 929       unsigned int getArgCount() const { return descTable[target].argc; }
 930       unsigned int getDim() const { return descTable[target].dim; }
 931       int isArray() const { return descTable[target].array ? 1 : 0; }
 932       int isCube() const { return descTable[target].cube ? 1 : 0; }
 933       int isShadow() const { return descTable[target].shadow ? 1 : 0; }
 934       int isMS() const {
 935         return target == TEX_TARGET_2D_MS || target == TEX_TARGET_2D_MS_ARRAY; }
 936       void clearMS() {
 937          if (isMS()) {
 938             if (isArray())
 939                target = TEX_TARGET_2D_ARRAY;
 940             else
 941                target = TEX_TARGET_2D;
 942          }
 943       }
 944
 945       Target& operator=(TexTarget targ)
 946       {
 947          assert(targ < TEX_TARGET_COUNT);
 948          target = targ;
 949          return *this;
 950       }
 951
 952       inline bool operator==(TexTarget targ) const { return target == targ; }
 953       inline bool operator!=(TexTarget targ) const { return target != targ; }
 954
 955       enum TexTarget getEnum() const { return target; }
 956
 957    private:
 958       struct Desc
 959       {
 960          char name[19];
 961          uint8_t dim;
 962          uint8_t argc;
 963          bool array;
 964          bool cube;
 965          bool shadow;
 966       };
 967
 968       static const struct Desc descTable[TEX_TARGET_COUNT];
 969
 970    private:
 971       enum TexTarget target;
 972    };
 973
 974 public:
 975    struct ImgFormatDesc
 976    {
 977       char name[19];
 978       uint8_t components;
 979       uint8_t bits[4];
 980       ImgType type;
 981       bool bgra;
 982    };
 983
 984    static const struct ImgFormatDesc formatTable[IMG_FORMAT_COUNT];
 985
 986 public:
 987    TexInstruction(Function *, operation);
 988    virtual ~TexInstruction();
 989
 990    virtual TexInstruction *clone(ClonePolicy<Function>&,
 991                                  Instruction * = NULL) const;
 992
 993    inline void setTexture(Target targ, uint8_t r, uint8_t s)
 994    {
 995       tex.r = r;
 996       tex.s = s;
 997       tex.target = targ;
 998    }
 999
1000    void setIndirectR(Value *);
1001    void setIndirectS(Value *);
1002    inline Value *getIndirectR() const;
1003    inline Value *getIndirectS() const;
1004
1005 public:
1006    struct {
1007       Target target;
1008
1009       uint16_t r;
1010       uint16_t s;
1011       int8_t rIndirectSrc;
1012       int8_t sIndirectSrc;
1013
1014       uint8_t mask;
1015       uint8_t gatherComp;
1016
1017       bool liveOnly; // only execute on live pixels of a quad (optimization)
1018       bool levelZero;
1019       bool derivAll;
1020
1021       int8_t useOffsets; // 0, 1, or 4 for textureGatherOffsets
1022       int8_t offset[3]; // only used on nv50
1023
1024       enum TexQuery query;
1025       const struct ImgFormatDesc *format;
1026    } tex;
1027
1028    ValueRef dPdx[3];
1029    ValueRef dPdy[3];
1030    ValueRef offset[4][3];
1031 };
1032
1033 class CmpInstruction : public Instruction
1034 {
1035 public:
1036    CmpInstruction(Function *, operation);
1037
1038    virtual CmpInstruction *clone(ClonePolicy<Function>&,
1039                                  Instruction * = NULL) const;
1040
1041    void setCondition(CondCode cond) { setCond = cond; }
1042    CondCode getCondition() const { return setCond; }
1043
1044 public:
1045    CondCode setCond;
1046 };
1047
1048 class FlowInstruction : public Instruction
1049 {
1050 public:
1051    FlowInstruction(Function *, operation, void *target);
1052
1053    virtual FlowInstruction *clone(ClonePolicy<Function>&,
1054                                   Instruction * = NULL) const;
1055
1056 public:
1057    unsigned allWarp  : 1;
1058    unsigned absolute : 1;
1059    unsigned limit    : 1;
1060    unsigned builtin  : 1; // true for calls to emulation code
1061    unsigned indirect : 1; // target in src(0)
1062
1063    union {
1064       BasicBlock *bb;
1065       int builtin;
1066       Function *fn;
1067    } target;
1068 };
1069
1070 class BasicBlock
1071 {
1072 public:
1073    BasicBlock(Function *);
1074    ~BasicBlock();
1075
1076    BasicBlock *clone(ClonePolicy<Function>&) const;
1077
1078    inline int getId() const { return id; }
1079    inline unsigned int getInsnCount() const { return numInsns; }
1080    inline bool isTerminated() const { return exit && exit->terminator; }
1081
1082    bool dominatedBy(BasicBlock *bb);
1083    inline bool reachableBy(const BasicBlock *by, const BasicBlock *term);
1084
1085    // returns mask of conditional out blocks
1086    // e.g. 3 for IF { .. } ELSE { .. } ENDIF, 1 for IF { .. } ENDIF
1087    unsigned int initiatesSimpleConditional() const;
1088
1089 public:
1090    Function *getFunction() const { return func; }
1091    Program *getProgram() const { return program; }
1092
1093    Instruction *getEntry() const { return entry; } // first non-phi instruction
1094    Instruction *getPhi() const { return phi; }
1095    Instruction *getFirst() const { return phi ? phi : entry; }
1096    Instruction *getExit() const { return exit; }
1097
1098    void insertHead(Instruction *);
1099    void insertTail(Instruction *);
1100    void insertBefore(Instruction *, Instruction *);
1101    void insertAfter(Instruction *, Instruction *);
1102    void remove(Instruction *);
1103    void permuteAdjacent(Instruction *, Instruction *);
1104
1105    BasicBlock *idom() const;
1106
1107    // NOTE: currently does not rebuild the dominator tree
1108    BasicBlock *splitBefore(Instruction *, bool attach = true);
1109    BasicBlock *splitAfter(Instruction *, bool attach = true);
1110
1111    DLList& getDF() { return df; }
1112    DLList::Iterator iterDF() { return df.iterator(); }
1113
1114    static inline BasicBlock *get(Iterator&);
1115    static inline BasicBlock *get(Graph::Node *);
1116
1117 public:
1118    Graph::Node cfg; // first edge is branch *taken* (the ELSE branch)
1119    Graph::Node dom;
1120
1121    BitSet liveSet;
1122    BitSet defSet;
1123
1124    uint32_t binPos;
1125    uint32_t binSize;
1126
1127    Instruction *joinAt; // for quick reference
1128
1129    bool explicitCont; // loop headers: true if loop contains continue stmts
1130
1131 private:
1132    int id;
1133    DLList df;
1134
1135    Instruction *phi;
1136    Instruction *entry;
1137    Instruction *exit;
1138
1139    unsigned int numInsns;
1140
1141 private:
1142    Function *func;
1143    Program *program;
1144
1145    void splitCommon(Instruction *, BasicBlock *, bool attach);
1146 };
1147
1148 class Function
1149 {
1150 public:
1151    Function(Program *, const char *name, uint32_t label);
1152    ~Function();
1153
1154    static inline Function *get(Graph::Node *node);
1155
1156    inline Program *getProgram() const { return prog; }
1157    inline const char *getName() const { return name; }
1158    inline int getId() const { return id; }
1159    inline uint32_t getLabel() const { return label; }
1160
1161    void print();
1162    void printLiveIntervals() const;
1163    void printCFGraph(const char *filePath);
1164
1165    bool setEntry(BasicBlock *);
1166    bool setExit(BasicBlock *);
1167
1168    unsigned int orderInstructions(ArrayList&);
1169
1170    inline void add(BasicBlock *bb, int& id) { allBBlocks.insert(bb, id); }
1171    inline void add(Instruction *insn, int& id) { allInsns.insert(insn, id); }
1172    inline void add(LValue *lval, int& id) { allLValues.insert(lval, id); }
1173
1174    inline LValue *getLValue(int id);
1175
1176    void buildLiveSets();
1177    void buildDefSets();
1178    bool convertToSSA();
1179
1180 public:
1181    std::deque<ValueDef> ins;
1182    std::deque<ValueRef> outs;
1183    std::deque<Value *> clobbers;
1184
1185    Graph cfg;
1186    Graph::Node *cfgExit;
1187    Graph *domTree;
1188    Graph::Node call; // node in the call graph
1189
1190    BasicBlock **bbArray; // BBs in emission order
1191    int bbCount;
1192
1193    unsigned int loopNestingBound;
1194    int regClobberMax;
1195
1196    uint32_t binPos;
1197    uint32_t binSize;
1198
1199    Value *stackPtr;
1200
1201    uint32_t tlsBase; // base address for l[] space (if no stack pointer is used)
1202    uint32_t tlsSize;
1203
1204    ArrayList allBBlocks;
1205    ArrayList allInsns;
1206    ArrayList allLValues;
1207
1208 private:
1209    void buildLiveSetsPreSSA(BasicBlock *, const int sequence);
1210    void buildDefSetsPreSSA(BasicBlock *bb, const int seq);
1211
1212 private:
1213    uint32_t label;
1214    int id;
1215    const char *const name;
1216    Program *prog;
1217 };
1218
1219 enum CGStage
1220 {
1221    CG_STAGE_PRE_SSA,
1222    CG_STAGE_SSA, // expected directly before register allocation
1223    CG_STAGE_POST_RA
1224 };
1225
1226 class Program
1227 {
1228 public:
1229    enum Type
1230    {
1231       TYPE_VERTEX,
1232       TYPE_TESSELLATION_CONTROL,
1233       TYPE_TESSELLATION_EVAL,
1234       TYPE_GEOMETRY,
1235       TYPE_FRAGMENT,
1236       TYPE_COMPUTE
1237    };
1238
1239    Program(Type type, Target *targ);
1240    ~Program();
1241
1242    void print();
1243
1244    Type getType() const { return progType; }
1245
1246    inline void add(Function *fn, int& id) { allFuncs.insert(fn, id); }
1247    inline void del(Function *fn, int& id) { allFuncs.remove(id); }
1248    inline void add(Value *rval, int& id) { allRValues.insert(rval, id); }
1249
1250    bool makeFromTGSI(struct nv50_ir_prog_info *);
1251    bool makeFromSM4(struct nv50_ir_prog_info *);
1252    bool convertToSSA();
1253    bool optimizeSSA(int level);
1254    bool optimizePostRA(int level);
1255    bool registerAllocation();
1256    bool emitBinary(struct nv50_ir_prog_info *);
1257
1258    const Target *getTarget() const { return target; }
1259
1260 private:
1261    void emitSymbolTable(struct nv50_ir_prog_info *);
1262
1263    Type progType;
1264    Target *target;
1265
1266 public:
1267    Function *main;
1268    Graph calls;
1269
1270    ArrayList allFuncs;
1271    ArrayList allRValues;
1272
1273    uint32_t *code;
1274    uint32_t binSize;
1275    uint32_t tlsSize; // size required for FILE_MEMORY_LOCAL
1276
1277    int maxGPR;
1278
1279    MemoryPool mem_Instruction;
1280    MemoryPool mem_CmpInstruction;
1281    MemoryPool mem_TexInstruction;
1282    MemoryPool mem_FlowInstruction;
1283    MemoryPool mem_LValue;
1284    MemoryPool mem_Symbol;
1285    MemoryPool mem_ImmediateValue;
1286
1287    uint32_t dbgFlags;
1288    uint8_t  optLevel;
1289
1290    void *targetPriv; // e.g. to carry information between passes
1291
1292    const struct nv50_ir_prog_info *driver; // for driver configuration
1293
1294    void releaseInstruction(Instruction *);
1295    void releaseValue(Value *);
1296 };
1297
1298 // TODO: add const version
1299 class Pass
1300 {
1301 public:
1302    bool run(Program *, bool ordered = false, bool skipPhi = false);
1303    bool run(Function *, bool ordered = false, bool skipPhi = false);
1304
1305 private:
1306    // return false to continue with next entity on next higher level
1307    virtual bool visit(Function *) { return true; }
1308    virtual bool visit(BasicBlock *) { return true; }
1309    virtual bool visit(Instruction *) { return false; }
1310
1311    bool doRun(Program *, bool ordered, bool skipPhi);
1312    bool doRun(Function *, bool ordered, bool skipPhi);
1313
1314 protected:
1315    bool err;
1316    Function *func;
1317    Program *prog;
1318 };
1319
1320 // =============================================================================
1321
1322 #include "codegen/nv50_ir_inlines.h"
1323
1324 } // namespace nv50_ir
1325
1326 #endif // __NV50_IR_H__