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