src/gallium/drivers/nouveau/codegen/nv50_ir.h

   1 /*
   2  * Copyright 2011 Christoph Bumiller
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice shall be included in
  12  * all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
  18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  20  * OTHER DEALINGS IN THE SOFTWARE.
  21  */
  22
  23 #ifndef __NV50_IR_H__
  24 #define __NV50_IR_H__
  25
  26 #include <stdio.h>
  27 #include <stdlib.h>
  28 #include <stdint.h>
  29 #include <deque>
  30 #include <list>
  31 #include <vector>
  32
  33 #include "codegen/nv50_ir_util.h"
  34 #include "codegen/nv50_ir_graph.h"
  35
  36 #include "codegen/nv50_ir_driver.h"
  37
  38 namespace nv50_ir {
  39
  40 enum operation
  41 {
  42    OP_NOP = 0,
  43    OP_PHI,
  44    OP_UNION, // unify a new definition and several source values
  45    OP_SPLIT, // $r0d -> { $r0, $r1 } ($r0d and $r0/$r1 will be coalesced)
  46    OP_MERGE, // opposite of split, e.g. combine 2 32 bit into a 64 bit value
  47    OP_CONSTRAINT, // copy values into consecutive registers
  48    OP_MOV, // simple copy, no modifiers allowed
  49    OP_LOAD,
  50    OP_STORE,
  51    OP_ADD, // NOTE: add u64 + u32 is legal for targets w/o 64-bit integer adds
  52    OP_SUB,
  53    OP_MUL,
  54    OP_DIV,
  55    OP_MOD,
  56    OP_MAD,
  57    OP_FMA,
  58    OP_SAD, // abs(src0 - src1) + src2
  59    OP_ABS,
  60    OP_NEG,
  61    OP_NOT,
  62    OP_AND,
  63    OP_OR,
  64    OP_XOR,
  65    OP_SHL,
  66    OP_SHR,
  67    OP_MAX,
  68    OP_MIN,
  69    OP_SAT, // CLAMP(f32, 0.0, 1.0)
  70    OP_CEIL,
  71    OP_FLOOR,
  72    OP_TRUNC,
  73    OP_CVT,
  74    OP_SET_AND, // dst = (src0 CMP src1) & src2
  75    OP_SET_OR,
  76    OP_SET_XOR,
  77    OP_SET,
  78    OP_SELP, // dst = src2 ? src0 : src1
  79    OP_SLCT, // dst = (src2 CMP 0) ? src0 : src1
  80    OP_RCP,
  81    OP_RSQ,
  82    OP_LG2,
  83    OP_SIN,
  84    OP_COS,
  85    OP_EX2,
  86    OP_EXP, // exponential (base M_E)
  87    OP_LOG, // natural logarithm
  88    OP_PRESIN,
  89    OP_PREEX2,
  90    OP_SQRT,
  91    OP_POW,
  92    OP_BRA,
  93    OP_CALL,
  94    OP_RET,
  95    OP_CONT,
  96    OP_BREAK,
  97    OP_PRERET,
  98    OP_PRECONT,
  99    OP_PREBREAK,
 100    OP_BRKPT,     // breakpoint (not related to loops)
 101    OP_JOINAT,    // push control flow convergence point
 102    OP_JOIN,      // converge
 103    OP_DISCARD,
 104    OP_EXIT,
 105    OP_MEMBAR, // memory barrier (mfence, lfence, sfence)
 106    OP_VFETCH, // indirection 0 in attribute space, indirection 1 is vertex base
 107    OP_PFETCH, // fetch base address of vertex src0 (immediate) [+ src1]
 108    OP_EXPORT,
 109    OP_LINTERP,
 110    OP_PINTERP,
 111    OP_EMIT,    // emit vertex
 112    OP_RESTART, // restart primitive
 113    OP_TEX,
 114    OP_TXB, // texture bias
 115    OP_TXL, // texure lod
 116    OP_TXF, // texel fetch
 117    OP_TXQ, // texture size query
 118    OP_TXD, // texture derivatives
 119    OP_TXG, // texture gather
 120    OP_TXLQ, // texture query lod
 121    OP_TEXCSAA, // texture op for coverage sampling
 122    OP_TEXPREP, // turn cube map array into 2d array coordinates
 123    OP_SULDB, // surface load (raw)
 124    OP_SULDP, // surface load (formatted)
 125    OP_SUSTB, // surface store (raw)
 126    OP_SUSTP, // surface store (formatted)
 127    OP_SUREDB,
 128    OP_SUREDP, // surface reduction (atomic op)
 129    OP_SULEA,   // surface load effective address
 130    OP_SUBFM,   // surface bitfield manipulation
 131    OP_SUCLAMP, // clamp surface coordinates
 132    OP_SUEAU,   // surface effective address
 133    OP_MADSP,   // special integer multiply-add
 134    OP_TEXBAR, // texture dependency barrier
 135    OP_DFDX,
 136    OP_DFDY,
 137    OP_RDSV, // read system value
 138    OP_WRSV, // write system value
 139    OP_PIXLD, // get info about raster object or surfaces
 140    OP_QUADOP,
 141    OP_QUADON,
 142    OP_QUADPOP,
 143    OP_POPCNT, // bitcount(src0 & src1)
 144    OP_INSBF,  // insert first src1[8:15] bits of src0 into src2 at src1[0:7]
 145    OP_EXTBF,  // place bits [K,K+N) of src0 into dst, src1 = 0xNNKK
 146    OP_BFIND,  // find highest/lowest set bit
 147    OP_PERMT,  // dst = bytes from src2,src0 selected by src1 (nvc0's src order)
 148    OP_ATOM,
 149    OP_BAR,    // execution barrier, sources = { id, thread count, predicate }
 150    OP_VADD,   // byte/word vector operations
 151    OP_VAVG,
 152    OP_VMIN,
 153    OP_VMAX,
 154    OP_VSAD,
 155    OP_VSET,
 156    OP_VSHR,
 157    OP_VSHL,
 158    OP_VSEL,
 159    OP_CCTL, // cache control
 160    OP_LAST
 161 };
 162
 163 // various instruction-specific modifier definitions Instruction::subOp
 164 // MOV_FINAL marks a MOV originating from an EXPORT (used for placing TEXBARs)
 165 #define NV50_IR_SUBOP_MUL_HIGH     1
 166 #define NV50_IR_SUBOP_EMIT_RESTART 1
 167 #define NV50_IR_SUBOP_LDC_IL       1
 168 #define NV50_IR_SUBOP_LDC_IS       2
 169 #define NV50_IR_SUBOP_LDC_ISL      3
 170 #define NV50_IR_SUBOP_SHIFT_WRAP   1
 171 #define NV50_IR_SUBOP_EMU_PRERET   1
 172 #define NV50_IR_SUBOP_TEXBAR(n)    n
 173 #define NV50_IR_SUBOP_MOV_FINAL    1
 174 #define NV50_IR_SUBOP_EXTBF_REV    1
 175 #define NV50_IR_SUBOP_BFIND_SAMT   1
 176 #define NV50_IR_SUBOP_PERMT_F4E    1
 177 #define NV50_IR_SUBOP_PERMT_B4E    2
 178 #define NV50_IR_SUBOP_PERMT_RC8    3
 179 #define NV50_IR_SUBOP_PERMT_ECL    4
 180 #define NV50_IR_SUBOP_PERMT_ECR    5
 181 #define NV50_IR_SUBOP_PERMT_RC16   6
 182 #define NV50_IR_SUBOP_BAR_SYNC     0
 183 #define NV50_IR_SUBOP_BAR_ARRIVE   1
 184 #define NV50_IR_SUBOP_BAR_RED_AND  2
 185 #define NV50_IR_SUBOP_BAR_RED_OR   3
 186 #define NV50_IR_SUBOP_BAR_RED_POPC 4
 187 #define NV50_IR_SUBOP_MEMBAR_L     1
 188 #define NV50_IR_SUBOP_MEMBAR_S     2
 189 #define NV50_IR_SUBOP_MEMBAR_M     3
 190 #define NV50_IR_SUBOP_MEMBAR_CTA  (0 << 2)
 191 #define NV50_IR_SUBOP_MEMBAR_GL   (1 << 2)
 192 #define NV50_IR_SUBOP_MEMBAR_SYS  (2 << 2)
 193 #define NV50_IR_SUBOP_MEMBAR_DIR(m)   ((m) & 0x3)
 194 #define NV50_IR_SUBOP_MEMBAR_SCOPE(m) ((m) & ~0x3)
 195 #define NV50_IR_SUBOP_MEMBAR(d,s) \
 196    (NV50_IR_SUBOP_MEMBAR_##d | NV50_IR_SUBOP_MEMBAR_##s)
 197 #define NV50_IR_SUBOP_ATOM_ADD      0
 198 #define NV50_IR_SUBOP_ATOM_MIN      1
 199 #define NV50_IR_SUBOP_ATOM_MAX      2
 200 #define NV50_IR_SUBOP_ATOM_INC      3
 201 #define NV50_IR_SUBOP_ATOM_DEC      4
 202 #define NV50_IR_SUBOP_ATOM_AND      5
 203 #define NV50_IR_SUBOP_ATOM_OR       6
 204 #define NV50_IR_SUBOP_ATOM_XOR      7
 205 #define NV50_IR_SUBOP_ATOM_CAS      8
 206 #define NV50_IR_SUBOP_ATOM_EXCH     9
 207 #define NV50_IR_SUBOP_CCTL_IV      5
 208 #define NV50_IR_SUBOP_CCTL_IVALL   6
 209 #define NV50_IR_SUBOP_SUST_IGN     0
 210 #define NV50_IR_SUBOP_SUST_TRAP    1
 211 #define NV50_IR_SUBOP_SUST_SDCL    3
 212 #define NV50_IR_SUBOP_SULD_ZERO    0
 213 #define NV50_IR_SUBOP_SULD_TRAP    1
 214 #define NV50_IR_SUBOP_SULD_SDCL    3
 215 #define NV50_IR_SUBOP_SUBFM_3D     1
 216 #define NV50_IR_SUBOP_SUCLAMP_2D   0x10
 217 #define NV50_IR_SUBOP_SUCLAMP_SD(r, d) (( 0 + (r)) | ((d == 2) ? 0x10 : 0))
 218 #define NV50_IR_SUBOP_SUCLAMP_PL(r, d) (( 5 + (r)) | ((d == 2) ? 0x10 : 0))
 219 #define NV50_IR_SUBOP_SUCLAMP_BL(r, d) ((10 + (r)) | ((d == 2) ? 0x10 : 0))
 220 #define NV50_IR_SUBOP_PIXLD_COUNT       0
 221 #define NV50_IR_SUBOP_PIXLD_COVMASK     1
 222 #define NV50_IR_SUBOP_PIXLD_COVERED     2
 223 #define NV50_IR_SUBOP_PIXLD_OFFSET      3
 224 #define NV50_IR_SUBOP_PIXLD_CENT_OFFSET 4
 225 #define NV50_IR_SUBOP_PIXLD_SAMPLEID    5
 226 #define NV50_IR_SUBOP_MADSP_SD     0xffff
 227 // Yes, we could represent those with DataType.
 228 // Or put the type into operation and have a couple 1000 values in that enum.
 229 // This will have to do for now.
 230 // The bitfields are supposed to correspond to nve4 ISA.
 231 #define NV50_IR_SUBOP_MADSP(a,b,c) (((c) << 8) | ((b) << 4) | (a))
 232 #define NV50_IR_SUBOP_V1(d,a,b)    (((d) << 10) | ((b) << 5) | (a) | 0x0000)
 233 #define NV50_IR_SUBOP_V2(d,a,b)    (((d) << 10) | ((b) << 5) | (a) | 0x4000)
 234 #define NV50_IR_SUBOP_V4(d,a,b)    (((d) << 10) | ((b) << 5) | (a) | 0x8000)
 235 #define NV50_IR_SUBOP_Vn(n)        ((n) >> 14)
 236
 237 enum DataType
 238 {
 239    TYPE_NONE,
 240    TYPE_U8,
 241    TYPE_S8,
 242    TYPE_U16,
 243    TYPE_S16,
 244    TYPE_U32,
 245    TYPE_S32,
 246    TYPE_U64, // 64 bit operations are only lowered after register allocation
 247    TYPE_S64,
 248    TYPE_F16,
 249    TYPE_F32,
 250    TYPE_F64,
 251    TYPE_B96,
 252    TYPE_B128
 253 };
 254
 255 enum CondCode
 256 {
 257    CC_FL = 0,
 258    CC_NEVER = CC_FL, // when used with FILE_FLAGS
 259    CC_LT = 1,
 260    CC_EQ = 2,
 261    CC_NOT_P = CC_EQ, // when used with FILE_PREDICATE
 262    CC_LE = 3,
 263    CC_GT = 4,
 264    CC_NE = 5,
 265    CC_P  = CC_NE,
 266    CC_GE = 6,
 267    CC_TR = 7,
 268    CC_ALWAYS = CC_TR,
 269    CC_U  = 8,
 270    CC_LTU = 9,
 271    CC_EQU = 10,
 272    CC_LEU = 11,
 273    CC_GTU = 12,
 274    CC_NEU = 13,
 275    CC_GEU = 14,
 276    CC_NO = 0x10,
 277    CC_NC = 0x11,
 278    CC_NS = 0x12,
 279    CC_NA = 0x13,
 280    CC_A  = 0x14,
 281    CC_S  = 0x15,
 282    CC_C  = 0x16,
 283    CC_O  = 0x17
 284 };
 285
 286 enum RoundMode
 287 {
 288    ROUND_N, // nearest
 289    ROUND_M, // towards -inf
 290    ROUND_Z, // towards 0
 291    ROUND_P, // towards +inf
 292    ROUND_NI, // nearest integer
 293    ROUND_MI, // to integer towards -inf
 294    ROUND_ZI, // to integer towards 0
 295    ROUND_PI, // to integer towards +inf
 296 };
 297
 298 enum CacheMode
 299 {
 300    CACHE_CA,            // cache at all levels
 301    CACHE_WB = CACHE_CA, // cache write back
 302    CACHE_CG,            // cache at global level
 303    CACHE_CS,            // cache streaming
 304    CACHE_CV,            // cache as volatile
 305    CACHE_WT = CACHE_CV  // cache write-through
 306 };
 307
 308 enum DataFile
 309 {
 310    FILE_NULL = 0,
 311    FILE_GPR,
 312    FILE_PREDICATE,       // boolean predicate
 313    FILE_FLAGS,           // zero/sign/carry/overflow bits
 314    FILE_ADDRESS,
 315    LAST_REGISTER_FILE = FILE_ADDRESS,
 316    FILE_IMMEDIATE,
 317    FILE_MEMORY_CONST,
 318    FILE_SHADER_INPUT,
 319    FILE_SHADER_OUTPUT,
 320    FILE_MEMORY_GLOBAL,
 321    FILE_MEMORY_SHARED,
 322    FILE_MEMORY_LOCAL,
 323    FILE_SYSTEM_VALUE,
 324    DATA_FILE_COUNT
 325 };
 326
 327 enum TexTarget
 328 {
 329    TEX_TARGET_1D,
 330    TEX_TARGET_2D,
 331    TEX_TARGET_2D_MS,
 332    TEX_TARGET_3D,
 333    TEX_TARGET_CUBE,
 334    TEX_TARGET_1D_SHADOW,
 335    TEX_TARGET_2D_SHADOW,
 336    TEX_TARGET_CUBE_SHADOW,
 337    TEX_TARGET_1D_ARRAY,
 338    TEX_TARGET_2D_ARRAY,
 339    TEX_TARGET_2D_MS_ARRAY,
 340    TEX_TARGET_CUBE_ARRAY,
 341    TEX_TARGET_1D_ARRAY_SHADOW,
 342    TEX_TARGET_2D_ARRAY_SHADOW,
 343    TEX_TARGET_RECT,
 344    TEX_TARGET_RECT_SHADOW,
 345    TEX_TARGET_CUBE_ARRAY_SHADOW,
 346    TEX_TARGET_BUFFER,
 347    TEX_TARGET_COUNT
 348 };
 349
 350 enum SVSemantic
 351 {
 352    SV_POSITION, // WPOS
 353    SV_VERTEX_ID,
 354    SV_INSTANCE_ID,
 355    SV_INVOCATION_ID,
 356    SV_PRIMITIVE_ID,
 357    SV_VERTEX_COUNT, // gl_PatchVerticesIn
 358    SV_LAYER,
 359    SV_VIEWPORT_INDEX,
 360    SV_YDIR,
 361    SV_FACE,
 362    SV_POINT_SIZE,
 363    SV_POINT_COORD,
 364    SV_CLIP_DISTANCE,
 365    SV_SAMPLE_INDEX,
 366    SV_SAMPLE_POS,
 367    SV_SAMPLE_MASK,
 368    SV_TESS_FACTOR,
 369    SV_TESS_COORD,
 370    SV_TID,
 371    SV_CTAID,
 372    SV_NTID,
 373    SV_GRIDID,
 374    SV_NCTAID,
 375    SV_LANEID,
 376    SV_PHYSID,
 377    SV_NPHYSID,
 378    SV_CLOCK,
 379    SV_LBASE,
 380    SV_SBASE,
 381    SV_VERTEX_STRIDE,
 382    SV_UNDEFINED,
 383    SV_LAST
 384 };
 385
 386 class Program;
 387 class Function;
 388 class BasicBlock;
 389
 390 class Target;
 391
 392 class Instruction;
 393 class CmpInstruction;
 394 class TexInstruction;
 395 class FlowInstruction;
 396
 397 class Value;
 398 class LValue;
 399 class Symbol;
 400 class ImmediateValue;
 401
 402 struct Storage
 403 {
 404    DataFile file;
 405    int8_t fileIndex; // signed, may be indirect for CONST[]
 406    uint8_t size; // this should match the Instruction type's size
 407    DataType type; // mainly for pretty printing
 408    union {
 409       uint64_t u64;    // immediate values
 410       uint32_t u32;
 411       uint16_t u16;
 412       uint8_t u8;
 413       int64_t s64;
 414       int32_t s32;
 415       int16_t s16;
 416       int8_t s8;
 417       float f32;
 418       double f64;
 419       int32_t offset; // offset from 0 (base of address space)
 420       int32_t id;     // register id (< 0 if virtual/unassigned, in units <= 4)
 421       struct {
 422          SVSemantic sv;
 423          int index;
 424       } sv;
 425    } data;
 426 };
 427
 428 // precedence: NOT after SAT after NEG after ABS
 429 #define NV50_IR_MOD_ABS (1 << 0)
 430 #define NV50_IR_MOD_NEG (1 << 1)
 431 #define NV50_IR_MOD_SAT (1 << 2)
 432 #define NV50_IR_MOD_NOT (1 << 3)
 433 #define NV50_IR_MOD_NEG_ABS (NV50_IR_MOD_NEG | NV50_IR_MOD_ABS)
 434
 435 #define NV50_IR_INTERP_MODE_MASK   0x3
 436 #define NV50_IR_INTERP_LINEAR      (0 << 0)
 437 #define NV50_IR_INTERP_PERSPECTIVE (1 << 0)
 438 #define NV50_IR_INTERP_FLAT        (2 << 0)
 439 #define NV50_IR_INTERP_SC          (3 << 0) // what exactly is that ?
 440 #define NV50_IR_INTERP_SAMPLE_MASK 0xc
 441 #define NV50_IR_INTERP_DEFAULT     (0 << 2)
 442 #define NV50_IR_INTERP_CENTROID    (1 << 2)
 443 #define NV50_IR_INTERP_OFFSET      (2 << 2)
 444 #define NV50_IR_INTERP_SAMPLEID    (3 << 2)
 445
 446 // do we really want this to be a class ?
 447 class Modifier
 448 {
 449 public:
 450    Modifier() : bits(0) { }
 451    Modifier(unsigned int m) : bits(m) { }
 452    Modifier(operation op);
 453
 454    // @return new Modifier applying a after b (asserts if unrepresentable)
 455    Modifier operator*(const Modifier) const;
 456    Modifier operator*=(const Modifier m) { *this = *this * m; return *this; }
 457    Modifier operator==(const Modifier m) const { return m.bits == bits; }
 458    Modifier operator!=(const Modifier m) const { return m.bits != bits; }
 459
 460    inline Modifier operator&(const Modifier m) const { return bits & m.bits; }
 461    inline Modifier operator|(const Modifier m) const { return bits | m.bits; }
 462    inline Modifier operator^(const Modifier m) const { return bits ^ m.bits; }
 463
 464    operation getOp() const;
 465
 466    inline int neg() const { return (bits & NV50_IR_MOD_NEG) ? 1 : 0; }
 467    inline int abs() const { return (bits & NV50_IR_MOD_ABS) ? 1 : 0; }
 468
 469    inline operator bool() const { return bits ? true : false; }
 470
 471    void applyTo(ImmediateValue &imm) const;
 472
 473    int print(char *buf, size_t size) const;
 474
 475 private:
 476    uint8_t bits;
 477 };
 478
 479 class ValueRef
 480 {
 481 public:
 482    ValueRef(Value * = NULL);
 483    ValueRef(const ValueRef&);
 484    ~ValueRef();
 485
 486    inline bool exists() const { return value != NULL; }
 487
 488    void set(Value *);
 489    void set(const ValueRef&);
 490    inline Value *get() const { return value; }
 491    inline Value *rep() const;
 492
 493    inline Instruction *getInsn() const { return insn; }
 494    inline void setInsn(Instruction *inst) { insn = inst; }
 495
 496    inline bool isIndirect(int dim) const { return indirect[dim] >= 0; }
 497    inline const ValueRef *getIndirect(int dim) const;
 498
 499    inline DataFile getFile() const;
 500    inline unsigned getSize() const;
 501
 502    // SSA: return eventual (traverse MOVs) literal value, if it exists
 503    bool getImmediate(ImmediateValue&) const;
 504
 505 public:
 506    Modifier mod;
 507    int8_t indirect[2]; // >= 0 if relative to lvalue in insn->src(indirect[i])
 508    uint8_t swizzle;
 509
 510    bool usedAsPtr; // for printing
 511
 512 private:
 513    Value *value;
 514    Instruction *insn;
 515 };
 516
 517 class ValueDef
 518 {
 519 public:
 520    ValueDef(Value * = NULL);
 521    ValueDef(const ValueDef&);
 522    ~ValueDef();
 523
 524    inline bool exists() const { return value != NULL; }
 525
 526    inline Value *get() const { return value; }
 527    inline Value *rep() const;
 528    void set(Value *);
 529    bool mayReplace(const ValueRef &);
 530    void replace(const ValueRef &, bool doSet); // replace all uses of the old value
 531
 532    inline Instruction *getInsn() const { return insn; }
 533    inline void setInsn(Instruction *inst) { insn = inst; }
 534
 535    inline DataFile getFile() const;
 536    inline unsigned getSize() const;
 537
 538    inline void setSSA(LValue *);
 539    inline const LValue *preSSA() const;
 540
 541 private:
 542    Value *value;   // should make this LValue * ...
 543    LValue *origin; // pre SSA value
 544    Instruction *insn;
 545 };
 546
 547 class Value
 548 {
 549 public:
 550    Value();
 551    virtual ~Value() { }
 552
 553    virtual Value *clone(ClonePolicy<Function>&) const = 0;
 554
 555    virtual int print(char *, size_t, DataType ty = TYPE_NONE) const = 0;
 556
 557    virtual bool equals(const Value *, bool strict = false) const;
 558    virtual bool interfers(const Value *) const;
 559    virtual bool isUniform() const { return true; }
 560
 561    inline Value *rep() const { return join; }
 562
 563    inline Instruction *getUniqueInsn() const;
 564    inline Instruction *getInsn() const; // use when uniqueness is certain
 565
 566    inline int refCount() { return uses.size(); }
 567
 568    inline LValue *asLValue();
 569    inline Symbol *asSym();
 570    inline ImmediateValue *asImm();
 571    inline const Symbol *asSym() const;
 572    inline const ImmediateValue *asImm() const;
 573
 574    inline bool inFile(DataFile f) { return reg.file == f; }
 575
 576    static inline Value *get(Iterator&);
 577
 578    std::list<ValueRef *> uses;
 579    std::list<ValueDef *> defs;
 580    typedef std::list<ValueRef *>::iterator UseIterator;
 581    typedef std::list<ValueRef *>::const_iterator UseCIterator;
 582    typedef std::list<ValueDef *>::iterator DefIterator;
 583    typedef std::list<ValueDef *>::const_iterator DefCIterator;
 584
 585    int id;
 586    Storage reg;
 587
 588    // TODO: these should be in LValue:
 589    Interval livei;
 590    Value *join;
 591 };
 592
 593 class LValue : public Value
 594 {
 595 public:
 596    LValue(Function *, DataFile file);
 597    LValue(Function *, LValue *);
 598    ~LValue() { }
 599
 600    virtual bool isUniform() const;
 601
 602    virtual LValue *clone(ClonePolicy<Function>&) const;
 603
 604    virtual int print(char *, size_t, DataType ty = TYPE_NONE) const;
 605
 606 public:
 607    unsigned compMask : 8; // compound/component mask
 608    unsigned compound : 1; // used by RA, value involved in split/merge
 609    unsigned ssa      : 1;
 610    unsigned fixedReg : 1; // set & used by RA, earlier just use (id < 0)
 611    unsigned noSpill  : 1; // do not spill (e.g. if spill temporary already)
 612 };
 613
 614 class Symbol : public Value
 615 {
 616 public:
 617    Symbol(Program *, DataFile file = FILE_MEMORY_CONST, ubyte fileIdx = 0);
 618    ~Symbol() { }
 619
 620    virtual Symbol *clone(ClonePolicy<Function>&) const;
 621
 622    virtual bool equals(const Value *that, bool strict) const;
 623
 624    virtual bool isUniform() const;
 625
 626    virtual int print(char *, size_t, DataType ty = TYPE_NONE) const;
 627
 628    // print with indirect values
 629    int print(char *, size_t, Value *, Value *, DataType ty = TYPE_NONE) const;
 630
 631    inline void setFile(DataFile file, ubyte fileIndex = 0)
 632    {
 633       reg.file = file;
 634       reg.fileIndex = fileIndex;
 635    }
 636
 637    inline void setOffset(int32_t offset);
 638    inline void setAddress(Symbol *base, int32_t offset);
 639    inline void setSV(SVSemantic sv, uint32_t idx = 0);
 640
 641    inline const Symbol *getBase() const { return baseSym; }
 642
 643 private:
 644    Symbol *baseSym; // array base for Symbols representing array elements
 645 };
 646
 647 class ImmediateValue : public Value
 648 {
 649 public:
 650    ImmediateValue() { }
 651    ImmediateValue(Program *, uint32_t);
 652    ImmediateValue(Program *, float);
 653    ImmediateValue(Program *, double);
 654    // NOTE: not added to program with
 655    ImmediateValue(const ImmediateValue *, DataType ty);
 656    ~ImmediateValue() { };
 657
 658    virtual ImmediateValue *clone(ClonePolicy<Function>&) const;
 659
 660    virtual bool equals(const Value *that, bool strict) const;
 661
 662    // these only work if 'type' is valid (we mostly use untyped literals):
 663    bool isInteger(const int ival) const; // ival is cast to this' type
 664    bool isNegative() const;
 665    bool isPow2() const;
 666
 667    void applyLog2();
 668
 669    // for constant folding:
 670    ImmediateValue operator+(const ImmediateValue&) const;
 671    ImmediateValue operator-(const ImmediateValue&) const;
 672    ImmediateValue operator*(const ImmediateValue&) const;
 673    ImmediateValue operator/(const ImmediateValue&) const;
 674
 675    ImmediateValue& operator=(const ImmediateValue&); // only sets value !
 676
 677    bool compare(CondCode cc, float fval) const;
 678
 679    virtual int print(char *, size_t, DataType ty = TYPE_NONE) const;
 680 };
 681
 682 class Instruction
 683 {
 684 public:
 685    Instruction();
 686    Instruction(Function *, operation, DataType);
 687    virtual ~Instruction();
 688
 689    virtual Instruction *clone(ClonePolicy<Function>&,
 690                               Instruction * = NULL) const;
 691
 692    void setDef(int i, Value *);
 693    void setSrc(int s, Value *);
 694    void setSrc(int s, const ValueRef&);
 695    void swapSources(int a, int b);
 696    void moveSources(int s, int delta);
 697    bool setIndirect(int s, int dim, Value *);
 698
 699    inline ValueRef& src(int s) { return srcs[s]; }
 700    inline ValueDef& def(int s) { return defs[s]; }
 701    inline const ValueRef& src(int s) const { return srcs[s]; }
 702    inline const ValueDef& def(int s) const { return defs[s]; }
 703
 704    inline Value *getDef(int d) const { return defs[d].get(); }
 705    inline Value *getSrc(int s) const { return srcs[s].get(); }
 706    inline Value *getIndirect(int s, int dim) const;
 707
 708    inline bool defExists(unsigned d) const
 709    {
 710       return d < defs.size() && defs[d].exists();
 711    }
 712    inline bool srcExists(unsigned s) const
 713    {
 714       return s < srcs.size() && srcs[s].exists();
 715    }
 716
 717    inline bool constrainedDefs() const;
 718
 719    bool setPredicate(CondCode ccode, Value *);
 720    inline Value *getPredicate() const;
 721    bool writesPredicate() const;
 722    inline bool isPredicated() const { return predSrc >= 0; }
 723
 724    inline void setFlagsSrc(int s, Value *);
 725    inline void setFlagsDef(int d, Value *);
 726    inline bool usesFlags() const { return flagsSrc >= 0; }
 727
 728    unsigned int defCount() const { return defs.size(); };
 729    unsigned int defCount(unsigned int mask, bool singleFile = false) const;
 730    unsigned int srcCount() const { return srcs.size(); };
 731    unsigned int srcCount(unsigned int mask, bool singleFile = false) const;
 732
 733    // save & remove / set indirect[0,1] and predicate source
 734    void takeExtraSources(int s, Value *[3]);
 735    void putExtraSources(int s, Value *[3]);
 736
 737    inline void setType(DataType type) { dType = sType = type; }
 738
 739    inline void setType(DataType dtype, DataType stype)
 740    {
 741       dType = dtype;
 742       sType = stype;
 743    }
 744
 745    inline bool isPseudo() const { return op < OP_MOV; }
 746    bool isDead() const;
 747    bool isNop() const;
 748    bool isCommutationLegal(const Instruction *) const; // must be adjacent !
 749    bool isActionEqual(const Instruction *) const;
 750    bool isResultEqual(const Instruction *) const;
 751
 752    void print() const;
 753
 754    inline CmpInstruction *asCmp();
 755    inline TexInstruction *asTex();
 756    inline FlowInstruction *asFlow();
 757    inline const TexInstruction *asTex() const;
 758    inline const CmpInstruction *asCmp() const;
 759    inline const FlowInstruction *asFlow() const;
 760
 761 public:
 762    Instruction *next;
 763    Instruction *prev;
 764    int id;
 765    int serial; // CFG order
 766
 767    operation op;
 768    DataType dType; // destination or defining type
 769    DataType sType; // source or secondary type
 770    CondCode cc;
 771    RoundMode rnd;
 772    CacheMode cache;
 773
 774    uint16_t subOp; // quadop, 1 for mul-high, etc.
 775
 776    unsigned encSize    : 4; // encoding size in bytes
 777    unsigned saturate   : 1; // to [0.0f, 1.0f]
 778    unsigned join       : 1; // converge control flow (use OP_JOIN until end)
 779    unsigned fixed      : 1; // prevent dead code elimination
 780    unsigned terminator : 1; // end of basic block
 781    unsigned ftz        : 1; // flush denormal to zero
 782    unsigned dnz        : 1; // denormals, NaN are zero
 783    unsigned ipa        : 4; // interpolation mode
 784    unsigned lanes      : 4;
 785    unsigned perPatch   : 1;
 786    unsigned exit       : 1; // terminate program after insn
 787    unsigned mask       : 4; // for vector ops
 788
 789    int8_t postFactor; // MUL/DIV(if < 0) by 1 << postFactor
 790
 791    int8_t predSrc;
 792    int8_t flagsDef;
 793    int8_t flagsSrc;
 794
 795    uint8_t sched; // scheduling data (NOTE: maybe move to separate storage)
 796
 797    BasicBlock *bb;
 798
 799 protected:
 800    std::deque<ValueDef> defs; // no gaps !
 801    std::deque<ValueRef> srcs; // no gaps !
 802
 803    // instruction specific methods:
 804    // (don't want to subclass, would need more constructors and memory pools)
 805 public:
 806    inline void setInterpolate(unsigned int mode) { ipa = mode; }
 807
 808    unsigned int getInterpMode() const { return ipa & 0x3; }
 809    unsigned int getSampleMode() const { return ipa & 0xc; }
 810
 811 private:
 812    void init();
 813 };
 814
 815 enum TexQuery
 816 {
 817    TXQ_DIMS,
 818    TXQ_TYPE,
 819    TXQ_SAMPLE_POSITION,
 820    TXQ_FILTER,
 821    TXQ_LOD,
 822    TXQ_WRAP,
 823    TXQ_BORDER_COLOUR
 824 };
 825
 826 class TexInstruction : public Instruction
 827 {
 828 public:
 829    class Target
 830    {
 831    public:
 832       Target(TexTarget targ = TEX_TARGET_2D) : target(targ) { }
 833
 834       const char *getName() const { return descTable[target].name; }
 835       unsigned int getArgCount() const { return descTable[target].argc; }
 836       unsigned int getDim() const { return descTable[target].dim; }
 837       int isArray() const { return descTable[target].array ? 1 : 0; }
 838       int isCube() const { return descTable[target].cube ? 1 : 0; }
 839       int isShadow() const { return descTable[target].shadow ? 1 : 0; }
 840       int isMS() const {
 841         return target == TEX_TARGET_2D_MS || target == TEX_TARGET_2D_MS_ARRAY; }
 842       void clearMS() {
 843          if (isMS()) {
 844             if (isArray())
 845                target = TEX_TARGET_2D_ARRAY;
 846             else
 847                target = TEX_TARGET_2D;
 848          }
 849       }
 850
 851       Target& operator=(TexTarget targ)
 852       {
 853          assert(targ < TEX_TARGET_COUNT);
 854          target = targ;
 855          return *this;
 856       }
 857
 858       inline bool operator==(TexTarget targ) const { return target == targ; }
 859       inline bool operator!=(TexTarget targ) const { return target != targ; }
 860
 861       enum TexTarget getEnum() const { return target; }
 862
 863    private:
 864       struct Desc
 865       {
 866          char name[19];
 867          uint8_t dim;
 868          uint8_t argc;
 869          bool array;
 870          bool cube;
 871          bool shadow;
 872       };
 873
 874       static const struct Desc descTable[TEX_TARGET_COUNT];
 875
 876    private:
 877       enum TexTarget target;
 878    };
 879
 880 public:
 881    TexInstruction(Function *, operation);
 882    virtual ~TexInstruction();
 883
 884    virtual TexInstruction *clone(ClonePolicy<Function>&,
 885                                  Instruction * = NULL) const;
 886
 887    inline void setTexture(Target targ, uint8_t r, uint8_t s)
 888    {
 889       tex.r = r;
 890       tex.s = s;
 891       tex.target = targ;
 892    }
 893
 894    void setIndirectR(Value *);
 895    void setIndirectS(Value *);
 896    inline Value *getIndirectR() const;
 897    inline Value *getIndirectS() const;
 898
 899 public:
 900    struct {
 901       Target target;
 902
 903       uint16_t r;
 904       uint16_t s;
 905       int8_t rIndirectSrc;
 906       int8_t sIndirectSrc;
 907
 908       uint8_t mask;
 909       uint8_t gatherComp;
 910
 911       bool liveOnly; // only execute on live pixels of a quad (optimization)
 912       bool levelZero;
 913       bool derivAll;
 914
 915       int8_t useOffsets; // 0, 1, or 4 for textureGatherOffsets
 916       int8_t offset[4][3];
 917
 918       enum TexQuery query;
 919    } tex;
 920
 921    ValueRef dPdx[3];
 922    ValueRef dPdy[3];
 923 };
 924
 925 class CmpInstruction : public Instruction
 926 {
 927 public:
 928    CmpInstruction(Function *, operation);
 929
 930    virtual CmpInstruction *clone(ClonePolicy<Function>&,
 931                                  Instruction * = NULL) const;
 932
 933    void setCondition(CondCode cond) { setCond = cond; }
 934    CondCode getCondition() const { return setCond; }
 935
 936 public:
 937    CondCode setCond;
 938 };
 939
 940 class FlowInstruction : public Instruction
 941 {
 942 public:
 943    FlowInstruction(Function *, operation, void *target);
 944
 945    virtual FlowInstruction *clone(ClonePolicy<Function>&,
 946                                   Instruction * = NULL) const;
 947
 948 public:
 949    unsigned allWarp  : 1;
 950    unsigned absolute : 1;
 951    unsigned limit    : 1;
 952    unsigned builtin  : 1; // true for calls to emulation code
 953    unsigned indirect : 1; // target in src(0)
 954
 955    union {
 956       BasicBlock *bb;
 957       int builtin;
 958       Function *fn;
 959    } target;
 960 };
 961
 962 class BasicBlock
 963 {
 964 public:
 965    BasicBlock(Function *);
 966    ~BasicBlock();
 967
 968    BasicBlock *clone(ClonePolicy<Function>&) const;
 969
 970    inline int getId() const { return id; }
 971    inline unsigned int getInsnCount() const { return numInsns; }
 972    inline bool isTerminated() const { return exit && exit->terminator; }
 973
 974    bool dominatedBy(BasicBlock *bb);
 975    inline bool reachableBy(const BasicBlock *by, const BasicBlock *term);
 976
 977    // returns mask of conditional out blocks
 978    // e.g. 3 for IF { .. } ELSE { .. } ENDIF, 1 for IF { .. } ENDIF
 979    unsigned int initiatesSimpleConditional() const;
 980
 981 public:
 982    Function *getFunction() const { return func; }
 983    Program *getProgram() const { return program; }
 984
 985    Instruction *getEntry() const { return entry; } // first non-phi instruction
 986    Instruction *getPhi() const { return phi; }
 987    Instruction *getFirst() const { return phi ? phi : entry; }
 988    Instruction *getExit() const { return exit; }
 989
 990    void insertHead(Instruction *);
 991    void insertTail(Instruction *);
 992    void insertBefore(Instruction *, Instruction *);
 993    void insertAfter(Instruction *, Instruction *);
 994    void remove(Instruction *);
 995    void permuteAdjacent(Instruction *, Instruction *);
 996
 997    BasicBlock *idom() const;
 998
 999    // NOTE: currently does not rebuild the dominator tree
1000    BasicBlock *splitBefore(Instruction *, bool attach = true);
1001    BasicBlock *splitAfter(Instruction *, bool attach = true);
1002
1003    DLList& getDF() { return df; }
1004    DLList::Iterator iterDF() { return df.iterator(); }
1005
1006    static inline BasicBlock *get(Iterator&);
1007    static inline BasicBlock *get(Graph::Node *);
1008
1009 public:
1010    Graph::Node cfg; // first edge is branch *taken* (the ELSE branch)
1011    Graph::Node dom;
1012
1013    BitSet liveSet;
1014    BitSet defSet;
1015
1016    uint32_t binPos;
1017    uint32_t binSize;
1018
1019    Instruction *joinAt; // for quick reference
1020
1021    bool explicitCont; // loop headers: true if loop contains continue stmts
1022
1023 private:
1024    int id;
1025    DLList df;
1026
1027    Instruction *phi;
1028    Instruction *entry;
1029    Instruction *exit;
1030
1031    unsigned int numInsns;
1032
1033 private:
1034    Function *func;
1035    Program *program;
1036
1037    void splitCommon(Instruction *, BasicBlock *, bool attach);
1038 };
1039
1040 class Function
1041 {
1042 public:
1043    Function(Program *, const char *name, uint32_t label);
1044    ~Function();
1045
1046    static inline Function *get(Graph::Node *node);
1047
1048    inline Program *getProgram() const { return prog; }
1049    inline const char *getName() const { return name; }
1050    inline int getId() const { return id; }
1051    inline uint32_t getLabel() const { return label; }
1052
1053    void print();
1054    void printLiveIntervals() const;
1055    void printCFGraph(const char *filePath);
1056
1057    bool setEntry(BasicBlock *);
1058    bool setExit(BasicBlock *);
1059
1060    unsigned int orderInstructions(ArrayList&);
1061
1062    inline void add(BasicBlock *bb, int& id) { allBBlocks.insert(bb, id); }
1063    inline void add(Instruction *insn, int& id) { allInsns.insert(insn, id); }
1064    inline void add(LValue *lval, int& id) { allLValues.insert(lval, id); }
1065
1066    inline LValue *getLValue(int id);
1067
1068    void buildLiveSets();
1069    void buildDefSets();
1070    bool convertToSSA();
1071
1072 public:
1073    std::deque<ValueDef> ins;
1074    std::deque<ValueRef> outs;
1075    std::deque<Value *> clobbers;
1076
1077    Graph cfg;
1078    Graph::Node *cfgExit;
1079    Graph *domTree;
1080    Graph::Node call; // node in the call graph
1081
1082    BasicBlock **bbArray; // BBs in emission order
1083    int bbCount;
1084
1085    unsigned int loopNestingBound;
1086    int regClobberMax;
1087
1088    uint32_t binPos;
1089    uint32_t binSize;
1090
1091    Value *stackPtr;
1092
1093    uint32_t tlsBase; // base address for l[] space (if no stack pointer is used)
1094    uint32_t tlsSize;
1095
1096    ArrayList allBBlocks;
1097    ArrayList allInsns;
1098    ArrayList allLValues;
1099
1100 private:
1101    void buildLiveSetsPreSSA(BasicBlock *, const int sequence);
1102    void buildDefSetsPreSSA(BasicBlock *bb, const int seq);
1103
1104 private:
1105    uint32_t label;
1106    int id;
1107    const char *const name;
1108    Program *prog;
1109 };
1110
1111 enum CGStage
1112 {
1113    CG_STAGE_PRE_SSA,
1114    CG_STAGE_SSA, // expected directly before register allocation
1115    CG_STAGE_POST_RA
1116 };
1117
1118 class Program
1119 {
1120 public:
1121    enum Type
1122    {
1123       TYPE_VERTEX,
1124       TYPE_TESSELLATION_CONTROL,
1125       TYPE_TESSELLATION_EVAL,
1126       TYPE_GEOMETRY,
1127       TYPE_FRAGMENT,
1128       TYPE_COMPUTE
1129    };
1130
1131    Program(Type type, Target *targ);
1132    ~Program();
1133
1134    void print();
1135
1136    Type getType() const { return progType; }
1137
1138    inline void add(Function *fn, int& id) { allFuncs.insert(fn, id); }
1139    inline void del(Function *fn, int& id) { allFuncs.remove(id); }
1140    inline void add(Value *rval, int& id) { allRValues.insert(rval, id); }
1141
1142    bool makeFromTGSI(struct nv50_ir_prog_info *);
1143    bool makeFromSM4(struct nv50_ir_prog_info *);
1144    bool convertToSSA();
1145    bool optimizeSSA(int level);
1146    bool optimizePostRA(int level);
1147    bool registerAllocation();
1148    bool emitBinary(struct nv50_ir_prog_info *);
1149
1150    const Target *getTarget() const { return target; }
1151
1152 private:
1153    void emitSymbolTable(struct nv50_ir_prog_info *);
1154
1155    Type progType;
1156    Target *target;
1157
1158 public:
1159    Function *main;
1160    Graph calls;
1161
1162    ArrayList allFuncs;
1163    ArrayList allRValues;
1164
1165    uint32_t *code;
1166    uint32_t binSize;
1167    uint32_t tlsSize; // size required for FILE_MEMORY_LOCAL
1168
1169    int maxGPR;
1170
1171    MemoryPool mem_Instruction;
1172    MemoryPool mem_CmpInstruction;
1173    MemoryPool mem_TexInstruction;
1174    MemoryPool mem_FlowInstruction;
1175    MemoryPool mem_LValue;
1176    MemoryPool mem_Symbol;
1177    MemoryPool mem_ImmediateValue;
1178
1179    uint32_t dbgFlags;
1180    uint8_t  optLevel;
1181
1182    void *targetPriv; // e.g. to carry information between passes
1183
1184    const struct nv50_ir_prog_info *driver; // for driver configuration
1185
1186    void releaseInstruction(Instruction *);
1187    void releaseValue(Value *);
1188 };
1189
1190 // TODO: add const version
1191 class Pass
1192 {
1193 public:
1194    bool run(Program *, bool ordered = false, bool skipPhi = false);
1195    bool run(Function *, bool ordered = false, bool skipPhi = false);
1196
1197 private:
1198    // return false to continue with next entity on next higher level
1199    virtual bool visit(Function *) { return true; }
1200    virtual bool visit(BasicBlock *) { return true; }
1201    virtual bool visit(Instruction *) { return false; }
1202
1203    bool doRun(Program *, bool ordered, bool skipPhi);
1204    bool doRun(Function *, bool ordered, bool skipPhi);
1205
1206 protected:
1207    bool err;
1208    Function *func;
1209    Program *prog;
1210 };
1211
1212 // =============================================================================
1213
1214 #include "codegen/nv50_ir_inlines.h"
1215
1216 } // namespace nv50_ir
1217
1218 #endif // __NV50_IR_H__