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

   1 /*
   2  * Copyright 2011 Christoph Bumiller
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice shall be included in
  12  * all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
  18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  20  * OTHER DEALINGS IN THE SOFTWARE.
  21  */
  22
  23 #ifndef __NV50_IR_H__
  24 #define __NV50_IR_H__
  25
  26 #include <stdio.h>
  27 #include <stdlib.h>
  28 #include <stdint.h>
  29 #include <deque>
  30 #include <list>
  31 #include <vector>
  32
  33 #include "codegen/unordered_set.h"
  34 #include "codegen/nv50_ir_util.h"
  35 #include "codegen/nv50_ir_graph.h"
  36
  37 #include "codegen/nv50_ir_driver.h"
  38
  39 namespace nv50_ir {
  40
  41 enum operation
  42 {
  43    OP_NOP = 0,
  44    OP_PHI,
  45    OP_UNION, // unify a new definition and several source values
  46    OP_SPLIT, // $r0d -> { $r0, $r1 } ($r0d and $r0/$r1 will be coalesced)
  47    OP_MERGE, // opposite of split, e.g. combine 2 32 bit into a 64 bit value
  48    OP_CONSTRAINT, // copy values into consecutive registers
  49    OP_MOV, // simple copy, no modifiers allowed
  50    OP_LOAD,
  51    OP_STORE,
  52    OP_ADD, // NOTE: add u64 + u32 is legal for targets w/o 64-bit integer adds
  53    OP_SUB,
  54    OP_MUL,
  55    OP_DIV,
  56    OP_MOD,
  57    OP_MAD,
  58    OP_FMA,
  59    OP_SAD, // abs(src0 - src1) + src2
  60    OP_ABS,
  61    OP_NEG,
  62    OP_NOT,
  63    OP_AND,
  64    OP_OR,
  65    OP_XOR,
  66    OP_SHL,
  67    OP_SHR,
  68    OP_MAX,
  69    OP_MIN,
  70    OP_SAT, // CLAMP(f32, 0.0, 1.0)
  71    OP_CEIL,
  72    OP_FLOOR,
  73    OP_TRUNC,
  74    OP_CVT,
  75    OP_SET_AND, // dst = (src0 CMP src1) & src2
  76    OP_SET_OR,
  77    OP_SET_XOR,
  78    OP_SET,
  79    OP_SELP, // dst = src2 ? src0 : src1
  80    OP_SLCT, // dst = (src2 CMP 0) ? src0 : src1
  81    OP_RCP,
  82    OP_RSQ,
  83    OP_LG2,
  84    OP_SIN,
  85    OP_COS,
  86    OP_EX2,
  87    OP_EXP, // exponential (base M_E)
  88    OP_LOG, // natural logarithm
  89    OP_PRESIN,
  90    OP_PREEX2,
  91    OP_SQRT,
  92    OP_POW,
  93    OP_BRA,
  94    OP_CALL,
  95    OP_RET,
  96    OP_CONT,
  97    OP_BREAK,
  98    OP_PRERET,
  99    OP_PRECONT,
 100    OP_PREBREAK,
 101    OP_BRKPT,     // breakpoint (not related to loops)
 102    OP_JOINAT,    // push control flow convergence point
 103    OP_JOIN,      // converge
 104    OP_DISCARD,
 105    OP_EXIT,
 106    OP_MEMBAR, // memory barrier (mfence, lfence, sfence)
 107    OP_VFETCH, // indirection 0 in attribute space, indirection 1 is vertex base
 108    OP_PFETCH, // fetch base address of vertex src0 (immediate) [+ src1]
 109    OP_AFETCH, // fetch base address of shader input (a[%r1+0x10])
 110    OP_EXPORT,
 111    OP_LINTERP,
 112    OP_PINTERP,
 113    OP_EMIT,    // emit vertex
 114    OP_RESTART, // restart primitive
 115    OP_TEX,
 116    OP_TXB, // texture bias
 117    OP_TXL, // texure lod
 118    OP_TXF, // texel fetch
 119    OP_TXQ, // texture size query
 120    OP_TXD, // texture derivatives
 121    OP_TXG, // texture gather
 122    OP_TXLQ, // texture query lod
 123    OP_TEXCSAA, // texture op for coverage sampling
 124    OP_TEXPREP, // turn cube map array into 2d array coordinates
 125    OP_SULDB, // surface load (raw)
 126    OP_SULDP, // surface load (formatted)
 127    OP_SUSTB, // surface store (raw)
 128    OP_SUSTP, // surface store (formatted)
 129    OP_SUREDB,
 130    OP_SUREDP, // surface reduction (atomic op)
 131    OP_SULEA,   // surface load effective address
 132    OP_SUBFM,   // surface bitfield manipulation
 133    OP_SUCLAMP, // clamp surface coordinates
 134    OP_SUEAU,   // surface effective address
 135    OP_MADSP,   // special integer multiply-add
 136    OP_TEXBAR, // texture dependency barrier
 137    OP_DFDX,
 138    OP_DFDY,
 139    OP_RDSV, // read system value
 140    OP_WRSV, // write system value
 141    OP_PIXLD, // get info about raster object or surfaces
 142    OP_QUADOP,
 143    OP_QUADON,
 144    OP_QUADPOP,
 145    OP_POPCNT, // bitcount(src0 & src1)
 146    OP_INSBF,  // insert first src1[8:15] bits of src0 into src2 at src1[0:7]
 147    OP_EXTBF,  // place bits [K,K+N) of src0 into dst, src1 = 0xNNKK
 148    OP_BFIND,  // find highest/lowest set bit
 149    OP_PERMT,  // dst = bytes from src2,src0 selected by src1 (nvc0's src order)
 150    OP_ATOM,
 151    OP_BAR,    // execution barrier, sources = { id, thread count, predicate }
 152    OP_VADD,   // byte/word vector operations
 153    OP_VAVG,
 154    OP_VMIN,
 155    OP_VMAX,
 156    OP_VSAD,
 157    OP_VSET,
 158    OP_VSHR,
 159    OP_VSHL,
 160    OP_VSEL,
 161    OP_CCTL, // cache control
 162    OP_SHFL, // warp shuffle
 163    OP_LAST
 164 };
 165
 166 // various instruction-specific modifier definitions Instruction::subOp
 167 // MOV_FINAL marks a MOV originating from an EXPORT (used for placing TEXBARs)
 168 #define NV50_IR_SUBOP_MUL_HIGH     1
 169 #define NV50_IR_SUBOP_EMIT_RESTART 1
 170 #define NV50_IR_SUBOP_LDC_IL       1
 171 #define NV50_IR_SUBOP_LDC_IS       2
 172 #define NV50_IR_SUBOP_LDC_ISL      3
 173 #define NV50_IR_SUBOP_SHIFT_WRAP   1
 174 #define NV50_IR_SUBOP_EMU_PRERET   1
 175 #define NV50_IR_SUBOP_TEXBAR(n)    n
 176 #define NV50_IR_SUBOP_MOV_FINAL    1
 177 #define NV50_IR_SUBOP_EXTBF_REV    1
 178 #define NV50_IR_SUBOP_BFIND_SAMT   1
 179 #define NV50_IR_SUBOP_RCPRSQ_64H   1
 180 #define NV50_IR_SUBOP_PERMT_F4E    1
 181 #define NV50_IR_SUBOP_PERMT_B4E    2
 182 #define NV50_IR_SUBOP_PERMT_RC8    3
 183 #define NV50_IR_SUBOP_PERMT_ECL    4
 184 #define NV50_IR_SUBOP_PERMT_ECR    5
 185 #define NV50_IR_SUBOP_PERMT_RC16   6
 186 #define NV50_IR_SUBOP_BAR_SYNC     0
 187 #define NV50_IR_SUBOP_BAR_ARRIVE   1
 188 #define NV50_IR_SUBOP_BAR_RED_AND  2
 189 #define NV50_IR_SUBOP_BAR_RED_OR   3
 190 #define NV50_IR_SUBOP_BAR_RED_POPC 4
 191 #define NV50_IR_SUBOP_MEMBAR_L     1
 192 #define NV50_IR_SUBOP_MEMBAR_S     2
 193 #define NV50_IR_SUBOP_MEMBAR_M     3
 194 #define NV50_IR_SUBOP_MEMBAR_CTA  (0 << 2)
 195 #define NV50_IR_SUBOP_MEMBAR_GL   (1 << 2)
 196 #define NV50_IR_SUBOP_MEMBAR_SYS  (2 << 2)
 197 #define NV50_IR_SUBOP_MEMBAR_DIR(m)   ((m) & 0x3)
 198 #define NV50_IR_SUBOP_MEMBAR_SCOPE(m) ((m) & ~0x3)
 199 #define NV50_IR_SUBOP_MEMBAR(d,s) \
 200    (NV50_IR_SUBOP_MEMBAR_##d | NV50_IR_SUBOP_MEMBAR_##s)
 201 #define NV50_IR_SUBOP_ATOM_ADD      0
 202 #define NV50_IR_SUBOP_ATOM_MIN      1
 203 #define NV50_IR_SUBOP_ATOM_MAX      2
 204 #define NV50_IR_SUBOP_ATOM_INC      3
 205 #define NV50_IR_SUBOP_ATOM_DEC      4
 206 #define NV50_IR_SUBOP_ATOM_AND      5
 207 #define NV50_IR_SUBOP_ATOM_OR       6
 208 #define NV50_IR_SUBOP_ATOM_XOR      7
 209 #define NV50_IR_SUBOP_ATOM_CAS      8
 210 #define NV50_IR_SUBOP_ATOM_EXCH     9
 211 #define NV50_IR_SUBOP_CCTL_IV      5
 212 #define NV50_IR_SUBOP_CCTL_IVALL   6
 213 #define NV50_IR_SUBOP_SUST_IGN     0
 214 #define NV50_IR_SUBOP_SUST_TRAP    1
 215 #define NV50_IR_SUBOP_SUST_SDCL    3
 216 #define NV50_IR_SUBOP_SULD_ZERO    0
 217 #define NV50_IR_SUBOP_SULD_TRAP    1
 218 #define NV50_IR_SUBOP_SULD_SDCL    3
 219 #define NV50_IR_SUBOP_SUBFM_3D     1
 220 #define NV50_IR_SUBOP_SUCLAMP_2D   0x10
 221 #define NV50_IR_SUBOP_SUCLAMP_SD(r, d) (( 0 + (r)) | ((d == 2) ? 0x10 : 0))
 222 #define NV50_IR_SUBOP_SUCLAMP_PL(r, d) (( 5 + (r)) | ((d == 2) ? 0x10 : 0))
 223 #define NV50_IR_SUBOP_SUCLAMP_BL(r, d) ((10 + (r)) | ((d == 2) ? 0x10 : 0))
 224 #define NV50_IR_SUBOP_PIXLD_COUNT       0
 225 #define NV50_IR_SUBOP_PIXLD_COVMASK     1
 226 #define NV50_IR_SUBOP_PIXLD_COVERED     2
 227 #define NV50_IR_SUBOP_PIXLD_OFFSET      3
 228 #define NV50_IR_SUBOP_PIXLD_CENT_OFFSET 4
 229 #define NV50_IR_SUBOP_PIXLD_SAMPLEID    5
 230 #define NV50_IR_SUBOP_SHFL_IDX  0
 231 #define NV50_IR_SUBOP_SHFL_UP   1
 232 #define NV50_IR_SUBOP_SHFL_DOWN 2
 233 #define NV50_IR_SUBOP_SHFL_BFLY 3
 234 #define NV50_IR_SUBOP_MADSP_SD     0xffff
 235 // Yes, we could represent those with DataType.
 236 // Or put the type into operation and have a couple 1000 values in that enum.
 237 // This will have to do for now.
 238 // The bitfields are supposed to correspond to nve4 ISA.
 239 #define NV50_IR_SUBOP_MADSP(a,b,c) (((c) << 8) | ((b) << 4) | (a))
 240 #define NV50_IR_SUBOP_V1(d,a,b)    (((d) << 10) | ((b) << 5) | (a) | 0x0000)
 241 #define NV50_IR_SUBOP_V2(d,a,b)    (((d) << 10) | ((b) << 5) | (a) | 0x4000)
 242 #define NV50_IR_SUBOP_V4(d,a,b)    (((d) << 10) | ((b) << 5) | (a) | 0x8000)
 243 #define NV50_IR_SUBOP_Vn(n)        ((n) >> 14)
 244
 245 enum DataType
 246 {
 247    TYPE_NONE,
 248    TYPE_U8,
 249    TYPE_S8,
 250    TYPE_U16,
 251    TYPE_S16,
 252    TYPE_U32,
 253    TYPE_S32,
 254    TYPE_U64, // 64 bit operations are only lowered after register allocation
 255    TYPE_S64,
 256    TYPE_F16,
 257    TYPE_F32,
 258    TYPE_F64,
 259    TYPE_B96,
 260    TYPE_B128
 261 };
 262
 263 enum CondCode
 264 {
 265    CC_FL = 0,
 266    CC_NEVER = CC_FL, // when used with FILE_FLAGS
 267    CC_LT = 1,
 268    CC_EQ = 2,
 269    CC_NOT_P = CC_EQ, // when used with FILE_PREDICATE
 270    CC_LE = 3,
 271    CC_GT = 4,
 272    CC_NE = 5,
 273    CC_P  = CC_NE,
 274    CC_GE = 6,
 275    CC_TR = 7,
 276    CC_ALWAYS = CC_TR,
 277    CC_U  = 8,
 278    CC_LTU = 9,
 279    CC_EQU = 10,
 280    CC_LEU = 11,
 281    CC_GTU = 12,
 282    CC_NEU = 13,
 283    CC_GEU = 14,
 284    CC_NO = 0x10,
 285    CC_NC = 0x11,
 286    CC_NS = 0x12,
 287    CC_NA = 0x13,
 288    CC_A  = 0x14,
 289    CC_S  = 0x15,
 290    CC_C  = 0x16,
 291    CC_O  = 0x17
 292 };
 293
 294 enum RoundMode
 295 {
 296    ROUND_N, // nearest
 297    ROUND_M, // towards -inf
 298    ROUND_Z, // towards 0
 299    ROUND_P, // towards +inf
 300    ROUND_NI, // nearest integer
 301    ROUND_MI, // to integer towards -inf
 302    ROUND_ZI, // to integer towards 0
 303    ROUND_PI, // to integer towards +inf
 304 };
 305
 306 enum CacheMode
 307 {
 308    CACHE_CA,            // cache at all levels
 309    CACHE_WB = CACHE_CA, // cache write back
 310    CACHE_CG,            // cache at global level
 311    CACHE_CS,            // cache streaming
 312    CACHE_CV,            // cache as volatile
 313    CACHE_WT = CACHE_CV  // cache write-through
 314 };
 315
 316 enum DataFile
 317 {
 318    FILE_NULL = 0,
 319    FILE_GPR,
 320    FILE_PREDICATE,       // boolean predicate
 321    FILE_FLAGS,           // zero/sign/carry/overflow bits
 322    FILE_ADDRESS,
 323    LAST_REGISTER_FILE = FILE_ADDRESS,
 324    FILE_IMMEDIATE,
 325    FILE_MEMORY_CONST,
 326    FILE_SHADER_INPUT,
 327    FILE_SHADER_OUTPUT,
 328    FILE_MEMORY_GLOBAL,
 329    FILE_MEMORY_SHARED,
 330    FILE_MEMORY_LOCAL,
 331    FILE_SYSTEM_VALUE,
 332    DATA_FILE_COUNT
 333 };
 334
 335 enum TexTarget
 336 {
 337    TEX_TARGET_1D,
 338    TEX_TARGET_2D,
 339    TEX_TARGET_2D_MS,
 340    TEX_TARGET_3D,
 341    TEX_TARGET_CUBE,
 342    TEX_TARGET_1D_SHADOW,
 343    TEX_TARGET_2D_SHADOW,
 344    TEX_TARGET_CUBE_SHADOW,
 345    TEX_TARGET_1D_ARRAY,
 346    TEX_TARGET_2D_ARRAY,
 347    TEX_TARGET_2D_MS_ARRAY,
 348    TEX_TARGET_CUBE_ARRAY,
 349    TEX_TARGET_1D_ARRAY_SHADOW,
 350    TEX_TARGET_2D_ARRAY_SHADOW,
 351    TEX_TARGET_RECT,
 352    TEX_TARGET_RECT_SHADOW,
 353    TEX_TARGET_CUBE_ARRAY_SHADOW,
 354    TEX_TARGET_BUFFER,
 355    TEX_TARGET_COUNT
 356 };
 357
 358 enum SVSemantic
 359 {
 360    SV_POSITION, // WPOS
 361    SV_VERTEX_ID,
 362    SV_INSTANCE_ID,
 363    SV_INVOCATION_ID,
 364    SV_PRIMITIVE_ID,
 365    SV_VERTEX_COUNT, // gl_PatchVerticesIn
 366    SV_LAYER,
 367    SV_VIEWPORT_INDEX,
 368    SV_YDIR,
 369    SV_FACE,
 370    SV_POINT_SIZE,
 371    SV_POINT_COORD,
 372    SV_CLIP_DISTANCE,
 373    SV_SAMPLE_INDEX,
 374    SV_SAMPLE_POS,
 375    SV_SAMPLE_MASK,
 376    SV_TESS_OUTER,
 377    SV_TESS_INNER,
 378    SV_TESS_COORD,
 379    SV_TID,
 380    SV_CTAID,
 381    SV_NTID,
 382    SV_GRIDID,
 383    SV_NCTAID,
 384    SV_LANEID,
 385    SV_PHYSID,
 386    SV_NPHYSID,
 387    SV_CLOCK,
 388    SV_LBASE,
 389    SV_SBASE,
 390    SV_VERTEX_STRIDE,
 391    SV_INVOCATION_INFO,
 392    SV_UNDEFINED,
 393    SV_LAST
 394 };
 395
 396 class Program;
 397 class Function;
 398 class BasicBlock;
 399
 400 class Target;
 401
 402 class Instruction;
 403 class CmpInstruction;
 404 class TexInstruction;
 405 class FlowInstruction;
 406
 407 class Value;
 408 class LValue;
 409 class Symbol;
 410 class ImmediateValue;
 411
 412 struct Storage
 413 {
 414    DataFile file;
 415    int8_t fileIndex; // signed, may be indirect for CONST[]
 416    uint8_t size; // this should match the Instruction type's size
 417    DataType type; // mainly for pretty printing
 418    union {
 419       uint64_t u64;    // immediate values
 420       uint32_t u32;
 421       uint16_t u16;
 422       uint8_t u8;
 423       int64_t s64;
 424       int32_t s32;
 425       int16_t s16;
 426       int8_t s8;
 427       float f32;
 428       double f64;
 429       int32_t offset; // offset from 0 (base of address space)
 430       int32_t id;     // register id (< 0 if virtual/unassigned, in units <= 4)
 431       struct {
 432          SVSemantic sv;
 433          int index;
 434       } sv;
 435    } data;
 436 };
 437
 438 // precedence: NOT after SAT after NEG after ABS
 439 #define NV50_IR_MOD_ABS (1 << 0)
 440 #define NV50_IR_MOD_NEG (1 << 1)
 441 #define NV50_IR_MOD_SAT (1 << 2)
 442 #define NV50_IR_MOD_NOT (1 << 3)
 443 #define NV50_IR_MOD_NEG_ABS (NV50_IR_MOD_NEG | NV50_IR_MOD_ABS)
 444
 445 #define NV50_IR_INTERP_MODE_MASK   0x3
 446 #define NV50_IR_INTERP_LINEAR      (0 << 0)
 447 #define NV50_IR_INTERP_PERSPECTIVE (1 << 0)
 448 #define NV50_IR_INTERP_FLAT        (2 << 0)
 449 #define NV50_IR_INTERP_SC          (3 << 0) // what exactly is that ?
 450 #define NV50_IR_INTERP_SAMPLE_MASK 0xc
 451 #define NV50_IR_INTERP_DEFAULT     (0 << 2)
 452 #define NV50_IR_INTERP_CENTROID    (1 << 2)
 453 #define NV50_IR_INTERP_OFFSET      (2 << 2)
 454 #define NV50_IR_INTERP_SAMPLEID    (3 << 2)
 455
 456 // do we really want this to be a class ?
 457 class Modifier
 458 {
 459 public:
 460    Modifier() : bits(0) { }
 461    Modifier(unsigned int m) : bits(m) { }
 462    Modifier(operation op);
 463
 464    // @return new Modifier applying a after b (asserts if unrepresentable)
 465    Modifier operator*(const Modifier) const;
 466    Modifier operator*=(const Modifier m) { *this = *this * m; return *this; }
 467    Modifier operator==(const Modifier m) const { return m.bits == bits; }
 468    Modifier operator!=(const Modifier m) const { return m.bits != bits; }
 469
 470    inline Modifier operator&(const Modifier m) const { return bits & m.bits; }
 471    inline Modifier operator|(const Modifier m) const { return bits | m.bits; }
 472    inline Modifier operator^(const Modifier m) const { return bits ^ m.bits; }
 473
 474    operation getOp() const;
 475
 476    inline int neg() const { return (bits & NV50_IR_MOD_NEG) ? 1 : 0; }
 477    inline int abs() const { return (bits & NV50_IR_MOD_ABS) ? 1 : 0; }
 478
 479    inline operator bool() const { return bits ? true : false; }
 480
 481    void applyTo(ImmediateValue &imm) const;
 482
 483    int print(char *buf, size_t size) const;
 484
 485 private:
 486    uint8_t bits;
 487 };
 488
 489 class ValueRef
 490 {
 491 public:
 492    ValueRef(Value * = NULL);
 493    ValueRef(const ValueRef&);
 494    ~ValueRef();
 495
 496    inline bool exists() const { return value != NULL; }
 497
 498    void set(Value *);
 499    void set(const ValueRef&);
 500    inline Value *get() const { return value; }
 501    inline Value *rep() const;
 502
 503    inline Instruction *getInsn() const { return insn; }
 504    inline void setInsn(Instruction *inst) { insn = inst; }
 505
 506    inline bool isIndirect(int dim) const { return indirect[dim] >= 0; }
 507    inline const ValueRef *getIndirect(int dim) const;
 508
 509    inline DataFile getFile() const;
 510    inline unsigned getSize() const;
 511
 512    // SSA: return eventual (traverse MOVs) literal value, if it exists
 513    bool getImmediate(ImmediateValue&) const;
 514
 515 public:
 516    Modifier mod;
 517    int8_t indirect[2]; // >= 0 if relative to lvalue in insn->src(indirect[i])
 518    uint8_t swizzle;
 519
 520    bool usedAsPtr; // for printing
 521
 522 private:
 523    Value *value;
 524    Instruction *insn;
 525 };
 526
 527 class ValueDef
 528 {
 529 public:
 530    ValueDef(Value * = NULL);
 531    ValueDef(const ValueDef&);
 532    ~ValueDef();
 533
 534    inline bool exists() const { return value != NULL; }
 535
 536    inline Value *get() const { return value; }
 537    inline Value *rep() const;
 538    void set(Value *);
 539    bool mayReplace(const ValueRef &);
 540    void replace(const ValueRef &, bool doSet); // replace all uses of the old value
 541
 542    inline Instruction *getInsn() const { return insn; }
 543    inline void setInsn(Instruction *inst) { insn = inst; }
 544
 545    inline DataFile getFile() const;
 546    inline unsigned getSize() const;
 547
 548    inline void setSSA(LValue *);
 549    inline const LValue *preSSA() const;
 550
 551 private:
 552    Value *value;   // should make this LValue * ...
 553    LValue *origin; // pre SSA value
 554    Instruction *insn;
 555 };
 556
 557 class Value
 558 {
 559 public:
 560    Value();
 561    virtual ~Value() { }
 562
 563    virtual Value *clone(ClonePolicy<Function>&) const = 0;
 564
 565    virtual int print(char *, size_t, DataType ty = TYPE_NONE) const = 0;
 566
 567    virtual bool equals(const Value *, bool strict = false) const;
 568    virtual bool interfers(const Value *) const;
 569    virtual bool isUniform() const { return true; }
 570
 571    inline Value *rep() const { return join; }
 572
 573    inline Instruction *getUniqueInsn() const;
 574    inline Instruction *getInsn() const; // use when uniqueness is certain
 575
 576    inline int refCount() { return uses.size(); }
 577
 578    inline LValue *asLValue();
 579    inline Symbol *asSym();
 580    inline ImmediateValue *asImm();
 581    inline const Symbol *asSym() const;
 582    inline const ImmediateValue *asImm() const;
 583
 584    inline bool inFile(DataFile f) { return reg.file == f; }
 585
 586    static inline Value *get(Iterator&);
 587
 588    unordered_set<ValueRef *> uses;
 589    std::list<ValueDef *> defs;
 590    typedef unordered_set<ValueRef *>::iterator UseIterator;
 591    typedef unordered_set<ValueRef *>::const_iterator UseCIterator;
 592    typedef std::list<ValueDef *>::iterator DefIterator;
 593    typedef std::list<ValueDef *>::const_iterator DefCIterator;
 594
 595    int id;
 596    Storage reg;
 597
 598    // TODO: these should be in LValue:
 599    Interval livei;
 600    Value *join;
 601 };
 602
 603 class LValue : public Value
 604 {
 605 public:
 606    LValue(Function *, DataFile file);
 607    LValue(Function *, LValue *);
 608    ~LValue() { }
 609
 610    virtual bool isUniform() const;
 611
 612    virtual LValue *clone(ClonePolicy<Function>&) const;
 613
 614    virtual int print(char *, size_t, DataType ty = TYPE_NONE) const;
 615
 616 public:
 617    unsigned compMask : 8; // compound/component mask
 618    unsigned compound : 1; // used by RA, value involved in split/merge
 619    unsigned ssa      : 1;
 620    unsigned fixedReg : 1; // set & used by RA, earlier just use (id < 0)
 621    unsigned noSpill  : 1; // do not spill (e.g. if spill temporary already)
 622 };
 623
 624 class Symbol : public Value
 625 {
 626 public:
 627    Symbol(Program *, DataFile file = FILE_MEMORY_CONST, ubyte fileIdx = 0);
 628    ~Symbol() { }
 629
 630    virtual Symbol *clone(ClonePolicy<Function>&) const;
 631
 632    virtual bool equals(const Value *that, bool strict) const;
 633
 634    virtual bool isUniform() const;
 635
 636    virtual int print(char *, size_t, DataType ty = TYPE_NONE) const;
 637
 638    // print with indirect values
 639    int print(char *, size_t, Value *, Value *, DataType ty = TYPE_NONE) const;
 640
 641    inline void setFile(DataFile file, ubyte fileIndex = 0)
 642    {
 643       reg.file = file;
 644       reg.fileIndex = fileIndex;
 645    }
 646
 647    inline void setOffset(int32_t offset);
 648    inline void setAddress(Symbol *base, int32_t offset);
 649    inline void setSV(SVSemantic sv, uint32_t idx = 0);
 650
 651    inline const Symbol *getBase() const { return baseSym; }
 652
 653 private:
 654    Symbol *baseSym; // array base for Symbols representing array elements
 655 };
 656
 657 class ImmediateValue : public Value
 658 {
 659 public:
 660    ImmediateValue() { }
 661    ImmediateValue(Program *, uint32_t);
 662    ImmediateValue(Program *, float);
 663    ImmediateValue(Program *, double);
 664    // NOTE: not added to program with
 665    ImmediateValue(const ImmediateValue *, DataType ty);
 666    ~ImmediateValue() { };
 667
 668    virtual ImmediateValue *clone(ClonePolicy<Function>&) const;
 669
 670    virtual bool equals(const Value *that, bool strict) const;
 671
 672    // these only work if 'type' is valid (we mostly use untyped literals):
 673    bool isInteger(const int ival) const; // ival is cast to this' type
 674    bool isNegative() const;
 675    bool isPow2() const;
 676
 677    void applyLog2();
 678
 679    // for constant folding:
 680    ImmediateValue operator+(const ImmediateValue&) const;
 681    ImmediateValue operator-(const ImmediateValue&) const;
 682    ImmediateValue operator*(const ImmediateValue&) const;
 683    ImmediateValue operator/(const ImmediateValue&) const;
 684
 685    ImmediateValue& operator=(const ImmediateValue&); // only sets value !
 686
 687    bool compare(CondCode cc, float fval) const;
 688
 689    virtual int print(char *, size_t, DataType ty = TYPE_NONE) const;
 690 };
 691
 692 class Instruction
 693 {
 694 public:
 695    Instruction();
 696    Instruction(Function *, operation, DataType);
 697    virtual ~Instruction();
 698
 699    virtual Instruction *clone(ClonePolicy<Function>&,
 700                               Instruction * = NULL) const;
 701
 702    void setDef(int i, Value *);
 703    void setSrc(int s, Value *);
 704    void setSrc(int s, const ValueRef&);
 705    void swapSources(int a, int b);
 706    void moveSources(int s, int delta);
 707    bool setIndirect(int s, int dim, Value *);
 708
 709    inline ValueRef& src(int s) { return srcs[s]; }
 710    inline ValueDef& def(int s) { return defs[s]; }
 711    inline const ValueRef& src(int s) const { return srcs[s]; }
 712    inline const ValueDef& def(int s) const { return defs[s]; }
 713
 714    inline Value *getDef(int d) const { return defs[d].get(); }
 715    inline Value *getSrc(int s) const { return srcs[s].get(); }
 716    inline Value *getIndirect(int s, int dim) const;
 717
 718    inline bool defExists(unsigned d) const
 719    {
 720       return d < defs.size() && defs[d].exists();
 721    }
 722    inline bool srcExists(unsigned s) const
 723    {
 724       return s < srcs.size() && srcs[s].exists();
 725    }
 726
 727    inline bool constrainedDefs() const;
 728
 729    bool setPredicate(CondCode ccode, Value *);
 730    inline Value *getPredicate() const;
 731    bool writesPredicate() const;
 732    inline bool isPredicated() const { return predSrc >= 0; }
 733
 734    inline void setFlagsSrc(int s, Value *);
 735    inline void setFlagsDef(int d, Value *);
 736    inline bool usesFlags() const { return flagsSrc >= 0; }
 737
 738    unsigned int defCount() const { return defs.size(); };
 739    unsigned int defCount(unsigned int mask, bool singleFile = false) const;
 740    unsigned int srcCount() const { return srcs.size(); };
 741    unsigned int srcCount(unsigned int mask, bool singleFile = false) const;
 742
 743    // save & remove / set indirect[0,1] and predicate source
 744    void takeExtraSources(int s, Value *[3]);
 745    void putExtraSources(int s, Value *[3]);
 746
 747    inline void setType(DataType type) { dType = sType = type; }
 748
 749    inline void setType(DataType dtype, DataType stype)
 750    {
 751       dType = dtype;
 752       sType = stype;
 753    }
 754
 755    inline bool isPseudo() const { return op < OP_MOV; }
 756    bool isDead() const;
 757    bool isNop() const;
 758    bool isCommutationLegal(const Instruction *) const; // must be adjacent !
 759    bool isActionEqual(const Instruction *) const;
 760    bool isResultEqual(const Instruction *) const;
 761
 762    void print() const;
 763
 764    inline CmpInstruction *asCmp();
 765    inline TexInstruction *asTex();
 766    inline FlowInstruction *asFlow();
 767    inline const TexInstruction *asTex() const;
 768    inline const CmpInstruction *asCmp() const;
 769    inline const FlowInstruction *asFlow() const;
 770
 771 public:
 772    Instruction *next;
 773    Instruction *prev;
 774    int id;
 775    int serial; // CFG order
 776
 777    operation op;
 778    DataType dType; // destination or defining type
 779    DataType sType; // source or secondary type
 780    CondCode cc;
 781    RoundMode rnd;
 782    CacheMode cache;
 783
 784    uint16_t subOp; // quadop, 1 for mul-high, etc.
 785
 786    unsigned encSize    : 4; // encoding size in bytes
 787    unsigned saturate   : 1; // to [0.0f, 1.0f]
 788    unsigned join       : 1; // converge control flow (use OP_JOIN until end)
 789    unsigned fixed      : 1; // prevent dead code elimination
 790    unsigned terminator : 1; // end of basic block
 791    unsigned ftz        : 1; // flush denormal to zero
 792    unsigned dnz        : 1; // denormals, NaN are zero
 793    unsigned ipa        : 4; // interpolation mode
 794    unsigned lanes      : 4;
 795    unsigned perPatch   : 1;
 796    unsigned exit       : 1; // terminate program after insn
 797    unsigned mask       : 4; // for vector ops
 798
 799    int8_t postFactor; // MUL/DIV(if < 0) by 1 << postFactor
 800
 801    int8_t predSrc;
 802    int8_t flagsDef;
 803    int8_t flagsSrc;
 804
 805    uint32_t sched; // scheduling data (NOTE: maybe move to separate storage)
 806
 807    BasicBlock *bb;
 808
 809 protected:
 810    std::deque<ValueDef> defs; // no gaps !
 811    std::deque<ValueRef> srcs; // no gaps !
 812
 813    // instruction specific methods:
 814    // (don't want to subclass, would need more constructors and memory pools)
 815 public:
 816    inline void setInterpolate(unsigned int mode) { ipa = mode; }
 817
 818    unsigned int getInterpMode() const { return ipa & 0x3; }
 819    unsigned int getSampleMode() const { return ipa & 0xc; }
 820
 821 private:
 822    void init();
 823 };
 824
 825 enum TexQuery
 826 {
 827    TXQ_DIMS, /* x, y, z, levels */
 828    TXQ_TYPE, /* ?, ?, samples, ? */
 829    TXQ_SAMPLE_POSITION,
 830    TXQ_FILTER,
 831    TXQ_LOD,
 832    TXQ_WRAP,
 833    TXQ_BORDER_COLOUR
 834 };
 835
 836 class TexInstruction : public Instruction
 837 {
 838 public:
 839    class Target
 840    {
 841    public:
 842       Target(TexTarget targ = TEX_TARGET_2D) : target(targ) { }
 843
 844       const char *getName() const { return descTable[target].name; }
 845       unsigned int getArgCount() const { return descTable[target].argc; }
 846       unsigned int getDim() const { return descTable[target].dim; }
 847       int isArray() const { return descTable[target].array ? 1 : 0; }
 848       int isCube() const { return descTable[target].cube ? 1 : 0; }
 849       int isShadow() const { return descTable[target].shadow ? 1 : 0; }
 850       int isMS() const {
 851         return target == TEX_TARGET_2D_MS || target == TEX_TARGET_2D_MS_ARRAY; }
 852       void clearMS() {
 853          if (isMS()) {
 854             if (isArray())
 855                target = TEX_TARGET_2D_ARRAY;
 856             else
 857                target = TEX_TARGET_2D;
 858          }
 859       }
 860
 861       Target& operator=(TexTarget targ)
 862       {
 863          assert(targ < TEX_TARGET_COUNT);
 864          target = targ;
 865          return *this;
 866       }
 867
 868       inline bool operator==(TexTarget targ) const { return target == targ; }
 869       inline bool operator!=(TexTarget targ) const { return target != targ; }
 870
 871       enum TexTarget getEnum() const { return target; }
 872
 873    private:
 874       struct Desc
 875       {
 876          char name[19];
 877          uint8_t dim;
 878          uint8_t argc;
 879          bool array;
 880          bool cube;
 881          bool shadow;
 882       };
 883
 884       static const struct Desc descTable[TEX_TARGET_COUNT];
 885
 886    private:
 887       enum TexTarget target;
 888    };
 889
 890 public:
 891    TexInstruction(Function *, operation);
 892    virtual ~TexInstruction();
 893
 894    virtual TexInstruction *clone(ClonePolicy<Function>&,
 895                                  Instruction * = NULL) const;
 896
 897    inline void setTexture(Target targ, uint8_t r, uint8_t s)
 898    {
 899       tex.r = r;
 900       tex.s = s;
 901       tex.target = targ;
 902    }
 903
 904    void setIndirectR(Value *);
 905    void setIndirectS(Value *);
 906    inline Value *getIndirectR() const;
 907    inline Value *getIndirectS() const;
 908
 909 public:
 910    struct {
 911       Target target;
 912
 913       uint16_t r;
 914       uint16_t s;
 915       int8_t rIndirectSrc;
 916       int8_t sIndirectSrc;
 917
 918       uint8_t mask;
 919       uint8_t gatherComp;
 920
 921       bool liveOnly; // only execute on live pixels of a quad (optimization)
 922       bool levelZero;
 923       bool derivAll;
 924
 925       int8_t useOffsets; // 0, 1, or 4 for textureGatherOffsets
 926       int8_t offset[3]; // only used on nv50
 927
 928       enum TexQuery query;
 929    } tex;
 930
 931    ValueRef dPdx[3];
 932    ValueRef dPdy[3];
 933    ValueRef offset[4][3];
 934 };
 935
 936 class CmpInstruction : public Instruction
 937 {
 938 public:
 939    CmpInstruction(Function *, operation);
 940
 941    virtual CmpInstruction *clone(ClonePolicy<Function>&,
 942                                  Instruction * = NULL) const;
 943
 944    void setCondition(CondCode cond) { setCond = cond; }
 945    CondCode getCondition() const { return setCond; }
 946
 947 public:
 948    CondCode setCond;
 949 };
 950
 951 class FlowInstruction : public Instruction
 952 {
 953 public:
 954    FlowInstruction(Function *, operation, void *target);
 955
 956    virtual FlowInstruction *clone(ClonePolicy<Function>&,
 957                                   Instruction * = NULL) const;
 958
 959 public:
 960    unsigned allWarp  : 1;
 961    unsigned absolute : 1;
 962    unsigned limit    : 1;
 963    unsigned builtin  : 1; // true for calls to emulation code
 964    unsigned indirect : 1; // target in src(0)
 965
 966    union {
 967       BasicBlock *bb;
 968       int builtin;
 969       Function *fn;
 970    } target;
 971 };
 972
 973 class BasicBlock
 974 {
 975 public:
 976    BasicBlock(Function *);
 977    ~BasicBlock();
 978
 979    BasicBlock *clone(ClonePolicy<Function>&) const;
 980
 981    inline int getId() const { return id; }
 982    inline unsigned int getInsnCount() const { return numInsns; }
 983    inline bool isTerminated() const { return exit && exit->terminator; }
 984
 985    bool dominatedBy(BasicBlock *bb);
 986    inline bool reachableBy(const BasicBlock *by, const BasicBlock *term);
 987
 988    // returns mask of conditional out blocks
 989    // e.g. 3 for IF { .. } ELSE { .. } ENDIF, 1 for IF { .. } ENDIF
 990    unsigned int initiatesSimpleConditional() const;
 991
 992 public:
 993    Function *getFunction() const { return func; }
 994    Program *getProgram() const { return program; }
 995
 996    Instruction *getEntry() const { return entry; } // first non-phi instruction
 997    Instruction *getPhi() const { return phi; }
 998    Instruction *getFirst() const { return phi ? phi : entry; }
 999    Instruction *getExit() const { return exit; }
1000
1001    void insertHead(Instruction *);
1002    void insertTail(Instruction *);
1003    void insertBefore(Instruction *, Instruction *);
1004    void insertAfter(Instruction *, Instruction *);
1005    void remove(Instruction *);
1006    void permuteAdjacent(Instruction *, Instruction *);
1007
1008    BasicBlock *idom() const;
1009
1010    // NOTE: currently does not rebuild the dominator tree
1011    BasicBlock *splitBefore(Instruction *, bool attach = true);
1012    BasicBlock *splitAfter(Instruction *, bool attach = true);
1013
1014    DLList& getDF() { return df; }
1015    DLList::Iterator iterDF() { return df.iterator(); }
1016
1017    static inline BasicBlock *get(Iterator&);
1018    static inline BasicBlock *get(Graph::Node *);
1019
1020 public:
1021    Graph::Node cfg; // first edge is branch *taken* (the ELSE branch)
1022    Graph::Node dom;
1023
1024    BitSet liveSet;
1025    BitSet defSet;
1026
1027    uint32_t binPos;
1028    uint32_t binSize;
1029
1030    Instruction *joinAt; // for quick reference
1031
1032    bool explicitCont; // loop headers: true if loop contains continue stmts
1033
1034 private:
1035    int id;
1036    DLList df;
1037
1038    Instruction *phi;
1039    Instruction *entry;
1040    Instruction *exit;
1041
1042    unsigned int numInsns;
1043
1044 private:
1045    Function *func;
1046    Program *program;
1047
1048    void splitCommon(Instruction *, BasicBlock *, bool attach);
1049 };
1050
1051 class Function
1052 {
1053 public:
1054    Function(Program *, const char *name, uint32_t label);
1055    ~Function();
1056
1057    static inline Function *get(Graph::Node *node);
1058
1059    inline Program *getProgram() const { return prog; }
1060    inline const char *getName() const { return name; }
1061    inline int getId() const { return id; }
1062    inline uint32_t getLabel() const { return label; }
1063
1064    void print();
1065    void printLiveIntervals() const;
1066    void printCFGraph(const char *filePath);
1067
1068    bool setEntry(BasicBlock *);
1069    bool setExit(BasicBlock *);
1070
1071    unsigned int orderInstructions(ArrayList&);
1072
1073    inline void add(BasicBlock *bb, int& id) { allBBlocks.insert(bb, id); }
1074    inline void add(Instruction *insn, int& id) { allInsns.insert(insn, id); }
1075    inline void add(LValue *lval, int& id) { allLValues.insert(lval, id); }
1076
1077    inline LValue *getLValue(int id);
1078
1079    void buildLiveSets();
1080    void buildDefSets();
1081    bool convertToSSA();
1082
1083 public:
1084    std::deque<ValueDef> ins;
1085    std::deque<ValueRef> outs;
1086    std::deque<Value *> clobbers;
1087
1088    Graph cfg;
1089    Graph::Node *cfgExit;
1090    Graph *domTree;
1091    Graph::Node call; // node in the call graph
1092
1093    BasicBlock **bbArray; // BBs in emission order
1094    int bbCount;
1095
1096    unsigned int loopNestingBound;
1097    int regClobberMax;
1098
1099    uint32_t binPos;
1100    uint32_t binSize;
1101
1102    Value *stackPtr;
1103
1104    uint32_t tlsBase; // base address for l[] space (if no stack pointer is used)
1105    uint32_t tlsSize;
1106
1107    ArrayList allBBlocks;
1108    ArrayList allInsns;
1109    ArrayList allLValues;
1110
1111 private:
1112    void buildLiveSetsPreSSA(BasicBlock *, const int sequence);
1113    void buildDefSetsPreSSA(BasicBlock *bb, const int seq);
1114
1115 private:
1116    uint32_t label;
1117    int id;
1118    const char *const name;
1119    Program *prog;
1120 };
1121
1122 enum CGStage
1123 {
1124    CG_STAGE_PRE_SSA,
1125    CG_STAGE_SSA, // expected directly before register allocation
1126    CG_STAGE_POST_RA
1127 };
1128
1129 class Program
1130 {
1131 public:
1132    enum Type
1133    {
1134       TYPE_VERTEX,
1135       TYPE_TESSELLATION_CONTROL,
1136       TYPE_TESSELLATION_EVAL,
1137       TYPE_GEOMETRY,
1138       TYPE_FRAGMENT,
1139       TYPE_COMPUTE
1140    };
1141
1142    Program(Type type, Target *targ);
1143    ~Program();
1144
1145    void print();
1146
1147    Type getType() const { return progType; }
1148
1149    inline void add(Function *fn, int& id) { allFuncs.insert(fn, id); }
1150    inline void del(Function *fn, int& id) { allFuncs.remove(id); }
1151    inline void add(Value *rval, int& id) { allRValues.insert(rval, id); }
1152
1153    bool makeFromTGSI(struct nv50_ir_prog_info *);
1154    bool makeFromSM4(struct nv50_ir_prog_info *);
1155    bool convertToSSA();
1156    bool optimizeSSA(int level);
1157    bool optimizePostRA(int level);
1158    bool registerAllocation();
1159    bool emitBinary(struct nv50_ir_prog_info *);
1160
1161    const Target *getTarget() const { return target; }
1162
1163 private:
1164    void emitSymbolTable(struct nv50_ir_prog_info *);
1165
1166    Type progType;
1167    Target *target;
1168
1169 public:
1170    Function *main;
1171    Graph calls;
1172
1173    ArrayList allFuncs;
1174    ArrayList allRValues;
1175
1176    uint32_t *code;
1177    uint32_t binSize;
1178    uint32_t tlsSize; // size required for FILE_MEMORY_LOCAL
1179
1180    int maxGPR;
1181
1182    MemoryPool mem_Instruction;
1183    MemoryPool mem_CmpInstruction;
1184    MemoryPool mem_TexInstruction;
1185    MemoryPool mem_FlowInstruction;
1186    MemoryPool mem_LValue;
1187    MemoryPool mem_Symbol;
1188    MemoryPool mem_ImmediateValue;
1189
1190    uint32_t dbgFlags;
1191    uint8_t  optLevel;
1192
1193    void *targetPriv; // e.g. to carry information between passes
1194
1195    const struct nv50_ir_prog_info *driver; // for driver configuration
1196
1197    void releaseInstruction(Instruction *);
1198    void releaseValue(Value *);
1199 };
1200
1201 // TODO: add const version
1202 class Pass
1203 {
1204 public:
1205    bool run(Program *, bool ordered = false, bool skipPhi = false);
1206    bool run(Function *, bool ordered = false, bool skipPhi = false);
1207
1208 private:
1209    // return false to continue with next entity on next higher level
1210    virtual bool visit(Function *) { return true; }
1211    virtual bool visit(BasicBlock *) { return true; }
1212    virtual bool visit(Instruction *) { return false; }
1213
1214    bool doRun(Program *, bool ordered, bool skipPhi);
1215    bool doRun(Function *, bool ordered, bool skipPhi);
1216
1217 protected:
1218    bool err;
1219    Function *func;
1220    Program *prog;
1221 };
1222
1223 // =============================================================================
1224
1225 #include "codegen/nv50_ir_inlines.h"
1226
1227 } // namespace nv50_ir
1228
1229 #endif // __NV50_IR_H__