src/gallium/drivers/nouveau/codegen/nv50_ir_from_nir.cpp

   1 /*
   2  * Copyright 2017 Red Hat Inc.
   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  * Authors: Karol Herbst <kherbst@redhat.com>
  23  */
  24
  25 #include "compiler/nir/nir.h"
  26
  27 #include "util/u_debug.h"
  28
  29 #include "codegen/nv50_ir.h"
  30 #include "codegen/nv50_ir_from_common.h"
  31 #include "codegen/nv50_ir_lowering_helper.h"
  32 #include "codegen/nv50_ir_util.h"
  33
  34 #if __cplusplus >= 201103L
  35 #include <unordered_map>
  36 #else
  37 #include <tr1/unordered_map>
  38 #endif
  39 #include <cstring>
  40 #include <list>
  41 #include <vector>
  42
  43 namespace {
  44
  45 #if __cplusplus >= 201103L
  46 using std::hash;
  47 using std::unordered_map;
  48 #else
  49 using std::tr1::hash;
  50 using std::tr1::unordered_map;
  51 #endif
  52
  53 using namespace nv50_ir;
  54
  55 int
  56 type_size(const struct glsl_type *type, bool bindless)
  57 {
  58    return glsl_count_attribute_slots(type, false);
  59 }
  60
  61 class Converter : public ConverterCommon
  62 {
  63 public:
  64    Converter(Program *, nir_shader *, nv50_ir_prog_info *);
  65
  66    bool run();
  67 private:
  68    typedef std::vector<LValue*> LValues;
  69    typedef unordered_map<unsigned, LValues> NirDefMap;
  70    typedef unordered_map<unsigned, nir_load_const_instr*> ImmediateMap;
  71    typedef unordered_map<unsigned, uint32_t> NirArrayLMemOffsets;
  72    typedef unordered_map<unsigned, BasicBlock*> NirBlockMap;
  73
  74    CacheMode convert(enum gl_access_qualifier);
  75    TexTarget convert(glsl_sampler_dim, bool isArray, bool isShadow);
  76    LValues& convert(nir_alu_dest *);
  77    BasicBlock* convert(nir_block *);
  78    LValues& convert(nir_dest *);
  79    SVSemantic convert(nir_intrinsic_op);
  80    Value* convert(nir_load_const_instr*, uint8_t);
  81    LValues& convert(nir_register *);
  82    LValues& convert(nir_ssa_def *);
  83
  84    ImgFormat convertGLImgFormat(GLuint);
  85
  86    Value* getSrc(nir_alu_src *, uint8_t component = 0);
  87    Value* getSrc(nir_register *, uint8_t);
  88    Value* getSrc(nir_src *, uint8_t, bool indirect = false);
  89    Value* getSrc(nir_ssa_def *, uint8_t);
  90
  91    // returned value is the constant part of the given source (either the
  92    // nir_src or the selected source component of an intrinsic). Even though
  93    // this is mostly an optimization to be able to skip indirects in a few
  94    // cases, sometimes we require immediate values or set some fileds on
  95    // instructions (e.g. tex) in order for codegen to consume those.
  96    // If the found value has not a constant part, the Value gets returned
  97    // through the Value parameter.
  98    uint32_t getIndirect(nir_src *, uint8_t, Value *&);
  99    // isScalar indicates that the addressing is scalar, vec4 addressing is
 100    // assumed otherwise
 101    uint32_t getIndirect(nir_intrinsic_instr *, uint8_t s, uint8_t c, Value *&,
 102                         bool isScalar = false);
 103
 104    uint32_t getSlotAddress(nir_intrinsic_instr *, uint8_t idx, uint8_t slot);
 105
 106    void setInterpolate(nv50_ir_varying *,
 107                        uint8_t,
 108                        bool centroid,
 109                        unsigned semantics);
 110
 111    Instruction *loadFrom(DataFile, uint8_t, DataType, Value *def, uint32_t base,
 112                          uint8_t c, Value *indirect0 = NULL,
 113                          Value *indirect1 = NULL, bool patch = false);
 114    void storeTo(nir_intrinsic_instr *, DataFile, operation, DataType,
 115                 Value *src, uint8_t idx, uint8_t c, Value *indirect0 = NULL,
 116                 Value *indirect1 = NULL);
 117
 118    bool isFloatType(nir_alu_type);
 119    bool isSignedType(nir_alu_type);
 120    bool isResultFloat(nir_op);
 121    bool isResultSigned(nir_op);
 122
 123    DataType getDType(nir_alu_instr *);
 124    DataType getDType(nir_intrinsic_instr *);
 125    DataType getDType(nir_intrinsic_instr *, bool isSigned);
 126    DataType getDType(nir_op, uint8_t);
 127
 128    std::vector<DataType> getSTypes(nir_alu_instr *);
 129    DataType getSType(nir_src &, bool isFloat, bool isSigned);
 130
 131    operation getOperation(nir_intrinsic_op);
 132    operation getOperation(nir_op);
 133    operation getOperation(nir_texop);
 134    operation preOperationNeeded(nir_op);
 135
 136    int getSubOp(nir_intrinsic_op);
 137    int getSubOp(nir_op);
 138
 139    CondCode getCondCode(nir_op);
 140
 141    bool assignSlots();
 142    bool parseNIR();
 143
 144    bool visit(nir_alu_instr *);
 145    bool visit(nir_block *);
 146    bool visit(nir_cf_node *);
 147    bool visit(nir_deref_instr *);
 148    bool visit(nir_function *);
 149    bool visit(nir_if *);
 150    bool visit(nir_instr *);
 151    bool visit(nir_intrinsic_instr *);
 152    bool visit(nir_jump_instr *);
 153    bool visit(nir_load_const_instr*);
 154    bool visit(nir_loop *);
 155    bool visit(nir_ssa_undef_instr *);
 156    bool visit(nir_tex_instr *);
 157
 158    // tex stuff
 159    Value* applyProjection(Value *src, Value *proj);
 160    unsigned int getNIRArgCount(TexInstruction::Target&);
 161
 162    // image stuff
 163    uint16_t handleDeref(nir_deref_instr *, Value * & indirect, const nir_variable * &);
 164    CacheMode getCacheModeFromVar(const nir_variable *);
 165
 166    nir_shader *nir;
 167
 168    NirDefMap ssaDefs;
 169    NirDefMap regDefs;
 170    ImmediateMap immediates;
 171    NirArrayLMemOffsets regToLmemOffset;
 172    NirBlockMap blocks;
 173    unsigned int curLoopDepth;
 174
 175    BasicBlock *exit;
 176    Value *zero;
 177    Instruction *immInsertPos;
 178
 179    int clipVertexOutput;
 180
 181    union {
 182       struct {
 183          Value *position;
 184       } fp;
 185    };
 186 };
 187
 188 Converter::Converter(Program *prog, nir_shader *nir, nv50_ir_prog_info *info)
 189    : ConverterCommon(prog, info),
 190      nir(nir),
 191      curLoopDepth(0),
 192      clipVertexOutput(-1)
 193 {
 194    zero = mkImm((uint32_t)0);
 195 }
 196
 197 BasicBlock *
 198 Converter::convert(nir_block *block)
 199 {
 200    NirBlockMap::iterator it = blocks.find(block->index);
 201    if (it != blocks.end())
 202       return it->second;
 203
 204    BasicBlock *bb = new BasicBlock(func);
 205    blocks[block->index] = bb;
 206    return bb;
 207 }
 208
 209 bool
 210 Converter::isFloatType(nir_alu_type type)
 211 {
 212    return nir_alu_type_get_base_type(type) == nir_type_float;
 213 }
 214
 215 bool
 216 Converter::isSignedType(nir_alu_type type)
 217 {
 218    return nir_alu_type_get_base_type(type) == nir_type_int;
 219 }
 220
 221 bool
 222 Converter::isResultFloat(nir_op op)
 223 {
 224    const nir_op_info &info = nir_op_infos[op];
 225    if (info.output_type != nir_type_invalid)
 226       return isFloatType(info.output_type);
 227
 228    ERROR("isResultFloat not implemented for %s\n", nir_op_infos[op].name);
 229    assert(false);
 230    return true;
 231 }
 232
 233 bool
 234 Converter::isResultSigned(nir_op op)
 235 {
 236    switch (op) {
 237    // there is no umul and we get wrong results if we treat all muls as signed
 238    case nir_op_imul:
 239    case nir_op_inot:
 240       return false;
 241    default:
 242       const nir_op_info &info = nir_op_infos[op];
 243       if (info.output_type != nir_type_invalid)
 244          return isSignedType(info.output_type);
 245       ERROR("isResultSigned not implemented for %s\n", nir_op_infos[op].name);
 246       assert(false);
 247       return true;
 248    }
 249 }
 250
 251 DataType
 252 Converter::getDType(nir_alu_instr *insn)
 253 {
 254    if (insn->dest.dest.is_ssa)
 255       return getDType(insn->op, insn->dest.dest.ssa.bit_size);
 256    else
 257       return getDType(insn->op, insn->dest.dest.reg.reg->bit_size);
 258 }
 259
 260 DataType
 261 Converter::getDType(nir_intrinsic_instr *insn)
 262 {
 263    bool isSigned;
 264    switch (insn->intrinsic) {
 265    case nir_intrinsic_shared_atomic_imax:
 266    case nir_intrinsic_shared_atomic_imin:
 267    case nir_intrinsic_ssbo_atomic_imax:
 268    case nir_intrinsic_ssbo_atomic_imin:
 269       isSigned = true;
 270       break;
 271    default:
 272       isSigned = false;
 273       break;
 274    }
 275
 276    return getDType(insn, isSigned);
 277 }
 278
 279 DataType
 280 Converter::getDType(nir_intrinsic_instr *insn, bool isSigned)
 281 {
 282    if (insn->dest.is_ssa)
 283       return typeOfSize(insn->dest.ssa.bit_size / 8, false, isSigned);
 284    else
 285       return typeOfSize(insn->dest.reg.reg->bit_size / 8, false, isSigned);
 286 }
 287
 288 DataType
 289 Converter::getDType(nir_op op, uint8_t bitSize)
 290 {
 291    DataType ty = typeOfSize(bitSize / 8, isResultFloat(op), isResultSigned(op));
 292    if (ty == TYPE_NONE) {
 293       ERROR("couldn't get Type for op %s with bitSize %u\n", nir_op_infos[op].name, bitSize);
 294       assert(false);
 295    }
 296    return ty;
 297 }
 298
 299 std::vector<DataType>
 300 Converter::getSTypes(nir_alu_instr *insn)
 301 {
 302    const nir_op_info &info = nir_op_infos[insn->op];
 303    std::vector<DataType> res(info.num_inputs);
 304
 305    for (uint8_t i = 0; i < info.num_inputs; ++i) {
 306       if (info.input_types[i] != nir_type_invalid) {
 307          res[i] = getSType(insn->src[i].src, isFloatType(info.input_types[i]), isSignedType(info.input_types[i]));
 308       } else {
 309          ERROR("getSType not implemented for %s idx %u\n", info.name, i);
 310          assert(false);
 311          res[i] = TYPE_NONE;
 312          break;
 313       }
 314    }
 315
 316    return res;
 317 }
 318
 319 DataType
 320 Converter::getSType(nir_src &src, bool isFloat, bool isSigned)
 321 {
 322    uint8_t bitSize;
 323    if (src.is_ssa)
 324       bitSize = src.ssa->bit_size;
 325    else
 326       bitSize = src.reg.reg->bit_size;
 327
 328    DataType ty = typeOfSize(bitSize / 8, isFloat, isSigned);
 329    if (ty == TYPE_NONE) {
 330       const char *str;
 331       if (isFloat)
 332          str = "float";
 333       else if (isSigned)
 334          str = "int";
 335       else
 336          str = "uint";
 337       ERROR("couldn't get Type for %s with bitSize %u\n", str, bitSize);
 338       assert(false);
 339    }
 340    return ty;
 341 }
 342
 343 operation
 344 Converter::getOperation(nir_op op)
 345 {
 346    switch (op) {
 347    // basic ops with float and int variants
 348    case nir_op_fabs:
 349    case nir_op_iabs:
 350       return OP_ABS;
 351    case nir_op_fadd:
 352    case nir_op_iadd:
 353       return OP_ADD;
 354    case nir_op_iand:
 355       return OP_AND;
 356    case nir_op_ifind_msb:
 357    case nir_op_ufind_msb:
 358       return OP_BFIND;
 359    case nir_op_fceil:
 360       return OP_CEIL;
 361    case nir_op_fcos:
 362       return OP_COS;
 363    case nir_op_f2f32:
 364    case nir_op_f2f64:
 365    case nir_op_f2i32:
 366    case nir_op_f2i64:
 367    case nir_op_f2u32:
 368    case nir_op_f2u64:
 369    case nir_op_i2f32:
 370    case nir_op_i2f64:
 371    case nir_op_i2i32:
 372    case nir_op_i2i64:
 373    case nir_op_u2f32:
 374    case nir_op_u2f64:
 375    case nir_op_u2u32:
 376    case nir_op_u2u64:
 377       return OP_CVT;
 378    case nir_op_fddx:
 379    case nir_op_fddx_coarse:
 380    case nir_op_fddx_fine:
 381       return OP_DFDX;
 382    case nir_op_fddy:
 383    case nir_op_fddy_coarse:
 384    case nir_op_fddy_fine:
 385       return OP_DFDY;
 386    case nir_op_fdiv:
 387    case nir_op_idiv:
 388    case nir_op_udiv:
 389       return OP_DIV;
 390    case nir_op_fexp2:
 391       return OP_EX2;
 392    case nir_op_ffloor:
 393       return OP_FLOOR;
 394    case nir_op_ffma:
 395       return OP_FMA;
 396    case nir_op_flog2:
 397       return OP_LG2;
 398    case nir_op_fmax:
 399    case nir_op_imax:
 400    case nir_op_umax:
 401       return OP_MAX;
 402    case nir_op_pack_64_2x32_split:
 403       return OP_MERGE;
 404    case nir_op_fmin:
 405    case nir_op_imin:
 406    case nir_op_umin:
 407       return OP_MIN;
 408    case nir_op_fmod:
 409    case nir_op_imod:
 410    case nir_op_umod:
 411    case nir_op_frem:
 412    case nir_op_irem:
 413       return OP_MOD;
 414    case nir_op_fmul:
 415    case nir_op_imul:
 416    case nir_op_imul_high:
 417    case nir_op_umul_high:
 418       return OP_MUL;
 419    case nir_op_fneg:
 420    case nir_op_ineg:
 421       return OP_NEG;
 422    case nir_op_inot:
 423       return OP_NOT;
 424    case nir_op_ior:
 425       return OP_OR;
 426    case nir_op_fpow:
 427       return OP_POW;
 428    case nir_op_frcp:
 429       return OP_RCP;
 430    case nir_op_frsq:
 431       return OP_RSQ;
 432    case nir_op_fsat:
 433       return OP_SAT;
 434    case nir_op_feq32:
 435    case nir_op_ieq32:
 436    case nir_op_fge32:
 437    case nir_op_ige32:
 438    case nir_op_uge32:
 439    case nir_op_flt32:
 440    case nir_op_ilt32:
 441    case nir_op_ult32:
 442    case nir_op_fne32:
 443    case nir_op_ine32:
 444       return OP_SET;
 445    case nir_op_ishl:
 446       return OP_SHL;
 447    case nir_op_ishr:
 448    case nir_op_ushr:
 449       return OP_SHR;
 450    case nir_op_fsin:
 451       return OP_SIN;
 452    case nir_op_fsqrt:
 453       return OP_SQRT;
 454    case nir_op_fsub:
 455    case nir_op_isub:
 456       return OP_SUB;
 457    case nir_op_ftrunc:
 458       return OP_TRUNC;
 459    case nir_op_ixor:
 460       return OP_XOR;
 461    default:
 462       ERROR("couldn't get operation for op %s\n", nir_op_infos[op].name);
 463       assert(false);
 464       return OP_NOP;
 465    }
 466 }
 467
 468 operation
 469 Converter::getOperation(nir_texop op)
 470 {
 471    switch (op) {
 472    case nir_texop_tex:
 473       return OP_TEX;
 474    case nir_texop_lod:
 475       return OP_TXLQ;
 476    case nir_texop_txb:
 477       return OP_TXB;
 478    case nir_texop_txd:
 479       return OP_TXD;
 480    case nir_texop_txf:
 481    case nir_texop_txf_ms:
 482       return OP_TXF;
 483    case nir_texop_tg4:
 484       return OP_TXG;
 485    case nir_texop_txl:
 486       return OP_TXL;
 487    case nir_texop_query_levels:
 488    case nir_texop_texture_samples:
 489    case nir_texop_txs:
 490       return OP_TXQ;
 491    default:
 492       ERROR("couldn't get operation for nir_texop %u\n", op);
 493       assert(false);
 494       return OP_NOP;
 495    }
 496 }
 497
 498 operation
 499 Converter::getOperation(nir_intrinsic_op op)
 500 {
 501    switch (op) {
 502    case nir_intrinsic_emit_vertex:
 503       return OP_EMIT;
 504    case nir_intrinsic_end_primitive:
 505       return OP_RESTART;
 506    case nir_intrinsic_bindless_image_atomic_add:
 507    case nir_intrinsic_image_atomic_add:
 508    case nir_intrinsic_image_deref_atomic_add:
 509    case nir_intrinsic_bindless_image_atomic_and:
 510    case nir_intrinsic_image_atomic_and:
 511    case nir_intrinsic_image_deref_atomic_and:
 512    case nir_intrinsic_bindless_image_atomic_comp_swap:
 513    case nir_intrinsic_image_atomic_comp_swap:
 514    case nir_intrinsic_image_deref_atomic_comp_swap:
 515    case nir_intrinsic_bindless_image_atomic_exchange:
 516    case nir_intrinsic_image_atomic_exchange:
 517    case nir_intrinsic_image_deref_atomic_exchange:
 518    case nir_intrinsic_bindless_image_atomic_max:
 519    case nir_intrinsic_image_atomic_max:
 520    case nir_intrinsic_image_deref_atomic_max:
 521    case nir_intrinsic_bindless_image_atomic_min:
 522    case nir_intrinsic_image_atomic_min:
 523    case nir_intrinsic_image_deref_atomic_min:
 524    case nir_intrinsic_bindless_image_atomic_or:
 525    case nir_intrinsic_image_atomic_or:
 526    case nir_intrinsic_image_deref_atomic_or:
 527    case nir_intrinsic_bindless_image_atomic_xor:
 528    case nir_intrinsic_image_atomic_xor:
 529    case nir_intrinsic_image_deref_atomic_xor:
 530       return OP_SUREDP;
 531    case nir_intrinsic_bindless_image_load:
 532    case nir_intrinsic_image_load:
 533    case nir_intrinsic_image_deref_load:
 534       return OP_SULDP;
 535    case nir_intrinsic_bindless_image_samples:
 536    case nir_intrinsic_image_samples:
 537    case nir_intrinsic_image_deref_samples:
 538    case nir_intrinsic_bindless_image_size:
 539    case nir_intrinsic_image_size:
 540    case nir_intrinsic_image_deref_size:
 541       return OP_SUQ;
 542    case nir_intrinsic_bindless_image_store:
 543    case nir_intrinsic_image_store:
 544    case nir_intrinsic_image_deref_store:
 545       return OP_SUSTP;
 546    default:
 547       ERROR("couldn't get operation for nir_intrinsic_op %u\n", op);
 548       assert(false);
 549       return OP_NOP;
 550    }
 551 }
 552
 553 operation
 554 Converter::preOperationNeeded(nir_op op)
 555 {
 556    switch (op) {
 557    case nir_op_fcos:
 558    case nir_op_fsin:
 559       return OP_PRESIN;
 560    default:
 561       return OP_NOP;
 562    }
 563 }
 564
 565 int
 566 Converter::getSubOp(nir_op op)
 567 {
 568    switch (op) {
 569    case nir_op_imul_high:
 570    case nir_op_umul_high:
 571       return NV50_IR_SUBOP_MUL_HIGH;
 572    default:
 573       return 0;
 574    }
 575 }
 576
 577 int
 578 Converter::getSubOp(nir_intrinsic_op op)
 579 {
 580    switch (op) {
 581    case nir_intrinsic_bindless_image_atomic_add:
 582    case nir_intrinsic_image_atomic_add:
 583    case nir_intrinsic_image_deref_atomic_add:
 584    case nir_intrinsic_shared_atomic_add:
 585    case nir_intrinsic_ssbo_atomic_add:
 586       return  NV50_IR_SUBOP_ATOM_ADD;
 587    case nir_intrinsic_bindless_image_atomic_and:
 588    case nir_intrinsic_image_atomic_and:
 589    case nir_intrinsic_image_deref_atomic_and:
 590    case nir_intrinsic_shared_atomic_and:
 591    case nir_intrinsic_ssbo_atomic_and:
 592       return  NV50_IR_SUBOP_ATOM_AND;
 593    case nir_intrinsic_bindless_image_atomic_comp_swap:
 594    case nir_intrinsic_image_atomic_comp_swap:
 595    case nir_intrinsic_image_deref_atomic_comp_swap:
 596    case nir_intrinsic_shared_atomic_comp_swap:
 597    case nir_intrinsic_ssbo_atomic_comp_swap:
 598       return  NV50_IR_SUBOP_ATOM_CAS;
 599    case nir_intrinsic_bindless_image_atomic_exchange:
 600    case nir_intrinsic_image_atomic_exchange:
 601    case nir_intrinsic_image_deref_atomic_exchange:
 602    case nir_intrinsic_shared_atomic_exchange:
 603    case nir_intrinsic_ssbo_atomic_exchange:
 604       return  NV50_IR_SUBOP_ATOM_EXCH;
 605    case nir_intrinsic_bindless_image_atomic_or:
 606    case nir_intrinsic_image_atomic_or:
 607    case nir_intrinsic_image_deref_atomic_or:
 608    case nir_intrinsic_shared_atomic_or:
 609    case nir_intrinsic_ssbo_atomic_or:
 610       return  NV50_IR_SUBOP_ATOM_OR;
 611    case nir_intrinsic_bindless_image_atomic_max:
 612    case nir_intrinsic_image_atomic_max:
 613    case nir_intrinsic_image_deref_atomic_max:
 614    case nir_intrinsic_shared_atomic_imax:
 615    case nir_intrinsic_shared_atomic_umax:
 616    case nir_intrinsic_ssbo_atomic_imax:
 617    case nir_intrinsic_ssbo_atomic_umax:
 618       return  NV50_IR_SUBOP_ATOM_MAX;
 619    case nir_intrinsic_bindless_image_atomic_min:
 620    case nir_intrinsic_image_atomic_min:
 621    case nir_intrinsic_image_deref_atomic_min:
 622    case nir_intrinsic_shared_atomic_imin:
 623    case nir_intrinsic_shared_atomic_umin:
 624    case nir_intrinsic_ssbo_atomic_imin:
 625    case nir_intrinsic_ssbo_atomic_umin:
 626       return  NV50_IR_SUBOP_ATOM_MIN;
 627    case nir_intrinsic_bindless_image_atomic_xor:
 628    case nir_intrinsic_image_atomic_xor:
 629    case nir_intrinsic_image_deref_atomic_xor:
 630    case nir_intrinsic_shared_atomic_xor:
 631    case nir_intrinsic_ssbo_atomic_xor:
 632       return  NV50_IR_SUBOP_ATOM_XOR;
 633
 634    case nir_intrinsic_group_memory_barrier:
 635    case nir_intrinsic_memory_barrier:
 636    case nir_intrinsic_memory_barrier_atomic_counter:
 637    case nir_intrinsic_memory_barrier_buffer:
 638    case nir_intrinsic_memory_barrier_image:
 639       return NV50_IR_SUBOP_MEMBAR(M, GL);
 640    case nir_intrinsic_memory_barrier_shared:
 641       return NV50_IR_SUBOP_MEMBAR(M, CTA);
 642
 643    case nir_intrinsic_vote_all:
 644       return NV50_IR_SUBOP_VOTE_ALL;
 645    case nir_intrinsic_vote_any:
 646       return NV50_IR_SUBOP_VOTE_ANY;
 647    case nir_intrinsic_vote_ieq:
 648       return NV50_IR_SUBOP_VOTE_UNI;
 649    default:
 650       return 0;
 651    }
 652 }
 653
 654 CondCode
 655 Converter::getCondCode(nir_op op)
 656 {
 657    switch (op) {
 658    case nir_op_feq32:
 659    case nir_op_ieq32:
 660       return CC_EQ;
 661    case nir_op_fge32:
 662    case nir_op_ige32:
 663    case nir_op_uge32:
 664       return CC_GE;
 665    case nir_op_flt32:
 666    case nir_op_ilt32:
 667    case nir_op_ult32:
 668       return CC_LT;
 669    case nir_op_fne32:
 670       return CC_NEU;
 671    case nir_op_ine32:
 672       return CC_NE;
 673    default:
 674       ERROR("couldn't get CondCode for op %s\n", nir_op_infos[op].name);
 675       assert(false);
 676       return CC_FL;
 677    }
 678 }
 679
 680 Converter::LValues&
 681 Converter::convert(nir_alu_dest *dest)
 682 {
 683    return convert(&dest->dest);
 684 }
 685
 686 Converter::LValues&
 687 Converter::convert(nir_dest *dest)
 688 {
 689    if (dest->is_ssa)
 690       return convert(&dest->ssa);
 691    if (dest->reg.indirect) {
 692       ERROR("no support for indirects.");
 693       assert(false);
 694    }
 695    return convert(dest->reg.reg);
 696 }
 697
 698 Converter::LValues&
 699 Converter::convert(nir_register *reg)
 700 {
 701    NirDefMap::iterator it = regDefs.find(reg->index);
 702    if (it != regDefs.end())
 703       return it->second;
 704
 705    LValues newDef(reg->num_components);
 706    for (uint8_t i = 0; i < reg->num_components; i++)
 707       newDef[i] = getScratch(std::max(4, reg->bit_size / 8));
 708    return regDefs[reg->index] = newDef;
 709 }
 710
 711 Converter::LValues&
 712 Converter::convert(nir_ssa_def *def)
 713 {
 714    NirDefMap::iterator it = ssaDefs.find(def->index);
 715    if (it != ssaDefs.end())
 716       return it->second;
 717
 718    LValues newDef(def->num_components);
 719    for (uint8_t i = 0; i < def->num_components; i++)
 720       newDef[i] = getSSA(std::max(4, def->bit_size / 8));
 721    return ssaDefs[def->index] = newDef;
 722 }
 723
 724 Value*
 725 Converter::getSrc(nir_alu_src *src, uint8_t component)
 726 {
 727    if (src->abs || src->negate) {
 728       ERROR("modifiers currently not supported on nir_alu_src\n");
 729       assert(false);
 730    }
 731    return getSrc(&src->src, src->swizzle[component]);
 732 }
 733
 734 Value*
 735 Converter::getSrc(nir_register *reg, uint8_t idx)
 736 {
 737    NirDefMap::iterator it = regDefs.find(reg->index);
 738    if (it == regDefs.end())
 739       return convert(reg)[idx];
 740    return it->second[idx];
 741 }
 742
 743 Value*
 744 Converter::getSrc(nir_src *src, uint8_t idx, bool indirect)
 745 {
 746    if (src->is_ssa)
 747       return getSrc(src->ssa, idx);
 748
 749    if (src->reg.indirect) {
 750       if (indirect)
 751          return getSrc(src->reg.indirect, idx);
 752       ERROR("no support for indirects.");
 753       assert(false);
 754       return NULL;
 755    }
 756
 757    return getSrc(src->reg.reg, idx);
 758 }
 759
 760 Value*
 761 Converter::getSrc(nir_ssa_def *src, uint8_t idx)
 762 {
 763    ImmediateMap::iterator iit = immediates.find(src->index);
 764    if (iit != immediates.end())
 765       return convert((*iit).second, idx);
 766
 767    NirDefMap::iterator it = ssaDefs.find(src->index);
 768    if (it == ssaDefs.end()) {
 769       ERROR("SSA value %u not found\n", src->index);
 770       assert(false);
 771       return NULL;
 772    }
 773    return it->second[idx];
 774 }
 775
 776 uint32_t
 777 Converter::getIndirect(nir_src *src, uint8_t idx, Value *&indirect)
 778 {
 779    nir_const_value *offset = nir_src_as_const_value(*src);
 780
 781    if (offset) {
 782       indirect = NULL;
 783       return offset[0].u32;
 784    }
 785
 786    indirect = getSrc(src, idx, true);
 787    return 0;
 788 }
 789
 790 uint32_t
 791 Converter::getIndirect(nir_intrinsic_instr *insn, uint8_t s, uint8_t c, Value *&indirect, bool isScalar)
 792 {
 793    int32_t idx = nir_intrinsic_base(insn) + getIndirect(&insn->src[s], c, indirect);
 794    if (indirect && !isScalar)
 795       indirect = mkOp2v(OP_SHL, TYPE_U32, getSSA(4, FILE_ADDRESS), indirect, loadImm(NULL, 4));
 796    return idx;
 797 }
 798
 799 static void
 800 vert_attrib_to_tgsi_semantic(gl_vert_attrib slot, unsigned *name, unsigned *index)
 801 {
 802    assert(name && index);
 803
 804    if (slot >= VERT_ATTRIB_MAX) {
 805       ERROR("invalid varying slot %u\n", slot);
 806       assert(false);
 807       return;
 808    }
 809
 810    if (slot >= VERT_ATTRIB_GENERIC0 &&
 811        slot < VERT_ATTRIB_GENERIC0 + VERT_ATTRIB_GENERIC_MAX) {
 812       *name = TGSI_SEMANTIC_GENERIC;
 813       *index = slot - VERT_ATTRIB_GENERIC0;
 814       return;
 815    }
 816
 817    if (slot >= VERT_ATTRIB_TEX0 &&
 818        slot < VERT_ATTRIB_TEX0 + VERT_ATTRIB_TEX_MAX) {
 819       *name = TGSI_SEMANTIC_TEXCOORD;
 820       *index = slot - VERT_ATTRIB_TEX0;
 821       return;
 822    }
 823
 824    switch (slot) {
 825    case VERT_ATTRIB_COLOR0:
 826       *name = TGSI_SEMANTIC_COLOR;
 827       *index = 0;
 828       break;
 829    case VERT_ATTRIB_COLOR1:
 830       *name = TGSI_SEMANTIC_COLOR;
 831       *index = 1;
 832       break;
 833    case VERT_ATTRIB_EDGEFLAG:
 834       *name = TGSI_SEMANTIC_EDGEFLAG;
 835       *index = 0;
 836       break;
 837    case VERT_ATTRIB_FOG:
 838       *name = TGSI_SEMANTIC_FOG;
 839       *index = 0;
 840       break;
 841    case VERT_ATTRIB_NORMAL:
 842       *name = TGSI_SEMANTIC_NORMAL;
 843       *index = 0;
 844       break;
 845    case VERT_ATTRIB_POS:
 846       *name = TGSI_SEMANTIC_POSITION;
 847       *index = 0;
 848       break;
 849    case VERT_ATTRIB_POINT_SIZE:
 850       *name = TGSI_SEMANTIC_PSIZE;
 851       *index = 0;
 852       break;
 853    default:
 854       ERROR("unknown vert attrib slot %u\n", slot);
 855       assert(false);
 856       break;
 857    }
 858 }
 859
 860 static void
 861 varying_slot_to_tgsi_semantic(gl_varying_slot slot, unsigned *name, unsigned *index)
 862 {
 863    assert(name && index);
 864
 865    if (slot >= VARYING_SLOT_TESS_MAX) {
 866       ERROR("invalid varying slot %u\n", slot);
 867       assert(false);
 868       return;
 869    }
 870
 871    if (slot >= VARYING_SLOT_PATCH0) {
 872       *name = TGSI_SEMANTIC_PATCH;
 873       *index = slot - VARYING_SLOT_PATCH0;
 874       return;
 875    }
 876
 877    if (slot >= VARYING_SLOT_VAR0) {
 878       *name = TGSI_SEMANTIC_GENERIC;
 879       *index = slot - VARYING_SLOT_VAR0;
 880       return;
 881    }
 882
 883    if (slot >= VARYING_SLOT_TEX0 && slot <= VARYING_SLOT_TEX7) {
 884       *name = TGSI_SEMANTIC_TEXCOORD;
 885       *index = slot - VARYING_SLOT_TEX0;
 886       return;
 887    }
 888
 889    switch (slot) {
 890    case VARYING_SLOT_BFC0:
 891       *name = TGSI_SEMANTIC_BCOLOR;
 892       *index = 0;
 893       break;
 894    case VARYING_SLOT_BFC1:
 895       *name = TGSI_SEMANTIC_BCOLOR;
 896       *index = 1;
 897       break;
 898    case VARYING_SLOT_CLIP_DIST0:
 899       *name = TGSI_SEMANTIC_CLIPDIST;
 900       *index = 0;
 901       break;
 902    case VARYING_SLOT_CLIP_DIST1:
 903       *name = TGSI_SEMANTIC_CLIPDIST;
 904       *index = 1;
 905       break;
 906    case VARYING_SLOT_CLIP_VERTEX:
 907       *name = TGSI_SEMANTIC_CLIPVERTEX;
 908       *index = 0;
 909       break;
 910    case VARYING_SLOT_COL0:
 911       *name = TGSI_SEMANTIC_COLOR;
 912       *index = 0;
 913       break;
 914    case VARYING_SLOT_COL1:
 915       *name = TGSI_SEMANTIC_COLOR;
 916       *index = 1;
 917       break;
 918    case VARYING_SLOT_EDGE:
 919       *name = TGSI_SEMANTIC_EDGEFLAG;
 920       *index = 0;
 921       break;
 922    case VARYING_SLOT_FACE:
 923       *name = TGSI_SEMANTIC_FACE;
 924       *index = 0;
 925       break;
 926    case VARYING_SLOT_FOGC:
 927       *name = TGSI_SEMANTIC_FOG;
 928       *index = 0;
 929       break;
 930    case VARYING_SLOT_LAYER:
 931       *name = TGSI_SEMANTIC_LAYER;
 932       *index = 0;
 933       break;
 934    case VARYING_SLOT_PNTC:
 935       *name = TGSI_SEMANTIC_PCOORD;
 936       *index = 0;
 937       break;
 938    case VARYING_SLOT_POS:
 939       *name = TGSI_SEMANTIC_POSITION;
 940       *index = 0;
 941       break;
 942    case VARYING_SLOT_PRIMITIVE_ID:
 943       *name = TGSI_SEMANTIC_PRIMID;
 944       *index = 0;
 945       break;
 946    case VARYING_SLOT_PSIZ:
 947       *name = TGSI_SEMANTIC_PSIZE;
 948       *index = 0;
 949       break;
 950    case VARYING_SLOT_TESS_LEVEL_INNER:
 951       *name = TGSI_SEMANTIC_TESSINNER;
 952       *index = 0;
 953       break;
 954    case VARYING_SLOT_TESS_LEVEL_OUTER:
 955       *name = TGSI_SEMANTIC_TESSOUTER;
 956       *index = 0;
 957       break;
 958    case VARYING_SLOT_VIEWPORT:
 959       *name = TGSI_SEMANTIC_VIEWPORT_INDEX;
 960       *index = 0;
 961       break;
 962    default:
 963       ERROR("unknown varying slot %u\n", slot);
 964       assert(false);
 965       break;
 966    }
 967 }
 968
 969 static void
 970 frag_result_to_tgsi_semantic(unsigned slot, unsigned *name, unsigned *index)
 971 {
 972    if (slot >= FRAG_RESULT_DATA0) {
 973       *name = TGSI_SEMANTIC_COLOR;
 974       *index = slot - FRAG_RESULT_COLOR - 2; // intentional
 975       return;
 976    }
 977
 978    switch (slot) {
 979    case FRAG_RESULT_COLOR:
 980       *name = TGSI_SEMANTIC_COLOR;
 981       *index = 0;
 982       break;
 983    case FRAG_RESULT_DEPTH:
 984       *name = TGSI_SEMANTIC_POSITION;
 985       *index = 0;
 986       break;
 987    case FRAG_RESULT_SAMPLE_MASK:
 988       *name = TGSI_SEMANTIC_SAMPLEMASK;
 989       *index = 0;
 990       break;
 991    default:
 992       ERROR("unknown frag result slot %u\n", slot);
 993       assert(false);
 994       break;
 995    }
 996 }
 997
 998 // copy of _mesa_sysval_to_semantic
 999 static void
1000 system_val_to_tgsi_semantic(unsigned val, unsigned *name, unsigned *index)
1001 {
1002    *index = 0;
1003    switch (val) {
1004    // Vertex shader
1005    case SYSTEM_VALUE_VERTEX_ID:
1006       *name = TGSI_SEMANTIC_VERTEXID;
1007       break;
1008    case SYSTEM_VALUE_INSTANCE_ID:
1009       *name = TGSI_SEMANTIC_INSTANCEID;
1010       break;
1011    case SYSTEM_VALUE_VERTEX_ID_ZERO_BASE:
1012       *name = TGSI_SEMANTIC_VERTEXID_NOBASE;
1013       break;
1014    case SYSTEM_VALUE_BASE_VERTEX:
1015       *name = TGSI_SEMANTIC_BASEVERTEX;
1016       break;
1017    case SYSTEM_VALUE_BASE_INSTANCE:
1018       *name = TGSI_SEMANTIC_BASEINSTANCE;
1019       break;
1020    case SYSTEM_VALUE_DRAW_ID:
1021       *name = TGSI_SEMANTIC_DRAWID;
1022       break;
1023
1024    // Geometry shader
1025    case SYSTEM_VALUE_INVOCATION_ID:
1026       *name = TGSI_SEMANTIC_INVOCATIONID;
1027       break;
1028
1029    // Fragment shader
1030    case SYSTEM_VALUE_FRAG_COORD:
1031       *name = TGSI_SEMANTIC_POSITION;
1032       break;
1033    case SYSTEM_VALUE_FRONT_FACE:
1034       *name = TGSI_SEMANTIC_FACE;
1035       break;
1036    case SYSTEM_VALUE_SAMPLE_ID:
1037       *name = TGSI_SEMANTIC_SAMPLEID;
1038       break;
1039    case SYSTEM_VALUE_SAMPLE_POS:
1040       *name = TGSI_SEMANTIC_SAMPLEPOS;
1041       break;
1042    case SYSTEM_VALUE_SAMPLE_MASK_IN:
1043       *name = TGSI_SEMANTIC_SAMPLEMASK;
1044       break;
1045    case SYSTEM_VALUE_HELPER_INVOCATION:
1046       *name = TGSI_SEMANTIC_HELPER_INVOCATION;
1047       break;
1048
1049    // Tessellation shader
1050    case SYSTEM_VALUE_TESS_COORD:
1051       *name = TGSI_SEMANTIC_TESSCOORD;
1052       break;
1053    case SYSTEM_VALUE_VERTICES_IN:
1054       *name = TGSI_SEMANTIC_VERTICESIN;
1055       break;
1056    case SYSTEM_VALUE_PRIMITIVE_ID:
1057       *name = TGSI_SEMANTIC_PRIMID;
1058       break;
1059    case SYSTEM_VALUE_TESS_LEVEL_OUTER:
1060       *name = TGSI_SEMANTIC_TESSOUTER;
1061       break;
1062    case SYSTEM_VALUE_TESS_LEVEL_INNER:
1063       *name = TGSI_SEMANTIC_TESSINNER;
1064       break;
1065
1066    // Compute shader
1067    case SYSTEM_VALUE_LOCAL_INVOCATION_ID:
1068       *name = TGSI_SEMANTIC_THREAD_ID;
1069       break;
1070    case SYSTEM_VALUE_WORK_GROUP_ID:
1071       *name = TGSI_SEMANTIC_BLOCK_ID;
1072       break;
1073    case SYSTEM_VALUE_NUM_WORK_GROUPS:
1074       *name = TGSI_SEMANTIC_GRID_SIZE;
1075       break;
1076    case SYSTEM_VALUE_LOCAL_GROUP_SIZE:
1077       *name = TGSI_SEMANTIC_BLOCK_SIZE;
1078       break;
1079
1080    // ARB_shader_ballot
1081    case SYSTEM_VALUE_SUBGROUP_SIZE:
1082       *name = TGSI_SEMANTIC_SUBGROUP_SIZE;
1083       break;
1084    case SYSTEM_VALUE_SUBGROUP_INVOCATION:
1085       *name = TGSI_SEMANTIC_SUBGROUP_INVOCATION;
1086       break;
1087    case SYSTEM_VALUE_SUBGROUP_EQ_MASK:
1088       *name = TGSI_SEMANTIC_SUBGROUP_EQ_MASK;
1089       break;
1090    case SYSTEM_VALUE_SUBGROUP_GE_MASK:
1091       *name = TGSI_SEMANTIC_SUBGROUP_GE_MASK;
1092       break;
1093    case SYSTEM_VALUE_SUBGROUP_GT_MASK:
1094       *name = TGSI_SEMANTIC_SUBGROUP_GT_MASK;
1095       break;
1096    case SYSTEM_VALUE_SUBGROUP_LE_MASK:
1097       *name = TGSI_SEMANTIC_SUBGROUP_LE_MASK;
1098       break;
1099    case SYSTEM_VALUE_SUBGROUP_LT_MASK:
1100       *name = TGSI_SEMANTIC_SUBGROUP_LT_MASK;
1101       break;
1102
1103    default:
1104       ERROR("unknown system value %u\n", val);
1105       assert(false);
1106       break;
1107    }
1108 }
1109
1110 void
1111 Converter::setInterpolate(nv50_ir_varying *var,
1112                           uint8_t mode,
1113                           bool centroid,
1114                           unsigned semantic)
1115 {
1116    switch (mode) {
1117    case INTERP_MODE_FLAT:
1118       var->flat = 1;
1119       break;
1120    case INTERP_MODE_NONE:
1121       if (semantic == TGSI_SEMANTIC_COLOR)
1122          var->sc = 1;
1123       else if (semantic == TGSI_SEMANTIC_POSITION)
1124          var->linear = 1;
1125       break;
1126    case INTERP_MODE_NOPERSPECTIVE:
1127       var->linear = 1;
1128       break;
1129    case INTERP_MODE_SMOOTH:
1130       break;
1131    }
1132    var->centroid = centroid;
1133 }
1134
1135 static uint16_t
1136 calcSlots(const glsl_type *type, Program::Type stage, const shader_info &info,
1137           bool input, const nir_variable *var)
1138 {
1139    if (!type->is_array())
1140       return type->count_attribute_slots(false);
1141
1142    uint16_t slots;
1143    switch (stage) {
1144    case Program::TYPE_GEOMETRY:
1145       slots = type->uniform_locations();
1146       if (input)
1147          slots /= info.gs.vertices_in;
1148       break;
1149    case Program::TYPE_TESSELLATION_CONTROL:
1150    case Program::TYPE_TESSELLATION_EVAL:
1151       // remove first dimension
1152       if (var->data.patch || (!input && stage == Program::TYPE_TESSELLATION_EVAL))
1153          slots = type->uniform_locations();
1154       else
1155          slots = type->fields.array->uniform_locations();
1156       break;
1157    default:
1158       slots = type->count_attribute_slots(false);
1159       break;
1160    }
1161
1162    return slots;
1163 }
1164
1165 bool Converter::assignSlots() {
1166    unsigned name;
1167    unsigned index;
1168
1169    info->io.viewportId = -1;
1170    info->numInputs = 0;
1171    info->numOutputs = 0;
1172
1173    // we have to fixup the uniform locations for arrays
1174    unsigned numImages = 0;
1175    nir_foreach_variable(var, &nir->uniforms) {
1176       const glsl_type *type = var->type;
1177       if (!type->without_array()->is_image())
1178          continue;
1179       var->data.driver_location = numImages;
1180       numImages += type->is_array() ? type->arrays_of_arrays_size() : 1;
1181    }
1182
1183    info->numSysVals = 0;
1184    for (uint8_t i = 0; i < SYSTEM_VALUE_MAX; ++i) {
1185       if (!(nir->info.system_values_read & 1ull << i))
1186          continue;
1187
1188       system_val_to_tgsi_semantic(i, &name, &index);
1189       info->sv[info->numSysVals].sn = name;
1190       info->sv[info->numSysVals].si = index;
1191       info->sv[info->numSysVals].input = 0; // TODO inferSysValDirection(sn);
1192
1193       switch (i) {
1194       case SYSTEM_VALUE_INSTANCE_ID:
1195          info->io.instanceId = info->numSysVals;
1196          break;
1197       case SYSTEM_VALUE_TESS_LEVEL_INNER:
1198       case SYSTEM_VALUE_TESS_LEVEL_OUTER:
1199          info->sv[info->numSysVals].patch = 1;
1200          break;
1201       case SYSTEM_VALUE_VERTEX_ID:
1202          info->io.vertexId = info->numSysVals;
1203          break;
1204       default:
1205          break;
1206       }
1207
1208       info->numSysVals += 1;
1209    }
1210
1211    if (prog->getType() == Program::TYPE_COMPUTE)
1212       return true;
1213
1214    nir_foreach_variable(var, &nir->inputs) {
1215       const glsl_type *type = var->type;
1216       int slot = var->data.location;
1217       uint16_t slots = calcSlots(type, prog->getType(), nir->info, true, var);
1218       uint32_t comp = type->is_array() ? type->without_array()->component_slots()
1219                                        : type->component_slots();
1220       uint32_t frac = var->data.location_frac;
1221       uint32_t vary = var->data.driver_location;
1222
1223       if (glsl_base_type_is_64bit(type->without_array()->base_type)) {
1224          if (comp > 2)
1225             slots *= 2;
1226       }
1227
1228       assert(vary + slots <= PIPE_MAX_SHADER_INPUTS);
1229
1230       switch(prog->getType()) {
1231       case Program::TYPE_FRAGMENT:
1232          varying_slot_to_tgsi_semantic((gl_varying_slot)slot, &name, &index);
1233          for (uint16_t i = 0; i < slots; ++i) {
1234             setInterpolate(&info->in[vary + i], var->data.interpolation,
1235                            var->data.centroid | var->data.sample, name);
1236          }
1237          break;
1238       case Program::TYPE_GEOMETRY:
1239          varying_slot_to_tgsi_semantic((gl_varying_slot)slot, &name, &index);
1240          break;
1241       case Program::TYPE_TESSELLATION_CONTROL:
1242       case Program::TYPE_TESSELLATION_EVAL:
1243          varying_slot_to_tgsi_semantic((gl_varying_slot)slot, &name, &index);
1244          if (var->data.patch && name == TGSI_SEMANTIC_PATCH)
1245             info->numPatchConstants = MAX2(info->numPatchConstants, index + slots);
1246          break;
1247       case Program::TYPE_VERTEX:
1248          vert_attrib_to_tgsi_semantic((gl_vert_attrib)slot, &name, &index);
1249          switch (name) {
1250          case TGSI_SEMANTIC_EDGEFLAG:
1251             info->io.edgeFlagIn = vary;
1252             break;
1253          default:
1254             break;
1255          }
1256          break;
1257       default:
1258          ERROR("unknown shader type %u in assignSlots\n", prog->getType());
1259          return false;
1260       }
1261
1262       for (uint16_t i = 0u; i < slots; ++i, ++vary) {
1263          info->in[vary].id = vary;
1264          info->in[vary].patch = var->data.patch;
1265          info->in[vary].sn = name;
1266          info->in[vary].si = index + i;
1267          if (glsl_base_type_is_64bit(type->without_array()->base_type))
1268             if (i & 0x1)
1269                info->in[vary].mask |= (((1 << (comp * 2)) - 1) << (frac * 2) >> 0x4);
1270             else
1271                info->in[vary].mask |= (((1 << (comp * 2)) - 1) << (frac * 2) & 0xf);
1272          else
1273             info->in[vary].mask |= ((1 << comp) - 1) << frac;
1274       }
1275       info->numInputs = std::max<uint8_t>(info->numInputs, vary);
1276    }
1277
1278    nir_foreach_variable(var, &nir->outputs) {
1279       const glsl_type *type = var->type;
1280       int slot = var->data.location;
1281       uint16_t slots = calcSlots(type, prog->getType(), nir->info, false, var);
1282       uint32_t comp = type->is_array() ? type->without_array()->component_slots()
1283                                        : type->component_slots();
1284       uint32_t frac = var->data.location_frac;
1285       uint32_t vary = var->data.driver_location;
1286
1287       if (glsl_base_type_is_64bit(type->without_array()->base_type)) {
1288          if (comp > 2)
1289             slots *= 2;
1290       }
1291
1292       assert(vary < PIPE_MAX_SHADER_OUTPUTS);
1293
1294       switch(prog->getType()) {
1295       case Program::TYPE_FRAGMENT:
1296          frag_result_to_tgsi_semantic((gl_frag_result)slot, &name, &index);
1297          switch (name) {
1298          case TGSI_SEMANTIC_COLOR:
1299             if (!var->data.fb_fetch_output)
1300                info->prop.fp.numColourResults++;
1301             info->prop.fp.separateFragData = true;
1302             // sometimes we get FRAG_RESULT_DATAX with data.index 0
1303             // sometimes we get FRAG_RESULT_DATA0 with data.index X
1304             index = index == 0 ? var->data.index : index;
1305             break;
1306          case TGSI_SEMANTIC_POSITION:
1307             info->io.fragDepth = vary;
1308             info->prop.fp.writesDepth = true;
1309             break;
1310          case TGSI_SEMANTIC_SAMPLEMASK:
1311             info->io.sampleMask = vary;
1312             break;
1313          default:
1314             break;
1315          }
1316          break;
1317       case Program::TYPE_GEOMETRY:
1318       case Program::TYPE_TESSELLATION_CONTROL:
1319       case Program::TYPE_TESSELLATION_EVAL:
1320       case Program::TYPE_VERTEX:
1321          varying_slot_to_tgsi_semantic((gl_varying_slot)slot, &name, &index);
1322
1323          if (var->data.patch && name != TGSI_SEMANTIC_TESSINNER &&
1324              name != TGSI_SEMANTIC_TESSOUTER)
1325             info->numPatchConstants = MAX2(info->numPatchConstants, index + slots);
1326
1327          switch (name) {
1328          case TGSI_SEMANTIC_CLIPDIST:
1329             info->io.genUserClip = -1;
1330             break;
1331          case TGSI_SEMANTIC_CLIPVERTEX:
1332             clipVertexOutput = vary;
1333             break;
1334          case TGSI_SEMANTIC_EDGEFLAG:
1335             info->io.edgeFlagOut = vary;
1336             break;
1337          case TGSI_SEMANTIC_POSITION:
1338             if (clipVertexOutput < 0)
1339                clipVertexOutput = vary;
1340             break;
1341          default:
1342             break;
1343          }
1344          break;
1345       default:
1346          ERROR("unknown shader type %u in assignSlots\n", prog->getType());
1347          return false;
1348       }
1349
1350       for (uint16_t i = 0u; i < slots; ++i, ++vary) {
1351          info->out[vary].id = vary;
1352          info->out[vary].patch = var->data.patch;
1353          info->out[vary].sn = name;
1354          info->out[vary].si = index + i;
1355          if (glsl_base_type_is_64bit(type->without_array()->base_type))
1356             if (i & 0x1)
1357                info->out[vary].mask |= (((1 << (comp * 2)) - 1) << (frac * 2) >> 0x4);
1358             else
1359                info->out[vary].mask |= (((1 << (comp * 2)) - 1) << (frac * 2) & 0xf);
1360          else
1361             info->out[vary].mask |= ((1 << comp) - 1) << frac;
1362
1363          if (nir->info.outputs_read & 1ull << slot)
1364             info->out[vary].oread = 1;
1365       }
1366       info->numOutputs = std::max<uint8_t>(info->numOutputs, vary);
1367    }
1368
1369    if (info->io.genUserClip > 0) {
1370       info->io.clipDistances = info->io.genUserClip;
1371
1372       const unsigned int nOut = (info->io.genUserClip + 3) / 4;
1373
1374       for (unsigned int n = 0; n < nOut; ++n) {
1375          unsigned int i = info->numOutputs++;
1376          info->out[i].id = i;
1377          info->out[i].sn = TGSI_SEMANTIC_CLIPDIST;
1378          info->out[i].si = n;
1379          info->out[i].mask = ((1 << info->io.clipDistances) - 1) >> (n * 4);
1380       }
1381    }
1382
1383    return info->assignSlots(info) == 0;
1384 }
1385
1386 uint32_t
1387 Converter::getSlotAddress(nir_intrinsic_instr *insn, uint8_t idx, uint8_t slot)
1388 {
1389    DataType ty;
1390    int offset = nir_intrinsic_component(insn);
1391    bool input;
1392
1393    if (nir_intrinsic_infos[insn->intrinsic].has_dest)
1394       ty = getDType(insn);
1395    else
1396       ty = getSType(insn->src[0], false, false);
1397
1398    switch (insn->intrinsic) {
1399    case nir_intrinsic_load_input:
1400    case nir_intrinsic_load_interpolated_input:
1401    case nir_intrinsic_load_per_vertex_input:
1402       input = true;
1403       break;
1404    case nir_intrinsic_load_output:
1405    case nir_intrinsic_load_per_vertex_output:
1406    case nir_intrinsic_store_output:
1407    case nir_intrinsic_store_per_vertex_output:
1408       input = false;
1409       break;
1410    default:
1411       ERROR("unknown intrinsic in getSlotAddress %s",
1412             nir_intrinsic_infos[insn->intrinsic].name);
1413       input = false;
1414       assert(false);
1415       break;
1416    }
1417
1418    if (typeSizeof(ty) == 8) {
1419       slot *= 2;
1420       slot += offset;
1421       if (slot >= 4) {
1422          idx += 1;
1423          slot -= 4;
1424       }
1425    } else {
1426       slot += offset;
1427    }
1428
1429    assert(slot < 4);
1430    assert(!input || idx < PIPE_MAX_SHADER_INPUTS);
1431    assert(input || idx < PIPE_MAX_SHADER_OUTPUTS);
1432
1433    const nv50_ir_varying *vary = input ? info->in : info->out;
1434    return vary[idx].slot[slot] * 4;
1435 }
1436
1437 Instruction *
1438 Converter::loadFrom(DataFile file, uint8_t i, DataType ty, Value *def,
1439                     uint32_t base, uint8_t c, Value *indirect0,
1440                     Value *indirect1, bool patch)
1441 {
1442    unsigned int tySize = typeSizeof(ty);
1443
1444    if (tySize == 8 &&
1445        (file == FILE_MEMORY_CONST || file == FILE_MEMORY_BUFFER || indirect0)) {
1446       Value *lo = getSSA();
1447       Value *hi = getSSA();
1448
1449       Instruction *loi =
1450          mkLoad(TYPE_U32, lo,
1451                 mkSymbol(file, i, TYPE_U32, base + c * tySize),
1452                 indirect0);
1453       loi->setIndirect(0, 1, indirect1);
1454       loi->perPatch = patch;
1455
1456       Instruction *hii =
1457          mkLoad(TYPE_U32, hi,
1458                 mkSymbol(file, i, TYPE_U32, base + c * tySize + 4),
1459                 indirect0);
1460       hii->setIndirect(0, 1, indirect1);
1461       hii->perPatch = patch;
1462
1463       return mkOp2(OP_MERGE, ty, def, lo, hi);
1464    } else {
1465       Instruction *ld =
1466          mkLoad(ty, def, mkSymbol(file, i, ty, base + c * tySize), indirect0);
1467       ld->setIndirect(0, 1, indirect1);
1468       ld->perPatch = patch;
1469       return ld;
1470    }
1471 }
1472
1473 void
1474 Converter::storeTo(nir_intrinsic_instr *insn, DataFile file, operation op,
1475                    DataType ty, Value *src, uint8_t idx, uint8_t c,
1476                    Value *indirect0, Value *indirect1)
1477 {
1478    uint8_t size = typeSizeof(ty);
1479    uint32_t address = getSlotAddress(insn, idx, c);
1480
1481    if (size == 8 && indirect0) {
1482       Value *split[2];
1483       mkSplit(split, 4, src);
1484
1485       if (op == OP_EXPORT) {
1486          split[0] = mkMov(getSSA(), split[0], ty)->getDef(0);
1487          split[1] = mkMov(getSSA(), split[1], ty)->getDef(0);
1488       }
1489
1490       mkStore(op, TYPE_U32, mkSymbol(file, 0, TYPE_U32, address), indirect0,
1491               split[0])->perPatch = info->out[idx].patch;
1492       mkStore(op, TYPE_U32, mkSymbol(file, 0, TYPE_U32, address + 4), indirect0,
1493               split[1])->perPatch = info->out[idx].patch;
1494    } else {
1495       if (op == OP_EXPORT)
1496          src = mkMov(getSSA(size), src, ty)->getDef(0);
1497       mkStore(op, ty, mkSymbol(file, 0, ty, address), indirect0,
1498               src)->perPatch = info->out[idx].patch;
1499    }
1500 }
1501
1502 bool
1503 Converter::parseNIR()
1504 {
1505    info->bin.tlsSpace = 0;
1506    info->io.clipDistances = nir->info.clip_distance_array_size;
1507    info->io.cullDistances = nir->info.cull_distance_array_size;
1508
1509    switch(prog->getType()) {
1510    case Program::TYPE_COMPUTE:
1511       info->prop.cp.numThreads[0] = nir->info.cs.local_size[0];
1512       info->prop.cp.numThreads[1] = nir->info.cs.local_size[1];
1513       info->prop.cp.numThreads[2] = nir->info.cs.local_size[2];
1514       info->bin.smemSize = nir->info.cs.shared_size;
1515       break;
1516    case Program::TYPE_FRAGMENT:
1517       info->prop.fp.earlyFragTests = nir->info.fs.early_fragment_tests;
1518       info->prop.fp.persampleInvocation =
1519          (nir->info.system_values_read & SYSTEM_BIT_SAMPLE_ID) ||
1520          (nir->info.system_values_read & SYSTEM_BIT_SAMPLE_POS);
1521       info->prop.fp.postDepthCoverage = nir->info.fs.post_depth_coverage;
1522       info->prop.fp.readsSampleLocations =
1523          (nir->info.system_values_read & SYSTEM_BIT_SAMPLE_POS);
1524       info->prop.fp.usesDiscard = nir->info.fs.uses_discard;
1525       info->prop.fp.usesSampleMaskIn =
1526          !!(nir->info.system_values_read & SYSTEM_BIT_SAMPLE_MASK_IN);
1527       break;
1528    case Program::TYPE_GEOMETRY:
1529       info->prop.gp.inputPrim = nir->info.gs.input_primitive;
1530       info->prop.gp.instanceCount = nir->info.gs.invocations;
1531       info->prop.gp.maxVertices = nir->info.gs.vertices_out;
1532       info->prop.gp.outputPrim = nir->info.gs.output_primitive;
1533       break;
1534    case Program::TYPE_TESSELLATION_CONTROL:
1535    case Program::TYPE_TESSELLATION_EVAL:
1536       if (nir->info.tess.primitive_mode == GL_ISOLINES)
1537          info->prop.tp.domain = GL_LINES;
1538       else
1539          info->prop.tp.domain = nir->info.tess.primitive_mode;
1540       info->prop.tp.outputPatchSize = nir->info.tess.tcs_vertices_out;
1541       info->prop.tp.outputPrim =
1542          nir->info.tess.point_mode ? PIPE_PRIM_POINTS : PIPE_PRIM_TRIANGLES;
1543       info->prop.tp.partitioning = (nir->info.tess.spacing + 1) % 3;
1544       info->prop.tp.winding = !nir->info.tess.ccw;
1545       break;
1546    case Program::TYPE_VERTEX:
1547       info->prop.vp.usesDrawParameters =
1548          (nir->info.system_values_read & BITFIELD64_BIT(SYSTEM_VALUE_BASE_VERTEX)) ||
1549          (nir->info.system_values_read & BITFIELD64_BIT(SYSTEM_VALUE_BASE_INSTANCE)) ||
1550          (nir->info.system_values_read & BITFIELD64_BIT(SYSTEM_VALUE_DRAW_ID));
1551       break;
1552    default:
1553       break;
1554    }
1555
1556    return true;
1557 }
1558
1559 bool
1560 Converter::visit(nir_function *function)
1561 {
1562    assert(function->impl);
1563
1564    // usually the blocks will set everything up, but main is special
1565    BasicBlock *entry = new BasicBlock(prog->main);
1566    exit = new BasicBlock(prog->main);
1567    blocks[nir_start_block(function->impl)->index] = entry;
1568    prog->main->setEntry(entry);
1569    prog->main->setExit(exit);
1570
1571    setPosition(entry, true);
1572
1573    if (info->io.genUserClip > 0) {
1574       for (int c = 0; c < 4; ++c)
1575          clipVtx[c] = getScratch();
1576    }
1577
1578    switch (prog->getType()) {
1579    case Program::TYPE_TESSELLATION_CONTROL:
1580       outBase = mkOp2v(
1581          OP_SUB, TYPE_U32, getSSA(),
1582          mkOp1v(OP_RDSV, TYPE_U32, getSSA(), mkSysVal(SV_LANEID, 0)),
1583          mkOp1v(OP_RDSV, TYPE_U32, getSSA(), mkSysVal(SV_INVOCATION_ID, 0)));
1584       break;
1585    case Program::TYPE_FRAGMENT: {
1586       Symbol *sv = mkSysVal(SV_POSITION, 3);
1587       fragCoord[3] = mkOp1v(OP_RDSV, TYPE_F32, getSSA(), sv);
1588       fp.position = mkOp1v(OP_RCP, TYPE_F32, fragCoord[3], fragCoord[3]);
1589       break;
1590    }
1591    default:
1592       break;
1593    }
1594
1595    nir_foreach_register(reg, &function->impl->registers) {
1596       if (reg->num_array_elems) {
1597          // TODO: packed variables would be nice, but MemoryOpt fails
1598          // replace 4 with reg->num_components
1599          uint32_t size = 4 * reg->num_array_elems * (reg->bit_size / 8);
1600          regToLmemOffset[reg->index] = info->bin.tlsSpace;
1601          info->bin.tlsSpace += size;
1602       }
1603    }
1604
1605    nir_index_ssa_defs(function->impl);
1606    foreach_list_typed(nir_cf_node, node, node, &function->impl->body) {
1607       if (!visit(node))
1608          return false;
1609    }
1610
1611    bb->cfg.attach(&exit->cfg, Graph::Edge::TREE);
1612    setPosition(exit, true);
1613
1614    if ((prog->getType() == Program::TYPE_VERTEX ||
1615         prog->getType() == Program::TYPE_TESSELLATION_EVAL)
1616        && info->io.genUserClip > 0)
1617       handleUserClipPlanes();
1618
1619    // TODO: for non main function this needs to be a OP_RETURN
1620    mkOp(OP_EXIT, TYPE_NONE, NULL)->terminator = 1;
1621    return true;
1622 }
1623
1624 bool
1625 Converter::visit(nir_cf_node *node)
1626 {
1627    switch (node->type) {
1628    case nir_cf_node_block:
1629       return visit(nir_cf_node_as_block(node));
1630    case nir_cf_node_if:
1631       return visit(nir_cf_node_as_if(node));
1632    case nir_cf_node_loop:
1633       return visit(nir_cf_node_as_loop(node));
1634    default:
1635       ERROR("unknown nir_cf_node type %u\n", node->type);
1636       return false;
1637    }
1638 }
1639
1640 bool
1641 Converter::visit(nir_block *block)
1642 {
1643    if (!block->predecessors->entries && block->instr_list.is_empty())
1644       return true;
1645
1646    BasicBlock *bb = convert(block);
1647
1648    setPosition(bb, true);
1649    nir_foreach_instr(insn, block) {
1650       if (!visit(insn))
1651          return false;
1652    }
1653    return true;
1654 }
1655
1656 bool
1657 Converter::visit(nir_if *nif)
1658 {
1659    DataType sType = getSType(nif->condition, false, false);
1660    Value *src = getSrc(&nif->condition, 0);
1661
1662    nir_block *lastThen = nir_if_last_then_block(nif);
1663    nir_block *lastElse = nir_if_last_else_block(nif);
1664
1665    assert(!lastThen->successors[1]);
1666    assert(!lastElse->successors[1]);
1667
1668    BasicBlock *ifBB = convert(nir_if_first_then_block(nif));
1669    BasicBlock *elseBB = convert(nir_if_first_else_block(nif));
1670
1671    bb->cfg.attach(&ifBB->cfg, Graph::Edge::TREE);
1672    bb->cfg.attach(&elseBB->cfg, Graph::Edge::TREE);
1673
1674    // we only insert joinats, if both nodes end up at the end of the if again.
1675    // the reason for this to not happens are breaks/continues/ret/... which
1676    // have their own handling
1677    if (lastThen->successors[0] == lastElse->successors[0])
1678       bb->joinAt = mkFlow(OP_JOINAT, convert(lastThen->successors[0]),
1679                           CC_ALWAYS, NULL);
1680
1681    mkFlow(OP_BRA, elseBB, CC_EQ, src)->setType(sType);
1682
1683    foreach_list_typed(nir_cf_node, node, node, &nif->then_list) {
1684       if (!visit(node))
1685          return false;
1686    }
1687    setPosition(convert(lastThen), true);
1688    if (!bb->getExit() ||
1689        !bb->getExit()->asFlow() ||
1690         bb->getExit()->asFlow()->op == OP_JOIN) {
1691       BasicBlock *tailBB = convert(lastThen->successors[0]);
1692       mkFlow(OP_BRA, tailBB, CC_ALWAYS, NULL);
1693       bb->cfg.attach(&tailBB->cfg, Graph::Edge::FORWARD);
1694    }
1695
1696    foreach_list_typed(nir_cf_node, node, node, &nif->else_list) {
1697       if (!visit(node))
1698          return false;
1699    }
1700    setPosition(convert(lastElse), true);
1701    if (!bb->getExit() ||
1702        !bb->getExit()->asFlow() ||
1703         bb->getExit()->asFlow()->op == OP_JOIN) {
1704       BasicBlock *tailBB = convert(lastElse->successors[0]);
1705       mkFlow(OP_BRA, tailBB, CC_ALWAYS, NULL);
1706       bb->cfg.attach(&tailBB->cfg, Graph::Edge::FORWARD);
1707    }
1708
1709    if (lastThen->successors[0] == lastElse->successors[0]) {
1710       setPosition(convert(lastThen->successors[0]), true);
1711       mkFlow(OP_JOIN, NULL, CC_ALWAYS, NULL)->fixed = 1;
1712    }
1713
1714    return true;
1715 }
1716
1717 bool
1718 Converter::visit(nir_loop *loop)
1719 {
1720    curLoopDepth += 1;
1721    func->loopNestingBound = std::max(func->loopNestingBound, curLoopDepth);
1722
1723    BasicBlock *loopBB = convert(nir_loop_first_block(loop));
1724    BasicBlock *tailBB =
1725       convert(nir_cf_node_as_block(nir_cf_node_next(&loop->cf_node)));
1726    bb->cfg.attach(&loopBB->cfg, Graph::Edge::TREE);
1727
1728    mkFlow(OP_PREBREAK, tailBB, CC_ALWAYS, NULL);
1729    setPosition(loopBB, false);
1730    mkFlow(OP_PRECONT, loopBB, CC_ALWAYS, NULL);
1731
1732    foreach_list_typed(nir_cf_node, node, node, &loop->body) {
1733       if (!visit(node))
1734          return false;
1735    }
1736    Instruction *insn = bb->getExit();
1737    if (bb->cfg.incidentCount() != 0) {
1738       if (!insn || !insn->asFlow()) {
1739          mkFlow(OP_CONT, loopBB, CC_ALWAYS, NULL);
1740          bb->cfg.attach(&loopBB->cfg, Graph::Edge::BACK);
1741       } else if (insn && insn->op == OP_BRA && !insn->getPredicate() &&
1742                  tailBB->cfg.incidentCount() == 0) {
1743          // RA doesn't like having blocks around with no incident edge,
1744          // so we create a fake one to make it happy
1745          bb->cfg.attach(&tailBB->cfg, Graph::Edge::TREE);
1746       }
1747    }
1748
1749    curLoopDepth -= 1;
1750
1751    return true;
1752 }
1753
1754 bool
1755 Converter::visit(nir_instr *insn)
1756 {
1757    // we need an insertion point for on the fly generated immediate loads
1758    immInsertPos = bb->getExit();
1759    switch (insn->type) {
1760    case nir_instr_type_alu:
1761       return visit(nir_instr_as_alu(insn));
1762    case nir_instr_type_deref:
1763       return visit(nir_instr_as_deref(insn));
1764    case nir_instr_type_intrinsic:
1765       return visit(nir_instr_as_intrinsic(insn));
1766    case nir_instr_type_jump:
1767       return visit(nir_instr_as_jump(insn));
1768    case nir_instr_type_load_const:
1769       return visit(nir_instr_as_load_const(insn));
1770    case nir_instr_type_ssa_undef:
1771       return visit(nir_instr_as_ssa_undef(insn));
1772    case nir_instr_type_tex:
1773       return visit(nir_instr_as_tex(insn));
1774    default:
1775       ERROR("unknown nir_instr type %u\n", insn->type);
1776       return false;
1777    }
1778    return true;
1779 }
1780
1781 SVSemantic
1782 Converter::convert(nir_intrinsic_op intr)
1783 {
1784    switch (intr) {
1785    case nir_intrinsic_load_base_vertex:
1786       return SV_BASEVERTEX;
1787    case nir_intrinsic_load_base_instance:
1788       return SV_BASEINSTANCE;
1789    case nir_intrinsic_load_draw_id:
1790       return SV_DRAWID;
1791    case nir_intrinsic_load_front_face:
1792       return SV_FACE;
1793    case nir_intrinsic_load_helper_invocation:
1794       return SV_THREAD_KILL;
1795    case nir_intrinsic_load_instance_id:
1796       return SV_INSTANCE_ID;
1797    case nir_intrinsic_load_invocation_id:
1798       return SV_INVOCATION_ID;
1799    case nir_intrinsic_load_local_group_size:
1800       return SV_NTID;
1801    case nir_intrinsic_load_local_invocation_id:
1802       return SV_TID;
1803    case nir_intrinsic_load_num_work_groups:
1804       return SV_NCTAID;
1805    case nir_intrinsic_load_patch_vertices_in:
1806       return SV_VERTEX_COUNT;
1807    case nir_intrinsic_load_primitive_id:
1808       return SV_PRIMITIVE_ID;
1809    case nir_intrinsic_load_sample_id:
1810       return SV_SAMPLE_INDEX;
1811    case nir_intrinsic_load_sample_mask_in:
1812       return SV_SAMPLE_MASK;
1813    case nir_intrinsic_load_sample_pos:
1814       return SV_SAMPLE_POS;
1815    case nir_intrinsic_load_subgroup_eq_mask:
1816       return SV_LANEMASK_EQ;
1817    case nir_intrinsic_load_subgroup_ge_mask:
1818       return SV_LANEMASK_GE;
1819    case nir_intrinsic_load_subgroup_gt_mask:
1820       return SV_LANEMASK_GT;
1821    case nir_intrinsic_load_subgroup_le_mask:
1822       return SV_LANEMASK_LE;
1823    case nir_intrinsic_load_subgroup_lt_mask:
1824       return SV_LANEMASK_LT;
1825    case nir_intrinsic_load_subgroup_invocation:
1826       return SV_LANEID;
1827    case nir_intrinsic_load_tess_coord:
1828       return SV_TESS_COORD;
1829    case nir_intrinsic_load_tess_level_inner:
1830       return SV_TESS_INNER;
1831    case nir_intrinsic_load_tess_level_outer:
1832       return SV_TESS_OUTER;
1833    case nir_intrinsic_load_vertex_id:
1834       return SV_VERTEX_ID;
1835    case nir_intrinsic_load_work_group_id:
1836       return SV_CTAID;
1837    default:
1838       ERROR("unknown SVSemantic for nir_intrinsic_op %s\n",
1839             nir_intrinsic_infos[intr].name);
1840       assert(false);
1841       return SV_LAST;
1842    }
1843 }
1844
1845 ImgFormat
1846 Converter::convertGLImgFormat(GLuint format)
1847 {
1848 #define FMT_CASE(a, b) \
1849   case GL_ ## a: return nv50_ir::FMT_ ## b
1850
1851    switch (format) {
1852    FMT_CASE(NONE, NONE);
1853
1854    FMT_CASE(RGBA32F, RGBA32F);
1855    FMT_CASE(RGBA16F, RGBA16F);
1856    FMT_CASE(RG32F, RG32F);
1857    FMT_CASE(RG16F, RG16F);
1858    FMT_CASE(R11F_G11F_B10F, R11G11B10F);
1859    FMT_CASE(R32F, R32F);
1860    FMT_CASE(R16F, R16F);
1861
1862    FMT_CASE(RGBA32UI, RGBA32UI);
1863    FMT_CASE(RGBA16UI, RGBA16UI);
1864    FMT_CASE(RGB10_A2UI, RGB10A2UI);
1865    FMT_CASE(RGBA8UI, RGBA8UI);
1866    FMT_CASE(RG32UI, RG32UI);
1867    FMT_CASE(RG16UI, RG16UI);
1868    FMT_CASE(RG8UI, RG8UI);
1869    FMT_CASE(R32UI, R32UI);
1870    FMT_CASE(R16UI, R16UI);
1871    FMT_CASE(R8UI, R8UI);
1872
1873    FMT_CASE(RGBA32I, RGBA32I);
1874    FMT_CASE(RGBA16I, RGBA16I);
1875    FMT_CASE(RGBA8I, RGBA8I);
1876    FMT_CASE(RG32I, RG32I);
1877    FMT_CASE(RG16I, RG16I);
1878    FMT_CASE(RG8I, RG8I);
1879    FMT_CASE(R32I, R32I);
1880    FMT_CASE(R16I, R16I);
1881    FMT_CASE(R8I, R8I);
1882
1883    FMT_CASE(RGBA16, RGBA16);
1884    FMT_CASE(RGB10_A2, RGB10A2);
1885    FMT_CASE(RGBA8, RGBA8);
1886    FMT_CASE(RG16, RG16);
1887    FMT_CASE(RG8, RG8);
1888    FMT_CASE(R16, R16);
1889    FMT_CASE(R8, R8);
1890
1891    FMT_CASE(RGBA16_SNORM, RGBA16_SNORM);
1892    FMT_CASE(RGBA8_SNORM, RGBA8_SNORM);
1893    FMT_CASE(RG16_SNORM, RG16_SNORM);
1894    FMT_CASE(RG8_SNORM, RG8_SNORM);
1895    FMT_CASE(R16_SNORM, R16_SNORM);
1896    FMT_CASE(R8_SNORM, R8_SNORM);
1897
1898    FMT_CASE(BGRA_INTEGER, BGRA8);
1899    default:
1900       ERROR("unknown format %x\n", format);
1901       assert(false);
1902       return nv50_ir::FMT_NONE;
1903    }
1904 #undef FMT_CASE
1905 }
1906
1907 bool
1908 Converter::visit(nir_intrinsic_instr *insn)
1909 {
1910    nir_intrinsic_op op = insn->intrinsic;
1911    const nir_intrinsic_info &opInfo = nir_intrinsic_infos[op];
1912
1913    switch (op) {
1914    case nir_intrinsic_load_uniform: {
1915       LValues &newDefs = convert(&insn->dest);
1916       const DataType dType = getDType(insn);
1917       Value *indirect;
1918       uint32_t coffset = getIndirect(insn, 0, 0, indirect);
1919       for (uint8_t i = 0; i < insn->num_components; ++i) {
1920          loadFrom(FILE_MEMORY_CONST, 0, dType, newDefs[i], 16 * coffset, i, indirect);
1921       }
1922       break;
1923    }
1924    case nir_intrinsic_store_output:
1925    case nir_intrinsic_store_per_vertex_output: {
1926       Value *indirect;
1927       DataType dType = getSType(insn->src[0], false, false);
1928       uint32_t idx = getIndirect(insn, op == nir_intrinsic_store_output ? 1 : 2, 0, indirect);
1929
1930       for (uint8_t i = 0u; i < insn->num_components; ++i) {
1931          if (!((1u << i) & nir_intrinsic_write_mask(insn)))
1932             continue;
1933
1934          uint8_t offset = 0;
1935          Value *src = getSrc(&insn->src[0], i);
1936          switch (prog->getType()) {
1937          case Program::TYPE_FRAGMENT: {
1938             if (info->out[idx].sn == TGSI_SEMANTIC_POSITION) {
1939                // TGSI uses a different interface than NIR, TGSI stores that
1940                // value in the z component, NIR in X
1941                offset += 2;
1942                src = mkOp1v(OP_SAT, TYPE_F32, getScratch(), src);
1943             }
1944             break;
1945          }
1946          case Program::TYPE_GEOMETRY:
1947          case Program::TYPE_VERTEX: {
1948             if (info->io.genUserClip > 0 && idx == clipVertexOutput) {
1949                mkMov(clipVtx[i], src);
1950                src = clipVtx[i];
1951             }
1952             break;
1953          }
1954          default:
1955             break;
1956          }
1957
1958          storeTo(insn, FILE_SHADER_OUTPUT, OP_EXPORT, dType, src, idx, i + offset, indirect);
1959       }
1960       break;
1961    }
1962    case nir_intrinsic_load_input:
1963    case nir_intrinsic_load_interpolated_input:
1964    case nir_intrinsic_load_output: {
1965       LValues &newDefs = convert(&insn->dest);
1966
1967       // FBFetch
1968       if (prog->getType() == Program::TYPE_FRAGMENT &&
1969           op == nir_intrinsic_load_output) {
1970          std::vector<Value*> defs, srcs;
1971          uint8_t mask = 0;
1972
1973          srcs.push_back(getSSA());
1974          srcs.push_back(getSSA());
1975          Value *x = mkOp1v(OP_RDSV, TYPE_F32, getSSA(), mkSysVal(SV_POSITION, 0));
1976          Value *y = mkOp1v(OP_RDSV, TYPE_F32, getSSA(), mkSysVal(SV_POSITION, 1));
1977          mkCvt(OP_CVT, TYPE_U32, srcs[0], TYPE_F32, x)->rnd = ROUND_Z;
1978          mkCvt(OP_CVT, TYPE_U32, srcs[1], TYPE_F32, y)->rnd = ROUND_Z;
1979
1980          srcs.push_back(mkOp1v(OP_RDSV, TYPE_U32, getSSA(), mkSysVal(SV_LAYER, 0)));
1981          srcs.push_back(mkOp1v(OP_RDSV, TYPE_U32, getSSA(), mkSysVal(SV_SAMPLE_INDEX, 0)));
1982
1983          for (uint8_t i = 0u; i < insn->num_components; ++i) {
1984             defs.push_back(newDefs[i]);
1985             mask |= 1 << i;
1986          }
1987
1988          TexInstruction *texi = mkTex(OP_TXF, TEX_TARGET_2D_MS_ARRAY, 0, 0, defs, srcs);
1989          texi->tex.levelZero = 1;
1990          texi->tex.mask = mask;
1991          texi->tex.useOffsets = 0;
1992          texi->tex.r = 0xffff;
1993          texi->tex.s = 0xffff;
1994
1995          info->prop.fp.readsFramebuffer = true;
1996          break;
1997       }
1998
1999       const DataType dType = getDType(insn);
2000       Value *indirect;
2001       bool input = op != nir_intrinsic_load_output;
2002       operation nvirOp;
2003       uint32_t mode = 0;
2004
2005       uint32_t idx = getIndirect(insn, op == nir_intrinsic_load_interpolated_input ? 1 : 0, 0, indirect);
2006       nv50_ir_varying& vary = input ? info->in[idx] : info->out[idx];
2007
2008       // see load_barycentric_* handling
2009       if (prog->getType() == Program::TYPE_FRAGMENT) {
2010          mode = translateInterpMode(&vary, nvirOp);
2011          if (op == nir_intrinsic_load_interpolated_input) {
2012             ImmediateValue immMode;
2013             if (getSrc(&insn->src[0], 1)->getUniqueInsn()->src(0).getImmediate(immMode))
2014                mode |= immMode.reg.data.u32;
2015          }
2016       }
2017
2018       for (uint8_t i = 0u; i < insn->num_components; ++i) {
2019          uint32_t address = getSlotAddress(insn, idx, i);
2020          Symbol *sym = mkSymbol(input ? FILE_SHADER_INPUT : FILE_SHADER_OUTPUT, 0, dType, address);
2021          if (prog->getType() == Program::TYPE_FRAGMENT) {
2022             int s = 1;
2023             if (typeSizeof(dType) == 8) {
2024                Value *lo = getSSA();
2025                Value *hi = getSSA();
2026                Instruction *interp;
2027
2028                interp = mkOp1(nvirOp, TYPE_U32, lo, sym);
2029                if (nvirOp == OP_PINTERP)
2030                   interp->setSrc(s++, fp.position);
2031                if (mode & NV50_IR_INTERP_OFFSET)
2032                   interp->setSrc(s++, getSrc(&insn->src[0], 0));
2033                interp->setInterpolate(mode);
2034                interp->setIndirect(0, 0, indirect);
2035
2036                Symbol *sym1 = mkSymbol(input ? FILE_SHADER_INPUT : FILE_SHADER_OUTPUT, 0, dType, address + 4);
2037                interp = mkOp1(nvirOp, TYPE_U32, hi, sym1);
2038                if (nvirOp == OP_PINTERP)
2039                   interp->setSrc(s++, fp.position);
2040                if (mode & NV50_IR_INTERP_OFFSET)
2041                   interp->setSrc(s++, getSrc(&insn->src[0], 0));
2042                interp->setInterpolate(mode);
2043                interp->setIndirect(0, 0, indirect);
2044
2045                mkOp2(OP_MERGE, dType, newDefs[i], lo, hi);
2046             } else {
2047                Instruction *interp = mkOp1(nvirOp, dType, newDefs[i], sym);
2048                if (nvirOp == OP_PINTERP)
2049                   interp->setSrc(s++, fp.position);
2050                if (mode & NV50_IR_INTERP_OFFSET)
2051                   interp->setSrc(s++, getSrc(&insn->src[0], 0));
2052                interp->setInterpolate(mode);
2053                interp->setIndirect(0, 0, indirect);
2054             }
2055          } else {
2056             mkLoad(dType, newDefs[i], sym, indirect)->perPatch = vary.patch;
2057          }
2058       }
2059       break;
2060    }
2061    case nir_intrinsic_load_kernel_input: {
2062       assert(prog->getType() == Program::TYPE_COMPUTE);
2063       assert(insn->num_components == 1);
2064
2065       LValues &newDefs = convert(&insn->dest);
2066       const DataType dType = getDType(insn);
2067       Value *indirect;
2068       uint32_t idx = getIndirect(insn, 0, 0, indirect, true);
2069
2070       mkLoad(dType, newDefs[0], mkSymbol(FILE_SHADER_INPUT, 0, dType, idx), indirect);
2071       break;
2072    }
2073    case nir_intrinsic_load_barycentric_at_offset:
2074    case nir_intrinsic_load_barycentric_at_sample:
2075    case nir_intrinsic_load_barycentric_centroid:
2076    case nir_intrinsic_load_barycentric_pixel:
2077    case nir_intrinsic_load_barycentric_sample: {
2078       LValues &newDefs = convert(&insn->dest);
2079       uint32_t mode;
2080
2081       if (op == nir_intrinsic_load_barycentric_centroid ||
2082           op == nir_intrinsic_load_barycentric_sample) {
2083          mode = NV50_IR_INTERP_CENTROID;
2084       } else if (op == nir_intrinsic_load_barycentric_at_offset) {
2085          Value *offs[2];
2086          for (uint8_t c = 0; c < 2; c++) {
2087             offs[c] = getScratch();
2088             mkOp2(OP_MIN, TYPE_F32, offs[c], getSrc(&insn->src[0], c), loadImm(NULL, 0.4375f));
2089             mkOp2(OP_MAX, TYPE_F32, offs[c], offs[c], loadImm(NULL, -0.5f));
2090             mkOp2(OP_MUL, TYPE_F32, offs[c], offs[c], loadImm(NULL, 4096.0f));
2091             mkCvt(OP_CVT, TYPE_S32, offs[c], TYPE_F32, offs[c]);
2092          }
2093          mkOp3v(OP_INSBF, TYPE_U32, newDefs[0], offs[1], mkImm(0x1010), offs[0]);
2094
2095          mode = NV50_IR_INTERP_OFFSET;
2096       } else if (op == nir_intrinsic_load_barycentric_pixel) {
2097          mode = NV50_IR_INTERP_DEFAULT;
2098       } else if (op == nir_intrinsic_load_barycentric_at_sample) {
2099          info->prop.fp.readsSampleLocations = true;
2100          mkOp1(OP_PIXLD, TYPE_U32, newDefs[0], getSrc(&insn->src[0], 0))->subOp = NV50_IR_SUBOP_PIXLD_OFFSET;
2101          mode = NV50_IR_INTERP_OFFSET;
2102       } else {
2103          unreachable("all intrinsics already handled above");
2104       }
2105
2106       loadImm(newDefs[1], mode);
2107       break;
2108    }
2109    case nir_intrinsic_discard:
2110       mkOp(OP_DISCARD, TYPE_NONE, NULL);
2111       break;
2112    case nir_intrinsic_discard_if: {
2113       Value *pred = getSSA(1, FILE_PREDICATE);
2114       if (insn->num_components > 1) {
2115          ERROR("nir_intrinsic_discard_if only with 1 component supported!\n");
2116          assert(false);
2117          return false;
2118       }
2119       mkCmp(OP_SET, CC_NE, TYPE_U8, pred, TYPE_U32, getSrc(&insn->src[0], 0), zero);
2120       mkOp(OP_DISCARD, TYPE_NONE, NULL)->setPredicate(CC_P, pred);
2121       break;
2122    }
2123    case nir_intrinsic_load_base_vertex:
2124    case nir_intrinsic_load_base_instance:
2125    case nir_intrinsic_load_draw_id:
2126    case nir_intrinsic_load_front_face:
2127    case nir_intrinsic_load_helper_invocation:
2128    case nir_intrinsic_load_instance_id:
2129    case nir_intrinsic_load_invocation_id:
2130    case nir_intrinsic_load_local_group_size:
2131    case nir_intrinsic_load_local_invocation_id:
2132    case nir_intrinsic_load_num_work_groups:
2133    case nir_intrinsic_load_patch_vertices_in:
2134    case nir_intrinsic_load_primitive_id:
2135    case nir_intrinsic_load_sample_id:
2136    case nir_intrinsic_load_sample_mask_in:
2137    case nir_intrinsic_load_sample_pos:
2138    case nir_intrinsic_load_subgroup_eq_mask:
2139    case nir_intrinsic_load_subgroup_ge_mask:
2140    case nir_intrinsic_load_subgroup_gt_mask:
2141    case nir_intrinsic_load_subgroup_le_mask:
2142    case nir_intrinsic_load_subgroup_lt_mask:
2143    case nir_intrinsic_load_subgroup_invocation:
2144    case nir_intrinsic_load_tess_coord:
2145    case nir_intrinsic_load_tess_level_inner:
2146    case nir_intrinsic_load_tess_level_outer:
2147    case nir_intrinsic_load_vertex_id:
2148    case nir_intrinsic_load_work_group_id: {
2149       const DataType dType = getDType(insn);
2150       SVSemantic sv = convert(op);
2151       LValues &newDefs = convert(&insn->dest);
2152
2153       for (uint8_t i = 0u; i < insn->num_components; ++i) {
2154          Value *def;
2155          if (typeSizeof(dType) == 8)
2156             def = getSSA();
2157          else
2158             def = newDefs[i];
2159
2160          if (sv == SV_TID && info->prop.cp.numThreads[i] == 1) {
2161             loadImm(def, 0u);
2162          } else {
2163             Symbol *sym = mkSysVal(sv, i);
2164             Instruction *rdsv = mkOp1(OP_RDSV, TYPE_U32, def, sym);
2165             if (sv == SV_TESS_OUTER || sv == SV_TESS_INNER)
2166                rdsv->perPatch = 1;
2167          }
2168
2169          if (typeSizeof(dType) == 8)
2170             mkOp2(OP_MERGE, dType, newDefs[i], def, loadImm(getSSA(), 0u));
2171       }
2172       break;
2173    }
2174    // constants
2175    case nir_intrinsic_load_subgroup_size: {
2176       LValues &newDefs = convert(&insn->dest);
2177       loadImm(newDefs[0], 32u);
2178       break;
2179    }
2180    case nir_intrinsic_vote_all:
2181    case nir_intrinsic_vote_any:
2182    case nir_intrinsic_vote_ieq: {
2183       LValues &newDefs = convert(&insn->dest);
2184       Value *pred = getScratch(1, FILE_PREDICATE);
2185       mkCmp(OP_SET, CC_NE, TYPE_U32, pred, TYPE_U32, getSrc(&insn->src[0], 0), zero);
2186       mkOp1(OP_VOTE, TYPE_U32, pred, pred)->subOp = getSubOp(op);
2187       mkCvt(OP_CVT, TYPE_U32, newDefs[0], TYPE_U8, pred);
2188       break;
2189    }
2190    case nir_intrinsic_ballot: {
2191       LValues &newDefs = convert(&insn->dest);
2192       Value *pred = getSSA(1, FILE_PREDICATE);
2193       mkCmp(OP_SET, CC_NE, TYPE_U32, pred, TYPE_U32, getSrc(&insn->src[0], 0), zero);
2194       mkOp1(OP_VOTE, TYPE_U32, newDefs[0], pred)->subOp = NV50_IR_SUBOP_VOTE_ANY;
2195       break;
2196    }
2197    case nir_intrinsic_read_first_invocation:
2198    case nir_intrinsic_read_invocation: {
2199       LValues &newDefs = convert(&insn->dest);
2200       const DataType dType = getDType(insn);
2201       Value *tmp = getScratch();
2202
2203       if (op == nir_intrinsic_read_first_invocation) {
2204          mkOp1(OP_VOTE, TYPE_U32, tmp, mkImm(1))->subOp = NV50_IR_SUBOP_VOTE_ANY;
2205          mkOp2(OP_EXTBF, TYPE_U32, tmp, tmp, mkImm(0x2000))->subOp = NV50_IR_SUBOP_EXTBF_REV;
2206          mkOp1(OP_BFIND, TYPE_U32, tmp, tmp)->subOp = NV50_IR_SUBOP_BFIND_SAMT;
2207       } else
2208          tmp = getSrc(&insn->src[1], 0);
2209
2210       for (uint8_t i = 0; i < insn->num_components; ++i) {
2211          mkOp3(OP_SHFL, dType, newDefs[i], getSrc(&insn->src[0], i), tmp, mkImm(0x1f))
2212             ->subOp = NV50_IR_SUBOP_SHFL_IDX;
2213       }
2214       break;
2215    }
2216    case nir_intrinsic_load_per_vertex_input: {
2217       const DataType dType = getDType(insn);
2218       LValues &newDefs = convert(&insn->dest);
2219       Value *indirectVertex;
2220       Value *indirectOffset;
2221       uint32_t baseVertex = getIndirect(&insn->src[0], 0, indirectVertex);
2222       uint32_t idx = getIndirect(insn, 1, 0, indirectOffset);
2223
2224       Value *vtxBase = mkOp2v(OP_PFETCH, TYPE_U32, getSSA(4, FILE_ADDRESS),
2225                               mkImm(baseVertex), indirectVertex);
2226       for (uint8_t i = 0u; i < insn->num_components; ++i) {
2227          uint32_t address = getSlotAddress(insn, idx, i);
2228          loadFrom(FILE_SHADER_INPUT, 0, dType, newDefs[i], address, 0,
2229                   indirectOffset, vtxBase, info->in[idx].patch);
2230       }
2231       break;
2232    }
2233    case nir_intrinsic_load_per_vertex_output: {
2234       const DataType dType = getDType(insn);
2235       LValues &newDefs = convert(&insn->dest);
2236       Value *indirectVertex;
2237       Value *indirectOffset;
2238       uint32_t baseVertex = getIndirect(&insn->src[0], 0, indirectVertex);
2239       uint32_t idx = getIndirect(insn, 1, 0, indirectOffset);
2240       Value *vtxBase = NULL;
2241
2242       if (indirectVertex)
2243          vtxBase = indirectVertex;
2244       else
2245          vtxBase = loadImm(NULL, baseVertex);
2246
2247       vtxBase = mkOp2v(OP_ADD, TYPE_U32, getSSA(4, FILE_ADDRESS), outBase, vtxBase);
2248
2249       for (uint8_t i = 0u; i < insn->num_components; ++i) {
2250          uint32_t address = getSlotAddress(insn, idx, i);
2251          loadFrom(FILE_SHADER_OUTPUT, 0, dType, newDefs[i], address, 0,
2252                   indirectOffset, vtxBase, info->in[idx].patch);
2253       }
2254       break;
2255    }
2256    case nir_intrinsic_emit_vertex:
2257       if (info->io.genUserClip > 0)
2258          handleUserClipPlanes();
2259       // fallthrough
2260    case nir_intrinsic_end_primitive: {
2261       uint32_t idx = nir_intrinsic_stream_id(insn);
2262       mkOp1(getOperation(op), TYPE_U32, NULL, mkImm(idx))->fixed = 1;
2263       break;
2264    }
2265    case nir_intrinsic_load_ubo: {
2266       const DataType dType = getDType(insn);
2267       LValues &newDefs = convert(&insn->dest);
2268       Value *indirectIndex;
2269       Value *indirectOffset;
2270       uint32_t index = getIndirect(&insn->src[0], 0, indirectIndex) + 1;
2271       uint32_t offset = getIndirect(&insn->src[1], 0, indirectOffset);
2272
2273       for (uint8_t i = 0u; i < insn->num_components; ++i) {
2274          loadFrom(FILE_MEMORY_CONST, index, dType, newDefs[i], offset, i,
2275                   indirectOffset, indirectIndex);
2276       }
2277       break;
2278    }
2279    case nir_intrinsic_get_buffer_size: {
2280       LValues &newDefs = convert(&insn->dest);
2281       const DataType dType = getDType(insn);
2282       Value *indirectBuffer;
2283       uint32_t buffer = getIndirect(&insn->src[0], 0, indirectBuffer);
2284
2285       Symbol *sym = mkSymbol(FILE_MEMORY_BUFFER, buffer, dType, 0);
2286       mkOp1(OP_BUFQ, dType, newDefs[0], sym)->setIndirect(0, 0, indirectBuffer);
2287       break;
2288    }
2289    case nir_intrinsic_store_ssbo: {
2290       DataType sType = getSType(insn->src[0], false, false);
2291       Value *indirectBuffer;
2292       Value *indirectOffset;
2293       uint32_t buffer = getIndirect(&insn->src[1], 0, indirectBuffer);
2294       uint32_t offset = getIndirect(&insn->src[2], 0, indirectOffset);
2295
2296       for (uint8_t i = 0u; i < insn->num_components; ++i) {
2297          if (!((1u << i) & nir_intrinsic_write_mask(insn)))
2298             continue;
2299          Symbol *sym = mkSymbol(FILE_MEMORY_BUFFER, buffer, sType,
2300                                 offset + i * typeSizeof(sType));
2301          mkStore(OP_STORE, sType, sym, indirectOffset, getSrc(&insn->src[0], i))
2302             ->setIndirect(0, 1, indirectBuffer);
2303       }
2304       info->io.globalAccess |= 0x2;
2305       break;
2306    }
2307    case nir_intrinsic_load_ssbo: {
2308       const DataType dType = getDType(insn);
2309       LValues &newDefs = convert(&insn->dest);
2310       Value *indirectBuffer;
2311       Value *indirectOffset;
2312       uint32_t buffer = getIndirect(&insn->src[0], 0, indirectBuffer);
2313       uint32_t offset = getIndirect(&insn->src[1], 0, indirectOffset);
2314
2315       for (uint8_t i = 0u; i < insn->num_components; ++i)
2316          loadFrom(FILE_MEMORY_BUFFER, buffer, dType, newDefs[i], offset, i,
2317                   indirectOffset, indirectBuffer);
2318
2319       info->io.globalAccess |= 0x1;
2320       break;
2321    }
2322    case nir_intrinsic_shared_atomic_add:
2323    case nir_intrinsic_shared_atomic_and:
2324    case nir_intrinsic_shared_atomic_comp_swap:
2325    case nir_intrinsic_shared_atomic_exchange:
2326    case nir_intrinsic_shared_atomic_or:
2327    case nir_intrinsic_shared_atomic_imax:
2328    case nir_intrinsic_shared_atomic_imin:
2329    case nir_intrinsic_shared_atomic_umax:
2330    case nir_intrinsic_shared_atomic_umin:
2331    case nir_intrinsic_shared_atomic_xor: {
2332       const DataType dType = getDType(insn);
2333       LValues &newDefs = convert(&insn->dest);
2334       Value *indirectOffset;
2335       uint32_t offset = getIndirect(&insn->src[0], 0, indirectOffset);
2336       Symbol *sym = mkSymbol(FILE_MEMORY_SHARED, 0, dType, offset);
2337       Instruction *atom = mkOp2(OP_ATOM, dType, newDefs[0], sym, getSrc(&insn->src[1], 0));
2338       if (op == nir_intrinsic_shared_atomic_comp_swap)
2339          atom->setSrc(2, getSrc(&insn->src[2], 0));
2340       atom->setIndirect(0, 0, indirectOffset);
2341       atom->subOp = getSubOp(op);
2342       break;
2343    }
2344    case nir_intrinsic_ssbo_atomic_add:
2345    case nir_intrinsic_ssbo_atomic_and:
2346    case nir_intrinsic_ssbo_atomic_comp_swap:
2347    case nir_intrinsic_ssbo_atomic_exchange:
2348    case nir_intrinsic_ssbo_atomic_or:
2349    case nir_intrinsic_ssbo_atomic_imax:
2350    case nir_intrinsic_ssbo_atomic_imin:
2351    case nir_intrinsic_ssbo_atomic_umax:
2352    case nir_intrinsic_ssbo_atomic_umin:
2353    case nir_intrinsic_ssbo_atomic_xor: {
2354       const DataType dType = getDType(insn);
2355       LValues &newDefs = convert(&insn->dest);
2356       Value *indirectBuffer;
2357       Value *indirectOffset;
2358       uint32_t buffer = getIndirect(&insn->src[0], 0, indirectBuffer);
2359       uint32_t offset = getIndirect(&insn->src[1], 0, indirectOffset);
2360
2361       Symbol *sym = mkSymbol(FILE_MEMORY_BUFFER, buffer, dType, offset);
2362       Instruction *atom = mkOp2(OP_ATOM, dType, newDefs[0], sym,
2363                                 getSrc(&insn->src[2], 0));
2364       if (op == nir_intrinsic_ssbo_atomic_comp_swap)
2365          atom->setSrc(2, getSrc(&insn->src[3], 0));
2366       atom->setIndirect(0, 0, indirectOffset);
2367       atom->setIndirect(0, 1, indirectBuffer);
2368       atom->subOp = getSubOp(op);
2369
2370       info->io.globalAccess |= 0x2;
2371       break;
2372    }
2373    case nir_intrinsic_bindless_image_atomic_add:
2374    case nir_intrinsic_bindless_image_atomic_and:
2375    case nir_intrinsic_bindless_image_atomic_comp_swap:
2376    case nir_intrinsic_bindless_image_atomic_exchange:
2377    case nir_intrinsic_bindless_image_atomic_max:
2378    case nir_intrinsic_bindless_image_atomic_min:
2379    case nir_intrinsic_bindless_image_atomic_or:
2380    case nir_intrinsic_bindless_image_atomic_xor:
2381    case nir_intrinsic_bindless_image_load:
2382    case nir_intrinsic_bindless_image_samples:
2383    case nir_intrinsic_bindless_image_size:
2384    case nir_intrinsic_bindless_image_store: {
2385       std::vector<Value*> srcs, defs;
2386       Value *indirect = getSrc(&insn->src[0], 0);
2387       DataType ty;
2388
2389       uint32_t mask = 0;
2390       TexInstruction::Target target =
2391          convert(nir_intrinsic_image_dim(insn), !!nir_intrinsic_image_array(insn), false);
2392       unsigned int argCount = getNIRArgCount(target);
2393       uint16_t location = 0;
2394
2395       if (opInfo.has_dest) {
2396          LValues &newDefs = convert(&insn->dest);
2397          for (uint8_t i = 0u; i < newDefs.size(); ++i) {
2398             defs.push_back(newDefs[i]);
2399             mask |= 1 << i;
2400          }
2401       }
2402
2403       switch (op) {
2404       case nir_intrinsic_bindless_image_atomic_add:
2405       case nir_intrinsic_bindless_image_atomic_and:
2406       case nir_intrinsic_bindless_image_atomic_comp_swap:
2407       case nir_intrinsic_bindless_image_atomic_exchange:
2408       case nir_intrinsic_bindless_image_atomic_max:
2409       case nir_intrinsic_bindless_image_atomic_min:
2410       case nir_intrinsic_bindless_image_atomic_or:
2411       case nir_intrinsic_bindless_image_atomic_xor:
2412          ty = getDType(insn);
2413          mask = 0x1;
2414          info->io.globalAccess |= 0x2;
2415          break;
2416       case nir_intrinsic_bindless_image_load:
2417          ty = TYPE_U32;
2418          info->io.globalAccess |= 0x1;
2419          break;
2420       case nir_intrinsic_bindless_image_store:
2421          ty = TYPE_U32;
2422          mask = 0xf;
2423          info->io.globalAccess |= 0x2;
2424          break;
2425       case nir_intrinsic_bindless_image_samples:
2426          mask = 0x8;
2427          ty = TYPE_U32;
2428          break;
2429       case nir_intrinsic_bindless_image_size:
2430          ty = TYPE_U32;
2431          break;
2432       default:
2433          unreachable("unhandled image opcode");
2434          break;
2435       }
2436
2437       // coords
2438       if (opInfo.num_srcs >= 2)
2439          for (unsigned int i = 0u; i < argCount; ++i)
2440             srcs.push_back(getSrc(&insn->src[1], i));
2441
2442       // the sampler is just another src added after coords
2443       if (opInfo.num_srcs >= 3 && target.isMS())
2444          srcs.push_back(getSrc(&insn->src[2], 0));
2445
2446       if (opInfo.num_srcs >= 4) {
2447          unsigned components = opInfo.src_components[3] ? opInfo.src_components[3] : insn->num_components;
2448          for (uint8_t i = 0u; i < components; ++i)
2449             srcs.push_back(getSrc(&insn->src[3], i));
2450       }
2451
2452       if (opInfo.num_srcs >= 5)
2453          // 1 for aotmic swap
2454          for (uint8_t i = 0u; i < opInfo.src_components[4]; ++i)
2455             srcs.push_back(getSrc(&insn->src[4], i));
2456
2457       TexInstruction *texi = mkTex(getOperation(op), target.getEnum(), location, 0, defs, srcs);
2458       texi->tex.bindless = false;
2459       texi->tex.format = &nv50_ir::TexInstruction::formatTable[convertGLImgFormat(nir_intrinsic_format(insn))];
2460       texi->tex.mask = mask;
2461       texi->tex.bindless = true;
2462       texi->cache = convert(nir_intrinsic_access(insn));
2463       texi->setType(ty);
2464       texi->subOp = getSubOp(op);
2465
2466       if (indirect)
2467          texi->setIndirectR(indirect);
2468
2469       break;
2470    }
2471    case nir_intrinsic_image_deref_atomic_add:
2472    case nir_intrinsic_image_deref_atomic_and:
2473    case nir_intrinsic_image_deref_atomic_comp_swap:
2474    case nir_intrinsic_image_deref_atomic_exchange:
2475    case nir_intrinsic_image_deref_atomic_max:
2476    case nir_intrinsic_image_deref_atomic_min:
2477    case nir_intrinsic_image_deref_atomic_or:
2478    case nir_intrinsic_image_deref_atomic_xor:
2479    case nir_intrinsic_image_deref_load:
2480    case nir_intrinsic_image_deref_samples:
2481    case nir_intrinsic_image_deref_size:
2482    case nir_intrinsic_image_deref_store: {
2483       const nir_variable *tex;
2484       std::vector<Value*> srcs, defs;
2485       Value *indirect;
2486       DataType ty;
2487
2488       uint32_t mask = 0;
2489       nir_deref_instr *deref = nir_src_as_deref(insn->src[0]);
2490       const glsl_type *type = deref->type;
2491       TexInstruction::Target target =
2492          convert((glsl_sampler_dim)type->sampler_dimensionality,
2493                  type->sampler_array, type->sampler_shadow);
2494       unsigned int argCount = getNIRArgCount(target);
2495       uint16_t location = handleDeref(deref, indirect, tex);
2496
2497       if (opInfo.has_dest) {
2498          LValues &newDefs = convert(&insn->dest);
2499          for (uint8_t i = 0u; i < newDefs.size(); ++i) {
2500             defs.push_back(newDefs[i]);
2501             mask |= 1 << i;
2502          }
2503       }
2504
2505       switch (op) {
2506       case nir_intrinsic_image_deref_atomic_add:
2507       case nir_intrinsic_image_deref_atomic_and:
2508       case nir_intrinsic_image_deref_atomic_comp_swap:
2509       case nir_intrinsic_image_deref_atomic_exchange:
2510       case nir_intrinsic_image_deref_atomic_max:
2511       case nir_intrinsic_image_deref_atomic_min:
2512       case nir_intrinsic_image_deref_atomic_or:
2513       case nir_intrinsic_image_deref_atomic_xor:
2514          ty = getDType(insn);
2515          mask = 0x1;
2516          info->io.globalAccess |= 0x2;
2517          break;
2518       case nir_intrinsic_image_deref_load:
2519          ty = TYPE_U32;
2520          info->io.globalAccess |= 0x1;
2521          break;
2522       case nir_intrinsic_image_deref_store:
2523          ty = TYPE_U32;
2524          mask = 0xf;
2525          info->io.globalAccess |= 0x2;
2526          break;
2527       case nir_intrinsic_image_deref_samples:
2528          mask = 0x8;
2529          ty = TYPE_U32;
2530          break;
2531       case nir_intrinsic_image_deref_size:
2532          ty = TYPE_U32;
2533          break;
2534       default:
2535          unreachable("unhandled image opcode");
2536          break;
2537       }
2538
2539       // coords
2540       if (opInfo.num_srcs >= 2)
2541          for (unsigned int i = 0u; i < argCount; ++i)
2542             srcs.push_back(getSrc(&insn->src[1], i));
2543
2544       // the sampler is just another src added after coords
2545       if (opInfo.num_srcs >= 3 && target.isMS())
2546          srcs.push_back(getSrc(&insn->src[2], 0));
2547
2548       if (opInfo.num_srcs >= 4) {
2549          unsigned components = opInfo.src_components[3] ? opInfo.src_components[3] : insn->num_components;
2550          for (uint8_t i = 0u; i < components; ++i)
2551             srcs.push_back(getSrc(&insn->src[3], i));
2552       }
2553
2554       if (opInfo.num_srcs >= 5)
2555          // 1 for aotmic swap
2556          for (uint8_t i = 0u; i < opInfo.src_components[4]; ++i)
2557             srcs.push_back(getSrc(&insn->src[4], i));
2558
2559       TexInstruction *texi = mkTex(getOperation(op), target.getEnum(), location, 0, defs, srcs);
2560       texi->tex.bindless = false;
2561       texi->tex.format = &nv50_ir::TexInstruction::formatTable[convertGLImgFormat(tex->data.image.format)];
2562       texi->tex.mask = mask;
2563       texi->cache = getCacheModeFromVar(tex);
2564       texi->setType(ty);
2565       texi->subOp = getSubOp(op);
2566
2567       if (indirect)
2568          texi->setIndirectR(indirect);
2569
2570       break;
2571    }
2572    case nir_intrinsic_store_shared: {
2573       DataType sType = getSType(insn->src[0], false, false);
2574       Value *indirectOffset;
2575       uint32_t offset = getIndirect(&insn->src[1], 0, indirectOffset);
2576
2577       for (uint8_t i = 0u; i < insn->num_components; ++i) {
2578          if (!((1u << i) & nir_intrinsic_write_mask(insn)))
2579             continue;
2580          Symbol *sym = mkSymbol(FILE_MEMORY_SHARED, 0, sType, offset + i * typeSizeof(sType));
2581          mkStore(OP_STORE, sType, sym, indirectOffset, getSrc(&insn->src[0], i));
2582       }
2583       break;
2584    }
2585    case nir_intrinsic_load_shared: {
2586       const DataType dType = getDType(insn);
2587       LValues &newDefs = convert(&insn->dest);
2588       Value *indirectOffset;
2589       uint32_t offset = getIndirect(&insn->src[0], 0, indirectOffset);
2590
2591       for (uint8_t i = 0u; i < insn->num_components; ++i)
2592          loadFrom(FILE_MEMORY_SHARED, 0, dType, newDefs[i], offset, i, indirectOffset);
2593
2594       break;
2595    }
2596    case nir_intrinsic_barrier: {
2597       // TODO: add flag to shader_info
2598       info->numBarriers = 1;
2599       Instruction *bar = mkOp2(OP_BAR, TYPE_U32, NULL, mkImm(0), mkImm(0));
2600       bar->fixed = 1;
2601       bar->subOp = NV50_IR_SUBOP_BAR_SYNC;
2602       break;
2603    }
2604    case nir_intrinsic_group_memory_barrier:
2605    case nir_intrinsic_memory_barrier:
2606    case nir_intrinsic_memory_barrier_atomic_counter:
2607    case nir_intrinsic_memory_barrier_buffer:
2608    case nir_intrinsic_memory_barrier_image:
2609    case nir_intrinsic_memory_barrier_shared: {
2610       Instruction *bar = mkOp(OP_MEMBAR, TYPE_NONE, NULL);
2611       bar->fixed = 1;
2612       bar->subOp = getSubOp(op);
2613       break;
2614    }
2615    case nir_intrinsic_shader_clock: {
2616       const DataType dType = getDType(insn);
2617       LValues &newDefs = convert(&insn->dest);
2618
2619       loadImm(newDefs[0], 0u);
2620       mkOp1(OP_RDSV, dType, newDefs[1], mkSysVal(SV_CLOCK, 0))->fixed = 1;
2621       break;
2622    }
2623    case nir_intrinsic_load_global: {
2624       const DataType dType = getDType(insn);
2625       LValues &newDefs = convert(&insn->dest);
2626       Value *indirectOffset;
2627       uint32_t offset = getIndirect(&insn->src[0], 0, indirectOffset);
2628
2629       for (auto i = 0u; i < insn->num_components; ++i)
2630          loadFrom(FILE_MEMORY_GLOBAL, 0, dType, newDefs[i], offset, i, indirectOffset);
2631
2632       info->io.globalAccess |= 0x1;
2633       break;
2634    }
2635    case nir_intrinsic_store_global: {
2636       DataType sType = getSType(insn->src[0], false, false);
2637
2638       for (auto i = 0u; i < insn->num_components; ++i) {
2639          if (!((1u << i) & nir_intrinsic_write_mask(insn)))
2640             continue;
2641          if (typeSizeof(sType) == 8) {
2642             Value *split[2];
2643             mkSplit(split, 4, getSrc(&insn->src[0], i));
2644
2645             Symbol *sym = mkSymbol(FILE_MEMORY_GLOBAL, 0, TYPE_U32, i * typeSizeof(sType));
2646             mkStore(OP_STORE, TYPE_U32, sym, getSrc(&insn->src[1], 0), split[0]);
2647
2648             sym = mkSymbol(FILE_MEMORY_GLOBAL, 0, TYPE_U32, i * typeSizeof(sType) + 4);
2649             mkStore(OP_STORE, TYPE_U32, sym, getSrc(&insn->src[1], 0), split[1]);
2650          } else {
2651             Symbol *sym = mkSymbol(FILE_MEMORY_GLOBAL, 0, sType, i * typeSizeof(sType));
2652             mkStore(OP_STORE, sType, sym, getSrc(&insn->src[1], 0), getSrc(&insn->src[0], i));
2653          }
2654       }
2655
2656       info->io.globalAccess |= 0x2;
2657       break;
2658    }
2659    default:
2660       ERROR("unknown nir_intrinsic_op %s\n", nir_intrinsic_infos[op].name);
2661       return false;
2662    }
2663
2664    return true;
2665 }
2666
2667 bool
2668 Converter::visit(nir_jump_instr *insn)
2669 {
2670    switch (insn->type) {
2671    case nir_jump_return:
2672       // TODO: this only works in the main function
2673       mkFlow(OP_BRA, exit, CC_ALWAYS, NULL);
2674       bb->cfg.attach(&exit->cfg, Graph::Edge::CROSS);
2675       break;
2676    case nir_jump_break:
2677    case nir_jump_continue: {
2678       bool isBreak = insn->type == nir_jump_break;
2679       nir_block *block = insn->instr.block;
2680       assert(!block->successors[1]);
2681       BasicBlock *target = convert(block->successors[0]);
2682       mkFlow(isBreak ? OP_BREAK : OP_CONT, target, CC_ALWAYS, NULL);
2683       bb->cfg.attach(&target->cfg, isBreak ? Graph::Edge::CROSS : Graph::Edge::BACK);
2684       break;
2685    }
2686    default:
2687       ERROR("unknown nir_jump_type %u\n", insn->type);
2688       return false;
2689    }
2690
2691    return true;
2692 }
2693
2694 Value*
2695 Converter::convert(nir_load_const_instr *insn, uint8_t idx)
2696 {
2697    Value *val;
2698
2699    if (immInsertPos)
2700       setPosition(immInsertPos, true);
2701    else
2702       setPosition(bb, false);
2703
2704    switch (insn->def.bit_size) {
2705    case 64:
2706       val = loadImm(getSSA(8), insn->value[idx].u64);
2707       break;
2708    case 32:
2709       val = loadImm(getSSA(4), insn->value[idx].u32);
2710       break;
2711    case 16:
2712       val = loadImm(getSSA(2), insn->value[idx].u16);
2713       break;
2714    case 8:
2715       val = loadImm(getSSA(1), insn->value[idx].u8);
2716       break;
2717    default:
2718       unreachable("unhandled bit size!\n");
2719    }
2720    setPosition(bb, true);
2721    return val;
2722 }
2723
2724 bool
2725 Converter::visit(nir_load_const_instr *insn)
2726 {
2727    assert(insn->def.bit_size <= 64);
2728    immediates[insn->def.index] = insn;
2729    return true;
2730 }
2731
2732 #define DEFAULT_CHECKS \
2733       if (insn->dest.dest.ssa.num_components > 1) { \
2734          ERROR("nir_alu_instr only supported with 1 component!\n"); \
2735          return false; \
2736       } \
2737       if (insn->dest.write_mask != 1) { \
2738          ERROR("nir_alu_instr only with write_mask of 1 supported!\n"); \
2739          return false; \
2740       }
2741 bool
2742 Converter::visit(nir_alu_instr *insn)
2743 {
2744    const nir_op op = insn->op;
2745    const nir_op_info &info = nir_op_infos[op];
2746    DataType dType = getDType(insn);
2747    const std::vector<DataType> sTypes = getSTypes(insn);
2748
2749    Instruction *oldPos = this->bb->getExit();
2750
2751    switch (op) {
2752    case nir_op_fabs:
2753    case nir_op_iabs:
2754    case nir_op_fadd:
2755    case nir_op_iadd:
2756    case nir_op_iand:
2757    case nir_op_fceil:
2758    case nir_op_fcos:
2759    case nir_op_fddx:
2760    case nir_op_fddx_coarse:
2761    case nir_op_fddx_fine:
2762    case nir_op_fddy:
2763    case nir_op_fddy_coarse:
2764    case nir_op_fddy_fine:
2765    case nir_op_fdiv:
2766    case nir_op_idiv:
2767    case nir_op_udiv:
2768    case nir_op_fexp2:
2769    case nir_op_ffloor:
2770    case nir_op_ffma:
2771    case nir_op_flog2:
2772    case nir_op_fmax:
2773    case nir_op_imax:
2774    case nir_op_umax:
2775    case nir_op_fmin:
2776    case nir_op_imin:
2777    case nir_op_umin:
2778    case nir_op_fmod:
2779    case nir_op_imod:
2780    case nir_op_umod:
2781    case nir_op_fmul:
2782    case nir_op_imul:
2783    case nir_op_imul_high:
2784    case nir_op_umul_high:
2785    case nir_op_fneg:
2786    case nir_op_ineg:
2787    case nir_op_inot:
2788    case nir_op_ior:
2789    case nir_op_pack_64_2x32_split:
2790    case nir_op_fpow:
2791    case nir_op_frcp:
2792    case nir_op_frem:
2793    case nir_op_irem:
2794    case nir_op_frsq:
2795    case nir_op_fsat:
2796    case nir_op_ishr:
2797    case nir_op_ushr:
2798    case nir_op_fsin:
2799    case nir_op_fsqrt:
2800    case nir_op_fsub:
2801    case nir_op_isub:
2802    case nir_op_ftrunc:
2803    case nir_op_ishl:
2804    case nir_op_ixor: {
2805       DEFAULT_CHECKS;
2806       LValues &newDefs = convert(&insn->dest);
2807       operation preOp = preOperationNeeded(op);
2808       if (preOp != OP_NOP) {
2809          assert(info.num_inputs < 2);
2810          Value *tmp = getSSA(typeSizeof(dType));
2811          Instruction *i0 = mkOp(preOp, dType, tmp);
2812          Instruction *i1 = mkOp(getOperation(op), dType, newDefs[0]);
2813          if (info.num_inputs) {
2814             i0->setSrc(0, getSrc(&insn->src[0]));
2815             i1->setSrc(0, tmp);
2816          }
2817          i1->subOp = getSubOp(op);
2818       } else {
2819          Instruction *i = mkOp(getOperation(op), dType, newDefs[0]);
2820          for (unsigned s = 0u; s < info.num_inputs; ++s) {
2821             i->setSrc(s, getSrc(&insn->src[s]));
2822          }
2823          i->subOp = getSubOp(op);
2824       }
2825       break;
2826    }
2827    case nir_op_ifind_msb:
2828    case nir_op_ufind_msb: {
2829       DEFAULT_CHECKS;
2830       LValues &newDefs = convert(&insn->dest);
2831       dType = sTypes[0];
2832       mkOp1(getOperation(op), dType, newDefs[0], getSrc(&insn->src[0]));
2833       break;
2834    }
2835    case nir_op_fround_even: {
2836       DEFAULT_CHECKS;
2837       LValues &newDefs = convert(&insn->dest);
2838       mkCvt(OP_CVT, dType, newDefs[0], dType, getSrc(&insn->src[0]))->rnd = ROUND_NI;
2839       break;
2840    }
2841    // convert instructions
2842    case nir_op_f2f32:
2843    case nir_op_f2i32:
2844    case nir_op_f2u32:
2845    case nir_op_i2f32:
2846    case nir_op_i2i32:
2847    case nir_op_u2f32:
2848    case nir_op_u2u32:
2849    case nir_op_f2f64:
2850    case nir_op_f2i64:
2851    case nir_op_f2u64:
2852    case nir_op_i2f64:
2853    case nir_op_i2i64:
2854    case nir_op_u2f64:
2855    case nir_op_u2u64: {
2856       DEFAULT_CHECKS;
2857       LValues &newDefs = convert(&insn->dest);
2858       Instruction *i = mkOp1(getOperation(op), dType, newDefs[0], getSrc(&insn->src[0]));
2859       if (op == nir_op_f2i32 || op == nir_op_f2i64 || op == nir_op_f2u32 || op == nir_op_f2u64)
2860          i->rnd = ROUND_Z;
2861       i->sType = sTypes[0];
2862       break;
2863    }
2864    // compare instructions
2865    case nir_op_feq32:
2866    case nir_op_ieq32:
2867    case nir_op_fge32:
2868    case nir_op_ige32:
2869    case nir_op_uge32:
2870    case nir_op_flt32:
2871    case nir_op_ilt32:
2872    case nir_op_ult32:
2873    case nir_op_fne32:
2874    case nir_op_ine32: {
2875       DEFAULT_CHECKS;
2876       LValues &newDefs = convert(&insn->dest);
2877       Instruction *i = mkCmp(getOperation(op),
2878                              getCondCode(op),
2879                              dType,
2880                              newDefs[0],
2881                              dType,
2882                              getSrc(&insn->src[0]),
2883                              getSrc(&insn->src[1]));
2884       if (info.num_inputs == 3)
2885          i->setSrc(2, getSrc(&insn->src[2]));
2886       i->sType = sTypes[0];
2887       break;
2888    }
2889    // those are weird ALU ops and need special handling, because
2890    //   1. they are always componend based
2891    //   2. they basically just merge multiple values into one data type
2892    case nir_op_mov:
2893       if (!insn->dest.dest.is_ssa && insn->dest.dest.reg.reg->num_array_elems) {
2894          nir_reg_dest& reg = insn->dest.dest.reg;
2895          uint32_t goffset = regToLmemOffset[reg.reg->index];
2896          uint8_t comps = reg.reg->num_components;
2897          uint8_t size = reg.reg->bit_size / 8;
2898          uint8_t csize = 4 * size; // TODO after fixing MemoryOpts: comps * size;
2899          uint32_t aoffset = csize * reg.base_offset;
2900          Value *indirect = NULL;
2901
2902          if (reg.indirect)
2903             indirect = mkOp2v(OP_MUL, TYPE_U32, getSSA(4, FILE_ADDRESS),
2904                               getSrc(reg.indirect, 0), mkImm(csize));
2905
2906          for (uint8_t i = 0u; i < comps; ++i) {
2907             if (!((1u << i) & insn->dest.write_mask))
2908                continue;
2909
2910             Symbol *sym = mkSymbol(FILE_MEMORY_LOCAL, 0, dType, goffset + aoffset + i * size);
2911             mkStore(OP_STORE, dType, sym, indirect, getSrc(&insn->src[0], i));
2912          }
2913          break;
2914       } else if (!insn->src[0].src.is_ssa && insn->src[0].src.reg.reg->num_array_elems) {
2915          LValues &newDefs = convert(&insn->dest);
2916          nir_reg_src& reg = insn->src[0].src.reg;
2917          uint32_t goffset = regToLmemOffset[reg.reg->index];
2918          // uint8_t comps = reg.reg->num_components;
2919          uint8_t size = reg.reg->bit_size / 8;
2920          uint8_t csize = 4 * size; // TODO after fixing MemoryOpts: comps * size;
2921          uint32_t aoffset = csize * reg.base_offset;
2922          Value *indirect = NULL;
2923
2924          if (reg.indirect)
2925             indirect = mkOp2v(OP_MUL, TYPE_U32, getSSA(4, FILE_ADDRESS), getSrc(reg.indirect, 0), mkImm(csize));
2926
2927          for (uint8_t i = 0u; i < newDefs.size(); ++i)
2928             loadFrom(FILE_MEMORY_LOCAL, 0, dType, newDefs[i], goffset + aoffset, i, indirect);
2929
2930          break;
2931       } else {
2932          LValues &newDefs = convert(&insn->dest);
2933          for (LValues::size_type c = 0u; c < newDefs.size(); ++c) {
2934             mkMov(newDefs[c], getSrc(&insn->src[0], c), dType);
2935          }
2936       }
2937       break;
2938    case nir_op_vec2:
2939    case nir_op_vec3:
2940    case nir_op_vec4: {
2941       LValues &newDefs = convert(&insn->dest);
2942       for (LValues::size_type c = 0u; c < newDefs.size(); ++c) {
2943          mkMov(newDefs[c], getSrc(&insn->src[c]), dType);
2944       }
2945       break;
2946    }
2947    // (un)pack
2948    case nir_op_pack_64_2x32: {
2949       LValues &newDefs = convert(&insn->dest);
2950       Instruction *merge = mkOp(OP_MERGE, dType, newDefs[0]);
2951       merge->setSrc(0, getSrc(&insn->src[0], 0));
2952       merge->setSrc(1, getSrc(&insn->src[0], 1));
2953       break;
2954    }
2955    case nir_op_pack_half_2x16_split: {
2956       LValues &newDefs = convert(&insn->dest);
2957       Value *tmpH = getSSA();
2958       Value *tmpL = getSSA();
2959
2960       mkCvt(OP_CVT, TYPE_F16, tmpL, TYPE_F32, getSrc(&insn->src[0]));
2961       mkCvt(OP_CVT, TYPE_F16, tmpH, TYPE_F32, getSrc(&insn->src[1]));
2962       mkOp3(OP_INSBF, TYPE_U32, newDefs[0], tmpH, mkImm(0x1010), tmpL);
2963       break;
2964    }
2965    case nir_op_unpack_half_2x16_split_x:
2966    case nir_op_unpack_half_2x16_split_y: {
2967       LValues &newDefs = convert(&insn->dest);
2968       Instruction *cvt = mkCvt(OP_CVT, TYPE_F32, newDefs[0], TYPE_F16, getSrc(&insn->src[0]));
2969       if (op == nir_op_unpack_half_2x16_split_y)
2970          cvt->subOp = 1;
2971       break;
2972    }
2973    case nir_op_unpack_64_2x32: {
2974       LValues &newDefs = convert(&insn->dest);
2975       mkOp1(OP_SPLIT, dType, newDefs[0], getSrc(&insn->src[0]))->setDef(1, newDefs[1]);
2976       break;
2977    }
2978    case nir_op_unpack_64_2x32_split_x: {
2979       LValues &newDefs = convert(&insn->dest);
2980       mkOp1(OP_SPLIT, dType, newDefs[0], getSrc(&insn->src[0]))->setDef(1, getSSA());
2981       break;
2982    }
2983    case nir_op_unpack_64_2x32_split_y: {
2984       LValues &newDefs = convert(&insn->dest);
2985       mkOp1(OP_SPLIT, dType, getSSA(), getSrc(&insn->src[0]))->setDef(1, newDefs[0]);
2986       break;
2987    }
2988    // special instructions
2989    case nir_op_fsign:
2990    case nir_op_isign: {
2991       DEFAULT_CHECKS;
2992       DataType iType;
2993       if (::isFloatType(dType))
2994          iType = TYPE_F32;
2995       else
2996          iType = TYPE_S32;
2997
2998       LValues &newDefs = convert(&insn->dest);
2999       LValue *val0 = getScratch();
3000       LValue *val1 = getScratch();
3001       mkCmp(OP_SET, CC_GT, iType, val0, dType, getSrc(&insn->src[0]), zero);
3002       mkCmp(OP_SET, CC_LT, iType, val1, dType, getSrc(&insn->src[0]), zero);
3003
3004       if (dType == TYPE_F64) {
3005          mkOp2(OP_SUB, iType, val0, val0, val1);
3006          mkCvt(OP_CVT, TYPE_F64, newDefs[0], iType, val0);
3007       } else if (dType == TYPE_S64 || dType == TYPE_U64) {
3008          mkOp2(OP_SUB, iType, val0, val1, val0);
3009          mkOp2(OP_SHR, iType, val1, val0, loadImm(NULL, 31));
3010          mkOp2(OP_MERGE, dType, newDefs[0], val0, val1);
3011       } else if (::isFloatType(dType))
3012          mkOp2(OP_SUB, iType, newDefs[0], val0, val1);
3013       else
3014          mkOp2(OP_SUB, iType, newDefs[0], val1, val0);
3015       break;
3016    }
3017    case nir_op_fcsel:
3018    case nir_op_b32csel: {
3019       DEFAULT_CHECKS;
3020       LValues &newDefs = convert(&insn->dest);
3021       mkCmp(OP_SLCT, CC_NE, dType, newDefs[0], sTypes[0], getSrc(&insn->src[1]), getSrc(&insn->src[2]), getSrc(&insn->src[0]));
3022       break;
3023    }
3024    case nir_op_ibitfield_extract:
3025    case nir_op_ubitfield_extract: {
3026       DEFAULT_CHECKS;
3027       Value *tmp = getSSA();
3028       LValues &newDefs = convert(&insn->dest);
3029       mkOp3(OP_INSBF, dType, tmp, getSrc(&insn->src[2]), loadImm(NULL, 0x808), getSrc(&insn->src[1]));
3030       mkOp2(OP_EXTBF, dType, newDefs[0], getSrc(&insn->src[0]), tmp);
3031       break;
3032    }
3033    case nir_op_bfm: {
3034       DEFAULT_CHECKS;
3035       LValues &newDefs = convert(&insn->dest);
3036       mkOp3(OP_INSBF, dType, newDefs[0], getSrc(&insn->src[0]), loadImm(NULL, 0x808), getSrc(&insn->src[1]));
3037       break;
3038    }
3039    case nir_op_bitfield_insert: {
3040       DEFAULT_CHECKS;
3041       LValues &newDefs = convert(&insn->dest);
3042       LValue *temp = getSSA();
3043       mkOp3(OP_INSBF, TYPE_U32, temp, getSrc(&insn->src[3]), mkImm(0x808), getSrc(&insn->src[2]));
3044       mkOp3(OP_INSBF, dType, newDefs[0], getSrc(&insn->src[1]), temp, getSrc(&insn->src[0]));
3045       break;
3046    }
3047    case nir_op_bit_count: {
3048       DEFAULT_CHECKS;
3049       LValues &newDefs = convert(&insn->dest);
3050       mkOp2(OP_POPCNT, dType, newDefs[0], getSrc(&insn->src[0]), getSrc(&insn->src[0]));
3051       break;
3052    }
3053    case nir_op_bitfield_reverse: {
3054       DEFAULT_CHECKS;
3055       LValues &newDefs = convert(&insn->dest);
3056       mkOp2(OP_EXTBF, TYPE_U32, newDefs[0], getSrc(&insn->src[0]), mkImm(0x2000))->subOp = NV50_IR_SUBOP_EXTBF_REV;
3057       break;
3058    }
3059    case nir_op_find_lsb: {
3060       DEFAULT_CHECKS;
3061       LValues &newDefs = convert(&insn->dest);
3062       Value *tmp = getSSA();
3063       mkOp2(OP_EXTBF, TYPE_U32, tmp, getSrc(&insn->src[0]), mkImm(0x2000))->subOp = NV50_IR_SUBOP_EXTBF_REV;
3064       mkOp1(OP_BFIND, TYPE_U32, newDefs[0], tmp)->subOp = NV50_IR_SUBOP_BFIND_SAMT;
3065       break;
3066    }
3067    // boolean conversions
3068    case nir_op_b2f32: {
3069       DEFAULT_CHECKS;
3070       LValues &newDefs = convert(&insn->dest);
3071       mkOp2(OP_AND, TYPE_U32, newDefs[0], getSrc(&insn->src[0]), loadImm(NULL, 1.0f));
3072       break;
3073    }
3074    case nir_op_b2f64: {
3075       DEFAULT_CHECKS;
3076       LValues &newDefs = convert(&insn->dest);
3077       Value *tmp = getSSA(4);
3078       mkOp2(OP_AND, TYPE_U32, tmp, getSrc(&insn->src[0]), loadImm(NULL, 0x3ff00000));
3079       mkOp2(OP_MERGE, TYPE_U64, newDefs[0], loadImm(NULL, 0), tmp);
3080       break;
3081    }
3082    case nir_op_f2b32:
3083    case nir_op_i2b32: {
3084       DEFAULT_CHECKS;
3085       LValues &newDefs = convert(&insn->dest);
3086       Value *src1;
3087       if (typeSizeof(sTypes[0]) == 8) {
3088          src1 = loadImm(getSSA(8), 0.0);
3089       } else {
3090          src1 = zero;
3091       }
3092       CondCode cc = op == nir_op_f2b32 ? CC_NEU : CC_NE;
3093       mkCmp(OP_SET, cc, TYPE_U32, newDefs[0], sTypes[0], getSrc(&insn->src[0]), src1);
3094       break;
3095    }
3096    case nir_op_b2i32: {
3097       DEFAULT_CHECKS;
3098       LValues &newDefs = convert(&insn->dest);
3099       mkOp2(OP_AND, TYPE_U32, newDefs[0], getSrc(&insn->src[0]), loadImm(NULL, 1));
3100       break;
3101    }
3102    case nir_op_b2i64: {
3103       DEFAULT_CHECKS;
3104       LValues &newDefs = convert(&insn->dest);
3105       LValue *def = getScratch();
3106       mkOp2(OP_AND, TYPE_U32, def, getSrc(&insn->src[0]), loadImm(NULL, 1));
3107       mkOp2(OP_MERGE, TYPE_S64, newDefs[0], def, loadImm(NULL, 0));
3108       break;
3109    }
3110    default:
3111       ERROR("unknown nir_op %s\n", info.name);
3112       return false;
3113    }
3114
3115    if (!oldPos) {
3116       oldPos = this->bb->getEntry();
3117       oldPos->precise = insn->exact;
3118    }
3119
3120    if (unlikely(!oldPos))
3121       return true;
3122
3123    while (oldPos->next) {
3124       oldPos = oldPos->next;
3125       oldPos->precise = insn->exact;
3126    }
3127    oldPos->saturate = insn->dest.saturate;
3128
3129    return true;
3130 }
3131 #undef DEFAULT_CHECKS
3132
3133 bool
3134 Converter::visit(nir_ssa_undef_instr *insn)
3135 {
3136    LValues &newDefs = convert(&insn->def);
3137    for (uint8_t i = 0u; i < insn->def.num_components; ++i) {
3138       mkOp(OP_NOP, TYPE_NONE, newDefs[i]);
3139    }
3140    return true;
3141 }
3142
3143 #define CASE_SAMPLER(ty) \
3144    case GLSL_SAMPLER_DIM_ ## ty : \
3145       if (isArray && !isShadow) \
3146          return TEX_TARGET_ ## ty ## _ARRAY; \
3147       else if (!isArray && isShadow) \
3148          return TEX_TARGET_## ty ## _SHADOW; \
3149       else if (isArray && isShadow) \
3150          return TEX_TARGET_## ty ## _ARRAY_SHADOW; \
3151       else \
3152          return TEX_TARGET_ ## ty
3153
3154 TexTarget
3155 Converter::convert(glsl_sampler_dim dim, bool isArray, bool isShadow)
3156 {
3157    switch (dim) {
3158    CASE_SAMPLER(1D);
3159    CASE_SAMPLER(2D);
3160    CASE_SAMPLER(CUBE);
3161    case GLSL_SAMPLER_DIM_3D:
3162       return TEX_TARGET_3D;
3163    case GLSL_SAMPLER_DIM_MS:
3164       if (isArray)
3165          return TEX_TARGET_2D_MS_ARRAY;
3166       return TEX_TARGET_2D_MS;
3167    case GLSL_SAMPLER_DIM_RECT:
3168       if (isShadow)
3169          return TEX_TARGET_RECT_SHADOW;
3170       return TEX_TARGET_RECT;
3171    case GLSL_SAMPLER_DIM_BUF:
3172       return TEX_TARGET_BUFFER;
3173    case GLSL_SAMPLER_DIM_EXTERNAL:
3174       return TEX_TARGET_2D;
3175    default:
3176       ERROR("unknown glsl_sampler_dim %u\n", dim);
3177       assert(false);
3178       return TEX_TARGET_COUNT;
3179    }
3180 }
3181 #undef CASE_SAMPLER
3182
3183 Value*
3184 Converter::applyProjection(Value *src, Value *proj)
3185 {
3186    if (!proj)
3187       return src;
3188    return mkOp2v(OP_MUL, TYPE_F32, getScratch(), src, proj);
3189 }
3190
3191 unsigned int
3192 Converter::getNIRArgCount(TexInstruction::Target& target)
3193 {
3194    unsigned int result = target.getArgCount();
3195    if (target.isCube() && target.isArray())
3196       result--;
3197    if (target.isMS())
3198       result--;
3199    return result;
3200 }
3201
3202 uint16_t
3203 Converter::handleDeref(nir_deref_instr *deref, Value * &indirect, const nir_variable * &tex)
3204 {
3205    typedef std::pair<uint32_t,Value*> DerefPair;
3206    std::list<DerefPair> derefs;
3207
3208    uint16_t result = 0;
3209    while (deref->deref_type != nir_deref_type_var) {
3210       switch (deref->deref_type) {
3211       case nir_deref_type_array: {
3212          Value *indirect;
3213          uint8_t size = type_size(deref->type, true);
3214          result += size * getIndirect(&deref->arr.index, 0, indirect);
3215
3216          if (indirect) {
3217             derefs.push_front(std::make_pair(size, indirect));
3218          }
3219
3220          break;
3221       }
3222       case nir_deref_type_struct: {
3223          result += nir_deref_instr_parent(deref)->type->struct_location_offset(deref->strct.index);
3224          break;
3225       }
3226       case nir_deref_type_var:
3227       default:
3228          unreachable("nir_deref_type_var reached in handleDeref!");
3229          break;
3230       }
3231       deref = nir_deref_instr_parent(deref);
3232    }
3233
3234    indirect = NULL;
3235    for (std::list<DerefPair>::const_iterator it = derefs.begin(); it != derefs.end(); ++it) {
3236       Value *offset = mkOp2v(OP_MUL, TYPE_U32, getSSA(), loadImm(getSSA(), it->first), it->second);
3237       if (indirect)
3238          indirect = mkOp2v(OP_ADD, TYPE_U32, getSSA(), indirect, offset);
3239       else
3240          indirect = offset;
3241    }
3242
3243    tex = nir_deref_instr_get_variable(deref);
3244    assert(tex);
3245
3246    return result + tex->data.driver_location;
3247 }
3248
3249 CacheMode
3250 Converter::convert(enum gl_access_qualifier access)
3251 {
3252    switch (access) {
3253    case ACCESS_VOLATILE:
3254       return CACHE_CV;
3255    case ACCESS_COHERENT:
3256       return CACHE_CG;
3257    default:
3258       return CACHE_CA;
3259    }
3260 }
3261
3262 CacheMode
3263 Converter::getCacheModeFromVar(const nir_variable *var)
3264 {
3265    return convert(var->data.image.access);
3266 }
3267
3268 bool
3269 Converter::visit(nir_tex_instr *insn)
3270 {
3271    switch (insn->op) {
3272    case nir_texop_lod:
3273    case nir_texop_query_levels:
3274    case nir_texop_tex:
3275    case nir_texop_texture_samples:
3276    case nir_texop_tg4:
3277    case nir_texop_txb:
3278    case nir_texop_txd:
3279    case nir_texop_txf:
3280    case nir_texop_txf_ms:
3281    case nir_texop_txl:
3282    case nir_texop_txs: {
3283       LValues &newDefs = convert(&insn->dest);
3284       std::vector<Value*> srcs;
3285       std::vector<Value*> defs;
3286       std::vector<nir_src*> offsets;
3287       uint8_t mask = 0;
3288       bool lz = false;
3289       Value *proj = NULL;
3290       TexInstruction::Target target = convert(insn->sampler_dim, insn->is_array, insn->is_shadow);
3291       operation op = getOperation(insn->op);
3292
3293       int r, s;
3294       int biasIdx = nir_tex_instr_src_index(insn, nir_tex_src_bias);
3295       int compIdx = nir_tex_instr_src_index(insn, nir_tex_src_comparator);
3296       int coordsIdx = nir_tex_instr_src_index(insn, nir_tex_src_coord);
3297       int ddxIdx = nir_tex_instr_src_index(insn, nir_tex_src_ddx);
3298       int ddyIdx = nir_tex_instr_src_index(insn, nir_tex_src_ddy);
3299       int msIdx = nir_tex_instr_src_index(insn, nir_tex_src_ms_index);
3300       int lodIdx = nir_tex_instr_src_index(insn, nir_tex_src_lod);
3301       int offsetIdx = nir_tex_instr_src_index(insn, nir_tex_src_offset);
3302       int projIdx = nir_tex_instr_src_index(insn, nir_tex_src_projector);
3303       int sampOffIdx = nir_tex_instr_src_index(insn, nir_tex_src_sampler_offset);
3304       int texOffIdx = nir_tex_instr_src_index(insn, nir_tex_src_texture_offset);
3305       int sampHandleIdx = nir_tex_instr_src_index(insn, nir_tex_src_sampler_handle);
3306       int texHandleIdx = nir_tex_instr_src_index(insn, nir_tex_src_texture_handle);
3307
3308       bool bindless = sampHandleIdx != -1 || texHandleIdx != -1;
3309       assert((sampHandleIdx != -1) == (texHandleIdx != -1));
3310
3311       if (projIdx != -1)
3312          proj = mkOp1v(OP_RCP, TYPE_F32, getScratch(), getSrc(&insn->src[projIdx].src, 0));
3313
3314       srcs.resize(insn->coord_components);
3315       for (uint8_t i = 0u; i < insn->coord_components; ++i)
3316          srcs[i] = applyProjection(getSrc(&insn->src[coordsIdx].src, i), proj);
3317
3318       // sometimes we get less args than target.getArgCount, but codegen expects the latter
3319       if (insn->coord_components) {
3320          uint32_t argCount = target.getArgCount();
3321
3322          if (target.isMS())
3323             argCount -= 1;
3324
3325          for (uint32_t i = 0u; i < (argCount - insn->coord_components); ++i)
3326             srcs.push_back(getSSA());
3327       }
3328
3329       if (insn->op == nir_texop_texture_samples)
3330          srcs.push_back(zero);
3331       else if (!insn->num_srcs)
3332          srcs.push_back(loadImm(NULL, 0));
3333       if (biasIdx != -1)
3334          srcs.push_back(getSrc(&insn->src[biasIdx].src, 0));
3335       if (lodIdx != -1)
3336          srcs.push_back(getSrc(&insn->src[lodIdx].src, 0));
3337       else if (op == OP_TXF)
3338          lz = true;
3339       if (msIdx != -1)
3340          srcs.push_back(getSrc(&insn->src[msIdx].src, 0));
3341       if (offsetIdx != -1)
3342          offsets.push_back(&insn->src[offsetIdx].src);
3343       if (compIdx != -1)
3344          srcs.push_back(applyProjection(getSrc(&insn->src[compIdx].src, 0), proj));
3345       if (texOffIdx != -1) {
3346          srcs.push_back(getSrc(&insn->src[texOffIdx].src, 0));
3347          texOffIdx = srcs.size() - 1;
3348       }
3349       if (sampOffIdx != -1) {
3350          srcs.push_back(getSrc(&insn->src[sampOffIdx].src, 0));
3351          sampOffIdx = srcs.size() - 1;
3352       }
3353       if (bindless) {
3354          // currently we use the lower bits
3355          Value *split[2];
3356          Value *handle = getSrc(&insn->src[sampHandleIdx].src, 0);
3357
3358          mkSplit(split, 4, handle);
3359
3360          srcs.push_back(split[0]);
3361          texOffIdx = srcs.size() - 1;
3362       }
3363
3364       r = bindless ? 0xff : insn->texture_index;
3365       s = bindless ? 0x1f : insn->sampler_index;
3366
3367       defs.resize(newDefs.size());
3368       for (uint8_t d = 0u; d < newDefs.size(); ++d) {
3369          defs[d] = newDefs[d];
3370          mask |= 1 << d;
3371       }
3372       if (target.isMS() || (op == OP_TEX && prog->getType() != Program::TYPE_FRAGMENT))
3373          lz = true;
3374
3375       TexInstruction *texi = mkTex(op, target.getEnum(), r, s, defs, srcs);
3376       texi->tex.levelZero = lz;
3377       texi->tex.mask = mask;
3378       texi->tex.bindless = bindless;
3379
3380       if (texOffIdx != -1)
3381          texi->tex.rIndirectSrc = texOffIdx;
3382       if (sampOffIdx != -1)
3383          texi->tex.sIndirectSrc = sampOffIdx;
3384
3385       switch (insn->op) {
3386       case nir_texop_tg4:
3387          if (!target.isShadow())
3388             texi->tex.gatherComp = insn->component;
3389          break;
3390       case nir_texop_txs:
3391          texi->tex.query = TXQ_DIMS;
3392          break;
3393       case nir_texop_texture_samples:
3394          texi->tex.mask = 0x4;
3395          texi->tex.query = TXQ_TYPE;
3396          break;
3397       case nir_texop_query_levels:
3398          texi->tex.mask = 0x8;
3399          texi->tex.query = TXQ_DIMS;
3400          break;
3401       default:
3402          break;
3403       }
3404
3405       texi->tex.useOffsets = offsets.size();
3406       if (texi->tex.useOffsets) {
3407          for (uint8_t s = 0; s < texi->tex.useOffsets; ++s) {
3408             for (uint32_t c = 0u; c < 3; ++c) {
3409                uint8_t s2 = std::min(c, target.getDim() - 1);
3410                texi->offset[s][c].set(getSrc(offsets[s], s2));
3411                texi->offset[s][c].setInsn(texi);
3412             }
3413          }
3414       }
3415
3416       if (op == OP_TXG && offsetIdx == -1) {
3417          if (nir_tex_instr_has_explicit_tg4_offsets(insn)) {
3418             texi->tex.useOffsets = 4;
3419             setPosition(texi, false);
3420             for (uint8_t i = 0; i < 4; ++i) {
3421                for (uint8_t j = 0; j < 2; ++j) {
3422                   texi->offset[i][j].set(loadImm(NULL, insn->tg4_offsets[i][j]));
3423                   texi->offset[i][j].setInsn(texi);
3424                }
3425             }
3426             setPosition(texi, true);
3427          }
3428       }
3429
3430       if (ddxIdx != -1 && ddyIdx != -1) {
3431          for (uint8_t c = 0u; c < target.getDim() + target.isCube(); ++c) {
3432             texi->dPdx[c].set(getSrc(&insn->src[ddxIdx].src, c));
3433             texi->dPdy[c].set(getSrc(&insn->src[ddyIdx].src, c));
3434          }
3435       }
3436
3437       break;
3438    }
3439    default:
3440       ERROR("unknown nir_texop %u\n", insn->op);
3441       return false;
3442    }
3443    return true;
3444 }
3445
3446 bool
3447 Converter::visit(nir_deref_instr *deref)
3448 {
3449    // we just ignore those, because images intrinsics are the only place where
3450    // we should end up with deref sources and those have to backtrack anyway
3451    // to get the nir_variable. This code just exists to handle some special
3452    // cases.
3453    switch (deref->deref_type) {
3454    case nir_deref_type_array:
3455    case nir_deref_type_struct:
3456    case nir_deref_type_var:
3457       break;
3458    default:
3459       ERROR("unknown nir_deref_instr %u\n", deref->deref_type);
3460       return false;
3461    }
3462    return true;
3463 }
3464
3465 bool
3466 Converter::run()
3467 {
3468    bool progress;
3469
3470    if (prog->dbgFlags & NV50_IR_DEBUG_VERBOSE)
3471       nir_print_shader(nir, stderr);
3472
3473    struct nir_lower_subgroups_options subgroup_options = {
3474       .subgroup_size = 32,
3475       .ballot_bit_size = 32,
3476    };
3477
3478    NIR_PASS_V(nir, nir_lower_io, nir_var_all, type_size, (nir_lower_io_options)0);
3479    NIR_PASS_V(nir, nir_lower_subgroups, &subgroup_options);
3480    NIR_PASS_V(nir, nir_lower_regs_to_ssa);
3481    NIR_PASS_V(nir, nir_lower_load_const_to_scalar);
3482    NIR_PASS_V(nir, nir_lower_vars_to_ssa);
3483    NIR_PASS_V(nir, nir_lower_alu_to_scalar, NULL);
3484    NIR_PASS_V(nir, nir_lower_phis_to_scalar);
3485
3486    do {
3487       progress = false;
3488       NIR_PASS(progress, nir, nir_copy_prop);
3489       NIR_PASS(progress, nir, nir_opt_remove_phis);
3490       NIR_PASS(progress, nir, nir_opt_trivial_continues);
3491       NIR_PASS(progress, nir, nir_opt_cse);
3492       NIR_PASS(progress, nir, nir_opt_algebraic);
3493       NIR_PASS(progress, nir, nir_opt_constant_folding);
3494       NIR_PASS(progress, nir, nir_copy_prop);
3495       NIR_PASS(progress, nir, nir_opt_dce);
3496       NIR_PASS(progress, nir, nir_opt_dead_cf);
3497    } while (progress);
3498
3499    NIR_PASS_V(nir, nir_lower_bool_to_int32);
3500    NIR_PASS_V(nir, nir_lower_locals_to_regs);
3501    NIR_PASS_V(nir, nir_remove_dead_variables, nir_var_function_temp);
3502    NIR_PASS_V(nir, nir_convert_from_ssa, true);
3503
3504    // Garbage collect dead instructions
3505    nir_sweep(nir);
3506
3507    if (!parseNIR()) {
3508       ERROR("Couldn't prase NIR!\n");
3509       return false;
3510    }
3511
3512    if (!assignSlots()) {
3513       ERROR("Couldn't assign slots!\n");
3514       return false;
3515    }
3516
3517    if (prog->dbgFlags & NV50_IR_DEBUG_BASIC)
3518       nir_print_shader(nir, stderr);
3519
3520    nir_foreach_function(function, nir) {
3521       if (!visit(function))
3522          return false;
3523    }
3524
3525    return true;
3526 }
3527
3528 } // unnamed namespace
3529
3530 namespace nv50_ir {
3531
3532 bool
3533 Program::makeFromNIR(struct nv50_ir_prog_info *info)
3534 {
3535    nir_shader *nir = (nir_shader*)info->bin.source;
3536    Converter converter(this, nir, info);
3537    bool result = converter.run();
3538    if (!result)
3539       return result;
3540    LoweringHelper lowering;
3541    lowering.run(this);
3542    tlsSize = info->bin.tlsSpace;
3543    return result;
3544 }
3545
3546 } // namespace nv50_ir