src/amd/compiler/aco_builder_h.py

   1
   2 template = """\
   3 /*
   4  * Copyright (c) 2019 Valve Corporation
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the "Software"),
   8  * to deal in the Software without restriction, including without limitation
   9  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  10  * and/or sell copies of the Software, and to permit persons to whom the
  11  * Software is furnished to do so, subject to the following conditions:
  12  *
  13  * The above copyright notice and this permission notice (including the next
  14  * paragraph) shall be included in all copies or substantial portions of the
  15  * Software.
  16  *
  17  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  18  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  19  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  20  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  21  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  22  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  23  * IN THE SOFTWARE.
  24  *
  25  * This file was generated by aco_builder_h.py
  26  */
  27
  28 #ifndef _ACO_BUILDER_
  29 #define _ACO_BUILDER_
  30
  31 #include "aco_ir.h"
  32 #include "util/u_math.h"
  33 #include "util/bitscan.h"
  34
  35 namespace aco {
  36 enum dpp_ctrl {
  37     _dpp_quad_perm = 0x000,
  38     _dpp_row_sl = 0x100,
  39     _dpp_row_sr = 0x110,
  40     _dpp_row_rr = 0x120,
  41     dpp_wf_sl1 = 0x130,
  42     dpp_wf_rl1 = 0x134,
  43     dpp_wf_sr1 = 0x138,
  44     dpp_wf_rr1 = 0x13C,
  45     dpp_row_mirror = 0x140,
  46     dpp_row_half_mirror = 0x141,
  47     dpp_row_bcast15 = 0x142,
  48     dpp_row_bcast31 = 0x143
  49 };
  50
  51 inline dpp_ctrl
  52 dpp_quad_perm(unsigned lane0, unsigned lane1, unsigned lane2, unsigned lane3)
  53 {
  54     assert(lane0 < 4 && lane1 < 4 && lane2 < 4 && lane3 < 4);
  55     return (dpp_ctrl)(lane0 | (lane1 << 2) | (lane2 << 4) | (lane3 << 6));
  56 }
  57
  58 inline dpp_ctrl
  59 dpp_row_sl(unsigned amount)
  60 {
  61     assert(amount > 0 && amount < 16);
  62     return (dpp_ctrl)(((unsigned) _dpp_row_sl) | amount);
  63 }
  64
  65 inline dpp_ctrl
  66 dpp_row_sr(unsigned amount)
  67 {
  68     assert(amount > 0 && amount < 16);
  69     return (dpp_ctrl)(((unsigned) _dpp_row_sr) | amount);
  70 }
  71
  72 inline unsigned
  73 ds_pattern_bitmode(unsigned and_mask, unsigned or_mask, unsigned xor_mask)
  74 {
  75     assert(and_mask < 32 && or_mask < 32 && xor_mask < 32);
  76     return and_mask | (or_mask << 5) | (xor_mask << 10);
  77 }
  78
  79 aco_ptr<Instruction> create_s_mov(Definition dst, Operand src);
  80
  81 enum sendmsg {
  82    sendmsg_none = 0,
  83    _sendmsg_gs = 2,
  84    _sendmsg_gs_done = 3,
  85    sendmsg_save_wave = 4,
  86    sendmsg_stall_wave_gen = 5,
  87    sendmsg_halt_waves = 6,
  88    sendmsg_ordered_ps_done = 7,
  89    sendmsg_early_prim_dealloc = 8,
  90    sendmsg_gs_alloc_req = 9,
  91    sendmsg_id_mask = 0xf,
  92 };
  93
  94 inline sendmsg
  95 sendmsg_gs(bool cut, bool emit, unsigned stream)
  96 {
  97     assert(stream < 4);
  98     return (sendmsg)((unsigned)_sendmsg_gs | (cut << 4) | (emit << 5) | (stream << 8));
  99 }
 100
 101 inline sendmsg
 102 sendmsg_gs_done(bool cut, bool emit, unsigned stream)
 103 {
 104     assert(stream < 4);
 105     return (sendmsg)((unsigned)_sendmsg_gs_done | (cut << 4) | (emit << 5) | (stream << 8));
 106 }
 107
 108 class Builder {
 109 public:
 110    struct Result {
 111       Instruction *instr;
 112
 113       Result(Instruction *instr) : instr(instr) {}
 114
 115       operator Instruction *() const {
 116          return instr;
 117       }
 118
 119       operator Temp() const {
 120          return instr->definitions[0].getTemp();
 121       }
 122
 123       operator Operand() const {
 124          return Operand((Temp)*this);
 125       }
 126
 127       Definition& def(unsigned index) const {
 128          return instr->definitions[index];
 129       }
 130
 131       aco_ptr<Instruction> get_ptr() const {
 132         return aco_ptr<Instruction>(instr);
 133       }
 134    };
 135
 136    struct Op {
 137       Operand op;
 138       Op(Temp tmp) : op(tmp) {}
 139       Op(Operand op_) : op(op_) {}
 140       Op(Result res) : op((Temp)res) {}
 141    };
 142
 143    enum WaveSpecificOpcode {
 144       s_cselect = (unsigned) aco_opcode::s_cselect_b64,
 145       s_cmp_lg = (unsigned) aco_opcode::s_cmp_lg_u64,
 146       s_and = (unsigned) aco_opcode::s_and_b64,
 147       s_andn2 = (unsigned) aco_opcode::s_andn2_b64,
 148       s_or = (unsigned) aco_opcode::s_or_b64,
 149       s_orn2 = (unsigned) aco_opcode::s_orn2_b64,
 150       s_not = (unsigned) aco_opcode::s_not_b64,
 151       s_mov = (unsigned) aco_opcode::s_mov_b64,
 152       s_wqm = (unsigned) aco_opcode::s_wqm_b64,
 153       s_and_saveexec = (unsigned) aco_opcode::s_and_saveexec_b64,
 154       s_or_saveexec = (unsigned) aco_opcode::s_or_saveexec_b64,
 155       s_xnor = (unsigned) aco_opcode::s_xnor_b64,
 156       s_xor = (unsigned) aco_opcode::s_xor_b64,
 157       s_bcnt1_i32 = (unsigned) aco_opcode::s_bcnt1_i32_b64,
 158       s_bitcmp1 = (unsigned) aco_opcode::s_bitcmp1_b64,
 159       s_ff1_i32 = (unsigned) aco_opcode::s_ff1_i32_b64,
 160    };
 161
 162    Program *program;
 163    bool use_iterator;
 164    bool start; // only when use_iterator == false
 165    RegClass lm;
 166
 167    std::vector<aco_ptr<Instruction>> *instructions;
 168    std::vector<aco_ptr<Instruction>>::iterator it;
 169
 170    Builder(Program *pgm) : program(pgm), use_iterator(false), start(false), lm(pgm->lane_mask), instructions(NULL) {}
 171    Builder(Program *pgm, Block *block) : program(pgm), use_iterator(false), start(false), lm(pgm ? pgm->lane_mask : s2), instructions(&block->instructions) {}
 172    Builder(Program *pgm, std::vector<aco_ptr<Instruction>> *instrs) : program(pgm), use_iterator(false), start(false), lm(pgm ? pgm->lane_mask : s2), instructions(instrs) {}
 173
 174    void moveEnd(Block *block) {
 175       instructions = &block->instructions;
 176    }
 177
 178    void reset() {
 179       use_iterator = false;
 180       start = false;
 181       instructions = NULL;
 182    }
 183
 184    void reset(Block *block) {
 185       use_iterator = false;
 186       start = false;
 187       instructions = &block->instructions;
 188    }
 189
 190    void reset(std::vector<aco_ptr<Instruction>> *instrs) {
 191       use_iterator = false;
 192       start = false;
 193       instructions = instrs;
 194    }
 195
 196    void reset(std::vector<aco_ptr<Instruction>> *instrs, std::vector<aco_ptr<Instruction>>::iterator instr_it) {
 197       use_iterator = true;
 198       start = false;
 199       instructions = instrs;
 200       it = instr_it;
 201    }
 202
 203    Result insert(aco_ptr<Instruction> instr) {
 204       Instruction *instr_ptr = instr.get();
 205       if (instructions) {
 206          if (use_iterator) {
 207             it = instructions->emplace(it, std::move(instr));
 208             it = std::next(it);
 209          } else if (!start) {
 210             instructions->emplace_back(std::move(instr));
 211          } else {
 212             instructions->emplace(instructions->begin(), std::move(instr));
 213          }
 214       }
 215       return Result(instr_ptr);
 216    }
 217
 218    Result insert(Instruction* instr) {
 219       if (instructions) {
 220          if (use_iterator) {
 221             it = instructions->emplace(it, aco_ptr<Instruction>(instr));
 222             it = std::next(it);
 223          } else if (!start) {
 224             instructions->emplace_back(aco_ptr<Instruction>(instr));
 225          } else {
 226             instructions->emplace(instructions->begin(), aco_ptr<Instruction>(instr));
 227          }
 228       }
 229       return Result(instr);
 230    }
 231
 232    Temp tmp(RegClass rc) {
 233       return (Temp){program->allocateId(), rc};
 234    }
 235
 236    Temp tmp(RegType type, unsigned size) {
 237       return (Temp){program->allocateId(), RegClass(type, size)};
 238    }
 239
 240    Definition def(RegClass rc) {
 241       return Definition((Temp){program->allocateId(), rc});
 242    }
 243
 244    Definition def(RegType type, unsigned size) {
 245       return Definition((Temp){program->allocateId(), RegClass(type, size)});
 246    }
 247
 248    Definition def(RegClass rc, PhysReg reg) {
 249       return Definition(program->allocateId(), reg, rc);
 250    }
 251
 252    inline aco_opcode w64or32(WaveSpecificOpcode opcode) const {
 253       if (program->wave_size == 64)
 254          return (aco_opcode) opcode;
 255
 256       switch (opcode) {
 257       case s_cselect:
 258          return aco_opcode::s_cselect_b32;
 259       case s_cmp_lg:
 260          return aco_opcode::s_cmp_lg_u32;
 261       case s_and:
 262          return aco_opcode::s_and_b32;
 263       case s_andn2:
 264          return aco_opcode::s_andn2_b32;
 265       case s_or:
 266          return aco_opcode::s_or_b32;
 267       case s_orn2:
 268          return aco_opcode::s_orn2_b32;
 269       case s_not:
 270          return aco_opcode::s_not_b32;
 271       case s_mov:
 272          return aco_opcode::s_mov_b32;
 273       case s_wqm:
 274          return aco_opcode::s_wqm_b32;
 275       case s_and_saveexec:
 276          return aco_opcode::s_and_saveexec_b32;
 277       case s_or_saveexec:
 278          return aco_opcode::s_or_saveexec_b32;
 279       case s_xnor:
 280          return aco_opcode::s_xnor_b32;
 281       case s_xor:
 282          return aco_opcode::s_xor_b32;
 283       case s_bcnt1_i32:
 284          return aco_opcode::s_bcnt1_i32_b32;
 285       case s_bitcmp1:
 286          return aco_opcode::s_bitcmp1_b32;
 287       case s_ff1_i32:
 288          return aco_opcode::s_ff1_i32_b32;
 289       default:
 290          unreachable("Unsupported wave specific opcode.");
 291       }
 292    }
 293
 294 % for fixed in ['m0', 'vcc', 'exec', 'scc']:
 295    Operand ${fixed}(Temp tmp) {
 296        % if fixed == 'vcc' or fixed == 'exec':
 297           assert(tmp.regClass() == lm);
 298        % endif
 299        Operand op(tmp);
 300        op.setFixed(aco::${fixed});
 301        return op;
 302    }
 303
 304    Definition ${fixed}(Definition def) {
 305        % if fixed == 'vcc' or fixed == 'exec':
 306           assert(def.regClass() == lm);
 307        % endif
 308        def.setFixed(aco::${fixed});
 309        return def;
 310    }
 311
 312    Definition hint_${fixed}(Definition def) {
 313        % if fixed == 'vcc' or fixed == 'exec':
 314           assert(def.regClass() == lm);
 315        % endif
 316        def.setHint(aco::${fixed});
 317        return def;
 318    }
 319
 320 % endfor
 321    /* hand-written helpers */
 322    Temp as_uniform(Op op)
 323    {
 324       assert(op.op.isTemp());
 325       if (op.op.getTemp().type() == RegType::vgpr)
 326          return pseudo(aco_opcode::p_as_uniform, def(RegType::sgpr, op.op.size()), op);
 327       else
 328          return op.op.getTemp();
 329    }
 330
 331    Result v_mul_imm(Definition dst, Temp tmp, uint32_t imm, bool bits24=false)
 332    {
 333       assert(tmp.type() == RegType::vgpr);
 334       if (imm == 0) {
 335          return vop1(aco_opcode::v_mov_b32, dst, Operand(0u));
 336       } else if (imm == 1) {
 337          return copy(dst, Operand(tmp));
 338       } else if (util_is_power_of_two_or_zero(imm)) {
 339          return vop2(aco_opcode::v_lshlrev_b32, dst, Operand((uint32_t)ffs(imm) - 1u), tmp);
 340       } else if (bits24) {
 341         return vop2(aco_opcode::v_mul_u32_u24, dst, Operand(imm), tmp);
 342       } else {
 343         Temp imm_tmp = copy(def(v1), Operand(imm));
 344         return vop3(aco_opcode::v_mul_lo_u32, dst, imm_tmp, tmp);
 345       }
 346    }
 347
 348    Result v_mul24_imm(Definition dst, Temp tmp, uint32_t imm)
 349    {
 350       return v_mul_imm(dst, tmp, imm, true);
 351    }
 352
 353    Result copy(Definition dst, Op op_) {
 354       Operand op = op_.op;
 355       assert(op.bytes() == dst.bytes());
 356       if (dst.regClass() == s1 && op.size() == 1 && op.isLiteral()) {
 357          uint32_t imm = op.constantValue();
 358          if (imm == 0x3e22f983) {
 359             if (program->chip_class >= GFX8)
 360                op.setFixed(PhysReg{248}); /* it can be an inline constant on GFX8+ */
 361          } else if (imm >= 0xffff8000 || imm <= 0x7fff) {
 362             return sopk(aco_opcode::s_movk_i32, dst, imm & 0xFFFFu);
 363          } else if (util_bitreverse(imm) <= 64 || util_bitreverse(imm) >= 0xFFFFFFF0) {
 364             uint32_t rev = util_bitreverse(imm);
 365             return dst.regClass() == v1 ?
 366                    vop1(aco_opcode::v_bfrev_b32, dst, Operand(rev)) :
 367                    sop1(aco_opcode::s_brev_b32, dst, Operand(rev));
 368          } else if (imm != 0) {
 369             unsigned start = (ffs(imm) - 1) & 0x1f;
 370             unsigned size = util_bitcount(imm) & 0x1f;
 371             if ((((1u << size) - 1u) << start) == imm)
 372                 return sop2(aco_opcode::s_bfm_b32, dst, Operand(size), Operand(start));
 373          }
 374       }
 375
 376       if (dst.regClass() == s1) {
 377         return sop1(aco_opcode::s_mov_b32, dst, op);
 378       } else if (dst.regClass() == s2) {
 379         return sop1(aco_opcode::s_mov_b64, dst, op);
 380       } else if (dst.regClass() == v1 || dst.regClass() == v1.as_linear()) {
 381         return vop1(aco_opcode::v_mov_b32, dst, op);
 382       } else if (op.bytes() > 2) {
 383          return pseudo(aco_opcode::p_create_vector, dst, op);
 384       } else if (dst.regClass().is_subdword()) {
 385         if (program->chip_class >= GFX8) {
 386             aco_ptr<SDWA_instruction> sdwa{create_instruction<SDWA_instruction>(aco_opcode::v_mov_b32, asSDWA(Format::VOP1), 1, 1)};
 387             sdwa->operands[0] = op;
 388             sdwa->definitions[0] = dst;
 389             sdwa->sel[0] = op.bytes() == 1 ? sdwa_ubyte : sdwa_uword;
 390             sdwa->dst_sel = dst.bytes() == 1 ? sdwa_ubyte : sdwa_uword;
 391             sdwa->dst_preserve = true;
 392             return insert(std::move(sdwa));
 393         } else {
 394             return vop1(aco_opcode::v_mov_b32, dst, op);
 395         }
 396       } else {
 397         unreachable("Unhandled case in bld.copy()");
 398       }
 399    }
 400
 401    Result vadd32(Definition dst, Op a, Op b, bool carry_out=false, Op carry_in=Op(Operand(s2)), bool post_ra=false) {
 402       if (!b.op.isTemp() || b.op.regClass().type() != RegType::vgpr)
 403          std::swap(a, b);
 404       assert((post_ra || b.op.hasRegClass()) && b.op.regClass().type() == RegType::vgpr);
 405
 406       if (!carry_in.op.isUndefined())
 407          return vop2(aco_opcode::v_addc_co_u32, Definition(dst), hint_vcc(def(lm)), a, b, carry_in);
 408       else if (program->chip_class >= GFX10 && carry_out)
 409          return vop3(aco_opcode::v_add_co_u32_e64, Definition(dst), def(lm), a, b);
 410       else if (program->chip_class < GFX9 || carry_out)
 411          return vop2(aco_opcode::v_add_co_u32, Definition(dst), hint_vcc(def(lm)), a, b);
 412       else
 413          return vop2(aco_opcode::v_add_u32, Definition(dst), a, b);
 414    }
 415
 416    Result vsub32(Definition dst, Op a, Op b, bool carry_out=false, Op borrow=Op(Operand(s2)))
 417    {
 418       if (!borrow.op.isUndefined() || program->chip_class < GFX9)
 419          carry_out = true;
 420
 421       bool reverse = !b.op.isTemp() || b.op.regClass().type() != RegType::vgpr;
 422       if (reverse)
 423          std::swap(a, b);
 424       assert(b.op.isTemp() && b.op.regClass().type() == RegType::vgpr);
 425
 426       aco_opcode op;
 427       Temp carry;
 428       if (carry_out) {
 429          carry = tmp(s2);
 430          if (borrow.op.isUndefined())
 431             op = reverse ? aco_opcode::v_subrev_co_u32 : aco_opcode::v_sub_co_u32;
 432          else
 433             op = reverse ? aco_opcode::v_subbrev_co_u32 : aco_opcode::v_subb_co_u32;
 434       } else {
 435          op = reverse ? aco_opcode::v_subrev_u32 : aco_opcode::v_sub_u32;
 436       }
 437       bool vop3 = false;
 438       if (program->chip_class >= GFX10 && op == aco_opcode::v_subrev_co_u32) {
 439         vop3 = true;
 440         op = aco_opcode::v_subrev_co_u32_e64;
 441       } else if (program->chip_class >= GFX10 && op == aco_opcode::v_sub_co_u32) {
 442         vop3 = true;
 443         op = aco_opcode::v_sub_co_u32_e64;
 444       }
 445
 446       int num_ops = borrow.op.isUndefined() ? 2 : 3;
 447       int num_defs = carry_out ? 2 : 1;
 448       aco_ptr<Instruction> sub;
 449       if (vop3)
 450         sub.reset(create_instruction<VOP3A_instruction>(op, Format::VOP3B, num_ops, num_defs));
 451       else
 452         sub.reset(create_instruction<VOP2_instruction>(op, Format::VOP2, num_ops, num_defs));
 453       sub->operands[0] = a.op;
 454       sub->operands[1] = b.op;
 455       if (!borrow.op.isUndefined())
 456          sub->operands[2] = borrow.op;
 457       sub->definitions[0] = dst;
 458       if (carry_out) {
 459          sub->definitions[1] = Definition(carry);
 460          sub->definitions[1].setHint(aco::vcc);
 461       }
 462       return insert(std::move(sub));
 463    }
 464
 465    Result readlane(Definition dst, Op vsrc, Op lane)
 466    {
 467       if (program->chip_class >= GFX8)
 468          return vop3(aco_opcode::v_readlane_b32_e64, dst, vsrc, lane);
 469       else
 470          return vop2(aco_opcode::v_readlane_b32, dst, vsrc, lane);
 471    }
 472    Result writelane(Definition dst, Op val, Op lane, Op vsrc) {
 473       if (program->chip_class >= GFX8)
 474          return vop3(aco_opcode::v_writelane_b32_e64, dst, val, lane, vsrc);
 475       else
 476          return vop2(aco_opcode::v_writelane_b32, dst, val, lane, vsrc);
 477    }
 478 <%
 479 import itertools
 480 formats = [("pseudo", [Format.PSEUDO], 'Pseudo_instruction', list(itertools.product(range(5), range(5))) + [(8, 1), (1, 8)]),
 481            ("sop1", [Format.SOP1], 'SOP1_instruction', [(1, 1), (2, 1), (3, 2)]),
 482            ("sop2", [Format.SOP2], 'SOP2_instruction', itertools.product([1, 2], [2, 3])),
 483            ("sopk", [Format.SOPK], 'SOPK_instruction', itertools.product([0, 1, 2], [0, 1])),
 484            ("sopp", [Format.SOPP], 'SOPP_instruction', [(0, 0), (0, 1)]),
 485            ("sopc", [Format.SOPC], 'SOPC_instruction', [(1, 2)]),
 486            ("smem", [Format.SMEM], 'SMEM_instruction', [(0, 4), (0, 3), (1, 0), (1, 3), (1, 2), (0, 0)]),
 487            ("ds", [Format.DS], 'DS_instruction', [(1, 1), (1, 2), (0, 3), (0, 4)]),
 488            ("mubuf", [Format.MUBUF], 'MUBUF_instruction', [(0, 4), (1, 3)]),
 489            ("mtbuf", [Format.MTBUF], 'MTBUF_instruction', [(0, 4), (1, 3)]),
 490            ("mimg", [Format.MIMG], 'MIMG_instruction', [(0, 3), (1, 3)]),
 491            ("exp", [Format.EXP], 'Export_instruction', [(0, 4)]),
 492            ("branch", [Format.PSEUDO_BRANCH], 'Pseudo_branch_instruction', itertools.product([0], [0, 1])),
 493            ("barrier", [Format.PSEUDO_BARRIER], 'Pseudo_barrier_instruction', [(0, 0)]),
 494            ("reduction", [Format.PSEUDO_REDUCTION], 'Pseudo_reduction_instruction', [(3, 2)]),
 495            ("vop1", [Format.VOP1], 'VOP1_instruction', [(1, 1), (2, 2)]),
 496            ("vop2", [Format.VOP2], 'VOP2_instruction', itertools.product([1, 2], [2, 3])),
 497            ("vop2_sdwa", [Format.VOP2, Format.SDWA], 'SDWA_instruction', itertools.product([1, 2], [2, 3])),
 498            ("vopc", [Format.VOPC], 'VOPC_instruction', itertools.product([1, 2], [2])),
 499            ("vop3", [Format.VOP3A], 'VOP3A_instruction', [(1, 3), (1, 2), (1, 1), (2, 2)]),
 500            ("vintrp", [Format.VINTRP], 'Interp_instruction', [(1, 2), (1, 3)]),
 501            ("vop1_dpp", [Format.VOP1, Format.DPP], 'DPP_instruction', [(1, 1)]),
 502            ("vop2_dpp", [Format.VOP2, Format.DPP], 'DPP_instruction', itertools.product([1, 2], [2, 3])),
 503            ("vopc_dpp", [Format.VOPC, Format.DPP], 'DPP_instruction', itertools.product([1, 2], [2])),
 504            ("vop1_e64", [Format.VOP1, Format.VOP3A], 'VOP3A_instruction', itertools.product([1], [1])),
 505            ("vop2_e64", [Format.VOP2, Format.VOP3A], 'VOP3A_instruction', itertools.product([1, 2], [2, 3])),
 506            ("vopc_e64", [Format.VOPC, Format.VOP3A], 'VOP3A_instruction', itertools.product([1, 2], [2])),
 507            ("flat", [Format.FLAT], 'FLAT_instruction', [(0, 3), (1, 2)]),
 508            ("global", [Format.GLOBAL], 'FLAT_instruction', [(0, 3), (1, 2)])]
 509 %>\\
 510 % for name, formats, struct, shapes in formats:
 511     % for num_definitions, num_operands in shapes:
 512         <%
 513         args = ['aco_opcode opcode']
 514         for i in range(num_definitions):
 515             args.append('Definition def%d' % i)
 516         for i in range(num_operands):
 517             args.append('Op op%d' % i)
 518         for f in formats:
 519             args += f.get_builder_field_decls()
 520         %>\\
 521
 522    Result ${name}(${', '.join(args)})
 523    {
 524       ${struct} *instr = create_instruction<${struct}>(opcode, (Format)(${'|'.join('(int)Format::%s' % f.name for f in formats)}), ${num_operands}, ${num_definitions});
 525         % for i in range(num_definitions):
 526             instr->definitions[${i}] = def${i};
 527         % endfor
 528         % for i in range(num_operands):
 529             instr->operands[${i}] = op${i}.op;
 530         % endfor
 531         % for f in formats:
 532             % for dest, field_name in zip(f.get_builder_field_dests(), f.get_builder_field_names()):
 533       instr->${dest} = ${field_name};
 534             % endfor
 535             ${f.get_builder_initialization(num_operands)}
 536         % endfor
 537       return insert(instr);
 538    }
 539
 540     % if name == 'sop1' or name == 'sop2' or name == 'sopc':
 541         <%
 542         args[0] = 'WaveSpecificOpcode opcode'
 543         params = []
 544         for i in range(num_definitions):
 545             params.append('def%d' % i)
 546         for i in range(num_operands):
 547             params.append('op%d' % i)
 548         %>\\
 549
 550    inline Result ${name}(${', '.join(args)})
 551    {
 552        return ${name}(w64or32(opcode), ${', '.join(params)});
 553    }
 554
 555     % endif
 556     % endfor
 557 % endfor
 558 };
 559
 560 }
 561 #endif /* _ACO_BUILDER_ */"""
 562
 563 from aco_opcodes import opcodes, Format
 564 from mako.template import Template
 565
 566 print(Template(template).render(opcodes=opcodes, Format=Format))