src/mesa/drivers/dri/i965/brw_vec4_vp.cpp

   1 /*
   2  * Copyright © 2012 Intel Corporation
   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 (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  */
  23
  24 /** @file brw_vec4_vp.cpp
  25  *
  26  * A translator from Mesa IR to the i965 driver's Vec4 IR, used to implement
  27  * ARB_vertex_program and fixed-function vertex processing.
  28  */
  29
  30 #include "brw_context.h"
  31 #include "brw_vec4.h"
  32 #include "brw_vs.h"
  33 extern "C" {
  34 #include "program/prog_parameter.h"
  35 #include "program/prog_print.h"
  36 }
  37 using namespace brw;
  38
  39 void
  40 vec4_visitor::emit_vp_sop(enum brw_conditional_mod conditional_mod,
  41                           dst_reg dst, src_reg src0, src_reg src1,
  42                           src_reg one)
  43 {
  44    vec4_instruction *inst;
  45
  46    inst = emit(CMP(dst_null_f(), src0, src1, conditional_mod));
  47
  48    inst = emit(BRW_OPCODE_SEL, dst, one, src_reg(0.0f));
  49    inst->predicate = BRW_PREDICATE_NORMAL;
  50 }
  51
  52 void
  53 vec4_vs_visitor::emit_program_code()
  54 {
  55    this->need_all_constants_in_pull_buffer = false;
  56
  57    setup_vp_regs();
  58
  59    /* Keep a reg with 1.0 around, for reuse by emit_vs_sop so that it can just
  60     * be:
  61     *
  62     * sel.f0 dst 1.0 0.0
  63     *
  64     * instead of
  65     *
  66     * mov    dst 0.0
  67     * mov.f0 dst 1.0
  68     */
  69    src_reg one = src_reg(this, glsl_type::float_type);
  70    emit(MOV(dst_reg(one), src_reg(1.0f)));
  71
  72    for (unsigned int insn = 0; insn < prog->NumInstructions; insn++) {
  73       const struct prog_instruction *vpi = &prog->Instructions[insn];
  74       base_ir = vpi;
  75
  76       dst_reg dst;
  77       src_reg src[3];
  78
  79       /* We always emit into a temporary destination register to avoid
  80        * aliasing issues.
  81        */
  82       dst = dst_reg(this, glsl_type::vec4_type);
  83
  84       for (int i = 0; i < 3; i++)
  85          src[i] = get_vp_src_reg(vpi->SrcReg[i]);
  86
  87       switch (vpi->Opcode) {
  88       case OPCODE_ABS:
  89          src[0].abs = true;
  90          src[0].negate = false;
  91          emit(MOV(dst, src[0]));
  92          break;
  93
  94       case OPCODE_ADD:
  95          emit(ADD(dst, src[0], src[1]));
  96          break;
  97
  98       case OPCODE_ARL:
  99          if (devinfo->gen >= 6) {
 100             dst.writemask = WRITEMASK_X;
 101             dst_reg dst_f = dst;
 102             dst_f.type = BRW_REGISTER_TYPE_F;
 103
 104             emit(RNDD(dst_f, src[0]));
 105             emit(MOV(dst, src_reg(dst_f)));
 106          } else {
 107             emit(RNDD(dst, src[0]));
 108          }
 109          break;
 110
 111       case OPCODE_DP3:
 112          emit(DP3(dst, src[0], src[1]));
 113          break;
 114       case OPCODE_DP4:
 115          emit(DP4(dst, src[0], src[1]));
 116          break;
 117       case OPCODE_DPH:
 118          emit(DPH(dst, src[0], src[1]));
 119          break;
 120
 121       case OPCODE_DST: {
 122          dst_reg t = dst;
 123          if (vpi->DstReg.WriteMask & WRITEMASK_X) {
 124             t.writemask = WRITEMASK_X;
 125             emit(MOV(t, src_reg(1.0f)));
 126          }
 127          if (vpi->DstReg.WriteMask & WRITEMASK_Y) {
 128             t.writemask = WRITEMASK_Y;
 129             emit(MUL(t, src[0], src[1]));
 130          }
 131          if (vpi->DstReg.WriteMask & WRITEMASK_Z) {
 132             t.writemask = WRITEMASK_Z;
 133             emit(MOV(t, src[0]));
 134          }
 135          if (vpi->DstReg.WriteMask & WRITEMASK_W) {
 136             t.writemask = WRITEMASK_W;
 137             emit(MOV(t, src[1]));
 138          }
 139          break;
 140       }
 141
 142       case OPCODE_EXP: {
 143          dst_reg result = dst;
 144          if (vpi->DstReg.WriteMask & WRITEMASK_X) {
 145             /* tmp_d = floor(src[0].x) */
 146             src_reg tmp_d = src_reg(this, glsl_type::ivec4_type);
 147             assert(tmp_d.type == BRW_REGISTER_TYPE_D);
 148             emit(RNDD(dst_reg(tmp_d), swizzle(src[0], BRW_SWIZZLE_XXXX)));
 149
 150             /* result[0] = 2.0 ^ tmp */
 151             /* Adjust exponent for floating point: exp += 127 */
 152             dst_reg tmp_d_x(GRF, tmp_d.reg, glsl_type::int_type, WRITEMASK_X);
 153             emit(ADD(tmp_d_x, tmp_d, src_reg(127)));
 154
 155             /* Install exponent and sign.  Excess drops off the edge: */
 156             dst_reg res_d_x(GRF, result.reg, glsl_type::int_type, WRITEMASK_X);
 157             emit(BRW_OPCODE_SHL, res_d_x, tmp_d, src_reg(23));
 158          }
 159          if (vpi->DstReg.WriteMask & WRITEMASK_Y) {
 160             result.writemask = WRITEMASK_Y;
 161             emit(FRC(result, src[0]));
 162          }
 163          if (vpi->DstReg.WriteMask & WRITEMASK_Z) {
 164             result.writemask = WRITEMASK_Z;
 165             emit_math(SHADER_OPCODE_EXP2, result, src[0]);
 166          }
 167          if (vpi->DstReg.WriteMask & WRITEMASK_W) {
 168             result.writemask = WRITEMASK_W;
 169             emit(MOV(result, src_reg(1.0f)));
 170          }
 171          break;
 172       }
 173
 174       case OPCODE_EX2:
 175          emit_math(SHADER_OPCODE_EXP2, dst, src[0]);
 176          break;
 177
 178       case OPCODE_FLR:
 179          emit(RNDD(dst, src[0]));
 180          break;
 181
 182       case OPCODE_FRC:
 183          emit(FRC(dst, src[0]));
 184          break;
 185
 186       case OPCODE_LG2:
 187          emit_math(SHADER_OPCODE_LOG2, dst, src[0]);
 188          break;
 189
 190       case OPCODE_LIT: {
 191          dst_reg result = dst;
 192          /* From the ARB_vertex_program spec:
 193           *
 194           *      tmp = VectorLoad(op0);
 195           *      if (tmp.x < 0) tmp.x = 0;
 196           *      if (tmp.y < 0) tmp.y = 0;
 197           *      if (tmp.w < -(128.0-epsilon)) tmp.w = -(128.0-epsilon);
 198           *      else if (tmp.w > 128-epsilon) tmp.w = 128-epsilon;
 199           *      result.x = 1.0;
 200           *      result.y = tmp.x;
 201           *      result.z = (tmp.x > 0) ? RoughApproxPower(tmp.y, tmp.w) : 0.0;
 202           *      result.w = 1.0;
 203           *
 204           * Note that we don't do the clamping to +/- 128.  We didn't in
 205           * brw_vs_emit.c either.
 206           */
 207          if (vpi->DstReg.WriteMask & WRITEMASK_XW) {
 208             result.writemask = WRITEMASK_XW;
 209             emit(MOV(result, src_reg(1.0f)));
 210          }
 211          if (vpi->DstReg.WriteMask & WRITEMASK_YZ) {
 212             result.writemask = WRITEMASK_YZ;
 213             emit(MOV(result, src_reg(0.0f)));
 214
 215             src_reg tmp_x = swizzle(src[0], BRW_SWIZZLE_XXXX);
 216
 217             emit(CMP(dst_null_d(), tmp_x, src_reg(0.0f), BRW_CONDITIONAL_G));
 218             emit(IF(BRW_PREDICATE_NORMAL));
 219
 220             if (vpi->DstReg.WriteMask & WRITEMASK_Y) {
 221                result.writemask = WRITEMASK_Y;
 222                emit(MOV(result, tmp_x));
 223             }
 224
 225             if (vpi->DstReg.WriteMask & WRITEMASK_Z) {
 226                /* if (tmp.y < 0) tmp.y = 0; */
 227                src_reg tmp_y = swizzle(src[0], BRW_SWIZZLE_YYYY);
 228                result.writemask = WRITEMASK_Z;
 229                emit_minmax(BRW_CONDITIONAL_GE, result, tmp_y, src_reg(0.0f));
 230
 231                src_reg clamped_y(result);
 232                clamped_y.swizzle = BRW_SWIZZLE_ZZZZ;
 233
 234                src_reg tmp_w = swizzle(src[0], BRW_SWIZZLE_WWWW);
 235
 236                emit_math(SHADER_OPCODE_POW, result, clamped_y, tmp_w);
 237             }
 238             emit(BRW_OPCODE_ENDIF);
 239          }
 240          break;
 241       }
 242
 243       case OPCODE_LOG: {
 244          dst_reg result = dst;
 245          result.type = BRW_REGISTER_TYPE_UD;
 246          src_reg result_src = src_reg(result);
 247
 248          src_reg arg0_ud = swizzle(src[0], BRW_SWIZZLE_XXXX);
 249          arg0_ud.type = BRW_REGISTER_TYPE_UD;
 250
 251          /* Perform mant = frexpf(fabsf(x), &exp), adjust exp and mnt
 252           * according to spec:
 253           *
 254           * These almost look likey they could be joined up, but not really
 255           * practical:
 256           *
 257           * result[0].f = (x.i & ((1<<31)-1) >> 23) - 127
 258           * result[1].i = (x.i & ((1<<23)-1)        + (127<<23)
 259           */
 260          if (vpi->DstReg.WriteMask & WRITEMASK_XZ) {
 261             result.writemask = WRITEMASK_X;
 262             emit(AND(result, arg0_ud, src_reg((1u << 31) - 1)));
 263             emit(BRW_OPCODE_SHR, result, result_src, src_reg(23u));
 264             src_reg result_d(result_src);
 265             result_d.type = BRW_REGISTER_TYPE_D; /* does it matter? */
 266             result.type = BRW_REGISTER_TYPE_F;
 267             emit(ADD(result, result_d, src_reg(-127)));
 268          }
 269
 270          if (vpi->DstReg.WriteMask & WRITEMASK_YZ) {
 271             result.writemask = WRITEMASK_Y;
 272             result.type = BRW_REGISTER_TYPE_UD;
 273             emit(AND(result, arg0_ud, src_reg((1u << 23) - 1)));
 274             emit(OR(result, result_src, src_reg(127u << 23)));
 275          }
 276
 277          if (vpi->DstReg.WriteMask & WRITEMASK_Z) {
 278             /* result[2] = result[0] + LOG2(result[1]); */
 279
 280             /* Why bother?  The above is just a hint how to do this with a
 281              * taylor series.  Maybe we *should* use a taylor series as by
 282              * the time all the above has been done it's almost certainly
 283              * quicker than calling the mathbox, even with low precision.
 284              *
 285              * Options are:
 286              *    - result[0] + mathbox.LOG2(result[1])
 287              *    - mathbox.LOG2(arg0.x)
 288              *    - result[0] + inline_taylor_approx(result[1])
 289              */
 290             result.type = BRW_REGISTER_TYPE_F;
 291             result.writemask = WRITEMASK_Z;
 292             src_reg result_x(result), result_y(result), result_z(result);
 293             result_x.swizzle = BRW_SWIZZLE_XXXX;
 294             result_y.swizzle = BRW_SWIZZLE_YYYY;
 295             result_z.swizzle = BRW_SWIZZLE_ZZZZ;
 296             emit_math(SHADER_OPCODE_LOG2, result, result_y);
 297             emit(ADD(result, result_z, result_x));
 298          }
 299
 300          if (vpi->DstReg.WriteMask & WRITEMASK_W) {
 301             result.type = BRW_REGISTER_TYPE_F;
 302             result.writemask = WRITEMASK_W;
 303             emit(MOV(result, src_reg(1.0f)));
 304          }
 305          break;
 306       }
 307
 308       case OPCODE_MAD: {
 309          src_reg temp = src_reg(this, glsl_type::vec4_type);
 310          emit(MUL(dst_reg(temp), src[0], src[1]));
 311          emit(ADD(dst, temp, src[2]));
 312          break;
 313       }
 314
 315       case OPCODE_MAX:
 316          emit_minmax(BRW_CONDITIONAL_GE, dst, src[0], src[1]);
 317          break;
 318
 319       case OPCODE_MIN:
 320          emit_minmax(BRW_CONDITIONAL_L, dst, src[0], src[1]);
 321          break;
 322
 323       case OPCODE_MOV:
 324          emit(MOV(dst, src[0]));
 325          break;
 326
 327       case OPCODE_MUL:
 328          emit(MUL(dst, src[0], src[1]));
 329          break;
 330
 331       case OPCODE_POW:
 332          emit_math(SHADER_OPCODE_POW, dst, src[0], src[1]);
 333          break;
 334
 335       case OPCODE_RCP:
 336          emit_math(SHADER_OPCODE_RCP, dst, src[0]);
 337          break;
 338
 339       case OPCODE_RSQ:
 340          emit_math(SHADER_OPCODE_RSQ, dst, src[0]);
 341          break;
 342
 343       case OPCODE_SGE:
 344          emit_vp_sop(BRW_CONDITIONAL_GE, dst, src[0], src[1], one);
 345          break;
 346
 347       case OPCODE_SLT:
 348          emit_vp_sop(BRW_CONDITIONAL_L, dst, src[0], src[1], one);
 349          break;
 350
 351       case OPCODE_SUB: {
 352          src_reg neg_src1 = src[1];
 353          neg_src1.negate = !src[1].negate;
 354          emit(ADD(dst, src[0], neg_src1));
 355          break;
 356       }
 357
 358       case OPCODE_SWZ:
 359          /* Note that SWZ's extended swizzles are handled in the general
 360           * get_src_reg() code.
 361           */
 362          emit(MOV(dst, src[0]));
 363          break;
 364
 365       case OPCODE_XPD: {
 366          src_reg t1 = src_reg(this, glsl_type::vec4_type);
 367          src_reg t2 = src_reg(this, glsl_type::vec4_type);
 368
 369          emit(MUL(dst_reg(t1),
 370                   swizzle(src[0], BRW_SWIZZLE_YZXW),
 371                   swizzle(src[1], BRW_SWIZZLE_ZXYW)));
 372          emit(MUL(dst_reg(t2),
 373                   swizzle(src[0], BRW_SWIZZLE_ZXYW),
 374                   swizzle(src[1], BRW_SWIZZLE_YZXW)));
 375          t2.negate = true;
 376          emit(ADD(dst, t1, t2));
 377          break;
 378       }
 379
 380       case OPCODE_END:
 381          break;
 382
 383       default:
 384          assert(!"Unsupported opcode in vertex program");
 385       }
 386
 387       /* Copy the temporary back into the actual destination register. */
 388       if (_mesa_num_inst_dst_regs(vpi->Opcode) != 0) {
 389          emit(MOV(get_vp_dst_reg(vpi->DstReg), src_reg(dst)));
 390       }
 391    }
 392
 393    /* If we used relative addressing, we need to upload all constants as
 394     * pull constants.  Do that now.
 395     */
 396    if (this->need_all_constants_in_pull_buffer) {
 397       const struct gl_program_parameter_list *params = vp->Base.Parameters;
 398       unsigned i;
 399       for (i = 0; i < params->NumParameters * 4; i++) {
 400          stage_prog_data->pull_param[i] =
 401             &params->ParameterValues[i / 4][i % 4];
 402       }
 403       stage_prog_data->nr_pull_params = i;
 404    }
 405 }
 406
 407 void
 408 vec4_vs_visitor::setup_vp_regs()
 409 {
 410    /* PROGRAM_TEMPORARY */
 411    int num_temp = prog->NumTemporaries;
 412    vp_temp_regs = rzalloc_array(mem_ctx, src_reg, num_temp);
 413    for (int i = 0; i < num_temp; i++)
 414       vp_temp_regs[i] = src_reg(this, glsl_type::vec4_type);
 415
 416    /* PROGRAM_STATE_VAR etc. */
 417    struct gl_program_parameter_list *plist = vp->Base.Parameters;
 418    for (unsigned p = 0; p < plist->NumParameters; p++) {
 419       unsigned components = plist->Parameters[p].Size;
 420
 421       /* Parameters should be either vec4 uniforms or single component
 422        * constants; matrices and other larger types should have been broken
 423        * down earlier.
 424        */
 425       assert(components <= 4);
 426
 427       this->uniform_size[this->uniforms] = 1; /* 1 vec4 */
 428       this->uniform_vector_size[this->uniforms] = components;
 429       for (unsigned i = 0; i < 4; i++) {
 430          stage_prog_data->param[this->uniforms * 4 + i] = i >= components
 431             ? 0 : &plist->ParameterValues[p][i];
 432       }
 433       this->uniforms++; /* counted in vec4 units */
 434    }
 435
 436    /* PROGRAM_OUTPUT */
 437    for (int slot = 0; slot < prog_data->vue_map.num_slots; slot++) {
 438       int varying = prog_data->vue_map.slot_to_varying[slot];
 439       if (varying == VARYING_SLOT_PSIZ)
 440          output_reg[varying] = dst_reg(this, glsl_type::float_type);
 441       else
 442          output_reg[varying] = dst_reg(this, glsl_type::vec4_type);
 443       assert(output_reg[varying].type == BRW_REGISTER_TYPE_F);
 444    }
 445
 446    /* PROGRAM_ADDRESS */
 447    this->vp_addr_reg = src_reg(this, glsl_type::int_type);
 448    assert(this->vp_addr_reg.type == BRW_REGISTER_TYPE_D);
 449 }
 450
 451 dst_reg
 452 vec4_vs_visitor::get_vp_dst_reg(const prog_dst_register &dst)
 453 {
 454    dst_reg result;
 455
 456    assert(!dst.RelAddr);
 457
 458    switch (dst.File) {
 459    case PROGRAM_TEMPORARY:
 460       result = dst_reg(vp_temp_regs[dst.Index]);
 461       break;
 462
 463    case PROGRAM_OUTPUT:
 464       result = output_reg[dst.Index];
 465       break;
 466
 467    case PROGRAM_ADDRESS: {
 468       assert(dst.Index == 0);
 469       result = dst_reg(this->vp_addr_reg);
 470       break;
 471    }
 472
 473    case PROGRAM_UNDEFINED:
 474       return dst_null_f();
 475
 476    default:
 477       unreachable("vec4_vp: bad destination register file");
 478    }
 479
 480    result.writemask = dst.WriteMask;
 481    return result;
 482 }
 483
 484 src_reg
 485 vec4_vs_visitor::get_vp_src_reg(const prog_src_register &src)
 486 {
 487    struct gl_program_parameter_list *plist = vp->Base.Parameters;
 488
 489    src_reg result;
 490
 491    assert(!src.Abs);
 492
 493    switch (src.File) {
 494    case PROGRAM_UNDEFINED:
 495       return src_reg(brw_null_reg());
 496
 497    case PROGRAM_TEMPORARY:
 498       result = vp_temp_regs[src.Index];
 499       break;
 500
 501    case PROGRAM_INPUT:
 502       result = src_reg(ATTR, src.Index, glsl_type::vec4_type);
 503       result.type = BRW_REGISTER_TYPE_F;
 504       break;
 505
 506    case PROGRAM_ADDRESS: {
 507       assert(src.Index == 0);
 508       result = this->vp_addr_reg;
 509       break;
 510    }
 511
 512    case PROGRAM_STATE_VAR:
 513    case PROGRAM_CONSTANT:
 514       /* From the ARB_vertex_program specification:
 515        * "Relative addressing can only be used for accessing program
 516        *  parameter arrays."
 517        */
 518       if (src.RelAddr) {
 519          /* Since we have no idea what the base of the array is, we need to
 520           * upload ALL constants as push constants.
 521           */
 522          this->need_all_constants_in_pull_buffer = true;
 523
 524          /* Add the small constant index to the address register */
 525          src_reg reladdr = src_reg(this, glsl_type::int_type);
 526
 527          dst_reg dst_reladdr = dst_reg(reladdr);
 528          dst_reladdr.writemask = WRITEMASK_X;
 529          emit(ADD(dst_reladdr, this->vp_addr_reg, src_reg(src.Index)));
 530
 531          if (devinfo->gen < 6)
 532             emit(MUL(dst_reladdr, reladdr, src_reg(16)));
 533
 534       #if 0
 535          assert(src.Index < this->uniforms);
 536          result = src_reg(dst_reg(UNIFORM, 0));
 537          result.type = BRW_REGISTER_TYPE_F;
 538          result.reladdr = new(mem_ctx) src_reg();
 539          memcpy(result.reladdr, &reladdr, sizeof(src_reg));
 540       #endif
 541
 542          result = src_reg(this, glsl_type::vec4_type);
 543          src_reg surf_index = src_reg(unsigned(prog_data->base.binding_table.pull_constants_start));
 544
 545          emit_pull_constant_load_reg(dst_reg(result),
 546                                      surf_index,
 547                                      reladdr,
 548                                      NULL, NULL /* before_block/inst */);
 549          break;
 550       }
 551
 552       /* We actually want to look at the type in the Parameters list for this,
 553        * because this lets us upload constant builtin uniforms as actual
 554        * constants.
 555        */
 556       switch (plist->Parameters[src.Index].Type) {
 557       case PROGRAM_CONSTANT:
 558          result = src_reg(this, glsl_type::vec4_type);
 559          for (int i = 0; i < 4; i++) {
 560             dst_reg t = dst_reg(result);
 561             t.writemask = 1 << i;
 562             emit(MOV(t, src_reg(plist->ParameterValues[src.Index][i].f)));
 563          }
 564          break;
 565
 566       case PROGRAM_STATE_VAR:
 567          assert(src.Index < this->uniforms);
 568          result = src_reg(dst_reg(UNIFORM, src.Index));
 569          result.type = BRW_REGISTER_TYPE_F;
 570          break;
 571
 572       default:
 573          assert(!"Bad uniform in src register file");
 574          return src_reg(this, glsl_type::vec4_type);
 575       }
 576       break;
 577
 578    default:
 579       assert(!"Bad src register file");
 580       return src_reg(this, glsl_type::vec4_type);
 581    }
 582
 583    if (src.Swizzle != SWIZZLE_NOOP || src.Negate) {
 584       unsigned short zeros_mask = 0;
 585       unsigned short ones_mask = 0;
 586       unsigned short src_mask = 0;
 587       unsigned short src_swiz[4];
 588
 589       for (int i = 0; i < 4; i++) {
 590          src_swiz[i] = 0; /* initialize for safety */
 591
 592          /* The ZERO, ONE, and Negate options are only used for OPCODE_SWZ,
 593           * but it's simplest to handle it here.
 594           */
 595          int s = GET_SWZ(src.Swizzle, i);
 596          switch (s) {
 597          case SWIZZLE_X:
 598          case SWIZZLE_Y:
 599          case SWIZZLE_Z:
 600          case SWIZZLE_W:
 601             src_mask |= 1 << i;
 602             src_swiz[i] = s;
 603             break;
 604          case SWIZZLE_ZERO:
 605             zeros_mask |= 1 << i;
 606             break;
 607          case SWIZZLE_ONE:
 608             ones_mask |= 1 << i;
 609             break;
 610          }
 611       }
 612
 613       result.swizzle =
 614          BRW_SWIZZLE4(src_swiz[0], src_swiz[1], src_swiz[2], src_swiz[3]);
 615
 616       /* The hardware doesn't natively handle the SWZ instruction's zero/one
 617        * swizzles or per-component negation, so we need to use a temporary.
 618        */
 619       if (zeros_mask || ones_mask || src.Negate) {
 620          src_reg temp_src(this, glsl_type::vec4_type);
 621          dst_reg temp(temp_src);
 622
 623          if (src_mask) {
 624             temp.writemask = src_mask;
 625             emit(MOV(temp, result));
 626          }
 627
 628          if (zeros_mask) {
 629             temp.writemask = zeros_mask;
 630             emit(MOV(temp, src_reg(0.0f)));
 631          }
 632
 633          if (ones_mask) {
 634             temp.writemask = ones_mask;
 635             emit(MOV(temp, src_reg(1.0f)));
 636          }
 637
 638          if (src.Negate) {
 639             temp.writemask = src.Negate;
 640             src_reg neg(temp_src);
 641             neg.negate = true;
 642             emit(MOV(temp, neg));
 643          }
 644          result = temp_src;
 645       }
 646    }
 647
 648    return result;
 649 }