src/gallium/auxiliary/gallivm/lp_bld_tgsi_aos.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2010 VMware, Inc.
   4  * All Rights Reserved.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the
   8  * "Software"), to deal in the Software without restriction, including
   9  * without limitation the rights to use, copy, modify, merge, publish,
  10  * distribute, sub license, and/or sell copies of the Software, and to
  11  * permit persons to whom the Software is furnished to do so, subject to
  12  * the following conditions:
  13  *
  14  * The above copyright notice and this permission notice (including the
  15  * next paragraph) shall be included in all copies or substantial portions
  16  * of the Software.
  17  *
  18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  19  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  20  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  21  * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
  22  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  23  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  24  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  25  *
  26  **************************************************************************/
  27
  28 /**
  29  * @file
  30  * TGSI to LLVM IR translation -- AoS.
  31  *
  32  * FIXME:
  33  * - No control flow support: the existing control flow code should be factored
  34  * out into from the SoA code into a common module and shared.
  35  * - No derivatives. Derivate logic should be pluggable, just like the samplers.
  36  *
  37  * @author Jose Fonseca <jfonseca@vmware.com>
  38  */
  39
  40 #include "pipe/p_config.h"
  41 #include "pipe/p_shader_tokens.h"
  42 #include "util/u_debug.h"
  43 #include "util/u_math.h"
  44 #include "util/u_memory.h"
  45 #include "tgsi/tgsi_dump.h"
  46 #include "tgsi/tgsi_info.h"
  47 #include "tgsi/tgsi_parse.h"
  48 #include "tgsi/tgsi_util.h"
  49 #include "tgsi/tgsi_scan.h"
  50 #include "lp_bld_type.h"
  51 #include "lp_bld_const.h"
  52 #include "lp_bld_arit.h"
  53 #include "lp_bld_logic.h"
  54 #include "lp_bld_swizzle.h"
  55 #include "lp_bld_flow.h"
  56 #include "lp_bld_quad.h"
  57 #include "lp_bld_tgsi.h"
  58 #include "lp_bld_debug.h"
  59 #include "lp_bld_sample.h"
  60
  61
  62 /**
  63  * Wrapper around lp_build_swizzle_aos which translates swizzles to another
  64  * ordering.
  65  */
  66 static LLVMValueRef
  67 swizzle_aos(struct lp_build_tgsi_context *bld_base,
  68             LLVMValueRef a,
  69             unsigned swizzle_x,
  70             unsigned swizzle_y,
  71             unsigned swizzle_z,
  72             unsigned swizzle_w)
  73 {
  74    unsigned char swizzles[4];
  75    struct lp_build_tgsi_aos_context *bld = lp_aos_context(bld_base);
  76
  77    assert(swizzle_x < 4);
  78    assert(swizzle_y < 4);
  79    assert(swizzle_z < 4);
  80    assert(swizzle_w < 4);
  81
  82    swizzles[bld->inv_swizzles[0]] = bld->swizzles[swizzle_x];
  83    swizzles[bld->inv_swizzles[1]] = bld->swizzles[swizzle_y];
  84    swizzles[bld->inv_swizzles[2]] = bld->swizzles[swizzle_z];
  85    swizzles[bld->inv_swizzles[3]] = bld->swizzles[swizzle_w];
  86
  87    return lp_build_swizzle_aos(&bld->bld_base.base, a, swizzles);
  88 }
  89
  90
  91 static LLVMValueRef
  92 swizzle_scalar_aos(struct lp_build_tgsi_aos_context *bld,
  93                    LLVMValueRef a,
  94                    unsigned chan)
  95 {
  96    chan = bld->swizzles[chan];
  97    return lp_build_swizzle_scalar_aos(&bld->bld_base.base, a, chan, 4);
  98 }
  99
 100
 101 static LLVMValueRef
 102 emit_fetch_constant(
 103    struct lp_build_tgsi_context * bld_base,
 104    const struct tgsi_full_src_register * reg,
 105    enum tgsi_opcode_type stype,
 106    unsigned swizzle)
 107 {
 108    struct lp_build_tgsi_aos_context * bld = lp_aos_context(bld_base);
 109    LLVMBuilderRef builder = bld_base->base.gallivm->builder;
 110    struct lp_type type = bld_base->base.type;
 111    LLVMValueRef res;
 112    unsigned chan;
 113
 114    assert(!reg->Register.Indirect);
 115
 116    /*
 117     * Get the constants components
 118     */
 119
 120    res = bld->bld_base.base.undef;
 121    for (chan = 0; chan < 4; ++chan) {
 122       LLVMValueRef index;
 123       LLVMValueRef scalar_ptr;
 124       LLVMValueRef scalar;
 125       LLVMValueRef swizzle;
 126
 127       index = lp_build_const_int32(bld->bld_base.base.gallivm,
 128                                    reg->Register.Index * 4 + chan);
 129
 130       scalar_ptr = LLVMBuildGEP(builder, bld->consts_ptr, &index, 1, "");
 131
 132       scalar = LLVMBuildLoad(builder, scalar_ptr, "");
 133
 134       lp_build_name(scalar, "const[%u].%c", reg->Register.Index, "xyzw"[chan]);
 135
 136       /*
 137        * NOTE: constants array is always assumed to be RGBA
 138        */
 139
 140       swizzle = lp_build_const_int32(bld->bld_base.base.gallivm,
 141                                      bld->swizzles[chan]);
 142
 143       res = LLVMBuildInsertElement(builder, res, scalar, swizzle, "");
 144    }
 145
 146    /*
 147     * Broadcast the first quaternion to all others.
 148     *
 149     * XXX: could be factored into a reusable function.
 150     */
 151
 152    if (type.length > 4) {
 153       LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
 154       unsigned i;
 155
 156       for (chan = 0; chan < 4; ++chan) {
 157          shuffles[chan] = lp_build_const_int32(bld->bld_base.base.gallivm, chan);
 158       }
 159
 160       for (i = 4; i < type.length; ++i) {
 161          shuffles[i] = shuffles[i % 4];
 162       }
 163
 164       res = LLVMBuildShuffleVector(builder,
 165                                    res, bld->bld_base.base.undef,
 166                                    LLVMConstVector(shuffles, type.length),
 167                                    "");
 168    }
 169    return res;
 170 }
 171
 172 static LLVMValueRef
 173 emit_fetch_immediate(
 174    struct lp_build_tgsi_context * bld_base,
 175    const struct tgsi_full_src_register * reg,
 176    enum tgsi_opcode_type stype,
 177    unsigned swizzle)
 178 {
 179    struct lp_build_tgsi_aos_context * bld = lp_aos_context(bld_base);
 180    LLVMValueRef res = bld->immediates[reg->Register.Index];
 181    assert(res);
 182    return res;
 183 }
 184
 185 static LLVMValueRef
 186 emit_fetch_input(
 187    struct lp_build_tgsi_context * bld_base,
 188    const struct tgsi_full_src_register * reg,
 189    enum tgsi_opcode_type stype,
 190    unsigned swizzle)
 191 {
 192    struct lp_build_tgsi_aos_context * bld = lp_aos_context(bld_base);
 193    LLVMValueRef res = bld->inputs[reg->Register.Index];
 194    assert(!reg->Register.Indirect);
 195    assert(res);
 196    return res;
 197 }
 198
 199 static LLVMValueRef
 200 emit_fetch_temporary(
 201    struct lp_build_tgsi_context * bld_base,
 202    const struct tgsi_full_src_register * reg,
 203    enum tgsi_opcode_type stype,
 204    unsigned swizzle)
 205 {
 206    struct lp_build_tgsi_aos_context * bld = lp_aos_context(bld_base);
 207    LLVMBuilderRef builder = bld_base->base.gallivm->builder;
 208    LLVMValueRef temp_ptr = bld->temps[reg->Register.Index];
 209    LLVMValueRef res = LLVMBuildLoad(builder, temp_ptr, "");
 210    assert(!reg->Register.Indirect);
 211    if (!res)
 212       return bld->bld_base.base.undef;
 213
 214    return res;
 215 }
 216
 217 /**
 218  * Register store.
 219  */
 220 void
 221 lp_emit_store_aos(
 222    struct lp_build_tgsi_aos_context *bld,
 223    const struct tgsi_full_instruction *inst,
 224    unsigned index,
 225    LLVMValueRef value)
 226 {
 227    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
 228    const struct tgsi_full_dst_register *reg = &inst->Dst[index];
 229    LLVMValueRef mask = NULL;
 230    LLVMValueRef ptr;
 231
 232    /*
 233     * Saturate the value
 234     */
 235
 236    switch (inst->Instruction.Saturate) {
 237    case TGSI_SAT_NONE:
 238       break;
 239
 240    case TGSI_SAT_ZERO_ONE:
 241       value = lp_build_max(&bld->bld_base.base, value, bld->bld_base.base.zero);
 242       value = lp_build_min(&bld->bld_base.base, value, bld->bld_base.base.one);
 243       break;
 244
 245    case TGSI_SAT_MINUS_PLUS_ONE:
 246       value = lp_build_max(&bld->bld_base.base, value, lp_build_const_vec(bld->bld_base.base.gallivm, bld->bld_base.base.type, -1.0));
 247       value = lp_build_min(&bld->bld_base.base, value, bld->bld_base.base.one);
 248       break;
 249
 250    default:
 251       assert(0);
 252    }
 253
 254    /*
 255     * Translate the register file
 256     */
 257
 258    assert(!reg->Register.Indirect);
 259
 260    switch (reg->Register.File) {
 261    case TGSI_FILE_OUTPUT:
 262       ptr = bld->outputs[reg->Register.Index];
 263       break;
 264
 265    case TGSI_FILE_TEMPORARY:
 266       ptr = bld->temps[reg->Register.Index];
 267       break;
 268
 269    case TGSI_FILE_ADDRESS:
 270       ptr = bld->addr[reg->Indirect.Index];
 271       break;
 272
 273    case TGSI_FILE_PREDICATE:
 274       ptr = bld->preds[reg->Register.Index];
 275       break;
 276
 277    default:
 278       assert(0);
 279       return;
 280    }
 281
 282    if (!ptr)
 283       return;
 284    /*
 285     * Predicate
 286     */
 287
 288    if (inst->Instruction.Predicate) {
 289       LLVMValueRef pred;
 290
 291       assert(inst->Predicate.Index < LP_MAX_TGSI_PREDS);
 292
 293       pred = LLVMBuildLoad(builder,
 294                            bld->preds[inst->Predicate.Index], "");
 295
 296       /*
 297        * Convert the value to an integer mask.
 298        */
 299       pred = lp_build_compare(bld->bld_base.base.gallivm,
 300                                bld->bld_base.base.type,
 301                                PIPE_FUNC_NOTEQUAL,
 302                                pred,
 303                                bld->bld_base.base.zero);
 304
 305       if (inst->Predicate.Negate) {
 306          pred = LLVMBuildNot(builder, pred, "");
 307       }
 308
 309       pred = bld->bld_base.emit_swizzle(&bld->bld_base, pred,
 310                          inst->Predicate.SwizzleX,
 311                          inst->Predicate.SwizzleY,
 312                          inst->Predicate.SwizzleZ,
 313                          inst->Predicate.SwizzleW);
 314
 315       if (mask) {
 316          mask = LLVMBuildAnd(builder, mask, pred, "");
 317       } else {
 318          mask = pred;
 319       }
 320    }
 321
 322    /*
 323     * Writemask
 324     */
 325
 326    if (reg->Register.WriteMask != TGSI_WRITEMASK_XYZW) {
 327       LLVMValueRef writemask;
 328
 329       writemask = lp_build_const_mask_aos_swizzled(bld->bld_base.base.gallivm,
 330                                                    bld->bld_base.base.type,
 331                                                    reg->Register.WriteMask,
 332                                                    TGSI_NUM_CHANNELS,
 333                                                    bld->swizzles);
 334
 335       if (mask) {
 336          mask = LLVMBuildAnd(builder, mask, writemask, "");
 337       } else {
 338          mask = writemask;
 339       }
 340    }
 341
 342    if (mask) {
 343       LLVMValueRef orig_value;
 344
 345       orig_value = LLVMBuildLoad(builder, ptr, "");
 346       value = lp_build_select(&bld->bld_base.base,
 347                               mask, value, orig_value);
 348    }
 349
 350    LLVMBuildStore(builder, value, ptr);
 351 }
 352
 353
 354 /**
 355  * High-level instruction translators.
 356  */
 357
 358 static LLVMValueRef
 359 emit_tex(struct lp_build_tgsi_aos_context *bld,
 360          const struct tgsi_full_instruction *inst,
 361          enum lp_build_tex_modifier modifier)
 362 {
 363    unsigned target;
 364    unsigned unit;
 365    LLVMValueRef coords;
 366    struct lp_derivatives derivs = { {NULL}, {NULL} };
 367
 368    if (!bld->sampler) {
 369       _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
 370       return bld->bld_base.base.undef;
 371    }
 372
 373    target = inst->Texture.Texture;
 374
 375    coords = lp_build_emit_fetch( &bld->bld_base, inst, 0 , LP_CHAN_ALL);
 376
 377    if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV) {
 378       /* probably not going to work */
 379       derivs.ddx[0] = lp_build_emit_fetch( &bld->bld_base, inst, 1 , LP_CHAN_ALL);
 380       derivs.ddy[0] = lp_build_emit_fetch( &bld->bld_base, inst, 2 , LP_CHAN_ALL);
 381       unit = inst->Src[3].Register.Index;
 382    }
 383    else {
 384       unit = inst->Src[1].Register.Index;
 385    }
 386    return bld->sampler->emit_fetch_texel(bld->sampler,
 387                                          &bld->bld_base.base,
 388                                          target, unit,
 389                                          coords, derivs,
 390                                          modifier);
 391 }
 392
 393
 394 static LLVMValueRef
 395 emit_sample(struct lp_build_tgsi_aos_context *bld,
 396             const struct tgsi_full_instruction *inst,
 397             enum lp_build_tex_modifier modifier)
 398 {
 399    unsigned target;
 400    unsigned unit;
 401    LLVMValueRef coords;
 402    struct lp_derivatives derivs = { {NULL}, {NULL} };
 403
 404    if (!bld->sampler) {
 405       _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
 406       return bld->bld_base.base.undef;
 407    }
 408
 409    coords = lp_build_emit_fetch( &bld->bld_base, inst, 0 , LP_CHAN_ALL);
 410
 411    /* ignore modifiers, can't handle different sampler / sampler view, etc... */
 412    unit = inst->Src[1].Register.Index;
 413    assert(inst->Src[2].Register.Index == unit);
 414
 415    target = bld->sv[unit].Resource;
 416
 417    return bld->sampler->emit_fetch_texel(bld->sampler,
 418                                          &bld->bld_base.base,
 419                                          target, unit,
 420                                          coords, derivs,
 421                                          modifier);
 422 }
 423
 424
 425 void
 426 lp_emit_declaration_aos(
 427    struct lp_build_tgsi_aos_context *bld,
 428    const struct tgsi_full_declaration *decl)
 429 {
 430    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
 431    LLVMTypeRef vec_type = lp_build_vec_type(bld->bld_base.base.gallivm, bld->bld_base.base.type);
 432
 433    unsigned first = decl->Range.First;
 434    unsigned last = decl->Range.Last;
 435    unsigned idx;
 436
 437    for (idx = first; idx <= last; ++idx) {
 438       switch (decl->Declaration.File) {
 439       case TGSI_FILE_TEMPORARY:
 440          assert(idx < LP_MAX_INLINED_TEMPS);
 441          if (bld->indirect_files & (1 << TGSI_FILE_TEMPORARY)) {
 442             LLVMValueRef array_size = lp_build_const_int32(gallivm, last + 1);
 443             bld->temps_array = lp_build_array_alloca(bld->bld_base.base.gallivm,
 444                                                      vec_type, array_size, "");
 445          } else {
 446             bld->temps[idx] = lp_build_alloca(gallivm, vec_type, "");
 447          }
 448          break;
 449
 450       case TGSI_FILE_OUTPUT:
 451          bld->outputs[idx] = lp_build_alloca(gallivm, vec_type, "");
 452          break;
 453
 454       case TGSI_FILE_ADDRESS:
 455          assert(idx < LP_MAX_TGSI_ADDRS);
 456          bld->addr[idx] = lp_build_alloca(gallivm, vec_type, "");
 457          break;
 458
 459       case TGSI_FILE_PREDICATE:
 460          assert(idx < LP_MAX_TGSI_PREDS);
 461          bld->preds[idx] = lp_build_alloca(gallivm, vec_type, "");
 462          break;
 463
 464       case TGSI_FILE_SAMPLER_VIEW:
 465          /*
 466           * The target stored here MUST match whatever there actually
 467           * is in the set sampler views (what about return type?).
 468           */
 469          assert(last < PIPE_MAX_SHADER_SAMPLER_VIEWS);
 470          for (idx = first; idx <= last; ++idx) {
 471             bld->sv[idx] = decl->SamplerView;
 472          }
 473          break;
 474
 475       default:
 476          /* don't need to declare other vars */
 477          break;
 478       }
 479    }
 480 }
 481
 482
 483 /**
 484  * Emit LLVM for one TGSI instruction.
 485  * \param return TRUE for success, FALSE otherwise
 486  */
 487 boolean
 488 lp_emit_instruction_aos(
 489    struct lp_build_tgsi_aos_context *bld,
 490    const struct tgsi_full_instruction *inst,
 491    const struct tgsi_opcode_info *info,
 492    int *pc)
 493 {
 494    LLVMValueRef src0, src1, src2;
 495    LLVMValueRef tmp0, tmp1;
 496    LLVMValueRef dst0 = NULL;
 497
 498    /*
 499     * Stores and write masks are handled in a general fashion after the long
 500     * instruction opcode switch statement.
 501     *
 502     * Although not stricitly necessary, we avoid generating instructions for
 503     * channels which won't be stored, in cases where's that easy. For some
 504     * complex instructions, like texture sampling, it is more convenient to
 505     * assume a full writemask and then let LLVM optimization passes eliminate
 506     * redundant code.
 507     */
 508
 509    (*pc)++;
 510
 511    assert(info->num_dst <= 1);
 512    if (info->num_dst) {
 513       dst0 = bld->bld_base.base.undef;
 514    }
 515
 516    switch (inst->Instruction.Opcode) {
 517    case TGSI_OPCODE_ARL:
 518       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 519       dst0 = lp_build_floor(&bld->bld_base.base, src0);
 520       break;
 521
 522    case TGSI_OPCODE_MOV:
 523       dst0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 524       break;
 525
 526    case TGSI_OPCODE_LIT:
 527       return FALSE;
 528
 529    case TGSI_OPCODE_RCP:
 530    /* TGSI_OPCODE_RECIP */
 531       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 532       dst0 = lp_build_rcp(&bld->bld_base.base, src0);
 533       break;
 534
 535    case TGSI_OPCODE_RSQ:
 536    /* TGSI_OPCODE_RECIPSQRT */
 537       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 538       tmp0 = lp_build_emit_llvm_unary(&bld->bld_base, TGSI_OPCODE_ABS, src0);
 539       dst0 = lp_build_rsqrt(&bld->bld_base.base, tmp0);
 540       break;
 541
 542    case TGSI_OPCODE_EXP:
 543       return FALSE;
 544
 545    case TGSI_OPCODE_LOG:
 546       return FALSE;
 547
 548    case TGSI_OPCODE_MUL:
 549       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 550       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 551       dst0 = lp_build_mul(&bld->bld_base.base, src0, src1);
 552       break;
 553
 554    case TGSI_OPCODE_ADD:
 555       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 556       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 557       dst0 = lp_build_add(&bld->bld_base.base, src0, src1);
 558       break;
 559
 560    case TGSI_OPCODE_DP3:
 561    /* TGSI_OPCODE_DOT3 */
 562       return FALSE;
 563
 564    case TGSI_OPCODE_DP4:
 565    /* TGSI_OPCODE_DOT4 */
 566       return FALSE;
 567
 568    case TGSI_OPCODE_DST:
 569       return FALSE;
 570
 571    case TGSI_OPCODE_MIN:
 572       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 573       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 574       dst0 = lp_build_min(&bld->bld_base.base, src0, src1);
 575       break;
 576
 577    case TGSI_OPCODE_MAX:
 578       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 579       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 580       dst0 = lp_build_max(&bld->bld_base.base, src0, src1);
 581       break;
 582
 583    case TGSI_OPCODE_SLT:
 584    /* TGSI_OPCODE_SETLT */
 585       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 586       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 587       tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_LESS, src0, src1);
 588       dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero);
 589       break;
 590
 591    case TGSI_OPCODE_SGE:
 592    /* TGSI_OPCODE_SETGE */
 593       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 594       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 595       tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_GEQUAL, src0, src1);
 596       dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero);
 597       break;
 598
 599    case TGSI_OPCODE_MAD:
 600    /* TGSI_OPCODE_MADD */
 601       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 602       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 603       src2 = lp_build_emit_fetch(&bld->bld_base, inst, 2, LP_CHAN_ALL);
 604       tmp0 = lp_build_mul(&bld->bld_base.base, src0, src1);
 605       dst0 = lp_build_add(&bld->bld_base.base, tmp0, src2);
 606       break;
 607
 608    case TGSI_OPCODE_SUB:
 609       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 610       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 611       dst0 = lp_build_sub(&bld->bld_base.base, src0, src1);
 612       break;
 613
 614    case TGSI_OPCODE_LRP:
 615       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 616       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 617       src2 = lp_build_emit_fetch(&bld->bld_base, inst, 2, LP_CHAN_ALL);
 618       tmp0 = lp_build_sub(&bld->bld_base.base, src1, src2);
 619       tmp0 = lp_build_mul(&bld->bld_base.base, src0, tmp0);
 620       dst0 = lp_build_add(&bld->bld_base.base, tmp0, src2);
 621       break;
 622
 623    case TGSI_OPCODE_CND:
 624       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 625       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 626       src2 = lp_build_emit_fetch(&bld->bld_base, inst, 2, LP_CHAN_ALL);
 627       tmp1 = lp_build_const_vec(bld->bld_base.base.gallivm, bld->bld_base.base.type, 0.5);
 628       tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_GREATER, src2, tmp1);
 629       dst0 = lp_build_select(&bld->bld_base.base, tmp0, src0, src1);
 630       break;
 631
 632    case TGSI_OPCODE_DP2A:
 633       return FALSE;
 634
 635    case TGSI_OPCODE_FRC:
 636       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 637       tmp0 = lp_build_floor(&bld->bld_base.base, src0);
 638       dst0 = lp_build_sub(&bld->bld_base.base, src0, tmp0);
 639       break;
 640
 641    case TGSI_OPCODE_CLAMP:
 642       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 643       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 644       src2 = lp_build_emit_fetch(&bld->bld_base, inst, 2, LP_CHAN_ALL);
 645       tmp0 = lp_build_max(&bld->bld_base.base, src0, src1);
 646       dst0 = lp_build_min(&bld->bld_base.base, tmp0, src2);
 647       break;
 648
 649    case TGSI_OPCODE_FLR:
 650       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 651       dst0 = lp_build_floor(&bld->bld_base.base, src0);
 652       break;
 653
 654    case TGSI_OPCODE_ROUND:
 655       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 656       dst0 = lp_build_round(&bld->bld_base.base, src0);
 657       break;
 658
 659    case TGSI_OPCODE_EX2:
 660       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 661       tmp0 = lp_build_swizzle_scalar_aos(&bld->bld_base.base, src0, TGSI_SWIZZLE_X, TGSI_NUM_CHANNELS);
 662       dst0 = lp_build_exp2(&bld->bld_base.base, tmp0);
 663       break;
 664
 665    case TGSI_OPCODE_LG2:
 666       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 667       tmp0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
 668       dst0 = lp_build_log2(&bld->bld_base.base, tmp0);
 669       break;
 670
 671    case TGSI_OPCODE_POW:
 672       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 673       src0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
 674       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 675       src1 = swizzle_scalar_aos(bld, src1, TGSI_SWIZZLE_X);
 676       dst0 = lp_build_pow(&bld->bld_base.base, src0, src1);
 677       break;
 678
 679    case TGSI_OPCODE_XPD:
 680       return FALSE;
 681
 682    case TGSI_OPCODE_DPH:
 683       return FALSE;
 684
 685    case TGSI_OPCODE_COS:
 686       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 687       tmp0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
 688       dst0 = lp_build_cos(&bld->bld_base.base, tmp0);
 689       break;
 690
 691    case TGSI_OPCODE_DDX:
 692       return FALSE;
 693
 694    case TGSI_OPCODE_DDY:
 695       return FALSE;
 696
 697    case TGSI_OPCODE_KILL:
 698       return FALSE;
 699
 700    case TGSI_OPCODE_KILL_IF:
 701       return FALSE;
 702
 703    case TGSI_OPCODE_PK2H:
 704       return FALSE;
 705       break;
 706
 707    case TGSI_OPCODE_PK2US:
 708       return FALSE;
 709       break;
 710
 711    case TGSI_OPCODE_PK4B:
 712       return FALSE;
 713       break;
 714
 715    case TGSI_OPCODE_PK4UB:
 716       return FALSE;
 717
 718    case TGSI_OPCODE_SEQ:
 719       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 720       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 721       tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_EQUAL, src0, src1);
 722       dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero);
 723       break;
 724
 725    case TGSI_OPCODE_SGT:
 726       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 727       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 728       tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_GREATER, src0, src1);
 729       dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero);
 730       break;
 731
 732    case TGSI_OPCODE_SIN:
 733       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 734       tmp0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
 735       dst0 = lp_build_sin(&bld->bld_base.base, tmp0);
 736       break;
 737
 738    case TGSI_OPCODE_SLE:
 739       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 740       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 741       tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_LEQUAL, src0, src1);
 742       dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero);
 743       break;
 744
 745    case TGSI_OPCODE_SNE:
 746       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 747       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 748       tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_NOTEQUAL, src0, src1);
 749       dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero);
 750       break;
 751
 752    case TGSI_OPCODE_TEX:
 753       dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_NONE);
 754       break;
 755
 756    case TGSI_OPCODE_TXD:
 757       dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV);
 758       break;
 759
 760    case TGSI_OPCODE_UP2H:
 761       /* deprecated */
 762       assert (0);
 763       return FALSE;
 764       break;
 765
 766    case TGSI_OPCODE_UP2US:
 767       /* deprecated */
 768       assert(0);
 769       return FALSE;
 770       break;
 771
 772    case TGSI_OPCODE_UP4B:
 773       /* deprecated */
 774       assert(0);
 775       return FALSE;
 776       break;
 777
 778    case TGSI_OPCODE_UP4UB:
 779       /* deprecated */
 780       assert(0);
 781       return FALSE;
 782       break;
 783
 784    case TGSI_OPCODE_ARR:
 785       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 786       dst0 = lp_build_round(&bld->bld_base.base, src0);
 787       break;
 788
 789    case TGSI_OPCODE_CAL:
 790       return FALSE;
 791
 792    case TGSI_OPCODE_RET:
 793       /* safe to ignore at end */
 794       break;
 795
 796    case TGSI_OPCODE_END:
 797       *pc = -1;
 798       break;
 799
 800    case TGSI_OPCODE_SSG:
 801    /* TGSI_OPCODE_SGN */
 802       tmp0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 803       dst0 = lp_build_sgn(&bld->bld_base.base, tmp0);
 804       break;
 805
 806    case TGSI_OPCODE_CMP:
 807       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 808       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 809       src2 = lp_build_emit_fetch(&bld->bld_base, inst, 2, LP_CHAN_ALL);
 810       tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_LESS, src0, bld->bld_base.base.zero);
 811       dst0 = lp_build_select(&bld->bld_base.base, tmp0, src1, src2);
 812       break;
 813
 814    case TGSI_OPCODE_SCS:
 815       return FALSE;
 816
 817    case TGSI_OPCODE_TXB:
 818       dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_LOD_BIAS);
 819       break;
 820
 821    case TGSI_OPCODE_DIV:
 822       assert(0);
 823       return FALSE;
 824       break;
 825
 826    case TGSI_OPCODE_DP2:
 827       return FALSE;
 828
 829    case TGSI_OPCODE_TXL:
 830       dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD);
 831       break;
 832
 833    case TGSI_OPCODE_TXP:
 834       dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_PROJECTED);
 835       break;
 836
 837    case TGSI_OPCODE_BRK:
 838       return FALSE;
 839
 840    case TGSI_OPCODE_IF:
 841    case TGSI_OPCODE_UIF:
 842       return FALSE;
 843
 844    case TGSI_OPCODE_BGNLOOP:
 845       return FALSE;
 846
 847    case TGSI_OPCODE_BGNSUB:
 848       return FALSE;
 849
 850    case TGSI_OPCODE_ELSE:
 851       return FALSE;
 852
 853    case TGSI_OPCODE_ENDIF:
 854       return FALSE;
 855
 856    case TGSI_OPCODE_ENDLOOP:
 857       return FALSE;
 858
 859    case TGSI_OPCODE_ENDSUB:
 860       return FALSE;
 861
 862    case TGSI_OPCODE_PUSHA:
 863       /* deprecated? */
 864       assert(0);
 865       return FALSE;
 866       break;
 867
 868    case TGSI_OPCODE_POPA:
 869       /* deprecated? */
 870       assert(0);
 871       return FALSE;
 872       break;
 873
 874    case TGSI_OPCODE_CEIL:
 875       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 876       dst0 = lp_build_ceil(&bld->bld_base.base, src0);
 877       break;
 878
 879    case TGSI_OPCODE_I2F:
 880       assert(0);
 881       return FALSE;
 882       break;
 883
 884    case TGSI_OPCODE_NOT:
 885       assert(0);
 886       return FALSE;
 887       break;
 888
 889    case TGSI_OPCODE_TRUNC:
 890       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 891       dst0 = lp_build_trunc(&bld->bld_base.base, src0);
 892       break;
 893
 894    case TGSI_OPCODE_SHL:
 895       assert(0);
 896       return FALSE;
 897       break;
 898
 899    case TGSI_OPCODE_ISHR:
 900       assert(0);
 901       return FALSE;
 902       break;
 903
 904    case TGSI_OPCODE_AND:
 905       assert(0);
 906       return FALSE;
 907       break;
 908
 909    case TGSI_OPCODE_OR:
 910       assert(0);
 911       return FALSE;
 912       break;
 913
 914    case TGSI_OPCODE_MOD:
 915       assert(0);
 916       return FALSE;
 917       break;
 918
 919    case TGSI_OPCODE_XOR:
 920       assert(0);
 921       return FALSE;
 922       break;
 923
 924    case TGSI_OPCODE_SAD:
 925       assert(0);
 926       return FALSE;
 927       break;
 928
 929    case TGSI_OPCODE_TXF:
 930       assert(0);
 931       return FALSE;
 932       break;
 933
 934    case TGSI_OPCODE_TXQ:
 935       assert(0);
 936       return FALSE;
 937       break;
 938
 939    case TGSI_OPCODE_CONT:
 940       return FALSE;
 941
 942    case TGSI_OPCODE_EMIT:
 943       return FALSE;
 944       break;
 945
 946    case TGSI_OPCODE_ENDPRIM:
 947       return FALSE;
 948       break;
 949
 950    case TGSI_OPCODE_NOP:
 951       break;
 952
 953    case TGSI_OPCODE_SAMPLE:
 954       dst0 = emit_sample(bld, inst, LP_BLD_TEX_MODIFIER_NONE);
 955       break;
 956
 957    default:
 958       return FALSE;
 959    }
 960
 961    if (info->num_dst) {
 962       lp_emit_store_aos(bld, inst, 0, dst0);
 963    }
 964
 965    return TRUE;
 966 }
 967
 968
 969 void
 970 lp_build_tgsi_aos(struct gallivm_state *gallivm,
 971                   const struct tgsi_token *tokens,
 972                   struct lp_type type,
 973                   const unsigned char swizzles[4],
 974                   LLVMValueRef consts_ptr,
 975                   const LLVMValueRef *inputs,
 976                   LLVMValueRef *outputs,
 977                   struct lp_build_sampler_aos *sampler,
 978                   const struct tgsi_shader_info *info)
 979 {
 980    struct lp_build_tgsi_aos_context bld;
 981    struct tgsi_parse_context parse;
 982    uint num_immediates = 0;
 983    unsigned chan;
 984    int pc = 0;
 985
 986    /* Setup build context */
 987    memset(&bld, 0, sizeof bld);
 988    lp_build_context_init(&bld.bld_base.base, gallivm, type);
 989    lp_build_context_init(&bld.bld_base.uint_bld, gallivm, lp_uint_type(type));
 990    lp_build_context_init(&bld.bld_base.int_bld, gallivm, lp_int_type(type));
 991    lp_build_context_init(&bld.int_bld, gallivm, lp_int_type(type));
 992
 993    for (chan = 0; chan < 4; ++chan) {
 994       bld.swizzles[chan] = swizzles[chan];
 995       bld.inv_swizzles[swizzles[chan]] = chan;
 996    }
 997
 998    bld.inputs = inputs;
 999    bld.outputs = outputs;
1000    bld.consts_ptr = consts_ptr;
1001    bld.sampler = sampler;
1002    bld.indirect_files = info->indirect_files;
1003    bld.bld_base.emit_swizzle = swizzle_aos;
1004    bld.bld_base.info = info;
1005
1006    bld.bld_base.emit_fetch_funcs[TGSI_FILE_CONSTANT] = emit_fetch_constant;
1007    bld.bld_base.emit_fetch_funcs[TGSI_FILE_IMMEDIATE] = emit_fetch_immediate;
1008    bld.bld_base.emit_fetch_funcs[TGSI_FILE_INPUT] = emit_fetch_input;
1009    bld.bld_base.emit_fetch_funcs[TGSI_FILE_TEMPORARY] = emit_fetch_temporary;
1010
1011    /* Set opcode actions */
1012    lp_set_default_actions_cpu(&bld.bld_base);
1013
1014    if (!lp_bld_tgsi_list_init(&bld.bld_base)) {
1015       return;
1016    }
1017
1018    tgsi_parse_init(&parse, tokens);
1019
1020    while (!tgsi_parse_end_of_tokens(&parse)) {
1021       tgsi_parse_token(&parse);
1022
1023       switch(parse.FullToken.Token.Type) {
1024       case TGSI_TOKEN_TYPE_DECLARATION:
1025          /* Inputs already interpolated */
1026          lp_emit_declaration_aos(&bld, &parse.FullToken.FullDeclaration);
1027          break;
1028
1029       case TGSI_TOKEN_TYPE_INSTRUCTION:
1030          /* save expanded instruction */
1031          lp_bld_tgsi_add_instruction(&bld.bld_base,
1032                                      &parse.FullToken.FullInstruction);
1033          break;
1034
1035       case TGSI_TOKEN_TYPE_IMMEDIATE:
1036          /* simply copy the immediate values into the next immediates[] slot */
1037          {
1038             const uint size = parse.FullToken.FullImmediate.Immediate.NrTokens - 1;
1039             float imm[4];
1040             assert(size <= 4);
1041             assert(num_immediates < LP_MAX_INLINED_IMMEDIATES);
1042             for (chan = 0; chan < 4; ++chan) {
1043                imm[chan] = 0.0f;
1044             }
1045             for (chan = 0; chan < size; ++chan) {
1046                unsigned swizzle = bld.swizzles[chan];
1047                imm[swizzle] = parse.FullToken.FullImmediate.u[chan].Float;
1048             }
1049             bld.immediates[num_immediates] =
1050                      lp_build_const_aos(gallivm, type,
1051                                         imm[0], imm[1], imm[2], imm[3],
1052                                         NULL);
1053             num_immediates++;
1054          }
1055          break;
1056
1057       case TGSI_TOKEN_TYPE_PROPERTY:
1058          break;
1059
1060       default:
1061          assert(0);
1062       }
1063    }
1064
1065    while (pc != -1) {
1066       struct tgsi_full_instruction *instr = bld.bld_base.instructions + pc;
1067       const struct tgsi_opcode_info *opcode_info =
1068          tgsi_get_opcode_info(instr->Instruction.Opcode);
1069       if (!lp_emit_instruction_aos(&bld, instr, opcode_info, &pc))
1070          _debug_printf("warning: failed to translate tgsi opcode %s to LLVM\n",
1071                        opcode_info->mnemonic);
1072    }
1073
1074    if (0) {
1075       LLVMBasicBlockRef block = LLVMGetInsertBlock(gallivm->builder);
1076       LLVMValueRef function = LLVMGetBasicBlockParent(block);
1077       debug_printf("11111111111111111111111111111 \n");
1078       tgsi_dump(tokens, 0);
1079       lp_debug_dump_value(function);
1080       debug_printf("2222222222222222222222222222 \n");
1081    }
1082    tgsi_parse_free(&parse);
1083    FREE(bld.bld_base.instructions);
1084
1085    if (0) {
1086       LLVMModuleRef module = LLVMGetGlobalParent(
1087          LLVMGetBasicBlockParent(LLVMGetInsertBlock(gallivm->builder)));
1088       LLVMDumpModule(module);
1089    }
1090
1091 }
1092