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    if (inst->Instruction.Saturate) {
 236       value = lp_build_max(&bld->bld_base.base, value, bld->bld_base.base.zero);
 237       value = lp_build_min(&bld->bld_base.base, value, bld->bld_base.base.one);
 238    }
 239
 240    /*
 241     * Translate the register file
 242     */
 243
 244    assert(!reg->Register.Indirect);
 245
 246    switch (reg->Register.File) {
 247    case TGSI_FILE_OUTPUT:
 248       ptr = bld->outputs[reg->Register.Index];
 249       break;
 250
 251    case TGSI_FILE_TEMPORARY:
 252       ptr = bld->temps[reg->Register.Index];
 253       break;
 254
 255    case TGSI_FILE_ADDRESS:
 256       ptr = bld->addr[reg->Indirect.Index];
 257       break;
 258
 259    case TGSI_FILE_PREDICATE:
 260       ptr = bld->preds[reg->Register.Index];
 261       break;
 262
 263    default:
 264       assert(0);
 265       return;
 266    }
 267
 268    if (!ptr)
 269       return;
 270    /*
 271     * Predicate
 272     */
 273
 274    if (inst->Instruction.Predicate) {
 275       LLVMValueRef pred;
 276
 277       assert(inst->Predicate.Index < LP_MAX_TGSI_PREDS);
 278
 279       pred = LLVMBuildLoad(builder,
 280                            bld->preds[inst->Predicate.Index], "");
 281
 282       /*
 283        * Convert the value to an integer mask.
 284        */
 285       pred = lp_build_compare(bld->bld_base.base.gallivm,
 286                                bld->bld_base.base.type,
 287                                PIPE_FUNC_NOTEQUAL,
 288                                pred,
 289                                bld->bld_base.base.zero);
 290
 291       if (inst->Predicate.Negate) {
 292          pred = LLVMBuildNot(builder, pred, "");
 293       }
 294
 295       pred = bld->bld_base.emit_swizzle(&bld->bld_base, pred,
 296                          inst->Predicate.SwizzleX,
 297                          inst->Predicate.SwizzleY,
 298                          inst->Predicate.SwizzleZ,
 299                          inst->Predicate.SwizzleW);
 300
 301       if (mask) {
 302          mask = LLVMBuildAnd(builder, mask, pred, "");
 303       } else {
 304          mask = pred;
 305       }
 306    }
 307
 308    /*
 309     * Writemask
 310     */
 311
 312    if (reg->Register.WriteMask != TGSI_WRITEMASK_XYZW) {
 313       LLVMValueRef writemask;
 314
 315       writemask = lp_build_const_mask_aos_swizzled(bld->bld_base.base.gallivm,
 316                                                    bld->bld_base.base.type,
 317                                                    reg->Register.WriteMask,
 318                                                    TGSI_NUM_CHANNELS,
 319                                                    bld->swizzles);
 320
 321       if (mask) {
 322          mask = LLVMBuildAnd(builder, mask, writemask, "");
 323       } else {
 324          mask = writemask;
 325       }
 326    }
 327
 328    if (mask) {
 329       LLVMValueRef orig_value;
 330
 331       orig_value = LLVMBuildLoad(builder, ptr, "");
 332       value = lp_build_select(&bld->bld_base.base,
 333                               mask, value, orig_value);
 334    }
 335
 336    LLVMBuildStore(builder, value, ptr);
 337 }
 338
 339
 340 /**
 341  * High-level instruction translators.
 342  */
 343
 344 static LLVMValueRef
 345 emit_tex(struct lp_build_tgsi_aos_context *bld,
 346          const struct tgsi_full_instruction *inst,
 347          enum lp_build_tex_modifier modifier)
 348 {
 349    unsigned target;
 350    unsigned unit;
 351    LLVMValueRef coords;
 352    struct lp_derivatives derivs = { {NULL}, {NULL} };
 353
 354    if (!bld->sampler) {
 355       _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
 356       return bld->bld_base.base.undef;
 357    }
 358
 359    target = inst->Texture.Texture;
 360
 361    coords = lp_build_emit_fetch( &bld->bld_base, inst, 0 , LP_CHAN_ALL);
 362
 363    if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV) {
 364       /* probably not going to work */
 365       derivs.ddx[0] = lp_build_emit_fetch( &bld->bld_base, inst, 1 , LP_CHAN_ALL);
 366       derivs.ddy[0] = lp_build_emit_fetch( &bld->bld_base, inst, 2 , LP_CHAN_ALL);
 367       unit = inst->Src[3].Register.Index;
 368    }
 369    else {
 370       unit = inst->Src[1].Register.Index;
 371    }
 372    return bld->sampler->emit_fetch_texel(bld->sampler,
 373                                          &bld->bld_base.base,
 374                                          target, unit,
 375                                          coords, derivs,
 376                                          modifier);
 377 }
 378
 379
 380 static LLVMValueRef
 381 emit_sample(struct lp_build_tgsi_aos_context *bld,
 382             const struct tgsi_full_instruction *inst,
 383             enum lp_build_tex_modifier modifier)
 384 {
 385    unsigned target;
 386    unsigned unit;
 387    LLVMValueRef coords;
 388    struct lp_derivatives derivs = { {NULL}, {NULL} };
 389
 390    if (!bld->sampler) {
 391       _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
 392       return bld->bld_base.base.undef;
 393    }
 394
 395    coords = lp_build_emit_fetch( &bld->bld_base, inst, 0 , LP_CHAN_ALL);
 396
 397    /* ignore modifiers, can't handle different sampler / sampler view, etc... */
 398    unit = inst->Src[1].Register.Index;
 399    assert(inst->Src[2].Register.Index == unit);
 400
 401    target = bld->sv[unit].Resource;
 402
 403    return bld->sampler->emit_fetch_texel(bld->sampler,
 404                                          &bld->bld_base.base,
 405                                          target, unit,
 406                                          coords, derivs,
 407                                          modifier);
 408 }
 409
 410
 411 void
 412 lp_emit_declaration_aos(
 413    struct lp_build_tgsi_aos_context *bld,
 414    const struct tgsi_full_declaration *decl)
 415 {
 416    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
 417    LLVMTypeRef vec_type = lp_build_vec_type(bld->bld_base.base.gallivm, bld->bld_base.base.type);
 418
 419    unsigned first = decl->Range.First;
 420    unsigned last = decl->Range.Last;
 421    unsigned idx;
 422
 423    for (idx = first; idx <= last; ++idx) {
 424       switch (decl->Declaration.File) {
 425       case TGSI_FILE_TEMPORARY:
 426          assert(idx < LP_MAX_INLINED_TEMPS);
 427          if (bld->indirect_files & (1 << TGSI_FILE_TEMPORARY)) {
 428             LLVMValueRef array_size = lp_build_const_int32(gallivm, last + 1);
 429             bld->temps_array = lp_build_array_alloca(bld->bld_base.base.gallivm,
 430                                                      vec_type, array_size, "");
 431          } else {
 432             bld->temps[idx] = lp_build_alloca(gallivm, vec_type, "");
 433          }
 434          break;
 435
 436       case TGSI_FILE_OUTPUT:
 437          bld->outputs[idx] = lp_build_alloca(gallivm, vec_type, "");
 438          break;
 439
 440       case TGSI_FILE_ADDRESS:
 441          assert(idx < LP_MAX_TGSI_ADDRS);
 442          bld->addr[idx] = lp_build_alloca(gallivm, vec_type, "");
 443          break;
 444
 445       case TGSI_FILE_PREDICATE:
 446          assert(idx < LP_MAX_TGSI_PREDS);
 447          bld->preds[idx] = lp_build_alloca(gallivm, vec_type, "");
 448          break;
 449
 450       case TGSI_FILE_SAMPLER_VIEW:
 451          /*
 452           * The target stored here MUST match whatever there actually
 453           * is in the set sampler views (what about return type?).
 454           */
 455          assert(last < PIPE_MAX_SHADER_SAMPLER_VIEWS);
 456          for (idx = first; idx <= last; ++idx) {
 457             bld->sv[idx] = decl->SamplerView;
 458          }
 459          break;
 460
 461       default:
 462          /* don't need to declare other vars */
 463          break;
 464       }
 465    }
 466 }
 467
 468
 469 /**
 470  * Emit LLVM for one TGSI instruction.
 471  * \param return TRUE for success, FALSE otherwise
 472  */
 473 boolean
 474 lp_emit_instruction_aos(
 475    struct lp_build_tgsi_aos_context *bld,
 476    const struct tgsi_full_instruction *inst,
 477    const struct tgsi_opcode_info *info,
 478    int *pc)
 479 {
 480    LLVMValueRef src0, src1, src2;
 481    LLVMValueRef tmp0;
 482    LLVMValueRef dst0 = NULL;
 483
 484    /*
 485     * Stores and write masks are handled in a general fashion after the long
 486     * instruction opcode switch statement.
 487     *
 488     * Although not stricitly necessary, we avoid generating instructions for
 489     * channels which won't be stored, in cases where's that easy. For some
 490     * complex instructions, like texture sampling, it is more convenient to
 491     * assume a full writemask and then let LLVM optimization passes eliminate
 492     * redundant code.
 493     */
 494
 495    (*pc)++;
 496
 497    assert(info->num_dst <= 1);
 498    if (info->num_dst) {
 499       dst0 = bld->bld_base.base.undef;
 500    }
 501
 502    switch (inst->Instruction.Opcode) {
 503    case TGSI_OPCODE_ARL:
 504       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 505       dst0 = lp_build_floor(&bld->bld_base.base, src0);
 506       break;
 507
 508    case TGSI_OPCODE_MOV:
 509       dst0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 510       break;
 511
 512    case TGSI_OPCODE_LIT:
 513       return FALSE;
 514
 515    case TGSI_OPCODE_RCP:
 516    /* TGSI_OPCODE_RECIP */
 517       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 518       dst0 = lp_build_rcp(&bld->bld_base.base, src0);
 519       break;
 520
 521    case TGSI_OPCODE_RSQ:
 522    /* TGSI_OPCODE_RECIPSQRT */
 523       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 524       tmp0 = lp_build_emit_llvm_unary(&bld->bld_base, TGSI_OPCODE_ABS, src0);
 525       dst0 = lp_build_rsqrt(&bld->bld_base.base, tmp0);
 526       break;
 527
 528    case TGSI_OPCODE_EXP:
 529       return FALSE;
 530
 531    case TGSI_OPCODE_LOG:
 532       return FALSE;
 533
 534    case TGSI_OPCODE_MUL:
 535       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 536       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 537       dst0 = lp_build_mul(&bld->bld_base.base, src0, src1);
 538       break;
 539
 540    case TGSI_OPCODE_ADD:
 541       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 542       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 543       dst0 = lp_build_add(&bld->bld_base.base, src0, src1);
 544       break;
 545
 546    case TGSI_OPCODE_DP3:
 547    /* TGSI_OPCODE_DOT3 */
 548       return FALSE;
 549
 550    case TGSI_OPCODE_DP4:
 551    /* TGSI_OPCODE_DOT4 */
 552       return FALSE;
 553
 554    case TGSI_OPCODE_DST:
 555       return FALSE;
 556
 557    case TGSI_OPCODE_MIN:
 558       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 559       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 560       dst0 = lp_build_min(&bld->bld_base.base, src0, src1);
 561       break;
 562
 563    case TGSI_OPCODE_MAX:
 564       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 565       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 566       dst0 = lp_build_max(&bld->bld_base.base, src0, src1);
 567       break;
 568
 569    case TGSI_OPCODE_SLT:
 570    /* TGSI_OPCODE_SETLT */
 571       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 572       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 573       tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_LESS, src0, src1);
 574       dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero);
 575       break;
 576
 577    case TGSI_OPCODE_SGE:
 578    /* TGSI_OPCODE_SETGE */
 579       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 580       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 581       tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_GEQUAL, src0, src1);
 582       dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero);
 583       break;
 584
 585    case TGSI_OPCODE_MAD:
 586    /* TGSI_OPCODE_MADD */
 587       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 588       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 589       src2 = lp_build_emit_fetch(&bld->bld_base, inst, 2, LP_CHAN_ALL);
 590       tmp0 = lp_build_mul(&bld->bld_base.base, src0, src1);
 591       dst0 = lp_build_add(&bld->bld_base.base, tmp0, src2);
 592       break;
 593
 594    case TGSI_OPCODE_SUB:
 595       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 596       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 597       dst0 = lp_build_sub(&bld->bld_base.base, src0, src1);
 598       break;
 599
 600    case TGSI_OPCODE_LRP:
 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_sub(&bld->bld_base.base, src1, src2);
 605       tmp0 = lp_build_mul(&bld->bld_base.base, src0, tmp0);
 606       dst0 = lp_build_add(&bld->bld_base.base, tmp0, src2);
 607       break;
 608
 609    case TGSI_OPCODE_DP2A:
 610       return FALSE;
 611
 612    case TGSI_OPCODE_FRC:
 613       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 614       tmp0 = lp_build_floor(&bld->bld_base.base, src0);
 615       dst0 = lp_build_sub(&bld->bld_base.base, src0, tmp0);
 616       break;
 617
 618    case TGSI_OPCODE_CLAMP:
 619       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 620       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 621       src2 = lp_build_emit_fetch(&bld->bld_base, inst, 2, LP_CHAN_ALL);
 622       tmp0 = lp_build_max(&bld->bld_base.base, src0, src1);
 623       dst0 = lp_build_min(&bld->bld_base.base, tmp0, src2);
 624       break;
 625
 626    case TGSI_OPCODE_FLR:
 627       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 628       dst0 = lp_build_floor(&bld->bld_base.base, src0);
 629       break;
 630
 631    case TGSI_OPCODE_ROUND:
 632       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 633       dst0 = lp_build_round(&bld->bld_base.base, src0);
 634       break;
 635
 636    case TGSI_OPCODE_EX2:
 637       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 638       tmp0 = lp_build_swizzle_scalar_aos(&bld->bld_base.base, src0, TGSI_SWIZZLE_X, TGSI_NUM_CHANNELS);
 639       dst0 = lp_build_exp2(&bld->bld_base.base, tmp0);
 640       break;
 641
 642    case TGSI_OPCODE_LG2:
 643       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 644       tmp0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
 645       dst0 = lp_build_log2(&bld->bld_base.base, tmp0);
 646       break;
 647
 648    case TGSI_OPCODE_POW:
 649       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 650       src0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
 651       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 652       src1 = swizzle_scalar_aos(bld, src1, TGSI_SWIZZLE_X);
 653       dst0 = lp_build_pow(&bld->bld_base.base, src0, src1);
 654       break;
 655
 656    case TGSI_OPCODE_XPD:
 657       return FALSE;
 658
 659    case TGSI_OPCODE_DPH:
 660       return FALSE;
 661
 662    case TGSI_OPCODE_COS:
 663       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 664       tmp0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
 665       dst0 = lp_build_cos(&bld->bld_base.base, tmp0);
 666       break;
 667
 668    case TGSI_OPCODE_DDX:
 669       return FALSE;
 670
 671    case TGSI_OPCODE_DDY:
 672       return FALSE;
 673
 674    case TGSI_OPCODE_KILL:
 675       return FALSE;
 676
 677    case TGSI_OPCODE_KILL_IF:
 678       return FALSE;
 679
 680    case TGSI_OPCODE_PK2H:
 681       return FALSE;
 682       break;
 683
 684    case TGSI_OPCODE_PK2US:
 685       return FALSE;
 686       break;
 687
 688    case TGSI_OPCODE_PK4B:
 689       return FALSE;
 690       break;
 691
 692    case TGSI_OPCODE_PK4UB:
 693       return FALSE;
 694
 695    case TGSI_OPCODE_SEQ:
 696       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 697       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 698       tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_EQUAL, src0, src1);
 699       dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero);
 700       break;
 701
 702    case TGSI_OPCODE_SGT:
 703       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 704       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 705       tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_GREATER, src0, src1);
 706       dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero);
 707       break;
 708
 709    case TGSI_OPCODE_SIN:
 710       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 711       tmp0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
 712       dst0 = lp_build_sin(&bld->bld_base.base, tmp0);
 713       break;
 714
 715    case TGSI_OPCODE_SLE:
 716       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 717       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 718       tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_LEQUAL, src0, src1);
 719       dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero);
 720       break;
 721
 722    case TGSI_OPCODE_SNE:
 723       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 724       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 725       tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_NOTEQUAL, src0, src1);
 726       dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero);
 727       break;
 728
 729    case TGSI_OPCODE_TEX:
 730       dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_NONE);
 731       break;
 732
 733    case TGSI_OPCODE_TXD:
 734       dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV);
 735       break;
 736
 737    case TGSI_OPCODE_UP2H:
 738       /* deprecated */
 739       assert (0);
 740       return FALSE;
 741       break;
 742
 743    case TGSI_OPCODE_UP2US:
 744       /* deprecated */
 745       assert(0);
 746       return FALSE;
 747       break;
 748
 749    case TGSI_OPCODE_UP4B:
 750       /* deprecated */
 751       assert(0);
 752       return FALSE;
 753       break;
 754
 755    case TGSI_OPCODE_UP4UB:
 756       /* deprecated */
 757       assert(0);
 758       return FALSE;
 759       break;
 760
 761    case TGSI_OPCODE_ARR:
 762       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 763       dst0 = lp_build_round(&bld->bld_base.base, src0);
 764       break;
 765
 766    case TGSI_OPCODE_CAL:
 767       return FALSE;
 768
 769    case TGSI_OPCODE_RET:
 770       /* safe to ignore at end */
 771       break;
 772
 773    case TGSI_OPCODE_END:
 774       *pc = -1;
 775       break;
 776
 777    case TGSI_OPCODE_SSG:
 778    /* TGSI_OPCODE_SGN */
 779       tmp0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 780       dst0 = lp_build_sgn(&bld->bld_base.base, tmp0);
 781       break;
 782
 783    case TGSI_OPCODE_CMP:
 784       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 785       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 786       src2 = lp_build_emit_fetch(&bld->bld_base, inst, 2, LP_CHAN_ALL);
 787       tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_LESS, src0, bld->bld_base.base.zero);
 788       dst0 = lp_build_select(&bld->bld_base.base, tmp0, src1, src2);
 789       break;
 790
 791    case TGSI_OPCODE_SCS:
 792       return FALSE;
 793
 794    case TGSI_OPCODE_TXB:
 795       dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_LOD_BIAS);
 796       break;
 797
 798    case TGSI_OPCODE_DIV:
 799       assert(0);
 800       return FALSE;
 801       break;
 802
 803    case TGSI_OPCODE_DP2:
 804       return FALSE;
 805
 806    case TGSI_OPCODE_TXL:
 807       dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD);
 808       break;
 809
 810    case TGSI_OPCODE_TXP:
 811       dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_PROJECTED);
 812       break;
 813
 814    case TGSI_OPCODE_BRK:
 815       return FALSE;
 816
 817    case TGSI_OPCODE_IF:
 818    case TGSI_OPCODE_UIF:
 819       return FALSE;
 820
 821    case TGSI_OPCODE_BGNLOOP:
 822       return FALSE;
 823
 824    case TGSI_OPCODE_BGNSUB:
 825       return FALSE;
 826
 827    case TGSI_OPCODE_ELSE:
 828       return FALSE;
 829
 830    case TGSI_OPCODE_ENDIF:
 831       return FALSE;
 832
 833    case TGSI_OPCODE_ENDLOOP:
 834       return FALSE;
 835
 836    case TGSI_OPCODE_ENDSUB:
 837       return FALSE;
 838
 839    case TGSI_OPCODE_PUSHA:
 840       /* deprecated? */
 841       assert(0);
 842       return FALSE;
 843       break;
 844
 845    case TGSI_OPCODE_POPA:
 846       /* deprecated? */
 847       assert(0);
 848       return FALSE;
 849       break;
 850
 851    case TGSI_OPCODE_CEIL:
 852       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 853       dst0 = lp_build_ceil(&bld->bld_base.base, src0);
 854       break;
 855
 856    case TGSI_OPCODE_I2F:
 857       assert(0);
 858       return FALSE;
 859       break;
 860
 861    case TGSI_OPCODE_NOT:
 862       assert(0);
 863       return FALSE;
 864       break;
 865
 866    case TGSI_OPCODE_TRUNC:
 867       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 868       dst0 = lp_build_trunc(&bld->bld_base.base, src0);
 869       break;
 870
 871    case TGSI_OPCODE_SHL:
 872       assert(0);
 873       return FALSE;
 874       break;
 875
 876    case TGSI_OPCODE_ISHR:
 877       assert(0);
 878       return FALSE;
 879       break;
 880
 881    case TGSI_OPCODE_AND:
 882       assert(0);
 883       return FALSE;
 884       break;
 885
 886    case TGSI_OPCODE_OR:
 887       assert(0);
 888       return FALSE;
 889       break;
 890
 891    case TGSI_OPCODE_MOD:
 892       assert(0);
 893       return FALSE;
 894       break;
 895
 896    case TGSI_OPCODE_XOR:
 897       assert(0);
 898       return FALSE;
 899       break;
 900
 901    case TGSI_OPCODE_SAD:
 902       assert(0);
 903       return FALSE;
 904       break;
 905
 906    case TGSI_OPCODE_TXF:
 907       assert(0);
 908       return FALSE;
 909       break;
 910
 911    case TGSI_OPCODE_TXQ:
 912       assert(0);
 913       return FALSE;
 914       break;
 915
 916    case TGSI_OPCODE_CONT:
 917       return FALSE;
 918
 919    case TGSI_OPCODE_EMIT:
 920       return FALSE;
 921       break;
 922
 923    case TGSI_OPCODE_ENDPRIM:
 924       return FALSE;
 925       break;
 926
 927    case TGSI_OPCODE_NOP:
 928       break;
 929
 930    case TGSI_OPCODE_SAMPLE:
 931       dst0 = emit_sample(bld, inst, LP_BLD_TEX_MODIFIER_NONE);
 932       break;
 933
 934    default:
 935       return FALSE;
 936    }
 937
 938    if (info->num_dst) {
 939       lp_emit_store_aos(bld, inst, 0, dst0);
 940    }
 941
 942    return TRUE;
 943 }
 944
 945
 946 void
 947 lp_build_tgsi_aos(struct gallivm_state *gallivm,
 948                   const struct tgsi_token *tokens,
 949                   struct lp_type type,
 950                   const unsigned char swizzles[4],
 951                   LLVMValueRef consts_ptr,
 952                   const LLVMValueRef *inputs,
 953                   LLVMValueRef *outputs,
 954                   struct lp_build_sampler_aos *sampler,
 955                   const struct tgsi_shader_info *info)
 956 {
 957    struct lp_build_tgsi_aos_context bld;
 958    struct tgsi_parse_context parse;
 959    uint num_immediates = 0;
 960    unsigned chan;
 961    int pc = 0;
 962
 963    /* Setup build context */
 964    memset(&bld, 0, sizeof bld);
 965    lp_build_context_init(&bld.bld_base.base, gallivm, type);
 966    lp_build_context_init(&bld.bld_base.uint_bld, gallivm, lp_uint_type(type));
 967    lp_build_context_init(&bld.bld_base.int_bld, gallivm, lp_int_type(type));
 968    lp_build_context_init(&bld.int_bld, gallivm, lp_int_type(type));
 969
 970    for (chan = 0; chan < 4; ++chan) {
 971       bld.swizzles[chan] = swizzles[chan];
 972       bld.inv_swizzles[swizzles[chan]] = chan;
 973    }
 974
 975    bld.inputs = inputs;
 976    bld.outputs = outputs;
 977    bld.consts_ptr = consts_ptr;
 978    bld.sampler = sampler;
 979    bld.indirect_files = info->indirect_files;
 980    bld.bld_base.emit_swizzle = swizzle_aos;
 981    bld.bld_base.info = info;
 982
 983    bld.bld_base.emit_fetch_funcs[TGSI_FILE_CONSTANT] = emit_fetch_constant;
 984    bld.bld_base.emit_fetch_funcs[TGSI_FILE_IMMEDIATE] = emit_fetch_immediate;
 985    bld.bld_base.emit_fetch_funcs[TGSI_FILE_INPUT] = emit_fetch_input;
 986    bld.bld_base.emit_fetch_funcs[TGSI_FILE_TEMPORARY] = emit_fetch_temporary;
 987
 988    /* Set opcode actions */
 989    lp_set_default_actions_cpu(&bld.bld_base);
 990
 991    if (!lp_bld_tgsi_list_init(&bld.bld_base)) {
 992       return;
 993    }
 994
 995    tgsi_parse_init(&parse, tokens);
 996
 997    while (!tgsi_parse_end_of_tokens(&parse)) {
 998       tgsi_parse_token(&parse);
 999
1000       switch(parse.FullToken.Token.Type) {
1001       case TGSI_TOKEN_TYPE_DECLARATION:
1002          /* Inputs already interpolated */
1003          lp_emit_declaration_aos(&bld, &parse.FullToken.FullDeclaration);
1004          break;
1005
1006       case TGSI_TOKEN_TYPE_INSTRUCTION:
1007          /* save expanded instruction */
1008          lp_bld_tgsi_add_instruction(&bld.bld_base,
1009                                      &parse.FullToken.FullInstruction);
1010          break;
1011
1012       case TGSI_TOKEN_TYPE_IMMEDIATE:
1013          /* simply copy the immediate values into the next immediates[] slot */
1014          {
1015             const uint size = parse.FullToken.FullImmediate.Immediate.NrTokens - 1;
1016             float imm[4];
1017             assert(size <= 4);
1018             assert(num_immediates < LP_MAX_INLINED_IMMEDIATES);
1019             for (chan = 0; chan < 4; ++chan) {
1020                imm[chan] = 0.0f;
1021             }
1022             for (chan = 0; chan < size; ++chan) {
1023                unsigned swizzle = bld.swizzles[chan];
1024                imm[swizzle] = parse.FullToken.FullImmediate.u[chan].Float;
1025             }
1026             bld.immediates[num_immediates] =
1027                      lp_build_const_aos(gallivm, type,
1028                                         imm[0], imm[1], imm[2], imm[3],
1029                                         NULL);
1030             num_immediates++;
1031          }
1032          break;
1033
1034       case TGSI_TOKEN_TYPE_PROPERTY:
1035          break;
1036
1037       default:
1038          assert(0);
1039       }
1040    }
1041
1042    while (pc != -1) {
1043       struct tgsi_full_instruction *instr = bld.bld_base.instructions + pc;
1044       const struct tgsi_opcode_info *opcode_info =
1045          tgsi_get_opcode_info(instr->Instruction.Opcode);
1046       if (!lp_emit_instruction_aos(&bld, instr, opcode_info, &pc))
1047          _debug_printf("warning: failed to translate tgsi opcode %s to LLVM\n",
1048                        opcode_info->mnemonic);
1049    }
1050
1051    if (0) {
1052       LLVMBasicBlockRef block = LLVMGetInsertBlock(gallivm->builder);
1053       LLVMValueRef function = LLVMGetBasicBlockParent(block);
1054       debug_printf("11111111111111111111111111111 \n");
1055       tgsi_dump(tokens, 0);
1056       lp_debug_dump_value(function);
1057       debug_printf("2222222222222222222222222222 \n");
1058    }
1059    tgsi_parse_free(&parse);
1060    FREE(bld.bld_base.instructions);
1061
1062    if (0) {
1063       LLVMModuleRef module = LLVMGetGlobalParent(
1064          LLVMGetBasicBlockParent(LLVMGetInsertBlock(gallivm->builder)));
1065       LLVMDumpModule(module);
1066    }
1067
1068 }
1069