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    default:
 260       assert(0);
 261       return;
 262    }
 263
 264    if (!ptr)
 265       return;
 266
 267    /*
 268     * Writemask
 269     */
 270
 271    if (reg->Register.WriteMask != TGSI_WRITEMASK_XYZW) {
 272       LLVMValueRef writemask;
 273
 274       writemask = lp_build_const_mask_aos_swizzled(bld->bld_base.base.gallivm,
 275                                                    bld->bld_base.base.type,
 276                                                    reg->Register.WriteMask,
 277                                                    TGSI_NUM_CHANNELS,
 278                                                    bld->swizzles);
 279
 280       if (mask) {
 281          mask = LLVMBuildAnd(builder, mask, writemask, "");
 282       } else {
 283          mask = writemask;
 284       }
 285    }
 286
 287    if (mask) {
 288       LLVMValueRef orig_value;
 289
 290       orig_value = LLVMBuildLoad(builder, ptr, "");
 291       value = lp_build_select(&bld->bld_base.base,
 292                               mask, value, orig_value);
 293    }
 294
 295    LLVMBuildStore(builder, value, ptr);
 296 }
 297
 298
 299 /**
 300  * High-level instruction translators.
 301  */
 302
 303 static LLVMValueRef
 304 emit_tex(struct lp_build_tgsi_aos_context *bld,
 305          const struct tgsi_full_instruction *inst,
 306          enum lp_build_tex_modifier modifier)
 307 {
 308    unsigned target;
 309    unsigned unit;
 310    LLVMValueRef coords;
 311    struct lp_derivatives derivs = { {NULL}, {NULL} };
 312
 313    if (!bld->sampler) {
 314       _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
 315       return bld->bld_base.base.undef;
 316    }
 317
 318    target = inst->Texture.Texture;
 319
 320    coords = lp_build_emit_fetch( &bld->bld_base, inst, 0 , LP_CHAN_ALL);
 321
 322    if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV) {
 323       /* probably not going to work */
 324       derivs.ddx[0] = lp_build_emit_fetch( &bld->bld_base, inst, 1 , LP_CHAN_ALL);
 325       derivs.ddy[0] = lp_build_emit_fetch( &bld->bld_base, inst, 2 , LP_CHAN_ALL);
 326       unit = inst->Src[3].Register.Index;
 327    }
 328    else {
 329       unit = inst->Src[1].Register.Index;
 330    }
 331    return bld->sampler->emit_fetch_texel(bld->sampler,
 332                                          &bld->bld_base.base,
 333                                          target, unit,
 334                                          coords, derivs,
 335                                          modifier);
 336 }
 337
 338
 339 static LLVMValueRef
 340 emit_sample(struct lp_build_tgsi_aos_context *bld,
 341             const struct tgsi_full_instruction *inst,
 342             enum lp_build_tex_modifier modifier)
 343 {
 344    unsigned target;
 345    unsigned unit;
 346    LLVMValueRef coords;
 347    struct lp_derivatives derivs = { {NULL}, {NULL} };
 348
 349    if (!bld->sampler) {
 350       _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
 351       return bld->bld_base.base.undef;
 352    }
 353
 354    coords = lp_build_emit_fetch( &bld->bld_base, inst, 0 , LP_CHAN_ALL);
 355
 356    /* ignore modifiers, can't handle different sampler / sampler view, etc... */
 357    unit = inst->Src[1].Register.Index;
 358    assert(inst->Src[2].Register.Index == unit);
 359
 360    target = bld->sv[unit].Resource;
 361
 362    return bld->sampler->emit_fetch_texel(bld->sampler,
 363                                          &bld->bld_base.base,
 364                                          target, unit,
 365                                          coords, derivs,
 366                                          modifier);
 367 }
 368
 369
 370 void
 371 lp_emit_declaration_aos(
 372    struct lp_build_tgsi_aos_context *bld,
 373    const struct tgsi_full_declaration *decl)
 374 {
 375    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
 376    LLVMTypeRef vec_type = lp_build_vec_type(bld->bld_base.base.gallivm, bld->bld_base.base.type);
 377
 378    unsigned first = decl->Range.First;
 379    unsigned last = decl->Range.Last;
 380    unsigned idx;
 381
 382    for (idx = first; idx <= last; ++idx) {
 383       switch (decl->Declaration.File) {
 384       case TGSI_FILE_TEMPORARY:
 385          assert(idx < LP_MAX_INLINED_TEMPS);
 386          if (bld->indirect_files & (1 << TGSI_FILE_TEMPORARY)) {
 387             LLVMValueRef array_size = lp_build_const_int32(gallivm, last + 1);
 388             bld->temps_array = lp_build_array_alloca(bld->bld_base.base.gallivm,
 389                                                      vec_type, array_size, "");
 390          } else {
 391             bld->temps[idx] = lp_build_alloca(gallivm, vec_type, "");
 392          }
 393          break;
 394
 395       case TGSI_FILE_OUTPUT:
 396          bld->outputs[idx] = lp_build_alloca(gallivm, vec_type, "");
 397          break;
 398
 399       case TGSI_FILE_ADDRESS:
 400          assert(idx < LP_MAX_TGSI_ADDRS);
 401          bld->addr[idx] = lp_build_alloca(gallivm, vec_type, "");
 402          break;
 403
 404       case TGSI_FILE_SAMPLER_VIEW:
 405          /*
 406           * The target stored here MUST match whatever there actually
 407           * is in the set sampler views (what about return type?).
 408           */
 409          assert(last < PIPE_MAX_SHADER_SAMPLER_VIEWS);
 410          for (idx = first; idx <= last; ++idx) {
 411             bld->sv[idx] = decl->SamplerView;
 412          }
 413          break;
 414
 415       default:
 416          /* don't need to declare other vars */
 417          break;
 418       }
 419    }
 420 }
 421
 422
 423 /**
 424  * Emit LLVM for one TGSI instruction.
 425  * \param return TRUE for success, FALSE otherwise
 426  */
 427 boolean
 428 lp_emit_instruction_aos(
 429    struct lp_build_tgsi_aos_context *bld,
 430    const struct tgsi_full_instruction *inst,
 431    const struct tgsi_opcode_info *info,
 432    int *pc)
 433 {
 434    LLVMValueRef src0, src1, src2;
 435    LLVMValueRef tmp0;
 436    LLVMValueRef dst0 = NULL;
 437
 438    /*
 439     * Stores and write masks are handled in a general fashion after the long
 440     * instruction opcode switch statement.
 441     *
 442     * Although not stricitly necessary, we avoid generating instructions for
 443     * channels which won't be stored, in cases where's that easy. For some
 444     * complex instructions, like texture sampling, it is more convenient to
 445     * assume a full writemask and then let LLVM optimization passes eliminate
 446     * redundant code.
 447     */
 448
 449    (*pc)++;
 450
 451    assert(info->num_dst <= 1);
 452    if (info->num_dst) {
 453       dst0 = bld->bld_base.base.undef;
 454    }
 455
 456    switch (inst->Instruction.Opcode) {
 457    case TGSI_OPCODE_ARL:
 458       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 459       dst0 = lp_build_floor(&bld->bld_base.base, src0);
 460       break;
 461
 462    case TGSI_OPCODE_MOV:
 463       dst0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 464       break;
 465
 466    case TGSI_OPCODE_LIT:
 467       return FALSE;
 468
 469    case TGSI_OPCODE_RCP:
 470    /* TGSI_OPCODE_RECIP */
 471       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 472       dst0 = lp_build_rcp(&bld->bld_base.base, src0);
 473       break;
 474
 475    case TGSI_OPCODE_RSQ:
 476    /* TGSI_OPCODE_RECIPSQRT */
 477       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 478       tmp0 = lp_build_abs(&bld->bld_base.base, src0);
 479       dst0 = lp_build_rsqrt(&bld->bld_base.base, tmp0);
 480       break;
 481
 482    case TGSI_OPCODE_EXP:
 483       return FALSE;
 484
 485    case TGSI_OPCODE_LOG:
 486       return FALSE;
 487
 488    case TGSI_OPCODE_MUL:
 489       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 490       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 491       dst0 = lp_build_mul(&bld->bld_base.base, src0, src1);
 492       break;
 493
 494    case TGSI_OPCODE_ADD:
 495       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 496       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 497       dst0 = lp_build_add(&bld->bld_base.base, src0, src1);
 498       break;
 499
 500    case TGSI_OPCODE_DP3:
 501    /* TGSI_OPCODE_DOT3 */
 502       return FALSE;
 503
 504    case TGSI_OPCODE_DP4:
 505    /* TGSI_OPCODE_DOT4 */
 506       return FALSE;
 507
 508    case TGSI_OPCODE_DST:
 509       return FALSE;
 510
 511    case TGSI_OPCODE_MIN:
 512       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 513       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 514       dst0 = lp_build_min(&bld->bld_base.base, src0, src1);
 515       break;
 516
 517    case TGSI_OPCODE_MAX:
 518       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 519       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 520       dst0 = lp_build_max(&bld->bld_base.base, src0, src1);
 521       break;
 522
 523    case TGSI_OPCODE_SLT:
 524    /* TGSI_OPCODE_SETLT */
 525       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 526       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 527       tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_LESS, src0, src1);
 528       dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero);
 529       break;
 530
 531    case TGSI_OPCODE_SGE:
 532    /* TGSI_OPCODE_SETGE */
 533       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 534       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 535       tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_GEQUAL, src0, src1);
 536       dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero);
 537       break;
 538
 539    case TGSI_OPCODE_MAD:
 540    /* TGSI_OPCODE_MADD */
 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       src2 = lp_build_emit_fetch(&bld->bld_base, inst, 2, LP_CHAN_ALL);
 544       tmp0 = lp_build_mul(&bld->bld_base.base, src0, src1);
 545       dst0 = lp_build_add(&bld->bld_base.base, tmp0, src2);
 546       break;
 547
 548    case TGSI_OPCODE_LRP:
 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       src2 = lp_build_emit_fetch(&bld->bld_base, inst, 2, LP_CHAN_ALL);
 552       tmp0 = lp_build_sub(&bld->bld_base.base, src1, src2);
 553       tmp0 = lp_build_mul(&bld->bld_base.base, src0, tmp0);
 554       dst0 = lp_build_add(&bld->bld_base.base, tmp0, src2);
 555       break;
 556
 557    case TGSI_OPCODE_FRC:
 558       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 559       tmp0 = lp_build_floor(&bld->bld_base.base, src0);
 560       dst0 = lp_build_sub(&bld->bld_base.base, src0, tmp0);
 561       break;
 562
 563    case TGSI_OPCODE_FLR:
 564       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 565       dst0 = lp_build_floor(&bld->bld_base.base, src0);
 566       break;
 567
 568    case TGSI_OPCODE_ROUND:
 569       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 570       dst0 = lp_build_round(&bld->bld_base.base, src0);
 571       break;
 572
 573    case TGSI_OPCODE_EX2:
 574       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 575       tmp0 = lp_build_swizzle_scalar_aos(&bld->bld_base.base, src0, TGSI_SWIZZLE_X, TGSI_NUM_CHANNELS);
 576       dst0 = lp_build_exp2(&bld->bld_base.base, tmp0);
 577       break;
 578
 579    case TGSI_OPCODE_LG2:
 580       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 581       tmp0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
 582       dst0 = lp_build_log2(&bld->bld_base.base, tmp0);
 583       break;
 584
 585    case TGSI_OPCODE_POW:
 586       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 587       src0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
 588       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 589       src1 = swizzle_scalar_aos(bld, src1, TGSI_SWIZZLE_X);
 590       dst0 = lp_build_pow(&bld->bld_base.base, src0, src1);
 591       break;
 592
 593    case TGSI_OPCODE_COS:
 594       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 595       tmp0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
 596       dst0 = lp_build_cos(&bld->bld_base.base, tmp0);
 597       break;
 598
 599    case TGSI_OPCODE_DDX:
 600       return FALSE;
 601
 602    case TGSI_OPCODE_DDY:
 603       return FALSE;
 604
 605    case TGSI_OPCODE_KILL:
 606       return FALSE;
 607
 608    case TGSI_OPCODE_KILL_IF:
 609       return FALSE;
 610
 611    case TGSI_OPCODE_PK2H:
 612       return FALSE;
 613       break;
 614
 615    case TGSI_OPCODE_PK2US:
 616       return FALSE;
 617       break;
 618
 619    case TGSI_OPCODE_PK4B:
 620       return FALSE;
 621       break;
 622
 623    case TGSI_OPCODE_PK4UB:
 624       return FALSE;
 625
 626    case TGSI_OPCODE_SEQ:
 627       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 628       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 629       tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_EQUAL, src0, src1);
 630       dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero);
 631       break;
 632
 633    case TGSI_OPCODE_SGT:
 634       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 635       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 636       tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_GREATER, src0, src1);
 637       dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero);
 638       break;
 639
 640    case TGSI_OPCODE_SIN:
 641       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 642       tmp0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
 643       dst0 = lp_build_sin(&bld->bld_base.base, tmp0);
 644       break;
 645
 646    case TGSI_OPCODE_SLE:
 647       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 648       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 649       tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_LEQUAL, src0, src1);
 650       dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero);
 651       break;
 652
 653    case TGSI_OPCODE_SNE:
 654       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 655       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 656       tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_NOTEQUAL, src0, src1);
 657       dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero);
 658       break;
 659
 660    case TGSI_OPCODE_TEX:
 661       dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_NONE);
 662       break;
 663
 664    case TGSI_OPCODE_TXD:
 665       dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV);
 666       break;
 667
 668    case TGSI_OPCODE_UP2H:
 669       /* deprecated */
 670       assert (0);
 671       return FALSE;
 672       break;
 673
 674    case TGSI_OPCODE_UP2US:
 675       /* deprecated */
 676       assert(0);
 677       return FALSE;
 678       break;
 679
 680    case TGSI_OPCODE_UP4B:
 681       /* deprecated */
 682       assert(0);
 683       return FALSE;
 684       break;
 685
 686    case TGSI_OPCODE_UP4UB:
 687       /* deprecated */
 688       assert(0);
 689       return FALSE;
 690       break;
 691
 692    case TGSI_OPCODE_ARR:
 693       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 694       dst0 = lp_build_round(&bld->bld_base.base, src0);
 695       break;
 696
 697    case TGSI_OPCODE_CAL:
 698       return FALSE;
 699
 700    case TGSI_OPCODE_RET:
 701       /* safe to ignore at end */
 702       break;
 703
 704    case TGSI_OPCODE_END:
 705       *pc = -1;
 706       break;
 707
 708    case TGSI_OPCODE_SSG:
 709    /* TGSI_OPCODE_SGN */
 710       tmp0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 711       dst0 = lp_build_sgn(&bld->bld_base.base, tmp0);
 712       break;
 713
 714    case TGSI_OPCODE_CMP:
 715       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 716       src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
 717       src2 = lp_build_emit_fetch(&bld->bld_base, inst, 2, LP_CHAN_ALL);
 718       tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_LESS, src0, bld->bld_base.base.zero);
 719       dst0 = lp_build_select(&bld->bld_base.base, tmp0, src1, src2);
 720       break;
 721
 722    case TGSI_OPCODE_TXB:
 723       dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_LOD_BIAS);
 724       break;
 725
 726    case TGSI_OPCODE_DIV:
 727       assert(0);
 728       return FALSE;
 729       break;
 730
 731    case TGSI_OPCODE_DP2:
 732       return FALSE;
 733
 734    case TGSI_OPCODE_TXL:
 735       dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD);
 736       break;
 737
 738    case TGSI_OPCODE_TXP:
 739       dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_PROJECTED);
 740       break;
 741
 742    case TGSI_OPCODE_BRK:
 743       return FALSE;
 744
 745    case TGSI_OPCODE_IF:
 746    case TGSI_OPCODE_UIF:
 747       return FALSE;
 748
 749    case TGSI_OPCODE_BGNLOOP:
 750       return FALSE;
 751
 752    case TGSI_OPCODE_BGNSUB:
 753       return FALSE;
 754
 755    case TGSI_OPCODE_ELSE:
 756       return FALSE;
 757
 758    case TGSI_OPCODE_ENDIF:
 759       return FALSE;
 760
 761    case TGSI_OPCODE_ENDLOOP:
 762       return FALSE;
 763
 764    case TGSI_OPCODE_ENDSUB:
 765       return FALSE;
 766
 767    case TGSI_OPCODE_CEIL:
 768       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 769       dst0 = lp_build_ceil(&bld->bld_base.base, src0);
 770       break;
 771
 772    case TGSI_OPCODE_I2F:
 773       assert(0);
 774       return FALSE;
 775       break;
 776
 777    case TGSI_OPCODE_NOT:
 778       assert(0);
 779       return FALSE;
 780       break;
 781
 782    case TGSI_OPCODE_TRUNC:
 783       src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
 784       dst0 = lp_build_trunc(&bld->bld_base.base, src0);
 785       break;
 786
 787    case TGSI_OPCODE_SHL:
 788       assert(0);
 789       return FALSE;
 790       break;
 791
 792    case TGSI_OPCODE_ISHR:
 793       assert(0);
 794       return FALSE;
 795       break;
 796
 797    case TGSI_OPCODE_AND:
 798       assert(0);
 799       return FALSE;
 800       break;
 801
 802    case TGSI_OPCODE_OR:
 803       assert(0);
 804       return FALSE;
 805       break;
 806
 807    case TGSI_OPCODE_MOD:
 808       assert(0);
 809       return FALSE;
 810       break;
 811
 812    case TGSI_OPCODE_XOR:
 813       assert(0);
 814       return FALSE;
 815       break;
 816
 817    case TGSI_OPCODE_TXF:
 818       assert(0);
 819       return FALSE;
 820       break;
 821
 822    case TGSI_OPCODE_TXQ:
 823       assert(0);
 824       return FALSE;
 825       break;
 826
 827    case TGSI_OPCODE_CONT:
 828       return FALSE;
 829
 830    case TGSI_OPCODE_EMIT:
 831       return FALSE;
 832       break;
 833
 834    case TGSI_OPCODE_ENDPRIM:
 835       return FALSE;
 836       break;
 837
 838    case TGSI_OPCODE_NOP:
 839       break;
 840
 841    case TGSI_OPCODE_SAMPLE:
 842       dst0 = emit_sample(bld, inst, LP_BLD_TEX_MODIFIER_NONE);
 843       break;
 844
 845    default:
 846       return FALSE;
 847    }
 848
 849    if (info->num_dst) {
 850       lp_emit_store_aos(bld, inst, 0, dst0);
 851    }
 852
 853    return TRUE;
 854 }
 855
 856
 857 void
 858 lp_build_tgsi_aos(struct gallivm_state *gallivm,
 859                   const struct tgsi_token *tokens,
 860                   struct lp_type type,
 861                   const unsigned char swizzles[4],
 862                   LLVMValueRef consts_ptr,
 863                   const LLVMValueRef *inputs,
 864                   LLVMValueRef *outputs,
 865                   const struct lp_build_sampler_aos *sampler,
 866                   const struct tgsi_shader_info *info)
 867 {
 868    struct lp_build_tgsi_aos_context bld;
 869    struct tgsi_parse_context parse;
 870    uint num_immediates = 0;
 871    unsigned chan;
 872    int pc = 0;
 873
 874    /* Setup build context */
 875    memset(&bld, 0, sizeof bld);
 876    lp_build_context_init(&bld.bld_base.base, gallivm, type);
 877    lp_build_context_init(&bld.bld_base.uint_bld, gallivm, lp_uint_type(type));
 878    lp_build_context_init(&bld.bld_base.int_bld, gallivm, lp_int_type(type));
 879    lp_build_context_init(&bld.int_bld, gallivm, lp_int_type(type));
 880
 881    for (chan = 0; chan < 4; ++chan) {
 882       bld.swizzles[chan] = swizzles[chan];
 883       bld.inv_swizzles[swizzles[chan]] = chan;
 884    }
 885
 886    bld.inputs = inputs;
 887    bld.outputs = outputs;
 888    bld.consts_ptr = consts_ptr;
 889    bld.sampler = sampler;
 890    bld.indirect_files = info->indirect_files;
 891    bld.bld_base.emit_swizzle = swizzle_aos;
 892    bld.bld_base.info = info;
 893
 894    bld.bld_base.emit_fetch_funcs[TGSI_FILE_CONSTANT] = emit_fetch_constant;
 895    bld.bld_base.emit_fetch_funcs[TGSI_FILE_IMMEDIATE] = emit_fetch_immediate;
 896    bld.bld_base.emit_fetch_funcs[TGSI_FILE_INPUT] = emit_fetch_input;
 897    bld.bld_base.emit_fetch_funcs[TGSI_FILE_TEMPORARY] = emit_fetch_temporary;
 898
 899    /* Set opcode actions */
 900    lp_set_default_actions_cpu(&bld.bld_base);
 901
 902    if (!lp_bld_tgsi_list_init(&bld.bld_base)) {
 903       return;
 904    }
 905
 906    tgsi_parse_init(&parse, tokens);
 907
 908    while (!tgsi_parse_end_of_tokens(&parse)) {
 909       tgsi_parse_token(&parse);
 910
 911       switch(parse.FullToken.Token.Type) {
 912       case TGSI_TOKEN_TYPE_DECLARATION:
 913          /* Inputs already interpolated */
 914          lp_emit_declaration_aos(&bld, &parse.FullToken.FullDeclaration);
 915          break;
 916
 917       case TGSI_TOKEN_TYPE_INSTRUCTION:
 918          /* save expanded instruction */
 919          lp_bld_tgsi_add_instruction(&bld.bld_base,
 920                                      &parse.FullToken.FullInstruction);
 921          break;
 922
 923       case TGSI_TOKEN_TYPE_IMMEDIATE:
 924          /* simply copy the immediate values into the next immediates[] slot */
 925          {
 926             const uint size = parse.FullToken.FullImmediate.Immediate.NrTokens - 1;
 927             float imm[4];
 928             assert(size <= 4);
 929             assert(num_immediates < LP_MAX_INLINED_IMMEDIATES);
 930             for (chan = 0; chan < 4; ++chan) {
 931                imm[chan] = 0.0f;
 932             }
 933             for (chan = 0; chan < size; ++chan) {
 934                unsigned swizzle = bld.swizzles[chan];
 935                imm[swizzle] = parse.FullToken.FullImmediate.u[chan].Float;
 936             }
 937             bld.immediates[num_immediates] =
 938                      lp_build_const_aos(gallivm, type,
 939                                         imm[0], imm[1], imm[2], imm[3],
 940                                         NULL);
 941             num_immediates++;
 942          }
 943          break;
 944
 945       case TGSI_TOKEN_TYPE_PROPERTY:
 946          break;
 947
 948       default:
 949          assert(0);
 950       }
 951    }
 952
 953    while (pc != -1) {
 954       struct tgsi_full_instruction *instr = bld.bld_base.instructions + pc;
 955       const struct tgsi_opcode_info *opcode_info =
 956          tgsi_get_opcode_info(instr->Instruction.Opcode);
 957       if (!lp_emit_instruction_aos(&bld, instr, opcode_info, &pc))
 958          _debug_printf("warning: failed to translate tgsi opcode %s to LLVM\n",
 959                        tgsi_get_opcode_name(instr->Instruction.Opcode));
 960    }
 961
 962    if (0) {
 963       LLVMBasicBlockRef block = LLVMGetInsertBlock(gallivm->builder);
 964       LLVMValueRef function = LLVMGetBasicBlockParent(block);
 965       debug_printf("11111111111111111111111111111 \n");
 966       tgsi_dump(tokens, 0);
 967       lp_debug_dump_value(function);
 968       debug_printf("2222222222222222222222222222 \n");
 969    }
 970    tgsi_parse_free(&parse);
 971    FREE(bld.bld_base.instructions);
 972
 973    if (0) {
 974       LLVMModuleRef module = LLVMGetGlobalParent(
 975          LLVMGetBasicBlockParent(LLVMGetInsertBlock(gallivm->builder)));
 976       LLVMDumpModule(module);
 977    }
 978
 979 }
 980