src/gallium/auxiliary/gallivm/lp_bld_tgsi_soa.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2009 VMware, Inc.
   4  * Copyright 2007-2008 Tungsten Graphics, Inc., Cedar Park, Texas.
   5  * All Rights Reserved.
   6  *
   7  * Permission is hereby granted, free of charge, to any person obtaining a
   8  * copy of this software and associated documentation files (the
   9  * "Software"), to deal in the Software without restriction, including
  10  * without limitation the rights to use, copy, modify, merge, publish,
  11  * distribute, sub license, and/or sell copies of the Software, and to
  12  * permit persons to whom the Software is furnished to do so, subject to
  13  * the following conditions:
  14  *
  15  * The above copyright notice and this permission notice (including the
  16  * next paragraph) shall be included in all copies or substantial portions
  17  * of the Software.
  18  *
  19  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  20  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  21  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  22  * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
  23  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  24  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  25  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  26  *
  27  **************************************************************************/
  28
  29 /**
  30  * @file
  31  * TGSI to LLVM IR translation -- SoA.
  32  *
  33  * @author Jose Fonseca <jfonseca@vmware.com>
  34  *
  35  * Based on tgsi_sse2.c code written by Michal Krol, Keith Whitwell,
  36  * Brian Paul, and others.
  37  */
  38
  39 #include "pipe/p_config.h"
  40 #include "pipe/p_shader_tokens.h"
  41 #include "util/u_debug.h"
  42 #include "util/u_math.h"
  43 #include "util/u_memory.h"
  44 #include "tgsi/tgsi_dump.h"
  45 #include "tgsi/tgsi_exec.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_tgsi_action.h"
  51 #include "lp_bld_type.h"
  52 #include "lp_bld_const.h"
  53 #include "lp_bld_arit.h"
  54 #include "lp_bld_bitarit.h"
  55 #include "lp_bld_gather.h"
  56 #include "lp_bld_init.h"
  57 #include "lp_bld_logic.h"
  58 #include "lp_bld_swizzle.h"
  59 #include "lp_bld_flow.h"
  60 #include "lp_bld_quad.h"
  61 #include "lp_bld_tgsi.h"
  62 #include "lp_bld_limits.h"
  63 #include "lp_bld_debug.h"
  64 #include "lp_bld_printf.h"
  65
  66
  67 static void lp_exec_mask_init(struct lp_exec_mask *mask, struct lp_build_context *bld)
  68 {
  69    mask->bld = bld;
  70    mask->has_mask = FALSE;
  71    mask->cond_stack_size = 0;
  72    mask->loop_stack_size = 0;
  73    mask->call_stack_size = 0;
  74
  75    mask->int_vec_type = lp_build_int_vec_type(bld->gallivm, mask->bld->type);
  76    mask->exec_mask = mask->ret_mask = mask->break_mask = mask->cont_mask = mask->cond_mask =
  77          LLVMConstAllOnes(mask->int_vec_type);
  78 }
  79
  80 static void lp_exec_mask_update(struct lp_exec_mask *mask)
  81 {
  82    LLVMBuilderRef builder = mask->bld->gallivm->builder;
  83
  84    if (mask->loop_stack_size) {
  85       /*for loops we need to update the entire mask at runtime */
  86       LLVMValueRef tmp;
  87       assert(mask->break_mask);
  88       tmp = LLVMBuildAnd(builder,
  89                          mask->cont_mask,
  90                          mask->break_mask,
  91                          "maskcb");
  92       mask->exec_mask = LLVMBuildAnd(builder,
  93                                      mask->cond_mask,
  94                                      tmp,
  95                                      "maskfull");
  96    } else
  97       mask->exec_mask = mask->cond_mask;
  98
  99    if (mask->call_stack_size) {
 100       mask->exec_mask = LLVMBuildAnd(builder,
 101                                      mask->exec_mask,
 102                                      mask->ret_mask,
 103                                      "callmask");
 104    }
 105
 106    mask->has_mask = (mask->cond_stack_size > 0 ||
 107                      mask->loop_stack_size > 0 ||
 108                      mask->call_stack_size > 0);
 109 }
 110
 111 static void lp_exec_mask_cond_push(struct lp_exec_mask *mask,
 112                                    LLVMValueRef val)
 113 {
 114    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 115
 116    assert(mask->cond_stack_size < LP_MAX_TGSI_NESTING);
 117    if (mask->cond_stack_size == 0) {
 118       assert(mask->cond_mask == LLVMConstAllOnes(mask->int_vec_type));
 119    }
 120    mask->cond_stack[mask->cond_stack_size++] = mask->cond_mask;
 121    assert(LLVMTypeOf(val) == mask->int_vec_type);
 122    mask->cond_mask = LLVMBuildAnd(builder,
 123                                   mask->cond_mask,
 124                                   val,
 125                                   "");
 126    lp_exec_mask_update(mask);
 127 }
 128
 129 static void lp_exec_mask_cond_invert(struct lp_exec_mask *mask)
 130 {
 131    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 132    LLVMValueRef prev_mask;
 133    LLVMValueRef inv_mask;
 134
 135    assert(mask->cond_stack_size);
 136    prev_mask = mask->cond_stack[mask->cond_stack_size - 1];
 137    if (mask->cond_stack_size == 1) {
 138       assert(prev_mask == LLVMConstAllOnes(mask->int_vec_type));
 139    }
 140
 141    inv_mask = LLVMBuildNot(builder, mask->cond_mask, "");
 142
 143    mask->cond_mask = LLVMBuildAnd(builder,
 144                                   inv_mask,
 145                                   prev_mask, "");
 146    lp_exec_mask_update(mask);
 147 }
 148
 149 static void lp_exec_mask_cond_pop(struct lp_exec_mask *mask)
 150 {
 151    assert(mask->cond_stack_size);
 152    mask->cond_mask = mask->cond_stack[--mask->cond_stack_size];
 153    lp_exec_mask_update(mask);
 154 }
 155
 156 static void lp_exec_bgnloop(struct lp_exec_mask *mask)
 157 {
 158    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 159
 160    if (mask->loop_stack_size == 0) {
 161       assert(mask->loop_block == NULL);
 162       assert(mask->cont_mask == LLVMConstAllOnes(mask->int_vec_type));
 163       assert(mask->break_mask == LLVMConstAllOnes(mask->int_vec_type));
 164       assert(mask->break_var == NULL);
 165    }
 166
 167    assert(mask->loop_stack_size < LP_MAX_TGSI_NESTING);
 168
 169    mask->loop_stack[mask->loop_stack_size].loop_block = mask->loop_block;
 170    mask->loop_stack[mask->loop_stack_size].cont_mask = mask->cont_mask;
 171    mask->loop_stack[mask->loop_stack_size].break_mask = mask->break_mask;
 172    mask->loop_stack[mask->loop_stack_size].break_var = mask->break_var;
 173    ++mask->loop_stack_size;
 174
 175    mask->break_var = lp_build_alloca(mask->bld->gallivm, mask->int_vec_type, "");
 176    LLVMBuildStore(builder, mask->break_mask, mask->break_var);
 177
 178    mask->loop_block = lp_build_insert_new_block(mask->bld->gallivm, "bgnloop");
 179    LLVMBuildBr(builder, mask->loop_block);
 180    LLVMPositionBuilderAtEnd(builder, mask->loop_block);
 181
 182    mask->break_mask = LLVMBuildLoad(builder, mask->break_var, "");
 183
 184    lp_exec_mask_update(mask);
 185 }
 186
 187 static void lp_exec_break(struct lp_exec_mask *mask)
 188 {
 189    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 190    LLVMValueRef exec_mask = LLVMBuildNot(builder,
 191                                          mask->exec_mask,
 192                                          "break");
 193
 194    mask->break_mask = LLVMBuildAnd(builder,
 195                                    mask->break_mask,
 196                                    exec_mask, "break_full");
 197
 198    lp_exec_mask_update(mask);
 199 }
 200
 201 static void lp_exec_continue(struct lp_exec_mask *mask)
 202 {
 203    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 204    LLVMValueRef exec_mask = LLVMBuildNot(builder,
 205                                          mask->exec_mask,
 206                                          "");
 207
 208    mask->cont_mask = LLVMBuildAnd(builder,
 209                                   mask->cont_mask,
 210                                   exec_mask, "");
 211
 212    lp_exec_mask_update(mask);
 213 }
 214
 215
 216 static void lp_exec_endloop(struct gallivm_state *gallivm,
 217                             struct lp_exec_mask *mask)
 218 {
 219    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 220    LLVMBasicBlockRef endloop;
 221    LLVMTypeRef reg_type = LLVMIntTypeInContext(gallivm->context,
 222                                                mask->bld->type.width *
 223                                                mask->bld->type.length);
 224    LLVMValueRef i1cond;
 225
 226    assert(mask->break_mask);
 227
 228    /*
 229     * Restore the cont_mask, but don't pop
 230     */
 231    assert(mask->loop_stack_size);
 232    mask->cont_mask = mask->loop_stack[mask->loop_stack_size - 1].cont_mask;
 233    lp_exec_mask_update(mask);
 234
 235    /*
 236     * Unlike the continue mask, the break_mask must be preserved across loop
 237     * iterations
 238     */
 239    LLVMBuildStore(builder, mask->break_mask, mask->break_var);
 240
 241    /* i1cond = (mask == 0) */
 242    i1cond = LLVMBuildICmp(
 243       builder,
 244       LLVMIntNE,
 245       LLVMBuildBitCast(builder, mask->exec_mask, reg_type, ""),
 246       LLVMConstNull(reg_type), "");
 247
 248    endloop = lp_build_insert_new_block(mask->bld->gallivm, "endloop");
 249
 250    LLVMBuildCondBr(builder,
 251                    i1cond, mask->loop_block, endloop);
 252
 253    LLVMPositionBuilderAtEnd(builder, endloop);
 254
 255    assert(mask->loop_stack_size);
 256    --mask->loop_stack_size;
 257    mask->loop_block = mask->loop_stack[mask->loop_stack_size].loop_block;
 258    mask->cont_mask = mask->loop_stack[mask->loop_stack_size].cont_mask;
 259    mask->break_mask = mask->loop_stack[mask->loop_stack_size].break_mask;
 260    mask->break_var = mask->loop_stack[mask->loop_stack_size].break_var;
 261
 262    lp_exec_mask_update(mask);
 263 }
 264
 265 /* stores val into an address pointed to by dst.
 266  * mask->exec_mask is used to figure out which bits of val
 267  * should be stored into the address
 268  * (0 means don't store this bit, 1 means do store).
 269  */
 270 static void lp_exec_mask_store(struct lp_exec_mask *mask,
 271                                struct lp_build_context *bld_store,
 272                                LLVMValueRef pred,
 273                                LLVMValueRef val,
 274                                LLVMValueRef dst)
 275 {
 276    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 277
 278    /* Mix the predicate and execution mask */
 279    if (mask->has_mask) {
 280       if (pred) {
 281          pred = LLVMBuildAnd(builder, pred, mask->exec_mask, "");
 282       } else {
 283          pred = mask->exec_mask;
 284       }
 285    }
 286
 287    if (pred) {
 288       LLVMValueRef real_val, dst_val;
 289
 290       dst_val = LLVMBuildLoad(builder, dst, "");
 291       real_val = lp_build_select(bld_store,
 292                                  pred,
 293                                  val, dst_val);
 294
 295       LLVMBuildStore(builder, real_val, dst);
 296    } else
 297       LLVMBuildStore(builder, val, dst);
 298 }
 299
 300 static void lp_exec_mask_call(struct lp_exec_mask *mask,
 301                               int func,
 302                               int *pc)
 303 {
 304    assert(mask->call_stack_size < LP_MAX_TGSI_NESTING);
 305    mask->call_stack[mask->call_stack_size].pc = *pc;
 306    mask->call_stack[mask->call_stack_size].ret_mask = mask->ret_mask;
 307    mask->call_stack_size++;
 308    *pc = func;
 309 }
 310
 311 static void lp_exec_mask_ret(struct lp_exec_mask *mask, int *pc)
 312 {
 313    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 314    LLVMValueRef exec_mask;
 315
 316    if (mask->call_stack_size == 0) {
 317       /* returning from main() */
 318       *pc = -1;
 319       return;
 320    }
 321    exec_mask = LLVMBuildNot(builder,
 322                             mask->exec_mask,
 323                             "ret");
 324
 325    mask->ret_mask = LLVMBuildAnd(builder,
 326                                  mask->ret_mask,
 327                                  exec_mask, "ret_full");
 328
 329    lp_exec_mask_update(mask);
 330 }
 331
 332 static void lp_exec_mask_bgnsub(struct lp_exec_mask *mask)
 333 {
 334 }
 335
 336 static void lp_exec_mask_endsub(struct lp_exec_mask *mask, int *pc)
 337 {
 338    assert(mask->call_stack_size);
 339    mask->call_stack_size--;
 340    *pc = mask->call_stack[mask->call_stack_size].pc;
 341    mask->ret_mask = mask->call_stack[mask->call_stack_size].ret_mask;
 342    lp_exec_mask_update(mask);
 343 }
 344
 345
 346 /**
 347  * Return pointer to a temporary register channel (src or dest).
 348  * Note that indirect addressing cannot be handled here.
 349  * \param index  which temporary register
 350  * \param chan  which channel of the temp register.
 351  */
 352 LLVMValueRef
 353 lp_get_temp_ptr_soa(struct lp_build_tgsi_soa_context *bld,
 354              unsigned index,
 355              unsigned chan)
 356 {
 357    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
 358    assert(chan < 4);
 359    if (bld->indirect_files & (1 << TGSI_FILE_TEMPORARY)) {
 360       LLVMValueRef lindex = lp_build_const_int32(bld->bld_base.base.gallivm, index * 4 + chan);
 361       return LLVMBuildGEP(builder, bld->temps_array, &lindex, 1, "");
 362    }
 363    else {
 364       return bld->temps[index][chan];
 365    }
 366 }
 367
 368 /**
 369  * Return pointer to a output register channel (src or dest).
 370  * Note that indirect addressing cannot be handled here.
 371  * \param index  which output register
 372  * \param chan  which channel of the output register.
 373  */
 374 LLVMValueRef
 375 lp_get_output_ptr(struct lp_build_tgsi_soa_context *bld,
 376                unsigned index,
 377                unsigned chan)
 378 {
 379    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
 380    assert(chan < 4);
 381    if (bld->indirect_files & (1 << TGSI_FILE_OUTPUT)) {
 382       LLVMValueRef lindex = lp_build_const_int32(bld->bld_base.base.gallivm,
 383                                                  index * 4 + chan);
 384       return LLVMBuildGEP(builder, bld->outputs_array, &lindex, 1, "");
 385    }
 386    else {
 387       return bld->outputs[index][chan];
 388    }
 389 }
 390
 391 /**
 392  * Gather vector.
 393  * XXX the lp_build_gather() function should be capable of doing this
 394  * with a little work.
 395  */
 396 static LLVMValueRef
 397 build_gather(struct lp_build_context *bld,
 398              LLVMValueRef base_ptr,
 399              LLVMValueRef indexes)
 400 {
 401    LLVMBuilderRef builder = bld->gallivm->builder;
 402    LLVMValueRef res = bld->undef;
 403    unsigned i;
 404
 405    /*
 406     * Loop over elements of index_vec, load scalar value, insert it into 'res'.
 407     */
 408    for (i = 0; i < bld->type.length; i++) {
 409       LLVMValueRef ii = lp_build_const_int32(bld->gallivm, i);
 410       LLVMValueRef index = LLVMBuildExtractElement(builder,
 411                                                    indexes, ii, "");
 412       LLVMValueRef scalar_ptr = LLVMBuildGEP(builder, base_ptr,
 413                                              &index, 1, "gather_ptr");
 414       LLVMValueRef scalar = LLVMBuildLoad(builder, scalar_ptr, "");
 415
 416       res = LLVMBuildInsertElement(builder, res, scalar, ii, "");
 417    }
 418
 419    return res;
 420 }
 421
 422
 423 /**
 424  * Scatter/store vector.
 425  */
 426 static void
 427 emit_mask_scatter(struct lp_build_tgsi_soa_context *bld,
 428                   LLVMValueRef base_ptr,
 429                   LLVMValueRef indexes,
 430                   LLVMValueRef values,
 431                   struct lp_exec_mask *mask,
 432                   LLVMValueRef pred)
 433 {
 434    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
 435    LLVMBuilderRef builder = gallivm->builder;
 436    unsigned i;
 437
 438    /* Mix the predicate and execution mask */
 439    if (mask->has_mask) {
 440       if (pred) {
 441          pred = LLVMBuildAnd(builder, pred, mask->exec_mask, "");
 442       }
 443       else {
 444          pred = mask->exec_mask;
 445       }
 446    }
 447
 448    /*
 449     * Loop over elements of index_vec, store scalar value.
 450     */
 451    for (i = 0; i < bld->bld_base.base.type.length; i++) {
 452       LLVMValueRef ii = lp_build_const_int32(gallivm, i);
 453       LLVMValueRef index = LLVMBuildExtractElement(builder, indexes, ii, "");
 454       LLVMValueRef scalar_ptr = LLVMBuildGEP(builder, base_ptr, &index, 1, "scatter_ptr");
 455       LLVMValueRef val = LLVMBuildExtractElement(builder, values, ii, "scatter_val");
 456       LLVMValueRef scalar_pred = pred ?
 457          LLVMBuildExtractElement(builder, pred, ii, "scatter_pred") : NULL;
 458
 459       if (0)
 460          lp_build_printf(gallivm, "scatter %d: val %f at %d %p\n",
 461                          ii, val, index, scalar_ptr);
 462
 463       if (scalar_pred) {
 464          LLVMValueRef real_val, dst_val;
 465          dst_val = LLVMBuildLoad(builder, scalar_ptr, "");
 466          real_val = lp_build_select(&bld->elem_bld, scalar_pred, val, dst_val);
 467          LLVMBuildStore(builder, real_val, scalar_ptr);
 468       }
 469       else {
 470          LLVMBuildStore(builder, val, scalar_ptr);
 471       }
 472    }
 473 }
 474
 475
 476 /**
 477  * Read the current value of the ADDR register, convert the floats to
 478  * ints, add the base index and return the vector of offsets.
 479  * The offsets will be used to index into the constant buffer or
 480  * temporary register file.
 481  */
 482 static LLVMValueRef
 483 get_indirect_index(struct lp_build_tgsi_soa_context *bld,
 484                    unsigned reg_file, unsigned reg_index,
 485                    const struct tgsi_src_register *indirect_reg)
 486 {
 487    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
 488    struct lp_build_context *uint_bld = &bld->bld_base.uint_bld;
 489    /* always use X component of address register */
 490    unsigned swizzle = indirect_reg->SwizzleX;
 491    LLVMValueRef base;
 492    LLVMValueRef rel;
 493    LLVMValueRef max_index;
 494    LLVMValueRef index;
 495
 496    assert(bld->indirect_files & (1 << reg_file));
 497
 498    base = lp_build_const_int_vec(bld->bld_base.base.gallivm, uint_bld->type, reg_index);
 499
 500    assert(swizzle < 4);
 501    rel = LLVMBuildLoad(builder,
 502                         bld->addr[indirect_reg->Index][swizzle],
 503                         "load addr reg");
 504
 505    index = lp_build_add(uint_bld, base, rel);
 506
 507    max_index = lp_build_const_int_vec(bld->bld_base.base.gallivm,
 508                                       uint_bld->type,
 509                                       bld->bld_base.info->file_max[reg_file]);
 510
 511    assert(!uint_bld->type.sign);
 512    index = lp_build_min(uint_bld, index, max_index);
 513
 514    return index;
 515 }
 516
 517 static struct lp_build_context *
 518 stype_to_fetch(struct lp_build_tgsi_context * bld_base,
 519                enum tgsi_opcode_type stype)
 520 {
 521    struct lp_build_context *bld_fetch;
 522
 523    switch (stype) {
 524    case TGSI_TYPE_FLOAT:
 525    case TGSI_TYPE_UNTYPED:
 526       bld_fetch = &bld_base->base;
 527       break;
 528    case TGSI_TYPE_UNSIGNED:
 529       bld_fetch = &bld_base->uint_bld;
 530       break;
 531    case TGSI_TYPE_SIGNED:
 532       bld_fetch = &bld_base->int_bld;
 533       break;
 534    case TGSI_TYPE_VOID:
 535    case TGSI_TYPE_DOUBLE:
 536    default:
 537       assert(0);
 538       bld_fetch = NULL;
 539       break;
 540    }
 541    return bld_fetch;
 542 }
 543
 544 static LLVMValueRef
 545 emit_fetch_constant(
 546    struct lp_build_tgsi_context * bld_base,
 547    const struct tgsi_full_src_register * reg,
 548    enum tgsi_opcode_type stype,
 549    unsigned swizzle)
 550 {
 551    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
 552    struct gallivm_state *gallivm = bld_base->base.gallivm;
 553    LLVMBuilderRef builder = gallivm->builder;
 554    struct lp_build_context *uint_bld = &bld_base->uint_bld;
 555    LLVMValueRef indirect_index = NULL;
 556    struct lp_build_context *bld_fetch = stype_to_fetch(bld_base, stype);
 557
 558    /* XXX: Handle fetching xyzw components as a vector */
 559    assert(swizzle != ~0);
 560
 561    if (reg->Register.Indirect) {
 562       indirect_index = get_indirect_index(bld,
 563                                           reg->Register.File,
 564                                           reg->Register.Index,
 565                                           &reg->Indirect);
 566    }
 567
 568    if (reg->Register.Indirect) {
 569       LLVMValueRef swizzle_vec =
 570          lp_build_const_int_vec(bld->bld_base.base.gallivm, uint_bld->type, swizzle);
 571       LLVMValueRef index_vec;  /* index into the const buffer */
 572
 573       /* index_vec = indirect_index * 4 + swizzle */
 574       index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
 575       index_vec = lp_build_add(uint_bld, index_vec, swizzle_vec);
 576
 577       /* Gather values from the constant buffer */
 578       return build_gather(bld_fetch, bld->consts_ptr, index_vec);
 579    }
 580    else {
 581       LLVMValueRef index;  /* index into the const buffer */
 582       LLVMValueRef scalar, scalar_ptr;
 583
 584       index = lp_build_const_int32(gallivm, reg->Register.Index*4 + swizzle);
 585
 586       scalar_ptr = LLVMBuildGEP(builder, bld->consts_ptr,
 587                                    &index, 1, "");
 588
 589       if (stype != TGSI_TYPE_FLOAT && stype != TGSI_TYPE_UNTYPED) {
 590          LLVMTypeRef ivtype = LLVMPointerType(LLVMInt32TypeInContext(gallivm->context), 0);
 591          LLVMValueRef temp_ptr;
 592          temp_ptr = LLVMBuildBitCast(builder, scalar_ptr, ivtype, "");
 593          scalar = LLVMBuildLoad(builder, temp_ptr, "");
 594       } else
 595          scalar = LLVMBuildLoad(builder, scalar_ptr, "");
 596
 597       return lp_build_broadcast_scalar(bld_fetch, scalar);
 598    }
 599 }
 600
 601 static LLVMValueRef
 602 emit_fetch_immediate(
 603    struct lp_build_tgsi_context * bld_base,
 604    const struct tgsi_full_src_register * reg,
 605    enum tgsi_opcode_type stype,
 606    unsigned swizzle)
 607 {
 608    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
 609    LLVMValueRef res = bld->immediates[reg->Register.Index][swizzle];
 610    assert(res);
 611
 612    if (stype == TGSI_TYPE_UNSIGNED) {
 613       res = LLVMConstBitCast(res, bld_base->uint_bld.vec_type);
 614    } else if (stype == TGSI_TYPE_SIGNED) {
 615       res = LLVMConstBitCast(res, bld_base->int_bld.vec_type);
 616    }
 617    return res;
 618 }
 619
 620 static LLVMValueRef
 621 emit_fetch_input(
 622    struct lp_build_tgsi_context * bld_base,
 623    const struct tgsi_full_src_register * reg,
 624    enum tgsi_opcode_type stype,
 625    unsigned swizzle)
 626 {
 627    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
 628    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
 629    LLVMBuilderRef builder = gallivm->builder;
 630    struct lp_build_context *uint_bld = &bld_base->uint_bld;
 631    LLVMValueRef indirect_index = NULL;
 632    LLVMValueRef res;
 633
 634    if (reg->Register.Indirect) {
 635       indirect_index = get_indirect_index(bld,
 636                                           reg->Register.File,
 637                                           reg->Register.Index,
 638                                           &reg->Indirect);
 639    }
 640
 641    if (reg->Register.Indirect) {
 642       LLVMValueRef swizzle_vec =
 643          lp_build_const_int_vec(gallivm, uint_bld->type, swizzle);
 644       LLVMValueRef length_vec =
 645          lp_build_const_int_vec(gallivm, uint_bld->type, bld->bld_base.base.type.length);
 646       LLVMValueRef index_vec;  /* index into the const buffer */
 647       LLVMValueRef inputs_array;
 648       LLVMTypeRef float4_ptr_type;
 649
 650       /* index_vec = (indirect_index * 4 + swizzle) * length */
 651       index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
 652       index_vec = lp_build_add(uint_bld, index_vec, swizzle_vec);
 653       index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
 654
 655       /* cast inputs_array pointer to float* */
 656       float4_ptr_type = LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
 657       inputs_array = LLVMBuildBitCast(builder, bld->inputs_array,
 658                                          float4_ptr_type, "");
 659
 660       /* Gather values from the temporary register array */
 661       res = build_gather(&bld_base->base, inputs_array, index_vec);
 662    } else {
 663       if (bld->indirect_files & (1 << TGSI_FILE_INPUT)) {
 664          LLVMValueRef lindex = lp_build_const_int32(gallivm,
 665                                         reg->Register.Index * 4 + swizzle);
 666          LLVMValueRef input_ptr =  LLVMBuildGEP(builder,
 667                                                 bld->inputs_array, &lindex, 1, "");
 668          res = LLVMBuildLoad(builder, input_ptr, "");
 669       }
 670       else {
 671          res = bld->inputs[reg->Register.Index][swizzle];
 672       }
 673    }
 674    assert(res);
 675    return res;
 676 }
 677
 678 static LLVMValueRef
 679 emit_fetch_temporary(
 680    struct lp_build_tgsi_context * bld_base,
 681    const struct tgsi_full_src_register * reg,
 682    enum tgsi_opcode_type stype,
 683    unsigned swizzle)
 684 {
 685    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
 686    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
 687    LLVMBuilderRef builder = gallivm->builder;
 688    struct lp_build_context *uint_bld = &bld_base->uint_bld;
 689    LLVMValueRef indirect_index = NULL;
 690    LLVMValueRef res;
 691
 692    if (reg->Register.Indirect) {
 693       indirect_index = get_indirect_index(bld,
 694                                           reg->Register.File,
 695                                           reg->Register.Index,
 696                                           &reg->Indirect);
 697    }
 698
 699    if (reg->Register.Indirect) {
 700       LLVMValueRef swizzle_vec =
 701          lp_build_const_int_vec(bld->bld_base.base.gallivm, uint_bld->type, swizzle);
 702       LLVMValueRef length_vec =
 703          lp_build_const_int_vec(bld->bld_base.base.gallivm, uint_bld->type,
 704                                 bld->bld_base.base.type.length);
 705       LLVMValueRef index_vec;  /* index into the const buffer */
 706       LLVMValueRef temps_array;
 707       LLVMTypeRef float4_ptr_type;
 708
 709       /* index_vec = (indirect_index * 4 + swizzle) * length */
 710       index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
 711       index_vec = lp_build_add(uint_bld, index_vec, swizzle_vec);
 712       index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
 713
 714       /* cast temps_array pointer to float* */
 715       float4_ptr_type = LLVMPointerType(LLVMFloatTypeInContext(bld->bld_base.base.gallivm->context), 0);
 716       temps_array = LLVMBuildBitCast(builder, bld->temps_array,
 717                                      float4_ptr_type, "");
 718
 719       /* Gather values from the temporary register array */
 720       res = build_gather(&bld_base->base, temps_array, index_vec);
 721    }
 722    else {
 723       LLVMValueRef temp_ptr;
 724       if (stype != TGSI_TYPE_FLOAT && stype != TGSI_TYPE_UNTYPED) {
 725          LLVMTypeRef itype = LLVMPointerType(LLVMVectorType(LLVMInt32TypeInContext(gallivm->context), 4), 0);
 726          LLVMValueRef tint_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index,
 727                                                      swizzle);
 728          temp_ptr = LLVMBuildBitCast(builder, tint_ptr, itype, "");
 729       } else
 730          temp_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index, swizzle);
 731       res = LLVMBuildLoad(builder, temp_ptr, "");
 732       if (!res)
 733          return bld->bld_base.base.undef;
 734    }
 735
 736    return res;
 737 }
 738
 739 static LLVMValueRef
 740 emit_fetch_system_value(
 741    struct lp_build_tgsi_context * bld_base,
 742    const struct tgsi_full_src_register * reg,
 743    enum tgsi_opcode_type stype,
 744    unsigned swizzle)
 745 {
 746    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
 747    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
 748    LLVMBuilderRef builder = gallivm->builder;
 749    LLVMValueRef index;  /* index into the system value array */
 750    LLVMValueRef scalar, scalar_ptr;
 751
 752    assert(!reg->Register.Indirect);
 753
 754    index = lp_build_const_int32(gallivm, reg->Register.Index * 4 + swizzle);
 755
 756    scalar_ptr = LLVMBuildGEP(builder, bld->system_values_array, &index, 1, "");
 757    scalar = LLVMBuildLoad(builder, scalar_ptr, "");
 758
 759    return lp_build_broadcast_scalar(&bld->bld_base.base, scalar);
 760 }
 761
 762 /**
 763  * Register fetch with derivatives.
 764  */
 765 static void
 766 emit_fetch_deriv(
 767    struct lp_build_tgsi_soa_context *bld,
 768    LLVMValueRef src,
 769    LLVMValueRef *res,
 770    LLVMValueRef *ddx,
 771    LLVMValueRef *ddy)
 772 {
 773    if(res)
 774       *res = src;
 775
 776    /* TODO: use interpolation coeffs for inputs */
 777
 778    if(ddx)
 779       *ddx = lp_build_ddx(&bld->bld_base.base, src);
 780
 781    if(ddy)
 782       *ddy = lp_build_ddy(&bld->bld_base.base, src);
 783 }
 784
 785
 786 /**
 787  * Predicate.
 788  */
 789 static void
 790 emit_fetch_predicate(
 791    struct lp_build_tgsi_soa_context *bld,
 792    const struct tgsi_full_instruction *inst,
 793    LLVMValueRef *pred)
 794 {
 795    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
 796    unsigned index;
 797    unsigned char swizzles[4];
 798    LLVMValueRef unswizzled[4] = {NULL, NULL, NULL, NULL};
 799    LLVMValueRef value;
 800    unsigned chan;
 801
 802    if (!inst->Instruction.Predicate) {
 803       TGSI_FOR_EACH_CHANNEL( chan ) {
 804          pred[chan] = NULL;
 805       }
 806       return;
 807    }
 808
 809    swizzles[0] = inst->Predicate.SwizzleX;
 810    swizzles[1] = inst->Predicate.SwizzleY;
 811    swizzles[2] = inst->Predicate.SwizzleZ;
 812    swizzles[3] = inst->Predicate.SwizzleW;
 813
 814    index = inst->Predicate.Index;
 815    assert(index < LP_MAX_TGSI_PREDS);
 816
 817    TGSI_FOR_EACH_CHANNEL( chan ) {
 818       unsigned swizzle = swizzles[chan];
 819
 820       /*
 821        * Only fetch the predicate register channels that are actually listed
 822        * in the swizzles
 823        */
 824       if (!unswizzled[swizzle]) {
 825          value = LLVMBuildLoad(builder,
 826                                bld->preds[index][swizzle], "");
 827
 828          /*
 829           * Convert the value to an integer mask.
 830           *
 831           * TODO: Short-circuit this comparison -- a D3D setp_xx instructions
 832           * is needlessly causing two comparisons due to storing the intermediate
 833           * result as float vector instead of an integer mask vector.
 834           */
 835          value = lp_build_compare(bld->bld_base.base.gallivm,
 836                                   bld->bld_base.base.type,
 837                                   PIPE_FUNC_NOTEQUAL,
 838                                   value,
 839                                   bld->bld_base.base.zero);
 840          if (inst->Predicate.Negate) {
 841             value = LLVMBuildNot(builder, value, "");
 842          }
 843
 844          unswizzled[swizzle] = value;
 845       } else {
 846          value = unswizzled[swizzle];
 847       }
 848
 849       pred[chan] = value;
 850    }
 851 }
 852
 853 /**
 854  * Register store.
 855  */
 856 static void
 857 emit_store_chan(
 858    struct lp_build_tgsi_context *bld_base,
 859    const struct tgsi_full_instruction *inst,
 860    unsigned index,
 861    unsigned chan_index,
 862    LLVMValueRef pred,
 863    LLVMValueRef value)
 864 {
 865    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
 866    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
 867    LLVMBuilderRef builder = gallivm->builder;
 868    const struct tgsi_full_dst_register *reg = &inst->Dst[index];
 869    struct lp_build_context *uint_bld = &bld_base->uint_bld;
 870    LLVMValueRef indirect_index = NULL;
 871    struct lp_build_context *bld_store;
 872    enum tgsi_opcode_type dtype = tgsi_opcode_infer_dst_type(inst->Instruction.Opcode);
 873
 874    switch (dtype) {
 875    default:
 876    case TGSI_TYPE_FLOAT:
 877    case TGSI_TYPE_UNTYPED:
 878       bld_store = &bld_base->base;
 879       break;
 880    case TGSI_TYPE_UNSIGNED:
 881       bld_store = &bld_base->uint_bld;
 882       break;
 883    case TGSI_TYPE_SIGNED:
 884       bld_store = &bld_base->int_bld;
 885       break;
 886    case TGSI_TYPE_DOUBLE:
 887    case TGSI_TYPE_VOID:
 888       assert(0);
 889       bld_store = NULL;
 890       break;
 891    }
 892
 893    switch( inst->Instruction.Saturate ) {
 894    case TGSI_SAT_NONE:
 895       break;
 896
 897    case TGSI_SAT_ZERO_ONE:
 898       value = lp_build_max(&bld->bld_base.base, value, bld->bld_base.base.zero);
 899       value = lp_build_min(&bld->bld_base.base, value, bld->bld_base.base.one);
 900       break;
 901
 902    case TGSI_SAT_MINUS_PLUS_ONE:
 903       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));
 904       value = lp_build_min(&bld->bld_base.base, value, bld->bld_base.base.one);
 905       break;
 906
 907    default:
 908       assert(0);
 909    }
 910
 911    if (reg->Register.Indirect) {
 912       indirect_index = get_indirect_index(bld,
 913                                           reg->Register.File,
 914                                           reg->Register.Index,
 915                                           &reg->Indirect);
 916    } else {
 917       assert(reg->Register.Index <=
 918                              bld->bld_base.info->file_max[reg->Register.File]);
 919    }
 920
 921    switch( reg->Register.File ) {
 922    case TGSI_FILE_OUTPUT:
 923       if (reg->Register.Indirect) {
 924          LLVMValueRef chan_vec =
 925             lp_build_const_int_vec(gallivm, uint_bld->type, chan_index);
 926          LLVMValueRef length_vec =
 927             lp_build_const_int_vec(gallivm, uint_bld->type, bld->bld_base.base.type.length);
 928          LLVMValueRef index_vec;  /* indexes into the temp registers */
 929          LLVMValueRef outputs_array;
 930          LLVMValueRef pixel_offsets;
 931          LLVMTypeRef float_ptr_type;
 932          int i;
 933
 934          /* build pixel offset vector: {0, 1, 2, 3, ...} */
 935          pixel_offsets = uint_bld->undef;
 936          for (i = 0; i < bld->bld_base.base.type.length; i++) {
 937             LLVMValueRef ii = lp_build_const_int32(gallivm, i);
 938             pixel_offsets = LLVMBuildInsertElement(builder, pixel_offsets,
 939                                                    ii, ii, "");
 940          }
 941
 942          /* index_vec = (indirect_index * 4 + chan_index) * length + offsets */
 943          index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
 944          index_vec = lp_build_add(uint_bld, index_vec, chan_vec);
 945          index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
 946          index_vec = lp_build_add(uint_bld, index_vec, pixel_offsets);
 947
 948          float_ptr_type =
 949             LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
 950          outputs_array = LLVMBuildBitCast(builder, bld->outputs_array,
 951                                           float_ptr_type, "");
 952
 953          /* Scatter store values into temp registers */
 954          emit_mask_scatter(bld, outputs_array, index_vec, value,
 955                            &bld->exec_mask, pred);
 956       }
 957       else {
 958          LLVMValueRef out_ptr = lp_get_output_ptr(bld, reg->Register.Index,
 959                                                chan_index);
 960          lp_exec_mask_store(&bld->exec_mask, bld_store, pred, value, out_ptr);
 961       }
 962       break;
 963
 964    case TGSI_FILE_TEMPORARY:
 965       if (reg->Register.Indirect) {
 966          LLVMValueRef chan_vec =
 967             lp_build_const_int_vec(gallivm, uint_bld->type, chan_index);
 968          LLVMValueRef length_vec =
 969             lp_build_const_int_vec(gallivm, uint_bld->type,
 970                                    bld->bld_base.base.type.length);
 971          LLVMValueRef index_vec;  /* indexes into the temp registers */
 972          LLVMValueRef temps_array;
 973          LLVMValueRef pixel_offsets;
 974          LLVMTypeRef float_ptr_type;
 975          int i;
 976
 977          /* build pixel offset vector: {0, 1, 2, 3, ...} */
 978          pixel_offsets = uint_bld->undef;
 979          for (i = 0; i < bld->bld_base.base.type.length; i++) {
 980             LLVMValueRef ii = lp_build_const_int32(gallivm, i);
 981             pixel_offsets = LLVMBuildInsertElement(builder, pixel_offsets,
 982                                                    ii, ii, "");
 983          }
 984
 985          /* index_vec = (indirect_index * 4 + chan_index) * length + offsets */
 986          index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
 987          index_vec = lp_build_add(uint_bld, index_vec, chan_vec);
 988          index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
 989          index_vec = lp_build_add(uint_bld, index_vec, pixel_offsets);
 990
 991          float_ptr_type =
 992             LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
 993          temps_array = LLVMBuildBitCast(builder, bld->temps_array,
 994                                         float_ptr_type, "");
 995
 996          /* Scatter store values into temp registers */
 997          emit_mask_scatter(bld, temps_array, index_vec, value,
 998                            &bld->exec_mask, pred);
 999       }
1000       else {
1001          LLVMValueRef temp_ptr;
1002
1003          switch (dtype) {
1004          case TGSI_TYPE_UNSIGNED:
1005          case TGSI_TYPE_SIGNED: {
1006             LLVMTypeRef itype = LLVMVectorType(LLVMInt32TypeInContext(gallivm->context), 4);
1007             LLVMTypeRef ivtype = LLVMPointerType(itype, 0);
1008             LLVMValueRef tint_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index,
1009                                                         chan_index);
1010             LLVMValueRef temp_value_ptr;
1011
1012             temp_ptr = LLVMBuildBitCast(builder, tint_ptr, ivtype, "");
1013             temp_value_ptr = LLVMBuildBitCast(builder, value, itype, "");
1014             value = temp_value_ptr;
1015             break;
1016          }
1017          default:
1018          case TGSI_TYPE_FLOAT:
1019          case TGSI_TYPE_UNTYPED:
1020             temp_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index,
1021                                            chan_index);
1022             break;
1023          }
1024
1025          lp_exec_mask_store(&bld->exec_mask, bld_store, pred, value, temp_ptr);
1026       }
1027       break;
1028
1029    case TGSI_FILE_ADDRESS:
1030       assert(dtype == TGSI_TYPE_SIGNED);
1031       assert(LLVMTypeOf(value) == bld_base->base.int_vec_type);
1032       lp_exec_mask_store(&bld->exec_mask, bld_store, pred, value,
1033                          bld->addr[reg->Register.Index][chan_index]);
1034       break;
1035
1036    case TGSI_FILE_PREDICATE:
1037       lp_exec_mask_store(&bld->exec_mask, bld_store, pred, value,
1038                          bld->preds[reg->Register.Index][chan_index]);
1039       break;
1040
1041    default:
1042       assert( 0 );
1043    }
1044 }
1045
1046 static void
1047 emit_store(
1048    struct lp_build_tgsi_context * bld_base,
1049    const struct tgsi_full_instruction * inst,
1050    const struct tgsi_opcode_info * info,
1051    LLVMValueRef dst[4])
1052
1053 {
1054    unsigned chan_index;
1055    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1056
1057    if(info->num_dst) {
1058       LLVMValueRef pred[TGSI_NUM_CHANNELS];
1059
1060       emit_fetch_predicate( bld, inst, pred );
1061
1062       TGSI_FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1063          emit_store_chan(bld_base, inst, 0, chan_index, pred[chan_index], dst[chan_index]);
1064       }
1065    }
1066 }
1067
1068 /**
1069  * High-level instruction translators.
1070  */
1071
1072 static void
1073 emit_tex( struct lp_build_tgsi_soa_context *bld,
1074           const struct tgsi_full_instruction *inst,
1075           enum lp_build_tex_modifier modifier,
1076           LLVMValueRef *texel)
1077 {
1078    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
1079    unsigned unit;
1080    LLVMValueRef lod_bias, explicit_lod;
1081    LLVMValueRef oow = NULL;
1082    LLVMValueRef coords[3];
1083    LLVMValueRef ddx[3];
1084    LLVMValueRef ddy[3];
1085    unsigned num_coords;
1086    unsigned i;
1087
1088    if (!bld->sampler) {
1089       _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
1090       for (i = 0; i < 4; i++) {
1091          texel[i] = bld->bld_base.base.undef;
1092       }
1093       return;
1094    }
1095
1096    switch (inst->Texture.Texture) {
1097    case TGSI_TEXTURE_1D:
1098       num_coords = 1;
1099       break;
1100    case TGSI_TEXTURE_1D_ARRAY:
1101    case TGSI_TEXTURE_2D:
1102    case TGSI_TEXTURE_RECT:
1103       num_coords = 2;
1104       break;
1105    case TGSI_TEXTURE_SHADOW1D:
1106    case TGSI_TEXTURE_SHADOW1D_ARRAY:
1107    case TGSI_TEXTURE_SHADOW2D:
1108    case TGSI_TEXTURE_SHADOWRECT:
1109    case TGSI_TEXTURE_2D_ARRAY:
1110    case TGSI_TEXTURE_3D:
1111    case TGSI_TEXTURE_CUBE:
1112       num_coords = 3;
1113       break;
1114    case TGSI_TEXTURE_SHADOW2D_ARRAY:
1115       num_coords = 4;
1116       break;
1117    default:
1118       assert(0);
1119       return;
1120    }
1121
1122    if (modifier == LP_BLD_TEX_MODIFIER_LOD_BIAS) {
1123       lod_bias = lp_build_emit_fetch( &bld->bld_base, inst, 0, 3 );
1124       explicit_lod = NULL;
1125    }
1126    else if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_LOD) {
1127       lod_bias = NULL;
1128       explicit_lod = lp_build_emit_fetch( &bld->bld_base, inst, 0, 3 );
1129    }
1130    else {
1131       lod_bias = NULL;
1132       explicit_lod = NULL;
1133    }
1134
1135    if (modifier == LP_BLD_TEX_MODIFIER_PROJECTED) {
1136       oow = lp_build_emit_fetch( &bld->bld_base, inst, 0, 3 );
1137       oow = lp_build_rcp(&bld->bld_base.base, oow);
1138    }
1139
1140    for (i = 0; i < num_coords; i++) {
1141       coords[i] = lp_build_emit_fetch( &bld->bld_base, inst, 0, i );
1142       if (modifier == LP_BLD_TEX_MODIFIER_PROJECTED)
1143          coords[i] = lp_build_mul(&bld->bld_base.base, coords[i], oow);
1144    }
1145    for (i = num_coords; i < 3; i++) {
1146       coords[i] = bld->bld_base.base.undef;
1147    }
1148
1149    if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV) {
1150       LLVMValueRef index0 = lp_build_const_int32(bld->bld_base.base.gallivm, 0);
1151       for (i = 0; i < num_coords; i++) {
1152          LLVMValueRef src1 = lp_build_emit_fetch( &bld->bld_base, inst, 1, i );
1153          LLVMValueRef src2 = lp_build_emit_fetch( &bld->bld_base, inst, 2, i );
1154          ddx[i] = LLVMBuildExtractElement(builder, src1, index0, "");
1155          ddy[i] = LLVMBuildExtractElement(builder, src2, index0, "");
1156       }
1157       unit = inst->Src[3].Register.Index;
1158    }  else {
1159       for (i = 0; i < num_coords; i++) {
1160          ddx[i] = lp_build_scalar_ddx( &bld->bld_base.base, coords[i] );
1161          ddy[i] = lp_build_scalar_ddy( &bld->bld_base.base, coords[i] );
1162       }
1163       unit = inst->Src[1].Register.Index;
1164    }
1165    for (i = num_coords; i < 3; i++) {
1166       ddx[i] = LLVMGetUndef(bld->bld_base.base.elem_type);
1167       ddy[i] = LLVMGetUndef(bld->bld_base.base.elem_type);
1168    }
1169
1170    bld->sampler->emit_fetch_texel(bld->sampler,
1171                                   bld->bld_base.base.gallivm,
1172                                   bld->bld_base.base.type,
1173                                   unit, num_coords, coords,
1174                                   ddx, ddy,
1175                                   lod_bias, explicit_lod,
1176                                   texel);
1177 }
1178
1179 static boolean
1180 near_end_of_shader(struct lp_build_tgsi_soa_context *bld,
1181                    int pc)
1182 {
1183    int i;
1184
1185    for (i = 0; i < 5; i++) {
1186       unsigned opcode;
1187
1188       if (pc + i >= bld->bld_base.info->num_instructions)
1189          return TRUE;
1190
1191       opcode = bld->bld_base.instructions[pc + i].Instruction.Opcode;
1192
1193       if (opcode == TGSI_OPCODE_END)
1194          return TRUE;
1195
1196       if (opcode == TGSI_OPCODE_TEX ||
1197           opcode == TGSI_OPCODE_TXP ||
1198           opcode == TGSI_OPCODE_TXD ||
1199           opcode == TGSI_OPCODE_TXB ||
1200           opcode == TGSI_OPCODE_TXL ||
1201           opcode == TGSI_OPCODE_TXF ||
1202           opcode == TGSI_OPCODE_TXQ ||
1203           opcode == TGSI_OPCODE_CAL ||
1204           opcode == TGSI_OPCODE_CALLNZ ||
1205           opcode == TGSI_OPCODE_IF ||
1206           opcode == TGSI_OPCODE_IFC ||
1207           opcode == TGSI_OPCODE_BGNLOOP ||
1208           opcode == TGSI_OPCODE_SWITCH)
1209          return FALSE;
1210    }
1211
1212    return TRUE;
1213 }
1214
1215
1216
1217 /**
1218  * Kill fragment if any of the src register values are negative.
1219  */
1220 static void
1221 emit_kil(
1222    struct lp_build_tgsi_soa_context *bld,
1223    const struct tgsi_full_instruction *inst,
1224    int pc)
1225 {
1226    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
1227    const struct tgsi_full_src_register *reg = &inst->Src[0];
1228    LLVMValueRef terms[TGSI_NUM_CHANNELS];
1229    LLVMValueRef mask;
1230    unsigned chan_index;
1231
1232    memset(&terms, 0, sizeof terms);
1233
1234    TGSI_FOR_EACH_CHANNEL( chan_index ) {
1235       unsigned swizzle;
1236
1237       /* Unswizzle channel */
1238       swizzle = tgsi_util_get_full_src_register_swizzle( reg, chan_index );
1239
1240       /* Check if the component has not been already tested. */
1241       assert(swizzle < TGSI_NUM_CHANNELS);
1242       if( !terms[swizzle] )
1243          /* TODO: change the comparison operator instead of setting the sign */
1244          terms[swizzle] =  lp_build_emit_fetch(&bld->bld_base, inst, 0, chan_index );
1245    }
1246
1247    mask = NULL;
1248    TGSI_FOR_EACH_CHANNEL( chan_index ) {
1249       if(terms[chan_index]) {
1250          LLVMValueRef chan_mask;
1251
1252          /*
1253           * If term < 0 then mask = 0 else mask = ~0.
1254           */
1255          chan_mask = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_GEQUAL, terms[chan_index], bld->bld_base.base.zero);
1256
1257          if(mask)
1258             mask = LLVMBuildAnd(builder, mask, chan_mask, "");
1259          else
1260             mask = chan_mask;
1261       }
1262    }
1263
1264    if(mask) {
1265       lp_build_mask_update(bld->mask, mask);
1266
1267       if (!near_end_of_shader(bld, pc))
1268          lp_build_mask_check(bld->mask);
1269    }
1270 }
1271
1272
1273 /**
1274  * Predicated fragment kill.
1275  * XXX Actually, we do an unconditional kill (as in tgsi_exec.c).
1276  * The only predication is the execution mask which will apply if
1277  * we're inside a loop or conditional.
1278  */
1279 static void
1280 emit_kilp(struct lp_build_tgsi_soa_context *bld,
1281           int pc)
1282 {
1283    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
1284    LLVMValueRef mask;
1285
1286    /* For those channels which are "alive", disable fragment shader
1287     * execution.
1288     */
1289    if (bld->exec_mask.has_mask) {
1290       mask = LLVMBuildNot(builder, bld->exec_mask.exec_mask, "kilp");
1291    }
1292    else {
1293       LLVMValueRef zero = LLVMConstNull(bld->bld_base.base.int_vec_type);
1294       mask = zero;
1295    }
1296
1297    lp_build_mask_update(bld->mask, mask);
1298
1299    if (!near_end_of_shader(bld, pc))
1300       lp_build_mask_check(bld->mask);
1301 }
1302
1303
1304 /**
1305  * Emit code which will dump the value of all the temporary registers
1306  * to stdout.
1307  */
1308 static void
1309 emit_dump_temps(struct lp_build_tgsi_soa_context *bld)
1310 {
1311    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
1312    LLVMBuilderRef builder = gallivm->builder;
1313    LLVMValueRef temp_ptr;
1314    LLVMValueRef i0 = lp_build_const_int32(gallivm, 0);
1315    LLVMValueRef i1 = lp_build_const_int32(gallivm, 1);
1316    LLVMValueRef i2 = lp_build_const_int32(gallivm, 2);
1317    LLVMValueRef i3 = lp_build_const_int32(gallivm, 3);
1318    int index;
1319    int n = bld->bld_base.info->file_max[TGSI_FILE_TEMPORARY];
1320
1321    for (index = 0; index < n; index++) {
1322       LLVMValueRef idx = lp_build_const_int32(gallivm, index);
1323       LLVMValueRef v[4][4], res;
1324       int chan;
1325
1326       lp_build_printf(gallivm, "TEMP[%d]:\n", idx);
1327
1328       for (chan = 0; chan < 4; chan++) {
1329          temp_ptr = lp_get_temp_ptr_soa(bld, index, chan);
1330          res = LLVMBuildLoad(builder, temp_ptr, "");
1331          v[chan][0] = LLVMBuildExtractElement(builder, res, i0, "");
1332          v[chan][1] = LLVMBuildExtractElement(builder, res, i1, "");
1333          v[chan][2] = LLVMBuildExtractElement(builder, res, i2, "");
1334          v[chan][3] = LLVMBuildExtractElement(builder, res, i3, "");
1335       }
1336
1337       lp_build_printf(gallivm, "  X: %f %f %f %f\n",
1338                       v[0][0], v[0][1], v[0][2], v[0][3]);
1339       lp_build_printf(gallivm, "  Y: %f %f %f %f\n",
1340                       v[1][0], v[1][1], v[1][2], v[1][3]);
1341       lp_build_printf(gallivm, "  Z: %f %f %f %f\n",
1342                       v[2][0], v[2][1], v[2][2], v[2][3]);
1343       lp_build_printf(gallivm, "  W: %f %f %f %f\n",
1344                       v[3][0], v[3][1], v[3][2], v[3][3]);
1345    }
1346 }
1347
1348
1349
1350 void
1351 lp_emit_declaration_soa(
1352    struct lp_build_tgsi_context *bld_base,
1353    const struct tgsi_full_declaration *decl)
1354 {
1355    struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
1356    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
1357    LLVMTypeRef vec_type = bld->bld_base.base.vec_type;
1358    const unsigned first = decl->Range.First;
1359    const unsigned last = decl->Range.Last;
1360    unsigned idx, i;
1361
1362    for (idx = first; idx <= last; ++idx) {
1363       assert(last <= bld->bld_base.info->file_max[decl->Declaration.File]);
1364       switch (decl->Declaration.File) {
1365       case TGSI_FILE_TEMPORARY:
1366          assert(idx < LP_MAX_TGSI_TEMPS);
1367          if (!(bld->indirect_files & (1 << TGSI_FILE_TEMPORARY))) {
1368             for (i = 0; i < TGSI_NUM_CHANNELS; i++)
1369                bld->temps[idx][i] = lp_build_alloca(gallivm, vec_type, "temp");
1370          }
1371          break;
1372
1373       case TGSI_FILE_OUTPUT:
1374          if (!(bld->indirect_files & (1 << TGSI_FILE_OUTPUT))) {
1375             for (i = 0; i < TGSI_NUM_CHANNELS; i++)
1376                bld->outputs[idx][i] = lp_build_alloca(gallivm,
1377                                                       vec_type, "output");
1378          }
1379          break;
1380
1381       case TGSI_FILE_ADDRESS:
1382          /* ADDR registers are the only allocated with an integer LLVM IR type,
1383           * as they are guaranteed to always have integers.
1384           * XXX: Not sure if this exception is worthwhile (or the whole idea of
1385           * an ADDR register for that matter).
1386           */
1387          assert(idx < LP_MAX_TGSI_ADDRS);
1388          for (i = 0; i < TGSI_NUM_CHANNELS; i++)
1389             bld->addr[idx][i] = lp_build_alloca(gallivm, bld_base->base.int_vec_type, "addr");
1390          break;
1391
1392       case TGSI_FILE_PREDICATE:
1393          assert(idx < LP_MAX_TGSI_PREDS);
1394          for (i = 0; i < TGSI_NUM_CHANNELS; i++)
1395             bld->preds[idx][i] = lp_build_alloca(gallivm, vec_type,
1396                                                  "predicate");
1397          break;
1398
1399       default:
1400          /* don't need to declare other vars */
1401          break;
1402       }
1403    }
1404 }
1405
1406
1407 void lp_emit_immediate_soa(
1408    struct lp_build_tgsi_context *bld_base,
1409    const struct tgsi_full_immediate *imm)
1410 {
1411    struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
1412    struct gallivm_state * gallivm = bld_base->base.gallivm;
1413
1414    /* simply copy the immediate values into the next immediates[] slot */
1415    unsigned i;
1416    const uint size = imm->Immediate.NrTokens - 1;
1417    assert(size <= 4);
1418    assert(bld->num_immediates < LP_MAX_TGSI_IMMEDIATES);
1419    switch (imm->Immediate.DataType) {
1420    case TGSI_IMM_FLOAT32:
1421       for( i = 0; i < size; ++i )
1422          bld->immediates[bld->num_immediates][i] =
1423             lp_build_const_vec(gallivm, bld_base->base.type, imm->u[i].Float);
1424
1425       break;
1426    case TGSI_IMM_UINT32:
1427       for( i = 0; i < size; ++i ) {
1428          LLVMValueRef tmp = lp_build_const_vec(gallivm, bld_base->uint_bld.type, imm->u[i].Uint);
1429          bld->immediates[bld->num_immediates][i] =
1430             LLVMConstBitCast(tmp, bld_base->base.vec_type);
1431       }
1432
1433       break;
1434    case TGSI_IMM_INT32:
1435       for( i = 0; i < size; ++i ) {
1436          LLVMValueRef tmp = lp_build_const_vec(gallivm, bld_base->int_bld.type, imm->u[i].Int);
1437          bld->immediates[bld->num_immediates][i] =
1438             LLVMConstBitCast(tmp, bld_base->base.vec_type);
1439       }
1440
1441       break;
1442    }
1443    for( i = size; i < 4; ++i )
1444       bld->immediates[bld->num_immediates][i] = bld_base->base.undef;
1445
1446    bld->num_immediates++;
1447 }
1448
1449 static void
1450 ddx_emit(
1451    const struct lp_build_tgsi_action * action,
1452    struct lp_build_tgsi_context * bld_base,
1453    struct lp_build_emit_data * emit_data)
1454 {
1455    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1456
1457    emit_fetch_deriv(bld, emit_data->args[0], NULL,
1458                     &emit_data->output[emit_data->chan], NULL);
1459 }
1460
1461 static void
1462 ddy_emit(
1463    const struct lp_build_tgsi_action * action,
1464    struct lp_build_tgsi_context * bld_base,
1465    struct lp_build_emit_data * emit_data)
1466 {
1467    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1468
1469    emit_fetch_deriv(bld, emit_data->args[0], NULL, NULL,
1470                     &emit_data->output[emit_data->chan]);
1471 }
1472
1473 static void
1474 kilp_emit(
1475    const struct lp_build_tgsi_action * action,
1476    struct lp_build_tgsi_context * bld_base,
1477    struct lp_build_emit_data * emit_data)
1478 {
1479    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1480
1481    emit_kilp(bld, bld_base->pc - 1);
1482 }
1483
1484 static void
1485 kil_emit(
1486    const struct lp_build_tgsi_action * action,
1487    struct lp_build_tgsi_context * bld_base,
1488    struct lp_build_emit_data * emit_data)
1489 {
1490    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1491
1492    emit_kil(bld, emit_data->inst, bld_base->pc - 1);
1493 }
1494
1495 static void
1496 tex_emit(
1497    const struct lp_build_tgsi_action * action,
1498    struct lp_build_tgsi_context * bld_base,
1499    struct lp_build_emit_data * emit_data)
1500 {
1501    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1502
1503    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_NONE, emit_data->output);
1504 }
1505
1506 static void
1507 txb_emit(
1508    const struct lp_build_tgsi_action * action,
1509    struct lp_build_tgsi_context * bld_base,
1510    struct lp_build_emit_data * emit_data)
1511 {
1512    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1513
1514    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_LOD_BIAS,
1515             emit_data->output);
1516 }
1517
1518 static void
1519 txd_emit(
1520    const struct lp_build_tgsi_action * action,
1521    struct lp_build_tgsi_context * bld_base,
1522    struct lp_build_emit_data * emit_data)
1523 {
1524    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1525
1526    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV,
1527             emit_data->output);
1528 }
1529
1530 static void
1531 txl_emit(
1532    const struct lp_build_tgsi_action * action,
1533    struct lp_build_tgsi_context * bld_base,
1534    struct lp_build_emit_data * emit_data)
1535 {
1536    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1537
1538    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD,
1539             emit_data->output);
1540 }
1541
1542 static void
1543 txp_emit(
1544    const struct lp_build_tgsi_action * action,
1545    struct lp_build_tgsi_context * bld_base,
1546    struct lp_build_emit_data * emit_data)
1547 {
1548    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1549
1550    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_PROJECTED,
1551             emit_data->output);
1552 }
1553
1554 static void
1555 cal_emit(
1556    const struct lp_build_tgsi_action * action,
1557    struct lp_build_tgsi_context * bld_base,
1558    struct lp_build_emit_data * emit_data)
1559 {
1560    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1561
1562    lp_exec_mask_call(&bld->exec_mask, emit_data->inst->Label.Label,
1563                      &bld_base->pc);
1564 }
1565
1566 static void
1567 ret_emit(
1568    const struct lp_build_tgsi_action * action,
1569    struct lp_build_tgsi_context * bld_base,
1570    struct lp_build_emit_data * emit_data)
1571 {
1572    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1573
1574    lp_exec_mask_ret(&bld->exec_mask, &bld_base->pc);
1575 }
1576
1577 static void
1578 brk_emit(
1579    const struct lp_build_tgsi_action * action,
1580    struct lp_build_tgsi_context * bld_base,
1581    struct lp_build_emit_data * emit_data)
1582 {
1583    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1584
1585    lp_exec_break(&bld->exec_mask);
1586 }
1587
1588 static void
1589 if_emit(
1590    const struct lp_build_tgsi_action * action,
1591    struct lp_build_tgsi_context * bld_base,
1592    struct lp_build_emit_data * emit_data)
1593 {
1594    LLVMValueRef tmp;
1595    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1596
1597    tmp = lp_build_cmp(&bld_base->base, PIPE_FUNC_NOTEQUAL,
1598                       emit_data->args[0], bld->bld_base.base.zero);
1599    lp_exec_mask_cond_push(&bld->exec_mask, tmp);
1600 }
1601
1602 static void
1603 bgnloop_emit(
1604    const struct lp_build_tgsi_action * action,
1605    struct lp_build_tgsi_context * bld_base,
1606    struct lp_build_emit_data * emit_data)
1607 {
1608    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1609
1610    lp_exec_bgnloop(&bld->exec_mask);
1611 }
1612
1613 static void
1614 bgnsub_emit(
1615    const struct lp_build_tgsi_action * action,
1616    struct lp_build_tgsi_context * bld_base,
1617    struct lp_build_emit_data * emit_data)
1618 {
1619    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1620
1621    lp_exec_mask_bgnsub(&bld->exec_mask);
1622 }
1623
1624 static void
1625 else_emit(
1626    const struct lp_build_tgsi_action * action,
1627    struct lp_build_tgsi_context * bld_base,
1628    struct lp_build_emit_data * emit_data)
1629 {
1630    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1631
1632    lp_exec_mask_cond_invert(&bld->exec_mask);
1633 }
1634
1635 static void
1636 endif_emit(
1637    const struct lp_build_tgsi_action * action,
1638    struct lp_build_tgsi_context * bld_base,
1639    struct lp_build_emit_data * emit_data)
1640 {
1641    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1642
1643    lp_exec_mask_cond_pop(&bld->exec_mask);
1644 }
1645
1646 static void
1647 endloop_emit(
1648    const struct lp_build_tgsi_action * action,
1649    struct lp_build_tgsi_context * bld_base,
1650    struct lp_build_emit_data * emit_data)
1651 {
1652    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1653
1654    lp_exec_endloop(bld_base->base.gallivm, &bld->exec_mask);
1655 }
1656
1657 static void
1658 endsub_emit(
1659    const struct lp_build_tgsi_action * action,
1660    struct lp_build_tgsi_context * bld_base,
1661    struct lp_build_emit_data * emit_data)
1662 {
1663    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1664
1665    lp_exec_mask_endsub(&bld->exec_mask, &bld_base->pc);
1666 }
1667
1668 static void
1669 cont_emit(
1670    const struct lp_build_tgsi_action * action,
1671    struct lp_build_tgsi_context * bld_base,
1672    struct lp_build_emit_data * emit_data)
1673 {
1674    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1675
1676    lp_exec_continue(&bld->exec_mask);
1677 }
1678
1679 /* XXX: Refactor and move it to lp_bld_tgsi_action.c
1680  *
1681  * XXX: What do the comments about xmm registers mean?  Maybe they are left over
1682  * from old code, but there is no garauntee that LLVM will use those registers
1683  * for this code.
1684  *
1685  * XXX: There should be no calls to lp_build_emit_fetch in this function.  This
1686  * should be handled by the emit_data->fetch_args function. */
1687 static void
1688 nrm_emit(
1689    const struct lp_build_tgsi_action * action,
1690    struct lp_build_tgsi_context * bld_base,
1691    struct lp_build_emit_data * emit_data)
1692 {
1693    LLVMValueRef tmp0, tmp1;
1694    LLVMValueRef tmp4 = NULL;
1695    LLVMValueRef tmp5 = NULL;
1696    LLVMValueRef tmp6 = NULL;
1697    LLVMValueRef tmp7 = NULL;
1698    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1699
1700    uint dims = (emit_data->inst->Instruction.Opcode == TGSI_OPCODE_NRM) ? 3 : 4;
1701
1702   if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_X) ||
1703       TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Y) ||
1704       TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Z) ||
1705       (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_W) && dims == 4)) {
1706
1707       /* NOTE: Cannot use xmm regs 2/3 here (see emit_rsqrt() above). */
1708
1709       /* xmm4 = src.x */
1710       /* xmm0 = src.x * src.x */
1711       tmp0 = lp_build_emit_fetch(&bld->bld_base, emit_data->inst, 0, TGSI_CHAN_X);
1712       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_X)) {
1713          tmp4 = tmp0;
1714       }
1715       tmp0 = lp_build_mul( &bld->bld_base.base, tmp0, tmp0);
1716
1717       /* xmm5 = src.y */
1718       /* xmm0 = xmm0 + src.y * src.y */
1719       tmp1 = lp_build_emit_fetch(&bld->bld_base, emit_data->inst, 0, TGSI_CHAN_Y);
1720       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Y)) {
1721          tmp5 = tmp1;
1722       }
1723       tmp1 = lp_build_mul( &bld->bld_base.base, tmp1, tmp1);
1724       tmp0 = lp_build_add( &bld->bld_base.base, tmp0, tmp1);
1725
1726       /* xmm6 = src.z */
1727       /* xmm0 = xmm0 + src.z * src.z */
1728       tmp1 = lp_build_emit_fetch(&bld->bld_base, emit_data->inst, 0, TGSI_CHAN_Z);
1729       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Z)) {
1730          tmp6 = tmp1;
1731       }
1732       tmp1 = lp_build_mul( &bld->bld_base.base, tmp1, tmp1);
1733       tmp0 = lp_build_add( &bld->bld_base.base, tmp0, tmp1);
1734
1735       if (dims == 4) {
1736          /* xmm7 = src.w */
1737          /* xmm0 = xmm0 + src.w * src.w */
1738          tmp1 = lp_build_emit_fetch(&bld->bld_base, emit_data->inst, 0, TGSI_CHAN_W);
1739          if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_W)) {
1740             tmp7 = tmp1;
1741          }
1742          tmp1 = lp_build_mul( &bld->bld_base.base, tmp1, tmp1);
1743          tmp0 = lp_build_add( &bld->bld_base.base, tmp0, tmp1);
1744       }
1745       /* xmm1 = 1 / sqrt(xmm0) */
1746       tmp1 = lp_build_rsqrt( &bld->bld_base.base, tmp0);
1747        /* dst.x = xmm1 * src.x */
1748       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_X)) {
1749          emit_data->output[TGSI_CHAN_X] = lp_build_mul( &bld->bld_base.base, tmp4, tmp1);
1750       }
1751       /* dst.y = xmm1 * src.y */
1752       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Y)) {
1753          emit_data->output[TGSI_CHAN_Y] = lp_build_mul( &bld->bld_base.base, tmp5, tmp1);
1754       }
1755
1756       /* dst.z = xmm1 * src.z */
1757       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Z)) {
1758          emit_data->output[TGSI_CHAN_Z] = lp_build_mul( &bld->bld_base.base, tmp6, tmp1);
1759       }
1760       /* dst.w = xmm1 * src.w */
1761       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_X) && dims == 4) {
1762          emit_data->output[TGSI_CHAN_W] = lp_build_mul( &bld->bld_base.base, tmp7, tmp1);
1763       }
1764    }
1765
1766    /* dst.w = 1.0 */
1767    if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_W) && dims == 3) {
1768        emit_data->output[TGSI_CHAN_W] = bld->bld_base.base.one;
1769    }
1770 }
1771
1772 static void emit_prologue(struct lp_build_tgsi_context * bld_base)
1773 {
1774    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1775    struct gallivm_state * gallivm = bld_base->base.gallivm;
1776
1777    if (bld->indirect_files & (1 << TGSI_FILE_TEMPORARY)) {
1778       LLVMValueRef array_size =
1779          lp_build_const_int32(gallivm,
1780                          bld_base->info->file_max[TGSI_FILE_TEMPORARY] * 4 + 4);
1781       bld->temps_array = lp_build_array_alloca(gallivm,
1782                                               bld_base->base.vec_type, array_size,
1783                                               "temp_array");
1784    }
1785
1786    if (bld->indirect_files & (1 << TGSI_FILE_OUTPUT)) {
1787       LLVMValueRef array_size =
1788          lp_build_const_int32(gallivm,
1789                             bld_base->info->file_max[TGSI_FILE_OUTPUT] * 4 + 4);
1790       bld->outputs_array = lp_build_array_alloca(gallivm,
1791                                                 bld_base->base.vec_type, array_size,
1792                                                 "output_array");
1793    }
1794
1795    /* If we have indirect addressing in inputs we need to copy them into
1796     * our alloca array to be able to iterate over them */
1797    if (bld->indirect_files & (1 << TGSI_FILE_INPUT)) {
1798       unsigned index, chan;
1799       LLVMTypeRef vec_type = bld_base->base.vec_type;
1800       LLVMValueRef array_size = lp_build_const_int32(gallivm,
1801             bld_base->info->file_max[TGSI_FILE_INPUT]*4 + 4);
1802       bld->inputs_array = lp_build_array_alloca(gallivm,
1803                                                vec_type, array_size,
1804                                                "input_array");
1805
1806       assert(bld_base->info->num_inputs
1807                         <= bld_base->info->file_max[TGSI_FILE_INPUT] + 1);
1808
1809       for (index = 0; index < bld_base->info->num_inputs; ++index) {
1810          for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
1811             LLVMValueRef lindex =
1812                lp_build_const_int32(gallivm, index * 4 + chan);
1813             LLVMValueRef input_ptr =
1814                LLVMBuildGEP(gallivm->builder, bld->inputs_array,
1815                             &lindex, 1, "");
1816             LLVMValueRef value = bld->inputs[index][chan];
1817             if (value)
1818                LLVMBuildStore(gallivm->builder, value, input_ptr);
1819          }
1820       }
1821    }
1822 }
1823
1824 static void emit_epilogue(struct lp_build_tgsi_context * bld_base)
1825 {
1826    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1827
1828    if (0) {
1829       /* for debugging */
1830       emit_dump_temps(bld);
1831    }
1832
1833    /* If we have indirect addressing in outputs we need to copy our alloca array
1834     * to the outputs slots specified by the called */
1835    if (bld->indirect_files & (1 << TGSI_FILE_OUTPUT)) {
1836       unsigned index, chan;
1837       assert(bld_base->info->num_outputs <=
1838                         bld_base->info->file_max[TGSI_FILE_OUTPUT] + 1);
1839       for (index = 0; index < bld_base->info->num_outputs; ++index) {
1840          for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
1841             bld->outputs[index][chan] = lp_get_output_ptr(bld, index, chan);
1842          }
1843       }
1844    }
1845 }
1846
1847 void
1848 lp_build_tgsi_soa(struct gallivm_state *gallivm,
1849                   const struct tgsi_token *tokens,
1850                   struct lp_type type,
1851                   struct lp_build_mask_context *mask,
1852                   LLVMValueRef consts_ptr,
1853                   LLVMValueRef system_values_array,
1854                   const LLVMValueRef *pos,
1855                   const LLVMValueRef (*inputs)[TGSI_NUM_CHANNELS],
1856                   LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1857                   struct lp_build_sampler_soa *sampler,
1858                   const struct tgsi_shader_info *info)
1859 {
1860    struct lp_build_tgsi_soa_context bld;
1861
1862    struct lp_type res_type;
1863
1864    assert(type.length <= LP_MAX_VECTOR_LENGTH);
1865    memset(&res_type, 0, sizeof res_type);
1866    res_type.width = type.width;
1867    res_type.length = type.length;
1868    res_type.sign = 1;
1869
1870    /* Setup build context */
1871    memset(&bld, 0, sizeof bld);
1872    lp_build_context_init(&bld.bld_base.base, gallivm, type);
1873    lp_build_context_init(&bld.bld_base.uint_bld, gallivm, lp_uint_type(type));
1874    lp_build_context_init(&bld.bld_base.int_bld, gallivm, lp_int_type(type));
1875    lp_build_context_init(&bld.elem_bld, gallivm, lp_elem_type(type));
1876    bld.mask = mask;
1877    bld.pos = pos;
1878    bld.inputs = inputs;
1879    bld.outputs = outputs;
1880    bld.consts_ptr = consts_ptr;
1881    bld.sampler = sampler;
1882    bld.bld_base.info = info;
1883    bld.indirect_files = info->indirect_files;
1884
1885    bld.bld_base.soa = TRUE;
1886    bld.bld_base.emit_fetch_funcs[TGSI_FILE_CONSTANT] = emit_fetch_constant;
1887    bld.bld_base.emit_fetch_funcs[TGSI_FILE_IMMEDIATE] = emit_fetch_immediate;
1888    bld.bld_base.emit_fetch_funcs[TGSI_FILE_INPUT] = emit_fetch_input;
1889    bld.bld_base.emit_fetch_funcs[TGSI_FILE_TEMPORARY] = emit_fetch_temporary;
1890    bld.bld_base.emit_fetch_funcs[TGSI_FILE_SYSTEM_VALUE] = emit_fetch_system_value;
1891    bld.bld_base.emit_store = emit_store;
1892
1893    bld.bld_base.emit_declaration = lp_emit_declaration_soa;
1894    bld.bld_base.emit_immediate = lp_emit_immediate_soa;
1895
1896    bld.bld_base.emit_prologue = emit_prologue;
1897    bld.bld_base.emit_epilogue = emit_epilogue;
1898
1899    /* Set opcode actions */
1900    lp_set_default_actions_cpu(&bld.bld_base);
1901
1902    bld.bld_base.op_actions[TGSI_OPCODE_BGNLOOP].emit = bgnloop_emit;
1903    bld.bld_base.op_actions[TGSI_OPCODE_BGNSUB].emit = bgnsub_emit;
1904    bld.bld_base.op_actions[TGSI_OPCODE_BRK].emit = brk_emit;
1905    bld.bld_base.op_actions[TGSI_OPCODE_CAL].emit = cal_emit;
1906    bld.bld_base.op_actions[TGSI_OPCODE_CONT].emit = cont_emit;
1907    bld.bld_base.op_actions[TGSI_OPCODE_DDX].emit = ddx_emit;
1908    bld.bld_base.op_actions[TGSI_OPCODE_DDY].emit = ddy_emit;
1909    bld.bld_base.op_actions[TGSI_OPCODE_ELSE].emit = else_emit;
1910    bld.bld_base.op_actions[TGSI_OPCODE_ENDIF].emit = endif_emit;
1911    bld.bld_base.op_actions[TGSI_OPCODE_ENDLOOP].emit = endloop_emit;
1912    bld.bld_base.op_actions[TGSI_OPCODE_ENDSUB].emit = endsub_emit;
1913    bld.bld_base.op_actions[TGSI_OPCODE_IF].emit = if_emit;
1914    bld.bld_base.op_actions[TGSI_OPCODE_KIL].emit = kil_emit;
1915    bld.bld_base.op_actions[TGSI_OPCODE_KILP].emit = kilp_emit;
1916    bld.bld_base.op_actions[TGSI_OPCODE_NRM].emit = nrm_emit;
1917    bld.bld_base.op_actions[TGSI_OPCODE_NRM4].emit = nrm_emit;
1918    bld.bld_base.op_actions[TGSI_OPCODE_RET].emit = ret_emit;
1919    bld.bld_base.op_actions[TGSI_OPCODE_TEX].emit = tex_emit;
1920    bld.bld_base.op_actions[TGSI_OPCODE_TXB].emit = txb_emit;
1921    bld.bld_base.op_actions[TGSI_OPCODE_TXD].emit = txd_emit;
1922    bld.bld_base.op_actions[TGSI_OPCODE_TXL].emit = txl_emit;
1923    bld.bld_base.op_actions[TGSI_OPCODE_TXP].emit = txp_emit;
1924
1925    lp_exec_mask_init(&bld.exec_mask, &bld.bld_base.base);
1926
1927
1928    bld.system_values_array = system_values_array;
1929
1930    lp_build_tgsi_llvm(&bld.bld_base, tokens);
1931
1932    if (0) {
1933       LLVMBasicBlockRef block = LLVMGetInsertBlock(gallivm->builder);
1934       LLVMValueRef function = LLVMGetBasicBlockParent(block);
1935       debug_printf("11111111111111111111111111111 \n");
1936       tgsi_dump(tokens, 0);
1937       lp_debug_dump_value(function);
1938       debug_printf("2222222222222222222222222222 \n");
1939    }
1940
1941    if (0) {
1942       LLVMModuleRef module = LLVMGetGlobalParent(
1943          LLVMGetBasicBlockParent(LLVMGetInsertBlock(gallivm->builder)));
1944       LLVMDumpModule(module);
1945
1946    }
1947 }
1948
1949
1950 /**
1951  * Build up the system values array out of individual values such as
1952  * the instance ID, front-face, primitive ID, etc.  The shader info is
1953  * used to determine which system values are needed and where to put
1954  * them in the system values array.
1955  *
1956  * XXX only instance ID is implemented at this time.
1957  *
1958  * The system values register file is similar to the constants buffer.
1959  * Example declaration:
1960  *    DCL SV[0], INSTANCEID
1961  * Example instruction:
1962  *    MOVE foo, SV[0].xxxx;
1963  *
1964  * \return  LLVM float array (interpreted as float [][4])
1965  */
1966 LLVMValueRef
1967 lp_build_system_values_array(struct gallivm_state *gallivm,
1968                              const struct tgsi_shader_info *info,
1969                              LLVMValueRef instance_id,
1970                              LLVMValueRef facing)
1971 {
1972    LLVMValueRef size = lp_build_const_int32(gallivm, 4 * info->num_system_values);
1973    LLVMTypeRef float_t = LLVMFloatTypeInContext(gallivm->context);
1974    LLVMValueRef array = lp_build_array_alloca(gallivm, float_t,
1975                                               size, "sysvals_array");
1976    unsigned i;
1977
1978    for (i = 0; i < info->num_system_values; i++) {
1979       LLVMValueRef index = lp_build_const_int32(gallivm, i * 4);
1980       LLVMValueRef ptr, value = 0;
1981
1982       switch (info->system_value_semantic_name[i]) {
1983       case TGSI_SEMANTIC_INSTANCEID:
1984          /* convert instance ID from int to float */
1985          value = LLVMBuildSIToFP(gallivm->builder, instance_id, float_t,
1986                                  "sysval_instanceid");
1987          break;
1988       case TGSI_SEMANTIC_FACE:
1989          /* fall-through */
1990       default:
1991          assert(0 && "unexpected semantic in build_system_values_array()");
1992       }
1993
1994       ptr = LLVMBuildGEP(gallivm->builder, array, &index, 1, "");
1995       LLVMBuildStore(gallivm->builder, value, ptr);
1996    }
1997
1998    return array;
1999 }