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    LLVMTypeRef int_type = LLVMInt32TypeInContext(bld->gallivm->context);
  70    LLVMBuilderRef builder = bld->gallivm->builder;
  71
  72    mask->bld = bld;
  73    mask->has_mask = FALSE;
  74    mask->cond_stack_size = 0;
  75    mask->loop_stack_size = 0;
  76    mask->call_stack_size = 0;
  77
  78    mask->int_vec_type = lp_build_int_vec_type(bld->gallivm, mask->bld->type);
  79    mask->exec_mask = mask->ret_mask = mask->break_mask = mask->cont_mask = mask->cond_mask =
  80          LLVMConstAllOnes(mask->int_vec_type);
  81
  82    mask->loop_limiter = lp_build_alloca(bld->gallivm, int_type, "looplimiter");
  83
  84    LLVMBuildStore(
  85       builder,
  86       LLVMConstInt(int_type, LP_MAX_TGSI_LOOP_ITERATIONS, false),
  87       mask->loop_limiter);
  88 }
  89
  90 static void lp_exec_mask_update(struct lp_exec_mask *mask)
  91 {
  92    LLVMBuilderRef builder = mask->bld->gallivm->builder;
  93
  94    if (mask->loop_stack_size) {
  95       /*for loops we need to update the entire mask at runtime */
  96       LLVMValueRef tmp;
  97       assert(mask->break_mask);
  98       tmp = LLVMBuildAnd(builder,
  99                          mask->cont_mask,
 100                          mask->break_mask,
 101                          "maskcb");
 102       mask->exec_mask = LLVMBuildAnd(builder,
 103                                      mask->cond_mask,
 104                                      tmp,
 105                                      "maskfull");
 106    } else
 107       mask->exec_mask = mask->cond_mask;
 108
 109    if (mask->call_stack_size) {
 110       mask->exec_mask = LLVMBuildAnd(builder,
 111                                      mask->exec_mask,
 112                                      mask->ret_mask,
 113                                      "callmask");
 114    }
 115
 116    mask->has_mask = (mask->cond_stack_size > 0 ||
 117                      mask->loop_stack_size > 0 ||
 118                      mask->call_stack_size > 0);
 119 }
 120
 121 static void lp_exec_mask_cond_push(struct lp_exec_mask *mask,
 122                                    LLVMValueRef val)
 123 {
 124    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 125
 126    assert(mask->cond_stack_size < LP_MAX_TGSI_NESTING);
 127    if (mask->cond_stack_size == 0) {
 128       assert(mask->cond_mask == LLVMConstAllOnes(mask->int_vec_type));
 129    }
 130    mask->cond_stack[mask->cond_stack_size++] = mask->cond_mask;
 131    assert(LLVMTypeOf(val) == mask->int_vec_type);
 132    mask->cond_mask = LLVMBuildAnd(builder,
 133                                   mask->cond_mask,
 134                                   val,
 135                                   "");
 136    lp_exec_mask_update(mask);
 137 }
 138
 139 static void lp_exec_mask_cond_invert(struct lp_exec_mask *mask)
 140 {
 141    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 142    LLVMValueRef prev_mask;
 143    LLVMValueRef inv_mask;
 144
 145    assert(mask->cond_stack_size);
 146    prev_mask = mask->cond_stack[mask->cond_stack_size - 1];
 147    if (mask->cond_stack_size == 1) {
 148       assert(prev_mask == LLVMConstAllOnes(mask->int_vec_type));
 149    }
 150
 151    inv_mask = LLVMBuildNot(builder, mask->cond_mask, "");
 152
 153    mask->cond_mask = LLVMBuildAnd(builder,
 154                                   inv_mask,
 155                                   prev_mask, "");
 156    lp_exec_mask_update(mask);
 157 }
 158
 159 static void lp_exec_mask_cond_pop(struct lp_exec_mask *mask)
 160 {
 161    assert(mask->cond_stack_size);
 162    mask->cond_mask = mask->cond_stack[--mask->cond_stack_size];
 163    lp_exec_mask_update(mask);
 164 }
 165
 166 static void lp_exec_bgnloop(struct lp_exec_mask *mask)
 167 {
 168    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 169
 170    if (mask->loop_stack_size == 0) {
 171       assert(mask->loop_block == NULL);
 172       assert(mask->cont_mask == LLVMConstAllOnes(mask->int_vec_type));
 173       assert(mask->break_mask == LLVMConstAllOnes(mask->int_vec_type));
 174       assert(mask->break_var == NULL);
 175    }
 176
 177    assert(mask->loop_stack_size < LP_MAX_TGSI_NESTING);
 178
 179    mask->loop_stack[mask->loop_stack_size].loop_block = mask->loop_block;
 180    mask->loop_stack[mask->loop_stack_size].cont_mask = mask->cont_mask;
 181    mask->loop_stack[mask->loop_stack_size].break_mask = mask->break_mask;
 182    mask->loop_stack[mask->loop_stack_size].break_var = mask->break_var;
 183    ++mask->loop_stack_size;
 184
 185    mask->break_var = lp_build_alloca(mask->bld->gallivm, mask->int_vec_type, "");
 186    LLVMBuildStore(builder, mask->break_mask, mask->break_var);
 187
 188    mask->loop_block = lp_build_insert_new_block(mask->bld->gallivm, "bgnloop");
 189
 190    LLVMBuildBr(builder, mask->loop_block);
 191    LLVMPositionBuilderAtEnd(builder, mask->loop_block);
 192
 193    mask->break_mask = LLVMBuildLoad(builder, mask->break_var, "");
 194
 195    lp_exec_mask_update(mask);
 196 }
 197
 198 static void lp_exec_break(struct lp_exec_mask *mask)
 199 {
 200    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 201    LLVMValueRef exec_mask = LLVMBuildNot(builder,
 202                                          mask->exec_mask,
 203                                          "break");
 204
 205    mask->break_mask = LLVMBuildAnd(builder,
 206                                    mask->break_mask,
 207                                    exec_mask, "break_full");
 208
 209    lp_exec_mask_update(mask);
 210 }
 211
 212 static void lp_exec_continue(struct lp_exec_mask *mask)
 213 {
 214    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 215    LLVMValueRef exec_mask = LLVMBuildNot(builder,
 216                                          mask->exec_mask,
 217                                          "");
 218
 219    mask->cont_mask = LLVMBuildAnd(builder,
 220                                   mask->cont_mask,
 221                                   exec_mask, "");
 222
 223    lp_exec_mask_update(mask);
 224 }
 225
 226
 227 static void lp_exec_endloop(struct gallivm_state *gallivm,
 228                             struct lp_exec_mask *mask)
 229 {
 230    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 231    LLVMBasicBlockRef endloop;
 232    LLVMTypeRef int_type = LLVMInt32TypeInContext(mask->bld->gallivm->context);
 233    LLVMTypeRef reg_type = LLVMIntTypeInContext(gallivm->context,
 234                                                mask->bld->type.width *
 235                                                mask->bld->type.length);
 236    LLVMValueRef i1cond, i2cond, icond, limiter;
 237
 238    assert(mask->break_mask);
 239
 240    /*
 241     * Restore the cont_mask, but don't pop
 242     */
 243    assert(mask->loop_stack_size);
 244    mask->cont_mask = mask->loop_stack[mask->loop_stack_size - 1].cont_mask;
 245    lp_exec_mask_update(mask);
 246
 247    /*
 248     * Unlike the continue mask, the break_mask must be preserved across loop
 249     * iterations
 250     */
 251    LLVMBuildStore(builder, mask->break_mask, mask->break_var);
 252
 253    /* Decrement the loop limiter */
 254    limiter = LLVMBuildLoad(builder, mask->loop_limiter, "");
 255
 256    limiter = LLVMBuildSub(
 257       builder,
 258       limiter,
 259       LLVMConstInt(int_type, 1, false),
 260       "");
 261
 262    LLVMBuildStore(builder, limiter, mask->loop_limiter);
 263
 264    /* i1cond = (mask != 0) */
 265    i1cond = LLVMBuildICmp(
 266       builder,
 267       LLVMIntNE,
 268       LLVMBuildBitCast(builder, mask->exec_mask, reg_type, ""),
 269       LLVMConstNull(reg_type), "");
 270
 271    /* i2cond = (looplimiter > 0) */
 272    i2cond = LLVMBuildICmp(
 273       builder,
 274       LLVMIntSGT,
 275       limiter,
 276       LLVMConstNull(int_type), "");
 277
 278    /* if( i1cond && i2cond ) */
 279    icond = LLVMBuildAnd(builder, i1cond, i2cond, "");
 280
 281    endloop = lp_build_insert_new_block(mask->bld->gallivm, "endloop");
 282
 283    LLVMBuildCondBr(builder,
 284                    icond, mask->loop_block, endloop);
 285
 286    LLVMPositionBuilderAtEnd(builder, endloop);
 287
 288    assert(mask->loop_stack_size);
 289    --mask->loop_stack_size;
 290    mask->loop_block = mask->loop_stack[mask->loop_stack_size].loop_block;
 291    mask->cont_mask = mask->loop_stack[mask->loop_stack_size].cont_mask;
 292    mask->break_mask = mask->loop_stack[mask->loop_stack_size].break_mask;
 293    mask->break_var = mask->loop_stack[mask->loop_stack_size].break_var;
 294
 295    lp_exec_mask_update(mask);
 296 }
 297
 298 /* stores val into an address pointed to by dst.
 299  * mask->exec_mask is used to figure out which bits of val
 300  * should be stored into the address
 301  * (0 means don't store this bit, 1 means do store).
 302  */
 303 static void lp_exec_mask_store(struct lp_exec_mask *mask,
 304                                struct lp_build_context *bld_store,
 305                                LLVMValueRef pred,
 306                                LLVMValueRef val,
 307                                LLVMValueRef dst)
 308 {
 309    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 310
 311    /* Mix the predicate and execution mask */
 312    if (mask->has_mask) {
 313       if (pred) {
 314          pred = LLVMBuildAnd(builder, pred, mask->exec_mask, "");
 315       } else {
 316          pred = mask->exec_mask;
 317       }
 318    }
 319
 320    if (pred) {
 321       LLVMValueRef real_val, dst_val;
 322
 323       dst_val = LLVMBuildLoad(builder, dst, "");
 324       real_val = lp_build_select(bld_store,
 325                                  pred,
 326                                  val, dst_val);
 327
 328       LLVMBuildStore(builder, real_val, dst);
 329    } else
 330       LLVMBuildStore(builder, val, dst);
 331 }
 332
 333 static void lp_exec_mask_call(struct lp_exec_mask *mask,
 334                               int func,
 335                               int *pc)
 336 {
 337    assert(mask->call_stack_size < LP_MAX_TGSI_NESTING);
 338    mask->call_stack[mask->call_stack_size].pc = *pc;
 339    mask->call_stack[mask->call_stack_size].ret_mask = mask->ret_mask;
 340    mask->call_stack_size++;
 341    *pc = func;
 342 }
 343
 344 static void lp_exec_mask_ret(struct lp_exec_mask *mask, int *pc)
 345 {
 346    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 347    LLVMValueRef exec_mask;
 348
 349    if (mask->call_stack_size == 0) {
 350       /* returning from main() */
 351       *pc = -1;
 352       return;
 353    }
 354    exec_mask = LLVMBuildNot(builder,
 355                             mask->exec_mask,
 356                             "ret");
 357
 358    mask->ret_mask = LLVMBuildAnd(builder,
 359                                  mask->ret_mask,
 360                                  exec_mask, "ret_full");
 361
 362    lp_exec_mask_update(mask);
 363 }
 364
 365 static void lp_exec_mask_bgnsub(struct lp_exec_mask *mask)
 366 {
 367 }
 368
 369 static void lp_exec_mask_endsub(struct lp_exec_mask *mask, int *pc)
 370 {
 371    assert(mask->call_stack_size);
 372    mask->call_stack_size--;
 373    *pc = mask->call_stack[mask->call_stack_size].pc;
 374    mask->ret_mask = mask->call_stack[mask->call_stack_size].ret_mask;
 375    lp_exec_mask_update(mask);
 376 }
 377
 378
 379 /**
 380  * Return pointer to a temporary register channel (src or dest).
 381  * Note that indirect addressing cannot be handled here.
 382  * \param index  which temporary register
 383  * \param chan  which channel of the temp register.
 384  */
 385 LLVMValueRef
 386 lp_get_temp_ptr_soa(struct lp_build_tgsi_soa_context *bld,
 387              unsigned index,
 388              unsigned chan)
 389 {
 390    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
 391    assert(chan < 4);
 392    if (bld->indirect_files & (1 << TGSI_FILE_TEMPORARY)) {
 393       LLVMValueRef lindex = lp_build_const_int32(bld->bld_base.base.gallivm, index * 4 + chan);
 394       return LLVMBuildGEP(builder, bld->temps_array, &lindex, 1, "");
 395    }
 396    else {
 397       return bld->temps[index][chan];
 398    }
 399 }
 400
 401 /**
 402  * Return pointer to a output register channel (src or dest).
 403  * Note that indirect addressing cannot be handled here.
 404  * \param index  which output register
 405  * \param chan  which channel of the output register.
 406  */
 407 LLVMValueRef
 408 lp_get_output_ptr(struct lp_build_tgsi_soa_context *bld,
 409                unsigned index,
 410                unsigned chan)
 411 {
 412    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
 413    assert(chan < 4);
 414    if (bld->indirect_files & (1 << TGSI_FILE_OUTPUT)) {
 415       LLVMValueRef lindex = lp_build_const_int32(bld->bld_base.base.gallivm,
 416                                                  index * 4 + chan);
 417       return LLVMBuildGEP(builder, bld->outputs_array, &lindex, 1, "");
 418    }
 419    else {
 420       return bld->outputs[index][chan];
 421    }
 422 }
 423
 424 /**
 425  * Gather vector.
 426  * XXX the lp_build_gather() function should be capable of doing this
 427  * with a little work.
 428  */
 429 static LLVMValueRef
 430 build_gather(struct lp_build_context *bld,
 431              LLVMValueRef base_ptr,
 432              LLVMValueRef indexes)
 433 {
 434    LLVMBuilderRef builder = bld->gallivm->builder;
 435    LLVMValueRef res = bld->undef;
 436    unsigned i;
 437
 438    /*
 439     * Loop over elements of index_vec, load scalar value, insert it into 'res'.
 440     */
 441    for (i = 0; i < bld->type.length; i++) {
 442       LLVMValueRef ii = lp_build_const_int32(bld->gallivm, i);
 443       LLVMValueRef index = LLVMBuildExtractElement(builder,
 444                                                    indexes, ii, "");
 445       LLVMValueRef scalar_ptr = LLVMBuildGEP(builder, base_ptr,
 446                                              &index, 1, "gather_ptr");
 447       LLVMValueRef scalar = LLVMBuildLoad(builder, scalar_ptr, "");
 448
 449       res = LLVMBuildInsertElement(builder, res, scalar, ii, "");
 450    }
 451
 452    return res;
 453 }
 454
 455
 456 /**
 457  * Scatter/store vector.
 458  */
 459 static void
 460 emit_mask_scatter(struct lp_build_tgsi_soa_context *bld,
 461                   LLVMValueRef base_ptr,
 462                   LLVMValueRef indexes,
 463                   LLVMValueRef values,
 464                   struct lp_exec_mask *mask,
 465                   LLVMValueRef pred)
 466 {
 467    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
 468    LLVMBuilderRef builder = gallivm->builder;
 469    unsigned i;
 470
 471    /* Mix the predicate and execution mask */
 472    if (mask->has_mask) {
 473       if (pred) {
 474          pred = LLVMBuildAnd(builder, pred, mask->exec_mask, "");
 475       }
 476       else {
 477          pred = mask->exec_mask;
 478       }
 479    }
 480
 481    /*
 482     * Loop over elements of index_vec, store scalar value.
 483     */
 484    for (i = 0; i < bld->bld_base.base.type.length; i++) {
 485       LLVMValueRef ii = lp_build_const_int32(gallivm, i);
 486       LLVMValueRef index = LLVMBuildExtractElement(builder, indexes, ii, "");
 487       LLVMValueRef scalar_ptr = LLVMBuildGEP(builder, base_ptr, &index, 1, "scatter_ptr");
 488       LLVMValueRef val = LLVMBuildExtractElement(builder, values, ii, "scatter_val");
 489       LLVMValueRef scalar_pred = pred ?
 490          LLVMBuildExtractElement(builder, pred, ii, "scatter_pred") : NULL;
 491
 492       if (0)
 493          lp_build_printf(gallivm, "scatter %d: val %f at %d %p\n",
 494                          ii, val, index, scalar_ptr);
 495
 496       if (scalar_pred) {
 497          LLVMValueRef real_val, dst_val;
 498          dst_val = LLVMBuildLoad(builder, scalar_ptr, "");
 499          real_val = lp_build_select(&bld->elem_bld, scalar_pred, val, dst_val);
 500          LLVMBuildStore(builder, real_val, scalar_ptr);
 501       }
 502       else {
 503          LLVMBuildStore(builder, val, scalar_ptr);
 504       }
 505    }
 506 }
 507
 508
 509 /**
 510  * Read the current value of the ADDR register, convert the floats to
 511  * ints, add the base index and return the vector of offsets.
 512  * The offsets will be used to index into the constant buffer or
 513  * temporary register file.
 514  */
 515 static LLVMValueRef
 516 get_indirect_index(struct lp_build_tgsi_soa_context *bld,
 517                    unsigned reg_file, unsigned reg_index,
 518                    const struct tgsi_src_register *indirect_reg)
 519 {
 520    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
 521    struct lp_build_context *uint_bld = &bld->bld_base.uint_bld;
 522    /* always use X component of address register */
 523    unsigned swizzle = indirect_reg->SwizzleX;
 524    LLVMValueRef base;
 525    LLVMValueRef rel;
 526    LLVMValueRef max_index;
 527    LLVMValueRef index;
 528
 529    assert(bld->indirect_files & (1 << reg_file));
 530
 531    base = lp_build_const_int_vec(bld->bld_base.base.gallivm, uint_bld->type, reg_index);
 532
 533    assert(swizzle < 4);
 534    rel = LLVMBuildLoad(builder,
 535                         bld->addr[indirect_reg->Index][swizzle],
 536                         "load addr reg");
 537
 538    index = lp_build_add(uint_bld, base, rel);
 539
 540    max_index = lp_build_const_int_vec(bld->bld_base.base.gallivm,
 541                                       uint_bld->type,
 542                                       bld->bld_base.info->file_max[reg_file]);
 543
 544    assert(!uint_bld->type.sign);
 545    index = lp_build_min(uint_bld, index, max_index);
 546
 547    return index;
 548 }
 549
 550 static struct lp_build_context *
 551 stype_to_fetch(struct lp_build_tgsi_context * bld_base,
 552                enum tgsi_opcode_type stype)
 553 {
 554    struct lp_build_context *bld_fetch;
 555
 556    switch (stype) {
 557    case TGSI_TYPE_FLOAT:
 558    case TGSI_TYPE_UNTYPED:
 559       bld_fetch = &bld_base->base;
 560       break;
 561    case TGSI_TYPE_UNSIGNED:
 562       bld_fetch = &bld_base->uint_bld;
 563       break;
 564    case TGSI_TYPE_SIGNED:
 565       bld_fetch = &bld_base->int_bld;
 566       break;
 567    case TGSI_TYPE_VOID:
 568    case TGSI_TYPE_DOUBLE:
 569    default:
 570       assert(0);
 571       bld_fetch = NULL;
 572       break;
 573    }
 574    return bld_fetch;
 575 }
 576
 577 static LLVMValueRef
 578 emit_fetch_constant(
 579    struct lp_build_tgsi_context * bld_base,
 580    const struct tgsi_full_src_register * reg,
 581    enum tgsi_opcode_type stype,
 582    unsigned swizzle)
 583 {
 584    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
 585    struct gallivm_state *gallivm = bld_base->base.gallivm;
 586    LLVMBuilderRef builder = gallivm->builder;
 587    struct lp_build_context *uint_bld = &bld_base->uint_bld;
 588    LLVMValueRef indirect_index = NULL;
 589    struct lp_build_context *bld_fetch = stype_to_fetch(bld_base, stype);
 590
 591    /* XXX: Handle fetching xyzw components as a vector */
 592    assert(swizzle != ~0);
 593
 594    if (reg->Register.Indirect) {
 595       indirect_index = get_indirect_index(bld,
 596                                           reg->Register.File,
 597                                           reg->Register.Index,
 598                                           &reg->Indirect);
 599    }
 600
 601    if (reg->Register.Indirect) {
 602       LLVMValueRef swizzle_vec =
 603          lp_build_const_int_vec(bld->bld_base.base.gallivm, uint_bld->type, swizzle);
 604       LLVMValueRef index_vec;  /* index into the const buffer */
 605
 606       /* index_vec = indirect_index * 4 + swizzle */
 607       index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
 608       index_vec = lp_build_add(uint_bld, index_vec, swizzle_vec);
 609
 610       /* Gather values from the constant buffer */
 611       return build_gather(bld_fetch, bld->consts_ptr, index_vec);
 612    }
 613    else {
 614       LLVMValueRef index;  /* index into the const buffer */
 615       LLVMValueRef scalar, scalar_ptr;
 616
 617       index = lp_build_const_int32(gallivm, reg->Register.Index*4 + swizzle);
 618
 619       scalar_ptr = LLVMBuildGEP(builder, bld->consts_ptr,
 620                                    &index, 1, "");
 621
 622       if (stype != TGSI_TYPE_FLOAT && stype != TGSI_TYPE_UNTYPED) {
 623          LLVMTypeRef ivtype = LLVMPointerType(LLVMInt32TypeInContext(gallivm->context), 0);
 624          LLVMValueRef temp_ptr;
 625          temp_ptr = LLVMBuildBitCast(builder, scalar_ptr, ivtype, "");
 626          scalar = LLVMBuildLoad(builder, temp_ptr, "");
 627       } else
 628          scalar = LLVMBuildLoad(builder, scalar_ptr, "");
 629
 630       return lp_build_broadcast_scalar(bld_fetch, scalar);
 631    }
 632 }
 633
 634 static LLVMValueRef
 635 emit_fetch_immediate(
 636    struct lp_build_tgsi_context * bld_base,
 637    const struct tgsi_full_src_register * reg,
 638    enum tgsi_opcode_type stype,
 639    unsigned swizzle)
 640 {
 641    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
 642    LLVMValueRef res = bld->immediates[reg->Register.Index][swizzle];
 643    assert(res);
 644
 645    if (stype == TGSI_TYPE_UNSIGNED) {
 646       res = LLVMConstBitCast(res, bld_base->uint_bld.vec_type);
 647    } else if (stype == TGSI_TYPE_SIGNED) {
 648       res = LLVMConstBitCast(res, bld_base->int_bld.vec_type);
 649    }
 650    return res;
 651 }
 652
 653 static LLVMValueRef
 654 emit_fetch_input(
 655    struct lp_build_tgsi_context * bld_base,
 656    const struct tgsi_full_src_register * reg,
 657    enum tgsi_opcode_type stype,
 658    unsigned swizzle)
 659 {
 660    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
 661    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
 662    LLVMBuilderRef builder = gallivm->builder;
 663    struct lp_build_context *uint_bld = &bld_base->uint_bld;
 664    LLVMValueRef indirect_index = NULL;
 665    LLVMValueRef res;
 666
 667    if (reg->Register.Indirect) {
 668       indirect_index = get_indirect_index(bld,
 669                                           reg->Register.File,
 670                                           reg->Register.Index,
 671                                           &reg->Indirect);
 672    }
 673
 674    if (reg->Register.Indirect) {
 675       LLVMValueRef swizzle_vec =
 676          lp_build_const_int_vec(gallivm, uint_bld->type, swizzle);
 677       LLVMValueRef length_vec =
 678          lp_build_const_int_vec(gallivm, uint_bld->type, bld->bld_base.base.type.length);
 679       LLVMValueRef index_vec;  /* index into the const buffer */
 680       LLVMValueRef inputs_array;
 681       LLVMTypeRef float4_ptr_type;
 682
 683       /* index_vec = (indirect_index * 4 + swizzle) * length */
 684       index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
 685       index_vec = lp_build_add(uint_bld, index_vec, swizzle_vec);
 686       index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
 687
 688       /* cast inputs_array pointer to float* */
 689       float4_ptr_type = LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
 690       inputs_array = LLVMBuildBitCast(builder, bld->inputs_array,
 691                                          float4_ptr_type, "");
 692
 693       /* Gather values from the temporary register array */
 694       res = build_gather(&bld_base->base, inputs_array, index_vec);
 695    } else {
 696       if (bld->indirect_files & (1 << TGSI_FILE_INPUT)) {
 697          LLVMValueRef lindex = lp_build_const_int32(gallivm,
 698                                         reg->Register.Index * 4 + swizzle);
 699          LLVMValueRef input_ptr =  LLVMBuildGEP(builder,
 700                                                 bld->inputs_array, &lindex, 1, "");
 701          res = LLVMBuildLoad(builder, input_ptr, "");
 702       }
 703       else {
 704          res = bld->inputs[reg->Register.Index][swizzle];
 705       }
 706    }
 707    assert(res);
 708    return res;
 709 }
 710
 711 static LLVMValueRef
 712 emit_fetch_temporary(
 713    struct lp_build_tgsi_context * bld_base,
 714    const struct tgsi_full_src_register * reg,
 715    enum tgsi_opcode_type stype,
 716    unsigned swizzle)
 717 {
 718    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
 719    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
 720    LLVMBuilderRef builder = gallivm->builder;
 721    struct lp_build_context *uint_bld = &bld_base->uint_bld;
 722    LLVMValueRef indirect_index = NULL;
 723    LLVMValueRef res;
 724
 725    if (reg->Register.Indirect) {
 726       indirect_index = get_indirect_index(bld,
 727                                           reg->Register.File,
 728                                           reg->Register.Index,
 729                                           &reg->Indirect);
 730    }
 731
 732    if (reg->Register.Indirect) {
 733       LLVMValueRef swizzle_vec =
 734          lp_build_const_int_vec(bld->bld_base.base.gallivm, uint_bld->type, swizzle);
 735       LLVMValueRef length_vec =
 736          lp_build_const_int_vec(bld->bld_base.base.gallivm, uint_bld->type,
 737                                 bld->bld_base.base.type.length);
 738       LLVMValueRef index_vec;  /* index into the const buffer */
 739       LLVMValueRef temps_array;
 740       LLVMTypeRef float4_ptr_type;
 741
 742       /* index_vec = (indirect_index * 4 + swizzle) * length */
 743       index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
 744       index_vec = lp_build_add(uint_bld, index_vec, swizzle_vec);
 745       index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
 746
 747       /* cast temps_array pointer to float* */
 748       float4_ptr_type = LLVMPointerType(LLVMFloatTypeInContext(bld->bld_base.base.gallivm->context), 0);
 749       temps_array = LLVMBuildBitCast(builder, bld->temps_array,
 750                                      float4_ptr_type, "");
 751
 752       /* Gather values from the temporary register array */
 753       res = build_gather(&bld_base->base, temps_array, index_vec);
 754    }
 755    else {
 756       LLVMValueRef temp_ptr;
 757       if (stype != TGSI_TYPE_FLOAT && stype != TGSI_TYPE_UNTYPED) {
 758          LLVMTypeRef itype = LLVMPointerType(LLVMVectorType(LLVMInt32TypeInContext(gallivm->context), 4), 0);
 759          LLVMValueRef tint_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index,
 760                                                      swizzle);
 761          temp_ptr = LLVMBuildBitCast(builder, tint_ptr, itype, "");
 762       } else
 763          temp_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index, swizzle);
 764       res = LLVMBuildLoad(builder, temp_ptr, "");
 765       if (!res)
 766          return bld->bld_base.base.undef;
 767    }
 768
 769    return res;
 770 }
 771
 772 static LLVMValueRef
 773 emit_fetch_system_value(
 774    struct lp_build_tgsi_context * bld_base,
 775    const struct tgsi_full_src_register * reg,
 776    enum tgsi_opcode_type stype,
 777    unsigned swizzle)
 778 {
 779    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
 780    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
 781    LLVMBuilderRef builder = gallivm->builder;
 782    LLVMValueRef index;  /* index into the system value array */
 783    LLVMValueRef scalar, scalar_ptr;
 784
 785    assert(!reg->Register.Indirect);
 786
 787    index = lp_build_const_int32(gallivm, reg->Register.Index * 4 + swizzle);
 788
 789    scalar_ptr = LLVMBuildGEP(builder, bld->system_values_array, &index, 1, "");
 790    scalar = LLVMBuildLoad(builder, scalar_ptr, "");
 791
 792    return lp_build_broadcast_scalar(&bld->bld_base.base, scalar);
 793 }
 794
 795 /**
 796  * Register fetch with derivatives.
 797  */
 798 static void
 799 emit_fetch_deriv(
 800    struct lp_build_tgsi_soa_context *bld,
 801    LLVMValueRef src,
 802    LLVMValueRef *res,
 803    LLVMValueRef *ddx,
 804    LLVMValueRef *ddy)
 805 {
 806    if(res)
 807       *res = src;
 808
 809    /* TODO: use interpolation coeffs for inputs */
 810
 811    if(ddx)
 812       *ddx = lp_build_ddx(&bld->bld_base.base, src);
 813
 814    if(ddy)
 815       *ddy = lp_build_ddy(&bld->bld_base.base, src);
 816 }
 817
 818
 819 /**
 820  * Predicate.
 821  */
 822 static void
 823 emit_fetch_predicate(
 824    struct lp_build_tgsi_soa_context *bld,
 825    const struct tgsi_full_instruction *inst,
 826    LLVMValueRef *pred)
 827 {
 828    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
 829    unsigned index;
 830    unsigned char swizzles[4];
 831    LLVMValueRef unswizzled[4] = {NULL, NULL, NULL, NULL};
 832    LLVMValueRef value;
 833    unsigned chan;
 834
 835    if (!inst->Instruction.Predicate) {
 836       TGSI_FOR_EACH_CHANNEL( chan ) {
 837          pred[chan] = NULL;
 838       }
 839       return;
 840    }
 841
 842    swizzles[0] = inst->Predicate.SwizzleX;
 843    swizzles[1] = inst->Predicate.SwizzleY;
 844    swizzles[2] = inst->Predicate.SwizzleZ;
 845    swizzles[3] = inst->Predicate.SwizzleW;
 846
 847    index = inst->Predicate.Index;
 848    assert(index < LP_MAX_TGSI_PREDS);
 849
 850    TGSI_FOR_EACH_CHANNEL( chan ) {
 851       unsigned swizzle = swizzles[chan];
 852
 853       /*
 854        * Only fetch the predicate register channels that are actually listed
 855        * in the swizzles
 856        */
 857       if (!unswizzled[swizzle]) {
 858          value = LLVMBuildLoad(builder,
 859                                bld->preds[index][swizzle], "");
 860
 861          /*
 862           * Convert the value to an integer mask.
 863           *
 864           * TODO: Short-circuit this comparison -- a D3D setp_xx instructions
 865           * is needlessly causing two comparisons due to storing the intermediate
 866           * result as float vector instead of an integer mask vector.
 867           */
 868          value = lp_build_compare(bld->bld_base.base.gallivm,
 869                                   bld->bld_base.base.type,
 870                                   PIPE_FUNC_NOTEQUAL,
 871                                   value,
 872                                   bld->bld_base.base.zero);
 873          if (inst->Predicate.Negate) {
 874             value = LLVMBuildNot(builder, value, "");
 875          }
 876
 877          unswizzled[swizzle] = value;
 878       } else {
 879          value = unswizzled[swizzle];
 880       }
 881
 882       pred[chan] = value;
 883    }
 884 }
 885
 886 /**
 887  * Register store.
 888  */
 889 static void
 890 emit_store_chan(
 891    struct lp_build_tgsi_context *bld_base,
 892    const struct tgsi_full_instruction *inst,
 893    unsigned index,
 894    unsigned chan_index,
 895    LLVMValueRef pred,
 896    LLVMValueRef value)
 897 {
 898    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
 899    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
 900    LLVMBuilderRef builder = gallivm->builder;
 901    const struct tgsi_full_dst_register *reg = &inst->Dst[index];
 902    struct lp_build_context *uint_bld = &bld_base->uint_bld;
 903    LLVMValueRef indirect_index = NULL;
 904    struct lp_build_context *bld_store;
 905    enum tgsi_opcode_type dtype = tgsi_opcode_infer_dst_type(inst->Instruction.Opcode);
 906
 907    switch (dtype) {
 908    default:
 909    case TGSI_TYPE_FLOAT:
 910    case TGSI_TYPE_UNTYPED:
 911       bld_store = &bld_base->base;
 912       break;
 913    case TGSI_TYPE_UNSIGNED:
 914       bld_store = &bld_base->uint_bld;
 915       break;
 916    case TGSI_TYPE_SIGNED:
 917       bld_store = &bld_base->int_bld;
 918       break;
 919    case TGSI_TYPE_DOUBLE:
 920    case TGSI_TYPE_VOID:
 921       assert(0);
 922       bld_store = NULL;
 923       break;
 924    }
 925
 926    switch( inst->Instruction.Saturate ) {
 927    case TGSI_SAT_NONE:
 928       break;
 929
 930    case TGSI_SAT_ZERO_ONE:
 931       value = lp_build_max(&bld->bld_base.base, value, bld->bld_base.base.zero);
 932       value = lp_build_min(&bld->bld_base.base, value, bld->bld_base.base.one);
 933       break;
 934
 935    case TGSI_SAT_MINUS_PLUS_ONE:
 936       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));
 937       value = lp_build_min(&bld->bld_base.base, value, bld->bld_base.base.one);
 938       break;
 939
 940    default:
 941       assert(0);
 942    }
 943
 944    if (reg->Register.Indirect) {
 945       indirect_index = get_indirect_index(bld,
 946                                           reg->Register.File,
 947                                           reg->Register.Index,
 948                                           &reg->Indirect);
 949    } else {
 950       assert(reg->Register.Index <=
 951                              bld->bld_base.info->file_max[reg->Register.File]);
 952    }
 953
 954    switch( reg->Register.File ) {
 955    case TGSI_FILE_OUTPUT:
 956       if (reg->Register.Indirect) {
 957          LLVMValueRef chan_vec =
 958             lp_build_const_int_vec(gallivm, uint_bld->type, chan_index);
 959          LLVMValueRef length_vec =
 960             lp_build_const_int_vec(gallivm, uint_bld->type, bld->bld_base.base.type.length);
 961          LLVMValueRef index_vec;  /* indexes into the temp registers */
 962          LLVMValueRef outputs_array;
 963          LLVMValueRef pixel_offsets;
 964          LLVMTypeRef float_ptr_type;
 965          int i;
 966
 967          /* build pixel offset vector: {0, 1, 2, 3, ...} */
 968          pixel_offsets = uint_bld->undef;
 969          for (i = 0; i < bld->bld_base.base.type.length; i++) {
 970             LLVMValueRef ii = lp_build_const_int32(gallivm, i);
 971             pixel_offsets = LLVMBuildInsertElement(builder, pixel_offsets,
 972                                                    ii, ii, "");
 973          }
 974
 975          /* index_vec = (indirect_index * 4 + chan_index) * length + offsets */
 976          index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
 977          index_vec = lp_build_add(uint_bld, index_vec, chan_vec);
 978          index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
 979          index_vec = lp_build_add(uint_bld, index_vec, pixel_offsets);
 980
 981          float_ptr_type =
 982             LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
 983          outputs_array = LLVMBuildBitCast(builder, bld->outputs_array,
 984                                           float_ptr_type, "");
 985
 986          /* Scatter store values into temp registers */
 987          emit_mask_scatter(bld, outputs_array, index_vec, value,
 988                            &bld->exec_mask, pred);
 989       }
 990       else {
 991          LLVMValueRef out_ptr = lp_get_output_ptr(bld, reg->Register.Index,
 992                                                chan_index);
 993          lp_exec_mask_store(&bld->exec_mask, bld_store, pred, value, out_ptr);
 994       }
 995       break;
 996
 997    case TGSI_FILE_TEMPORARY:
 998       if (reg->Register.Indirect) {
 999          LLVMValueRef chan_vec =
1000             lp_build_const_int_vec(gallivm, uint_bld->type, chan_index);
1001          LLVMValueRef length_vec =
1002             lp_build_const_int_vec(gallivm, uint_bld->type,
1003                                    bld->bld_base.base.type.length);
1004          LLVMValueRef index_vec;  /* indexes into the temp registers */
1005          LLVMValueRef temps_array;
1006          LLVMValueRef pixel_offsets;
1007          LLVMTypeRef float_ptr_type;
1008          int i;
1009
1010          /* build pixel offset vector: {0, 1, 2, 3, ...} */
1011          pixel_offsets = uint_bld->undef;
1012          for (i = 0; i < bld->bld_base.base.type.length; i++) {
1013             LLVMValueRef ii = lp_build_const_int32(gallivm, i);
1014             pixel_offsets = LLVMBuildInsertElement(builder, pixel_offsets,
1015                                                    ii, ii, "");
1016          }
1017
1018          /* index_vec = (indirect_index * 4 + chan_index) * length + offsets */
1019          index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
1020          index_vec = lp_build_add(uint_bld, index_vec, chan_vec);
1021          index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
1022          index_vec = lp_build_add(uint_bld, index_vec, pixel_offsets);
1023
1024          float_ptr_type =
1025             LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
1026          temps_array = LLVMBuildBitCast(builder, bld->temps_array,
1027                                         float_ptr_type, "");
1028
1029          /* Scatter store values into temp registers */
1030          emit_mask_scatter(bld, temps_array, index_vec, value,
1031                            &bld->exec_mask, pred);
1032       }
1033       else {
1034          LLVMValueRef temp_ptr;
1035
1036          switch (dtype) {
1037          case TGSI_TYPE_UNSIGNED:
1038          case TGSI_TYPE_SIGNED: {
1039             LLVMTypeRef itype = LLVMVectorType(LLVMInt32TypeInContext(gallivm->context), 4);
1040             LLVMTypeRef ivtype = LLVMPointerType(itype, 0);
1041             LLVMValueRef tint_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index,
1042                                                         chan_index);
1043             LLVMValueRef temp_value_ptr;
1044
1045             temp_ptr = LLVMBuildBitCast(builder, tint_ptr, ivtype, "");
1046             temp_value_ptr = LLVMBuildBitCast(builder, value, itype, "");
1047             value = temp_value_ptr;
1048             break;
1049          }
1050          default:
1051          case TGSI_TYPE_FLOAT:
1052          case TGSI_TYPE_UNTYPED:
1053             temp_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index,
1054                                            chan_index);
1055             break;
1056          }
1057
1058          lp_exec_mask_store(&bld->exec_mask, bld_store, pred, value, temp_ptr);
1059       }
1060       break;
1061
1062    case TGSI_FILE_ADDRESS:
1063       assert(dtype == TGSI_TYPE_SIGNED);
1064       assert(LLVMTypeOf(value) == bld_base->base.int_vec_type);
1065       lp_exec_mask_store(&bld->exec_mask, bld_store, pred, value,
1066                          bld->addr[reg->Register.Index][chan_index]);
1067       break;
1068
1069    case TGSI_FILE_PREDICATE:
1070       lp_exec_mask_store(&bld->exec_mask, bld_store, pred, value,
1071                          bld->preds[reg->Register.Index][chan_index]);
1072       break;
1073
1074    default:
1075       assert( 0 );
1076    }
1077 }
1078
1079 static void
1080 emit_store(
1081    struct lp_build_tgsi_context * bld_base,
1082    const struct tgsi_full_instruction * inst,
1083    const struct tgsi_opcode_info * info,
1084    LLVMValueRef dst[4])
1085
1086 {
1087    unsigned chan_index;
1088    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1089
1090    if(info->num_dst) {
1091       LLVMValueRef pred[TGSI_NUM_CHANNELS];
1092
1093       emit_fetch_predicate( bld, inst, pred );
1094
1095       TGSI_FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1096          emit_store_chan(bld_base, inst, 0, chan_index, pred[chan_index], dst[chan_index]);
1097       }
1098    }
1099 }
1100
1101 /**
1102  * High-level instruction translators.
1103  */
1104
1105 static void
1106 emit_tex( struct lp_build_tgsi_soa_context *bld,
1107           const struct tgsi_full_instruction *inst,
1108           enum lp_build_tex_modifier modifier,
1109           LLVMValueRef *texel)
1110 {
1111    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
1112    unsigned unit;
1113    LLVMValueRef lod_bias, explicit_lod;
1114    LLVMValueRef oow = NULL;
1115    LLVMValueRef coords[3];
1116    LLVMValueRef ddx[3];
1117    LLVMValueRef ddy[3];
1118    unsigned num_coords;
1119    unsigned i;
1120
1121    if (!bld->sampler) {
1122       _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
1123       for (i = 0; i < 4; i++) {
1124          texel[i] = bld->bld_base.base.undef;
1125       }
1126       return;
1127    }
1128
1129    switch (inst->Texture.Texture) {
1130    case TGSI_TEXTURE_1D:
1131       num_coords = 1;
1132       break;
1133    case TGSI_TEXTURE_1D_ARRAY:
1134    case TGSI_TEXTURE_2D:
1135    case TGSI_TEXTURE_RECT:
1136       num_coords = 2;
1137       break;
1138    case TGSI_TEXTURE_SHADOW1D:
1139    case TGSI_TEXTURE_SHADOW1D_ARRAY:
1140    case TGSI_TEXTURE_SHADOW2D:
1141    case TGSI_TEXTURE_SHADOWRECT:
1142    case TGSI_TEXTURE_2D_ARRAY:
1143    case TGSI_TEXTURE_3D:
1144    case TGSI_TEXTURE_CUBE:
1145       num_coords = 3;
1146       break;
1147    case TGSI_TEXTURE_SHADOW2D_ARRAY:
1148       num_coords = 4;
1149       break;
1150    default:
1151       assert(0);
1152       return;
1153    }
1154
1155    if (modifier == LP_BLD_TEX_MODIFIER_LOD_BIAS) {
1156       lod_bias = lp_build_emit_fetch( &bld->bld_base, inst, 0, 3 );
1157       explicit_lod = NULL;
1158    }
1159    else if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_LOD) {
1160       lod_bias = NULL;
1161       explicit_lod = lp_build_emit_fetch( &bld->bld_base, inst, 0, 3 );
1162    }
1163    else {
1164       lod_bias = NULL;
1165       explicit_lod = NULL;
1166    }
1167
1168    if (modifier == LP_BLD_TEX_MODIFIER_PROJECTED) {
1169       oow = lp_build_emit_fetch( &bld->bld_base, inst, 0, 3 );
1170       oow = lp_build_rcp(&bld->bld_base.base, oow);
1171    }
1172
1173    for (i = 0; i < num_coords; i++) {
1174       coords[i] = lp_build_emit_fetch( &bld->bld_base, inst, 0, i );
1175       if (modifier == LP_BLD_TEX_MODIFIER_PROJECTED)
1176          coords[i] = lp_build_mul(&bld->bld_base.base, coords[i], oow);
1177    }
1178    for (i = num_coords; i < 3; i++) {
1179       coords[i] = bld->bld_base.base.undef;
1180    }
1181
1182    if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV) {
1183       LLVMValueRef index0 = lp_build_const_int32(bld->bld_base.base.gallivm, 0);
1184       for (i = 0; i < num_coords; i++) {
1185          LLVMValueRef src1 = lp_build_emit_fetch( &bld->bld_base, inst, 1, i );
1186          LLVMValueRef src2 = lp_build_emit_fetch( &bld->bld_base, inst, 2, i );
1187          ddx[i] = LLVMBuildExtractElement(builder, src1, index0, "");
1188          ddy[i] = LLVMBuildExtractElement(builder, src2, index0, "");
1189       }
1190       unit = inst->Src[3].Register.Index;
1191    }  else {
1192       for (i = 0; i < num_coords; i++) {
1193          ddx[i] = lp_build_scalar_ddx( &bld->bld_base.base, coords[i] );
1194          ddy[i] = lp_build_scalar_ddy( &bld->bld_base.base, coords[i] );
1195       }
1196       unit = inst->Src[1].Register.Index;
1197    }
1198    for (i = num_coords; i < 3; i++) {
1199       ddx[i] = LLVMGetUndef(bld->bld_base.base.elem_type);
1200       ddy[i] = LLVMGetUndef(bld->bld_base.base.elem_type);
1201    }
1202
1203    bld->sampler->emit_fetch_texel(bld->sampler,
1204                                   bld->bld_base.base.gallivm,
1205                                   bld->bld_base.base.type,
1206                                   unit, num_coords, coords,
1207                                   ddx, ddy,
1208                                   lod_bias, explicit_lod,
1209                                   texel);
1210 }
1211
1212 static boolean
1213 near_end_of_shader(struct lp_build_tgsi_soa_context *bld,
1214                    int pc)
1215 {
1216    int i;
1217
1218    for (i = 0; i < 5; i++) {
1219       unsigned opcode;
1220
1221       if (pc + i >= bld->bld_base.info->num_instructions)
1222          return TRUE;
1223
1224       opcode = bld->bld_base.instructions[pc + i].Instruction.Opcode;
1225
1226       if (opcode == TGSI_OPCODE_END)
1227          return TRUE;
1228
1229       if (opcode == TGSI_OPCODE_TEX ||
1230           opcode == TGSI_OPCODE_TXP ||
1231           opcode == TGSI_OPCODE_TXD ||
1232           opcode == TGSI_OPCODE_TXB ||
1233           opcode == TGSI_OPCODE_TXL ||
1234           opcode == TGSI_OPCODE_TXF ||
1235           opcode == TGSI_OPCODE_TXQ ||
1236           opcode == TGSI_OPCODE_CAL ||
1237           opcode == TGSI_OPCODE_CALLNZ ||
1238           opcode == TGSI_OPCODE_IF ||
1239           opcode == TGSI_OPCODE_IFC ||
1240           opcode == TGSI_OPCODE_BGNLOOP ||
1241           opcode == TGSI_OPCODE_SWITCH)
1242          return FALSE;
1243    }
1244
1245    return TRUE;
1246 }
1247
1248
1249
1250 /**
1251  * Kill fragment if any of the src register values are negative.
1252  */
1253 static void
1254 emit_kil(
1255    struct lp_build_tgsi_soa_context *bld,
1256    const struct tgsi_full_instruction *inst,
1257    int pc)
1258 {
1259    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
1260    const struct tgsi_full_src_register *reg = &inst->Src[0];
1261    LLVMValueRef terms[TGSI_NUM_CHANNELS];
1262    LLVMValueRef mask;
1263    unsigned chan_index;
1264
1265    memset(&terms, 0, sizeof terms);
1266
1267    TGSI_FOR_EACH_CHANNEL( chan_index ) {
1268       unsigned swizzle;
1269
1270       /* Unswizzle channel */
1271       swizzle = tgsi_util_get_full_src_register_swizzle( reg, chan_index );
1272
1273       /* Check if the component has not been already tested. */
1274       assert(swizzle < TGSI_NUM_CHANNELS);
1275       if( !terms[swizzle] )
1276          /* TODO: change the comparison operator instead of setting the sign */
1277          terms[swizzle] =  lp_build_emit_fetch(&bld->bld_base, inst, 0, chan_index );
1278    }
1279
1280    mask = NULL;
1281    TGSI_FOR_EACH_CHANNEL( chan_index ) {
1282       if(terms[chan_index]) {
1283          LLVMValueRef chan_mask;
1284
1285          /*
1286           * If term < 0 then mask = 0 else mask = ~0.
1287           */
1288          chan_mask = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_GEQUAL, terms[chan_index], bld->bld_base.base.zero);
1289
1290          if(mask)
1291             mask = LLVMBuildAnd(builder, mask, chan_mask, "");
1292          else
1293             mask = chan_mask;
1294       }
1295    }
1296
1297    if(mask) {
1298       lp_build_mask_update(bld->mask, mask);
1299
1300       if (!near_end_of_shader(bld, pc))
1301          lp_build_mask_check(bld->mask);
1302    }
1303 }
1304
1305
1306 /**
1307  * Predicated fragment kill.
1308  * XXX Actually, we do an unconditional kill (as in tgsi_exec.c).
1309  * The only predication is the execution mask which will apply if
1310  * we're inside a loop or conditional.
1311  */
1312 static void
1313 emit_kilp(struct lp_build_tgsi_soa_context *bld,
1314           int pc)
1315 {
1316    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
1317    LLVMValueRef mask;
1318
1319    /* For those channels which are "alive", disable fragment shader
1320     * execution.
1321     */
1322    if (bld->exec_mask.has_mask) {
1323       mask = LLVMBuildNot(builder, bld->exec_mask.exec_mask, "kilp");
1324    }
1325    else {
1326       LLVMValueRef zero = LLVMConstNull(bld->bld_base.base.int_vec_type);
1327       mask = zero;
1328    }
1329
1330    lp_build_mask_update(bld->mask, mask);
1331
1332    if (!near_end_of_shader(bld, pc))
1333       lp_build_mask_check(bld->mask);
1334 }
1335
1336
1337 /**
1338  * Emit code which will dump the value of all the temporary registers
1339  * to stdout.
1340  */
1341 static void
1342 emit_dump_temps(struct lp_build_tgsi_soa_context *bld)
1343 {
1344    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
1345    LLVMBuilderRef builder = gallivm->builder;
1346    LLVMValueRef temp_ptr;
1347    LLVMValueRef i0 = lp_build_const_int32(gallivm, 0);
1348    LLVMValueRef i1 = lp_build_const_int32(gallivm, 1);
1349    LLVMValueRef i2 = lp_build_const_int32(gallivm, 2);
1350    LLVMValueRef i3 = lp_build_const_int32(gallivm, 3);
1351    int index;
1352    int n = bld->bld_base.info->file_max[TGSI_FILE_TEMPORARY];
1353
1354    for (index = 0; index < n; index++) {
1355       LLVMValueRef idx = lp_build_const_int32(gallivm, index);
1356       LLVMValueRef v[4][4], res;
1357       int chan;
1358
1359       lp_build_printf(gallivm, "TEMP[%d]:\n", idx);
1360
1361       for (chan = 0; chan < 4; chan++) {
1362          temp_ptr = lp_get_temp_ptr_soa(bld, index, chan);
1363          res = LLVMBuildLoad(builder, temp_ptr, "");
1364          v[chan][0] = LLVMBuildExtractElement(builder, res, i0, "");
1365          v[chan][1] = LLVMBuildExtractElement(builder, res, i1, "");
1366          v[chan][2] = LLVMBuildExtractElement(builder, res, i2, "");
1367          v[chan][3] = LLVMBuildExtractElement(builder, res, i3, "");
1368       }
1369
1370       lp_build_printf(gallivm, "  X: %f %f %f %f\n",
1371                       v[0][0], v[0][1], v[0][2], v[0][3]);
1372       lp_build_printf(gallivm, "  Y: %f %f %f %f\n",
1373                       v[1][0], v[1][1], v[1][2], v[1][3]);
1374       lp_build_printf(gallivm, "  Z: %f %f %f %f\n",
1375                       v[2][0], v[2][1], v[2][2], v[2][3]);
1376       lp_build_printf(gallivm, "  W: %f %f %f %f\n",
1377                       v[3][0], v[3][1], v[3][2], v[3][3]);
1378    }
1379 }
1380
1381
1382
1383 void
1384 lp_emit_declaration_soa(
1385    struct lp_build_tgsi_context *bld_base,
1386    const struct tgsi_full_declaration *decl)
1387 {
1388    struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
1389    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
1390    LLVMTypeRef vec_type = bld->bld_base.base.vec_type;
1391    const unsigned first = decl->Range.First;
1392    const unsigned last = decl->Range.Last;
1393    unsigned idx, i;
1394
1395    for (idx = first; idx <= last; ++idx) {
1396       assert(last <= bld->bld_base.info->file_max[decl->Declaration.File]);
1397       switch (decl->Declaration.File) {
1398       case TGSI_FILE_TEMPORARY:
1399          assert(idx < LP_MAX_TGSI_TEMPS);
1400          if (!(bld->indirect_files & (1 << TGSI_FILE_TEMPORARY))) {
1401             for (i = 0; i < TGSI_NUM_CHANNELS; i++)
1402                bld->temps[idx][i] = lp_build_alloca(gallivm, vec_type, "temp");
1403          }
1404          break;
1405
1406       case TGSI_FILE_OUTPUT:
1407          if (!(bld->indirect_files & (1 << TGSI_FILE_OUTPUT))) {
1408             for (i = 0; i < TGSI_NUM_CHANNELS; i++)
1409                bld->outputs[idx][i] = lp_build_alloca(gallivm,
1410                                                       vec_type, "output");
1411          }
1412          break;
1413
1414       case TGSI_FILE_ADDRESS:
1415          /* ADDR registers are the only allocated with an integer LLVM IR type,
1416           * as they are guaranteed to always have integers.
1417           * XXX: Not sure if this exception is worthwhile (or the whole idea of
1418           * an ADDR register for that matter).
1419           */
1420          assert(idx < LP_MAX_TGSI_ADDRS);
1421          for (i = 0; i < TGSI_NUM_CHANNELS; i++)
1422             bld->addr[idx][i] = lp_build_alloca(gallivm, bld_base->base.int_vec_type, "addr");
1423          break;
1424
1425       case TGSI_FILE_PREDICATE:
1426          assert(idx < LP_MAX_TGSI_PREDS);
1427          for (i = 0; i < TGSI_NUM_CHANNELS; i++)
1428             bld->preds[idx][i] = lp_build_alloca(gallivm, vec_type,
1429                                                  "predicate");
1430          break;
1431
1432       default:
1433          /* don't need to declare other vars */
1434          break;
1435       }
1436    }
1437 }
1438
1439
1440 void lp_emit_immediate_soa(
1441    struct lp_build_tgsi_context *bld_base,
1442    const struct tgsi_full_immediate *imm)
1443 {
1444    struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
1445    struct gallivm_state * gallivm = bld_base->base.gallivm;
1446
1447    /* simply copy the immediate values into the next immediates[] slot */
1448    unsigned i;
1449    const uint size = imm->Immediate.NrTokens - 1;
1450    assert(size <= 4);
1451    assert(bld->num_immediates < LP_MAX_TGSI_IMMEDIATES);
1452    switch (imm->Immediate.DataType) {
1453    case TGSI_IMM_FLOAT32:
1454       for( i = 0; i < size; ++i )
1455          bld->immediates[bld->num_immediates][i] =
1456             lp_build_const_vec(gallivm, bld_base->base.type, imm->u[i].Float);
1457
1458       break;
1459    case TGSI_IMM_UINT32:
1460       for( i = 0; i < size; ++i ) {
1461          LLVMValueRef tmp = lp_build_const_vec(gallivm, bld_base->uint_bld.type, imm->u[i].Uint);
1462          bld->immediates[bld->num_immediates][i] =
1463             LLVMConstBitCast(tmp, bld_base->base.vec_type);
1464       }
1465
1466       break;
1467    case TGSI_IMM_INT32:
1468       for( i = 0; i < size; ++i ) {
1469          LLVMValueRef tmp = lp_build_const_vec(gallivm, bld_base->int_bld.type, imm->u[i].Int);
1470          bld->immediates[bld->num_immediates][i] =
1471             LLVMConstBitCast(tmp, bld_base->base.vec_type);
1472       }
1473
1474       break;
1475    }
1476    for( i = size; i < 4; ++i )
1477       bld->immediates[bld->num_immediates][i] = bld_base->base.undef;
1478
1479    bld->num_immediates++;
1480 }
1481
1482 static void
1483 ddx_emit(
1484    const struct lp_build_tgsi_action * action,
1485    struct lp_build_tgsi_context * bld_base,
1486    struct lp_build_emit_data * emit_data)
1487 {
1488    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1489
1490    emit_fetch_deriv(bld, emit_data->args[0], NULL,
1491                     &emit_data->output[emit_data->chan], NULL);
1492 }
1493
1494 static void
1495 ddy_emit(
1496    const struct lp_build_tgsi_action * action,
1497    struct lp_build_tgsi_context * bld_base,
1498    struct lp_build_emit_data * emit_data)
1499 {
1500    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1501
1502    emit_fetch_deriv(bld, emit_data->args[0], NULL, NULL,
1503                     &emit_data->output[emit_data->chan]);
1504 }
1505
1506 static void
1507 kilp_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_kilp(bld, bld_base->pc - 1);
1515 }
1516
1517 static void
1518 kil_emit(
1519    const struct lp_build_tgsi_action * action,
1520    struct lp_build_tgsi_context * bld_base,
1521    struct lp_build_emit_data * emit_data)
1522 {
1523    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1524
1525    emit_kil(bld, emit_data->inst, bld_base->pc - 1);
1526 }
1527
1528 static void
1529 tex_emit(
1530    const struct lp_build_tgsi_action * action,
1531    struct lp_build_tgsi_context * bld_base,
1532    struct lp_build_emit_data * emit_data)
1533 {
1534    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1535
1536    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_NONE, emit_data->output);
1537 }
1538
1539 static void
1540 txb_emit(
1541    const struct lp_build_tgsi_action * action,
1542    struct lp_build_tgsi_context * bld_base,
1543    struct lp_build_emit_data * emit_data)
1544 {
1545    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1546
1547    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_LOD_BIAS,
1548             emit_data->output);
1549 }
1550
1551 static void
1552 txd_emit(
1553    const struct lp_build_tgsi_action * action,
1554    struct lp_build_tgsi_context * bld_base,
1555    struct lp_build_emit_data * emit_data)
1556 {
1557    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1558
1559    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV,
1560             emit_data->output);
1561 }
1562
1563 static void
1564 txl_emit(
1565    const struct lp_build_tgsi_action * action,
1566    struct lp_build_tgsi_context * bld_base,
1567    struct lp_build_emit_data * emit_data)
1568 {
1569    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1570
1571    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD,
1572             emit_data->output);
1573 }
1574
1575 static void
1576 txp_emit(
1577    const struct lp_build_tgsi_action * action,
1578    struct lp_build_tgsi_context * bld_base,
1579    struct lp_build_emit_data * emit_data)
1580 {
1581    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1582
1583    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_PROJECTED,
1584             emit_data->output);
1585 }
1586
1587 static void
1588 cal_emit(
1589    const struct lp_build_tgsi_action * action,
1590    struct lp_build_tgsi_context * bld_base,
1591    struct lp_build_emit_data * emit_data)
1592 {
1593    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1594
1595    lp_exec_mask_call(&bld->exec_mask, emit_data->inst->Label.Label,
1596                      &bld_base->pc);
1597 }
1598
1599 static void
1600 ret_emit(
1601    const struct lp_build_tgsi_action * action,
1602    struct lp_build_tgsi_context * bld_base,
1603    struct lp_build_emit_data * emit_data)
1604 {
1605    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1606
1607    lp_exec_mask_ret(&bld->exec_mask, &bld_base->pc);
1608 }
1609
1610 static void
1611 brk_emit(
1612    const struct lp_build_tgsi_action * action,
1613    struct lp_build_tgsi_context * bld_base,
1614    struct lp_build_emit_data * emit_data)
1615 {
1616    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1617
1618    lp_exec_break(&bld->exec_mask);
1619 }
1620
1621 static void
1622 if_emit(
1623    const struct lp_build_tgsi_action * action,
1624    struct lp_build_tgsi_context * bld_base,
1625    struct lp_build_emit_data * emit_data)
1626 {
1627    LLVMValueRef tmp;
1628    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1629
1630    tmp = lp_build_cmp(&bld_base->base, PIPE_FUNC_NOTEQUAL,
1631                       emit_data->args[0], bld->bld_base.base.zero);
1632    lp_exec_mask_cond_push(&bld->exec_mask, tmp);
1633 }
1634
1635 static void
1636 bgnloop_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_bgnloop(&bld->exec_mask);
1644 }
1645
1646 static void
1647 bgnsub_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_mask_bgnsub(&bld->exec_mask);
1655 }
1656
1657 static void
1658 else_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_cond_invert(&bld->exec_mask);
1666 }
1667
1668 static void
1669 endif_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_mask_cond_pop(&bld->exec_mask);
1677 }
1678
1679 static void
1680 endloop_emit(
1681    const struct lp_build_tgsi_action * action,
1682    struct lp_build_tgsi_context * bld_base,
1683    struct lp_build_emit_data * emit_data)
1684 {
1685    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1686
1687    lp_exec_endloop(bld_base->base.gallivm, &bld->exec_mask);
1688 }
1689
1690 static void
1691 endsub_emit(
1692    const struct lp_build_tgsi_action * action,
1693    struct lp_build_tgsi_context * bld_base,
1694    struct lp_build_emit_data * emit_data)
1695 {
1696    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1697
1698    lp_exec_mask_endsub(&bld->exec_mask, &bld_base->pc);
1699 }
1700
1701 static void
1702 cont_emit(
1703    const struct lp_build_tgsi_action * action,
1704    struct lp_build_tgsi_context * bld_base,
1705    struct lp_build_emit_data * emit_data)
1706 {
1707    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1708
1709    lp_exec_continue(&bld->exec_mask);
1710 }
1711
1712 /* XXX: Refactor and move it to lp_bld_tgsi_action.c
1713  *
1714  * XXX: What do the comments about xmm registers mean?  Maybe they are left over
1715  * from old code, but there is no garauntee that LLVM will use those registers
1716  * for this code.
1717  *
1718  * XXX: There should be no calls to lp_build_emit_fetch in this function.  This
1719  * should be handled by the emit_data->fetch_args function. */
1720 static void
1721 nrm_emit(
1722    const struct lp_build_tgsi_action * action,
1723    struct lp_build_tgsi_context * bld_base,
1724    struct lp_build_emit_data * emit_data)
1725 {
1726    LLVMValueRef tmp0, tmp1;
1727    LLVMValueRef tmp4 = NULL;
1728    LLVMValueRef tmp5 = NULL;
1729    LLVMValueRef tmp6 = NULL;
1730    LLVMValueRef tmp7 = NULL;
1731    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1732
1733    uint dims = (emit_data->inst->Instruction.Opcode == TGSI_OPCODE_NRM) ? 3 : 4;
1734
1735   if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_X) ||
1736       TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Y) ||
1737       TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Z) ||
1738       (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_W) && dims == 4)) {
1739
1740       /* NOTE: Cannot use xmm regs 2/3 here (see emit_rsqrt() above). */
1741
1742       /* xmm4 = src.x */
1743       /* xmm0 = src.x * src.x */
1744       tmp0 = lp_build_emit_fetch(&bld->bld_base, emit_data->inst, 0, TGSI_CHAN_X);
1745       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_X)) {
1746          tmp4 = tmp0;
1747       }
1748       tmp0 = lp_build_mul( &bld->bld_base.base, tmp0, tmp0);
1749
1750       /* xmm5 = src.y */
1751       /* xmm0 = xmm0 + src.y * src.y */
1752       tmp1 = lp_build_emit_fetch(&bld->bld_base, emit_data->inst, 0, TGSI_CHAN_Y);
1753       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Y)) {
1754          tmp5 = tmp1;
1755       }
1756       tmp1 = lp_build_mul( &bld->bld_base.base, tmp1, tmp1);
1757       tmp0 = lp_build_add( &bld->bld_base.base, tmp0, tmp1);
1758
1759       /* xmm6 = src.z */
1760       /* xmm0 = xmm0 + src.z * src.z */
1761       tmp1 = lp_build_emit_fetch(&bld->bld_base, emit_data->inst, 0, TGSI_CHAN_Z);
1762       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Z)) {
1763          tmp6 = tmp1;
1764       }
1765       tmp1 = lp_build_mul( &bld->bld_base.base, tmp1, tmp1);
1766       tmp0 = lp_build_add( &bld->bld_base.base, tmp0, tmp1);
1767
1768       if (dims == 4) {
1769          /* xmm7 = src.w */
1770          /* xmm0 = xmm0 + src.w * src.w */
1771          tmp1 = lp_build_emit_fetch(&bld->bld_base, emit_data->inst, 0, TGSI_CHAN_W);
1772          if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_W)) {
1773             tmp7 = tmp1;
1774          }
1775          tmp1 = lp_build_mul( &bld->bld_base.base, tmp1, tmp1);
1776          tmp0 = lp_build_add( &bld->bld_base.base, tmp0, tmp1);
1777       }
1778       /* xmm1 = 1 / sqrt(xmm0) */
1779       tmp1 = lp_build_rsqrt( &bld->bld_base.base, tmp0);
1780        /* dst.x = xmm1 * src.x */
1781       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_X)) {
1782          emit_data->output[TGSI_CHAN_X] = lp_build_mul( &bld->bld_base.base, tmp4, tmp1);
1783       }
1784       /* dst.y = xmm1 * src.y */
1785       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Y)) {
1786          emit_data->output[TGSI_CHAN_Y] = lp_build_mul( &bld->bld_base.base, tmp5, tmp1);
1787       }
1788
1789       /* dst.z = xmm1 * src.z */
1790       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Z)) {
1791          emit_data->output[TGSI_CHAN_Z] = lp_build_mul( &bld->bld_base.base, tmp6, tmp1);
1792       }
1793       /* dst.w = xmm1 * src.w */
1794       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_X) && dims == 4) {
1795          emit_data->output[TGSI_CHAN_W] = lp_build_mul( &bld->bld_base.base, tmp7, tmp1);
1796       }
1797    }
1798
1799    /* dst.w = 1.0 */
1800    if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_W) && dims == 3) {
1801        emit_data->output[TGSI_CHAN_W] = bld->bld_base.base.one;
1802    }
1803 }
1804
1805 static void emit_prologue(struct lp_build_tgsi_context * bld_base)
1806 {
1807    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1808    struct gallivm_state * gallivm = bld_base->base.gallivm;
1809
1810    if (bld->indirect_files & (1 << TGSI_FILE_TEMPORARY)) {
1811       LLVMValueRef array_size =
1812          lp_build_const_int32(gallivm,
1813                          bld_base->info->file_max[TGSI_FILE_TEMPORARY] * 4 + 4);
1814       bld->temps_array = lp_build_array_alloca(gallivm,
1815                                               bld_base->base.vec_type, array_size,
1816                                               "temp_array");
1817    }
1818
1819    if (bld->indirect_files & (1 << TGSI_FILE_OUTPUT)) {
1820       LLVMValueRef array_size =
1821          lp_build_const_int32(gallivm,
1822                             bld_base->info->file_max[TGSI_FILE_OUTPUT] * 4 + 4);
1823       bld->outputs_array = lp_build_array_alloca(gallivm,
1824                                                 bld_base->base.vec_type, array_size,
1825                                                 "output_array");
1826    }
1827
1828    /* If we have indirect addressing in inputs we need to copy them into
1829     * our alloca array to be able to iterate over them */
1830    if (bld->indirect_files & (1 << TGSI_FILE_INPUT)) {
1831       unsigned index, chan;
1832       LLVMTypeRef vec_type = bld_base->base.vec_type;
1833       LLVMValueRef array_size = lp_build_const_int32(gallivm,
1834             bld_base->info->file_max[TGSI_FILE_INPUT]*4 + 4);
1835       bld->inputs_array = lp_build_array_alloca(gallivm,
1836                                                vec_type, array_size,
1837                                                "input_array");
1838
1839       assert(bld_base->info->num_inputs
1840                         <= bld_base->info->file_max[TGSI_FILE_INPUT] + 1);
1841
1842       for (index = 0; index < bld_base->info->num_inputs; ++index) {
1843          for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
1844             LLVMValueRef lindex =
1845                lp_build_const_int32(gallivm, index * 4 + chan);
1846             LLVMValueRef input_ptr =
1847                LLVMBuildGEP(gallivm->builder, bld->inputs_array,
1848                             &lindex, 1, "");
1849             LLVMValueRef value = bld->inputs[index][chan];
1850             if (value)
1851                LLVMBuildStore(gallivm->builder, value, input_ptr);
1852          }
1853       }
1854    }
1855 }
1856
1857 static void emit_epilogue(struct lp_build_tgsi_context * bld_base)
1858 {
1859    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1860
1861    if (0) {
1862       /* for debugging */
1863       emit_dump_temps(bld);
1864    }
1865
1866    /* If we have indirect addressing in outputs we need to copy our alloca array
1867     * to the outputs slots specified by the called */
1868    if (bld->indirect_files & (1 << TGSI_FILE_OUTPUT)) {
1869       unsigned index, chan;
1870       assert(bld_base->info->num_outputs <=
1871                         bld_base->info->file_max[TGSI_FILE_OUTPUT] + 1);
1872       for (index = 0; index < bld_base->info->num_outputs; ++index) {
1873          for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
1874             bld->outputs[index][chan] = lp_get_output_ptr(bld, index, chan);
1875          }
1876       }
1877    }
1878 }
1879
1880 void
1881 lp_build_tgsi_soa(struct gallivm_state *gallivm,
1882                   const struct tgsi_token *tokens,
1883                   struct lp_type type,
1884                   struct lp_build_mask_context *mask,
1885                   LLVMValueRef consts_ptr,
1886                   LLVMValueRef system_values_array,
1887                   const LLVMValueRef *pos,
1888                   const LLVMValueRef (*inputs)[TGSI_NUM_CHANNELS],
1889                   LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1890                   struct lp_build_sampler_soa *sampler,
1891                   const struct tgsi_shader_info *info)
1892 {
1893    struct lp_build_tgsi_soa_context bld;
1894
1895    struct lp_type res_type;
1896
1897    assert(type.length <= LP_MAX_VECTOR_LENGTH);
1898    memset(&res_type, 0, sizeof res_type);
1899    res_type.width = type.width;
1900    res_type.length = type.length;
1901    res_type.sign = 1;
1902
1903    /* Setup build context */
1904    memset(&bld, 0, sizeof bld);
1905    lp_build_context_init(&bld.bld_base.base, gallivm, type);
1906    lp_build_context_init(&bld.bld_base.uint_bld, gallivm, lp_uint_type(type));
1907    lp_build_context_init(&bld.bld_base.int_bld, gallivm, lp_int_type(type));
1908    lp_build_context_init(&bld.elem_bld, gallivm, lp_elem_type(type));
1909    bld.mask = mask;
1910    bld.pos = pos;
1911    bld.inputs = inputs;
1912    bld.outputs = outputs;
1913    bld.consts_ptr = consts_ptr;
1914    bld.sampler = sampler;
1915    bld.bld_base.info = info;
1916    bld.indirect_files = info->indirect_files;
1917
1918    bld.bld_base.soa = TRUE;
1919    bld.bld_base.emit_fetch_funcs[TGSI_FILE_CONSTANT] = emit_fetch_constant;
1920    bld.bld_base.emit_fetch_funcs[TGSI_FILE_IMMEDIATE] = emit_fetch_immediate;
1921    bld.bld_base.emit_fetch_funcs[TGSI_FILE_INPUT] = emit_fetch_input;
1922    bld.bld_base.emit_fetch_funcs[TGSI_FILE_TEMPORARY] = emit_fetch_temporary;
1923    bld.bld_base.emit_fetch_funcs[TGSI_FILE_SYSTEM_VALUE] = emit_fetch_system_value;
1924    bld.bld_base.emit_store = emit_store;
1925
1926    bld.bld_base.emit_declaration = lp_emit_declaration_soa;
1927    bld.bld_base.emit_immediate = lp_emit_immediate_soa;
1928
1929    bld.bld_base.emit_prologue = emit_prologue;
1930    bld.bld_base.emit_epilogue = emit_epilogue;
1931
1932    /* Set opcode actions */
1933    lp_set_default_actions_cpu(&bld.bld_base);
1934
1935    bld.bld_base.op_actions[TGSI_OPCODE_BGNLOOP].emit = bgnloop_emit;
1936    bld.bld_base.op_actions[TGSI_OPCODE_BGNSUB].emit = bgnsub_emit;
1937    bld.bld_base.op_actions[TGSI_OPCODE_BRK].emit = brk_emit;
1938    bld.bld_base.op_actions[TGSI_OPCODE_CAL].emit = cal_emit;
1939    bld.bld_base.op_actions[TGSI_OPCODE_CONT].emit = cont_emit;
1940    bld.bld_base.op_actions[TGSI_OPCODE_DDX].emit = ddx_emit;
1941    bld.bld_base.op_actions[TGSI_OPCODE_DDY].emit = ddy_emit;
1942    bld.bld_base.op_actions[TGSI_OPCODE_ELSE].emit = else_emit;
1943    bld.bld_base.op_actions[TGSI_OPCODE_ENDIF].emit = endif_emit;
1944    bld.bld_base.op_actions[TGSI_OPCODE_ENDLOOP].emit = endloop_emit;
1945    bld.bld_base.op_actions[TGSI_OPCODE_ENDSUB].emit = endsub_emit;
1946    bld.bld_base.op_actions[TGSI_OPCODE_IF].emit = if_emit;
1947    bld.bld_base.op_actions[TGSI_OPCODE_KIL].emit = kil_emit;
1948    bld.bld_base.op_actions[TGSI_OPCODE_KILP].emit = kilp_emit;
1949    bld.bld_base.op_actions[TGSI_OPCODE_NRM].emit = nrm_emit;
1950    bld.bld_base.op_actions[TGSI_OPCODE_NRM4].emit = nrm_emit;
1951    bld.bld_base.op_actions[TGSI_OPCODE_RET].emit = ret_emit;
1952    bld.bld_base.op_actions[TGSI_OPCODE_TEX].emit = tex_emit;
1953    bld.bld_base.op_actions[TGSI_OPCODE_TXB].emit = txb_emit;
1954    bld.bld_base.op_actions[TGSI_OPCODE_TXD].emit = txd_emit;
1955    bld.bld_base.op_actions[TGSI_OPCODE_TXL].emit = txl_emit;
1956    bld.bld_base.op_actions[TGSI_OPCODE_TXP].emit = txp_emit;
1957
1958    lp_exec_mask_init(&bld.exec_mask, &bld.bld_base.base);
1959
1960
1961    bld.system_values_array = system_values_array;
1962
1963    lp_build_tgsi_llvm(&bld.bld_base, tokens);
1964
1965    if (0) {
1966       LLVMBasicBlockRef block = LLVMGetInsertBlock(gallivm->builder);
1967       LLVMValueRef function = LLVMGetBasicBlockParent(block);
1968       debug_printf("11111111111111111111111111111 \n");
1969       tgsi_dump(tokens, 0);
1970       lp_debug_dump_value(function);
1971       debug_printf("2222222222222222222222222222 \n");
1972    }
1973
1974    if (0) {
1975       LLVMModuleRef module = LLVMGetGlobalParent(
1976          LLVMGetBasicBlockParent(LLVMGetInsertBlock(gallivm->builder)));
1977       LLVMDumpModule(module);
1978
1979    }
1980 }
1981
1982
1983 /**
1984  * Build up the system values array out of individual values such as
1985  * the instance ID, front-face, primitive ID, etc.  The shader info is
1986  * used to determine which system values are needed and where to put
1987  * them in the system values array.
1988  *
1989  * XXX only instance ID is implemented at this time.
1990  *
1991  * The system values register file is similar to the constants buffer.
1992  * Example declaration:
1993  *    DCL SV[0], INSTANCEID
1994  * Example instruction:
1995  *    MOVE foo, SV[0].xxxx;
1996  *
1997  * \return  LLVM float array (interpreted as float [][4])
1998  */
1999 LLVMValueRef
2000 lp_build_system_values_array(struct gallivm_state *gallivm,
2001                              const struct tgsi_shader_info *info,
2002                              LLVMValueRef instance_id,
2003                              LLVMValueRef facing)
2004 {
2005    LLVMValueRef size = lp_build_const_int32(gallivm, 4 * info->num_system_values);
2006    LLVMTypeRef float_t = LLVMFloatTypeInContext(gallivm->context);
2007    LLVMValueRef array = lp_build_array_alloca(gallivm, float_t,
2008                                               size, "sysvals_array");
2009    unsigned i;
2010
2011    for (i = 0; i < info->num_system_values; i++) {
2012       LLVMValueRef index = lp_build_const_int32(gallivm, i * 4);
2013       LLVMValueRef ptr, value = 0;
2014
2015       switch (info->system_value_semantic_name[i]) {
2016       case TGSI_SEMANTIC_INSTANCEID:
2017          /* convert instance ID from int to float */
2018          value = LLVMBuildSIToFP(gallivm->builder, instance_id, float_t,
2019                                  "sysval_instanceid");
2020          break;
2021       case TGSI_SEMANTIC_FACE:
2022          /* fall-through */
2023       default:
2024          assert(0 && "unexpected semantic in build_system_values_array()");
2025       }
2026
2027       ptr = LLVMBuildGEP(gallivm->builder, array, &index, 1, "");
2028       LLVMBuildStore(gallivm->builder, value, ptr);
2029    }
2030
2031    return array;
2032 }