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 = LLVMBuildAlloca(
  83       builder,
  84       int_type,
  85       "looplimiter");
  86
  87    LLVMBuildStore(
  88       builder,
  89       LLVMConstInt(int_type, LP_MAX_TGSI_LOOP_ITERATIONS, false),
  90       mask->loop_limiter);
  91 }
  92
  93 static void lp_exec_mask_update(struct lp_exec_mask *mask)
  94 {
  95    LLVMBuilderRef builder = mask->bld->gallivm->builder;
  96
  97    if (mask->loop_stack_size) {
  98       /*for loops we need to update the entire mask at runtime */
  99       LLVMValueRef tmp;
 100       assert(mask->break_mask);
 101       tmp = LLVMBuildAnd(builder,
 102                          mask->cont_mask,
 103                          mask->break_mask,
 104                          "maskcb");
 105       mask->exec_mask = LLVMBuildAnd(builder,
 106                                      mask->cond_mask,
 107                                      tmp,
 108                                      "maskfull");
 109    } else
 110       mask->exec_mask = mask->cond_mask;
 111
 112    if (mask->call_stack_size) {
 113       mask->exec_mask = LLVMBuildAnd(builder,
 114                                      mask->exec_mask,
 115                                      mask->ret_mask,
 116                                      "callmask");
 117    }
 118
 119    mask->has_mask = (mask->cond_stack_size > 0 ||
 120                      mask->loop_stack_size > 0 ||
 121                      mask->call_stack_size > 0);
 122 }
 123
 124 static void lp_exec_mask_cond_push(struct lp_exec_mask *mask,
 125                                    LLVMValueRef val)
 126 {
 127    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 128
 129    assert(mask->cond_stack_size < LP_MAX_TGSI_NESTING);
 130    if (mask->cond_stack_size == 0) {
 131       assert(mask->cond_mask == LLVMConstAllOnes(mask->int_vec_type));
 132    }
 133    mask->cond_stack[mask->cond_stack_size++] = mask->cond_mask;
 134    assert(LLVMTypeOf(val) == mask->int_vec_type);
 135    mask->cond_mask = LLVMBuildAnd(builder,
 136                                   mask->cond_mask,
 137                                   val,
 138                                   "");
 139    lp_exec_mask_update(mask);
 140 }
 141
 142 static void lp_exec_mask_cond_invert(struct lp_exec_mask *mask)
 143 {
 144    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 145    LLVMValueRef prev_mask;
 146    LLVMValueRef inv_mask;
 147
 148    assert(mask->cond_stack_size);
 149    prev_mask = mask->cond_stack[mask->cond_stack_size - 1];
 150    if (mask->cond_stack_size == 1) {
 151       assert(prev_mask == LLVMConstAllOnes(mask->int_vec_type));
 152    }
 153
 154    inv_mask = LLVMBuildNot(builder, mask->cond_mask, "");
 155
 156    mask->cond_mask = LLVMBuildAnd(builder,
 157                                   inv_mask,
 158                                   prev_mask, "");
 159    lp_exec_mask_update(mask);
 160 }
 161
 162 static void lp_exec_mask_cond_pop(struct lp_exec_mask *mask)
 163 {
 164    assert(mask->cond_stack_size);
 165    mask->cond_mask = mask->cond_stack[--mask->cond_stack_size];
 166    lp_exec_mask_update(mask);
 167 }
 168
 169 static void lp_exec_bgnloop(struct lp_exec_mask *mask)
 170 {
 171    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 172
 173    if (mask->loop_stack_size == 0) {
 174       assert(mask->loop_block == NULL);
 175       assert(mask->cont_mask == LLVMConstAllOnes(mask->int_vec_type));
 176       assert(mask->break_mask == LLVMConstAllOnes(mask->int_vec_type));
 177       assert(mask->break_var == NULL);
 178    }
 179
 180    assert(mask->loop_stack_size < LP_MAX_TGSI_NESTING);
 181
 182    mask->loop_stack[mask->loop_stack_size].loop_block = mask->loop_block;
 183    mask->loop_stack[mask->loop_stack_size].cont_mask = mask->cont_mask;
 184    mask->loop_stack[mask->loop_stack_size].break_mask = mask->break_mask;
 185    mask->loop_stack[mask->loop_stack_size].break_var = mask->break_var;
 186    ++mask->loop_stack_size;
 187
 188    mask->break_var = lp_build_alloca(mask->bld->gallivm, mask->int_vec_type, "");
 189    LLVMBuildStore(builder, mask->break_mask, mask->break_var);
 190
 191    mask->loop_block = lp_build_insert_new_block(mask->bld->gallivm, "bgnloop");
 192
 193    LLVMBuildBr(builder, mask->loop_block);
 194    LLVMPositionBuilderAtEnd(builder, mask->loop_block);
 195
 196    mask->break_mask = LLVMBuildLoad(builder, mask->break_var, "");
 197
 198    lp_exec_mask_update(mask);
 199 }
 200
 201 static void lp_exec_break(struct lp_exec_mask *mask)
 202 {
 203    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 204    LLVMValueRef exec_mask = LLVMBuildNot(builder,
 205                                          mask->exec_mask,
 206                                          "break");
 207
 208    mask->break_mask = LLVMBuildAnd(builder,
 209                                    mask->break_mask,
 210                                    exec_mask, "break_full");
 211
 212    lp_exec_mask_update(mask);
 213 }
 214
 215 static void lp_exec_continue(struct lp_exec_mask *mask)
 216 {
 217    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 218    LLVMValueRef exec_mask = LLVMBuildNot(builder,
 219                                          mask->exec_mask,
 220                                          "");
 221
 222    mask->cont_mask = LLVMBuildAnd(builder,
 223                                   mask->cont_mask,
 224                                   exec_mask, "");
 225
 226    lp_exec_mask_update(mask);
 227 }
 228
 229
 230 static void lp_exec_endloop(struct gallivm_state *gallivm,
 231                             struct lp_exec_mask *mask)
 232 {
 233    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 234    LLVMBasicBlockRef endloop;
 235    LLVMTypeRef int_type = LLVMInt32TypeInContext(mask->bld->gallivm->context);
 236    LLVMTypeRef reg_type = LLVMIntTypeInContext(gallivm->context,
 237                                                mask->bld->type.width *
 238                                                mask->bld->type.length);
 239    LLVMValueRef i1cond, i2cond, icond, limiter;
 240
 241    assert(mask->break_mask);
 242
 243    /*
 244     * Restore the cont_mask, but don't pop
 245     */
 246    assert(mask->loop_stack_size);
 247    mask->cont_mask = mask->loop_stack[mask->loop_stack_size - 1].cont_mask;
 248    lp_exec_mask_update(mask);
 249
 250    /*
 251     * Unlike the continue mask, the break_mask must be preserved across loop
 252     * iterations
 253     */
 254    LLVMBuildStore(builder, mask->break_mask, mask->break_var);
 255
 256    /* Decrement the loop limiter */
 257    limiter = LLVMBuildLoad(builder, mask->loop_limiter, "");
 258
 259    limiter = LLVMBuildSub(
 260       builder,
 261       limiter,
 262       LLVMConstInt(int_type, 1, false),
 263       "");
 264
 265    LLVMBuildStore(builder, limiter, mask->loop_limiter);
 266
 267    /* i1cond = (mask != 0) */
 268    i1cond = LLVMBuildICmp(
 269       builder,
 270       LLVMIntNE,
 271       LLVMBuildBitCast(builder, mask->exec_mask, reg_type, ""),
 272       LLVMConstNull(reg_type), "");
 273
 274    /* i2cond = (looplimiter > 0) */
 275    i2cond = LLVMBuildICmp(
 276       builder,
 277       LLVMIntSGT,
 278       limiter,
 279       LLVMConstNull(int_type), "");
 280
 281    /* if( i1cond && i2cond ) */
 282    icond = LLVMBuildAnd(builder, i1cond, i2cond, "");
 283
 284    endloop = lp_build_insert_new_block(mask->bld->gallivm, "endloop");
 285
 286    LLVMBuildCondBr(builder,
 287                    icond, mask->loop_block, endloop);
 288
 289    LLVMPositionBuilderAtEnd(builder, endloop);
 290
 291    assert(mask->loop_stack_size);
 292    --mask->loop_stack_size;
 293    mask->loop_block = mask->loop_stack[mask->loop_stack_size].loop_block;
 294    mask->cont_mask = mask->loop_stack[mask->loop_stack_size].cont_mask;
 295    mask->break_mask = mask->loop_stack[mask->loop_stack_size].break_mask;
 296    mask->break_var = mask->loop_stack[mask->loop_stack_size].break_var;
 297
 298    lp_exec_mask_update(mask);
 299 }
 300
 301 /* stores val into an address pointed to by dst.
 302  * mask->exec_mask is used to figure out which bits of val
 303  * should be stored into the address
 304  * (0 means don't store this bit, 1 means do store).
 305  */
 306 static void lp_exec_mask_store(struct lp_exec_mask *mask,
 307                                struct lp_build_context *bld_store,
 308                                LLVMValueRef pred,
 309                                LLVMValueRef val,
 310                                LLVMValueRef dst)
 311 {
 312    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 313
 314    /* Mix the predicate and execution mask */
 315    if (mask->has_mask) {
 316       if (pred) {
 317          pred = LLVMBuildAnd(builder, pred, mask->exec_mask, "");
 318       } else {
 319          pred = mask->exec_mask;
 320       }
 321    }
 322
 323    if (pred) {
 324       LLVMValueRef real_val, dst_val;
 325
 326       dst_val = LLVMBuildLoad(builder, dst, "");
 327       real_val = lp_build_select(bld_store,
 328                                  pred,
 329                                  val, dst_val);
 330
 331       LLVMBuildStore(builder, real_val, dst);
 332    } else
 333       LLVMBuildStore(builder, val, dst);
 334 }
 335
 336 static void lp_exec_mask_call(struct lp_exec_mask *mask,
 337                               int func,
 338                               int *pc)
 339 {
 340    assert(mask->call_stack_size < LP_MAX_TGSI_NESTING);
 341    mask->call_stack[mask->call_stack_size].pc = *pc;
 342    mask->call_stack[mask->call_stack_size].ret_mask = mask->ret_mask;
 343    mask->call_stack_size++;
 344    *pc = func;
 345 }
 346
 347 static void lp_exec_mask_ret(struct lp_exec_mask *mask, int *pc)
 348 {
 349    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 350    LLVMValueRef exec_mask;
 351
 352    if (mask->call_stack_size == 0) {
 353       /* returning from main() */
 354       *pc = -1;
 355       return;
 356    }
 357    exec_mask = LLVMBuildNot(builder,
 358                             mask->exec_mask,
 359                             "ret");
 360
 361    mask->ret_mask = LLVMBuildAnd(builder,
 362                                  mask->ret_mask,
 363                                  exec_mask, "ret_full");
 364
 365    lp_exec_mask_update(mask);
 366 }
 367
 368 static void lp_exec_mask_bgnsub(struct lp_exec_mask *mask)
 369 {
 370 }
 371
 372 static void lp_exec_mask_endsub(struct lp_exec_mask *mask, int *pc)
 373 {
 374    assert(mask->call_stack_size);
 375    mask->call_stack_size--;
 376    *pc = mask->call_stack[mask->call_stack_size].pc;
 377    mask->ret_mask = mask->call_stack[mask->call_stack_size].ret_mask;
 378    lp_exec_mask_update(mask);
 379 }
 380
 381
 382 /**
 383  * Return pointer to a temporary register channel (src or dest).
 384  * Note that indirect addressing cannot be handled here.
 385  * \param index  which temporary register
 386  * \param chan  which channel of the temp register.
 387  */
 388 LLVMValueRef
 389 lp_get_temp_ptr_soa(struct lp_build_tgsi_soa_context *bld,
 390              unsigned index,
 391              unsigned chan)
 392 {
 393    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
 394    assert(chan < 4);
 395    if (bld->indirect_files & (1 << TGSI_FILE_TEMPORARY)) {
 396       LLVMValueRef lindex = lp_build_const_int32(bld->bld_base.base.gallivm, index * 4 + chan);
 397       return LLVMBuildGEP(builder, bld->temps_array, &lindex, 1, "");
 398    }
 399    else {
 400       return bld->temps[index][chan];
 401    }
 402 }
 403
 404 /**
 405  * Return pointer to a output register channel (src or dest).
 406  * Note that indirect addressing cannot be handled here.
 407  * \param index  which output register
 408  * \param chan  which channel of the output register.
 409  */
 410 LLVMValueRef
 411 lp_get_output_ptr(struct lp_build_tgsi_soa_context *bld,
 412                unsigned index,
 413                unsigned chan)
 414 {
 415    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
 416    assert(chan < 4);
 417    if (bld->indirect_files & (1 << TGSI_FILE_OUTPUT)) {
 418       LLVMValueRef lindex = lp_build_const_int32(bld->bld_base.base.gallivm,
 419                                                  index * 4 + chan);
 420       return LLVMBuildGEP(builder, bld->outputs_array, &lindex, 1, "");
 421    }
 422    else {
 423       return bld->outputs[index][chan];
 424    }
 425 }
 426
 427 /**
 428  * Gather vector.
 429  * XXX the lp_build_gather() function should be capable of doing this
 430  * with a little work.
 431  */
 432 static LLVMValueRef
 433 build_gather(struct lp_build_context *bld,
 434              LLVMValueRef base_ptr,
 435              LLVMValueRef indexes)
 436 {
 437    LLVMBuilderRef builder = bld->gallivm->builder;
 438    LLVMValueRef res = bld->undef;
 439    unsigned i;
 440
 441    /*
 442     * Loop over elements of index_vec, load scalar value, insert it into 'res'.
 443     */
 444    for (i = 0; i < bld->type.length; i++) {
 445       LLVMValueRef ii = lp_build_const_int32(bld->gallivm, i);
 446       LLVMValueRef index = LLVMBuildExtractElement(builder,
 447                                                    indexes, ii, "");
 448       LLVMValueRef scalar_ptr = LLVMBuildGEP(builder, base_ptr,
 449                                              &index, 1, "gather_ptr");
 450       LLVMValueRef scalar = LLVMBuildLoad(builder, scalar_ptr, "");
 451
 452       res = LLVMBuildInsertElement(builder, res, scalar, ii, "");
 453    }
 454
 455    return res;
 456 }
 457
 458
 459 /**
 460  * Scatter/store vector.
 461  */
 462 static void
 463 emit_mask_scatter(struct lp_build_tgsi_soa_context *bld,
 464                   LLVMValueRef base_ptr,
 465                   LLVMValueRef indexes,
 466                   LLVMValueRef values,
 467                   struct lp_exec_mask *mask,
 468                   LLVMValueRef pred)
 469 {
 470    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
 471    LLVMBuilderRef builder = gallivm->builder;
 472    unsigned i;
 473
 474    /* Mix the predicate and execution mask */
 475    if (mask->has_mask) {
 476       if (pred) {
 477          pred = LLVMBuildAnd(builder, pred, mask->exec_mask, "");
 478       }
 479       else {
 480          pred = mask->exec_mask;
 481       }
 482    }
 483
 484    /*
 485     * Loop over elements of index_vec, store scalar value.
 486     */
 487    for (i = 0; i < bld->bld_base.base.type.length; i++) {
 488       LLVMValueRef ii = lp_build_const_int32(gallivm, i);
 489       LLVMValueRef index = LLVMBuildExtractElement(builder, indexes, ii, "");
 490       LLVMValueRef scalar_ptr = LLVMBuildGEP(builder, base_ptr, &index, 1, "scatter_ptr");
 491       LLVMValueRef val = LLVMBuildExtractElement(builder, values, ii, "scatter_val");
 492       LLVMValueRef scalar_pred = pred ?
 493          LLVMBuildExtractElement(builder, pred, ii, "scatter_pred") : NULL;
 494
 495       if (0)
 496          lp_build_printf(gallivm, "scatter %d: val %f at %d %p\n",
 497                          ii, val, index, scalar_ptr);
 498
 499       if (scalar_pred) {
 500          LLVMValueRef real_val, dst_val;
 501          dst_val = LLVMBuildLoad(builder, scalar_ptr, "");
 502          real_val = lp_build_select(&bld->elem_bld, scalar_pred, val, dst_val);
 503          LLVMBuildStore(builder, real_val, scalar_ptr);
 504       }
 505       else {
 506          LLVMBuildStore(builder, val, scalar_ptr);
 507       }
 508    }
 509 }
 510
 511
 512 /**
 513  * Read the current value of the ADDR register, convert the floats to
 514  * ints, add the base index and return the vector of offsets.
 515  * The offsets will be used to index into the constant buffer or
 516  * temporary register file.
 517  */
 518 static LLVMValueRef
 519 get_indirect_index(struct lp_build_tgsi_soa_context *bld,
 520                    unsigned reg_file, unsigned reg_index,
 521                    const struct tgsi_src_register *indirect_reg)
 522 {
 523    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
 524    struct lp_build_context *uint_bld = &bld->bld_base.uint_bld;
 525    /* always use X component of address register */
 526    unsigned swizzle = indirect_reg->SwizzleX;
 527    LLVMValueRef base;
 528    LLVMValueRef rel;
 529    LLVMValueRef max_index;
 530    LLVMValueRef index;
 531
 532    assert(bld->indirect_files & (1 << reg_file));
 533
 534    base = lp_build_const_int_vec(bld->bld_base.base.gallivm, uint_bld->type, reg_index);
 535
 536    assert(swizzle < 4);
 537    rel = LLVMBuildLoad(builder,
 538                         bld->addr[indirect_reg->Index][swizzle],
 539                         "load addr reg");
 540
 541    index = lp_build_add(uint_bld, base, rel);
 542
 543    max_index = lp_build_const_int_vec(bld->bld_base.base.gallivm,
 544                                       uint_bld->type,
 545                                       bld->bld_base.info->file_max[reg_file]);
 546
 547    assert(!uint_bld->type.sign);
 548    index = lp_build_min(uint_bld, index, max_index);
 549
 550    return index;
 551 }
 552
 553 static struct lp_build_context *
 554 stype_to_fetch(struct lp_build_tgsi_context * bld_base,
 555                enum tgsi_opcode_type stype)
 556 {
 557    struct lp_build_context *bld_fetch;
 558
 559    switch (stype) {
 560    case TGSI_TYPE_FLOAT:
 561    case TGSI_TYPE_UNTYPED:
 562       bld_fetch = &bld_base->base;
 563       break;
 564    case TGSI_TYPE_UNSIGNED:
 565       bld_fetch = &bld_base->uint_bld;
 566       break;
 567    case TGSI_TYPE_SIGNED:
 568       bld_fetch = &bld_base->int_bld;
 569       break;
 570    case TGSI_TYPE_VOID:
 571    case TGSI_TYPE_DOUBLE:
 572    default:
 573       assert(0);
 574       bld_fetch = NULL;
 575       break;
 576    }
 577    return bld_fetch;
 578 }
 579
 580 static LLVMValueRef
 581 emit_fetch_constant(
 582    struct lp_build_tgsi_context * bld_base,
 583    const struct tgsi_full_src_register * reg,
 584    enum tgsi_opcode_type stype,
 585    unsigned swizzle)
 586 {
 587    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
 588    struct gallivm_state *gallivm = bld_base->base.gallivm;
 589    LLVMBuilderRef builder = gallivm->builder;
 590    struct lp_build_context *uint_bld = &bld_base->uint_bld;
 591    LLVMValueRef indirect_index = NULL;
 592    struct lp_build_context *bld_fetch = stype_to_fetch(bld_base, stype);
 593
 594    /* XXX: Handle fetching xyzw components as a vector */
 595    assert(swizzle != ~0);
 596
 597    if (reg->Register.Indirect) {
 598       indirect_index = get_indirect_index(bld,
 599                                           reg->Register.File,
 600                                           reg->Register.Index,
 601                                           &reg->Indirect);
 602    }
 603
 604    if (reg->Register.Indirect) {
 605       LLVMValueRef swizzle_vec =
 606          lp_build_const_int_vec(bld->bld_base.base.gallivm, uint_bld->type, swizzle);
 607       LLVMValueRef index_vec;  /* index into the const buffer */
 608
 609       /* index_vec = indirect_index * 4 + swizzle */
 610       index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
 611       index_vec = lp_build_add(uint_bld, index_vec, swizzle_vec);
 612
 613       /* Gather values from the constant buffer */
 614       return build_gather(bld_fetch, bld->consts_ptr, index_vec);
 615    }
 616    else {
 617       LLVMValueRef index;  /* index into the const buffer */
 618       LLVMValueRef scalar, scalar_ptr;
 619
 620       index = lp_build_const_int32(gallivm, reg->Register.Index*4 + swizzle);
 621
 622       scalar_ptr = LLVMBuildGEP(builder, bld->consts_ptr,
 623                                    &index, 1, "");
 624
 625       if (stype != TGSI_TYPE_FLOAT && stype != TGSI_TYPE_UNTYPED) {
 626          LLVMTypeRef ivtype = LLVMPointerType(LLVMInt32TypeInContext(gallivm->context), 0);
 627          LLVMValueRef temp_ptr;
 628          temp_ptr = LLVMBuildBitCast(builder, scalar_ptr, ivtype, "");
 629          scalar = LLVMBuildLoad(builder, temp_ptr, "");
 630       } else
 631          scalar = LLVMBuildLoad(builder, scalar_ptr, "");
 632
 633       return lp_build_broadcast_scalar(bld_fetch, scalar);
 634    }
 635 }
 636
 637 static LLVMValueRef
 638 emit_fetch_immediate(
 639    struct lp_build_tgsi_context * bld_base,
 640    const struct tgsi_full_src_register * reg,
 641    enum tgsi_opcode_type stype,
 642    unsigned swizzle)
 643 {
 644    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
 645    LLVMValueRef res = bld->immediates[reg->Register.Index][swizzle];
 646    assert(res);
 647
 648    if (stype == TGSI_TYPE_UNSIGNED) {
 649       res = LLVMConstBitCast(res, bld_base->uint_bld.vec_type);
 650    } else if (stype == TGSI_TYPE_SIGNED) {
 651       res = LLVMConstBitCast(res, bld_base->int_bld.vec_type);
 652    }
 653    return res;
 654 }
 655
 656 static LLVMValueRef
 657 emit_fetch_input(
 658    struct lp_build_tgsi_context * bld_base,
 659    const struct tgsi_full_src_register * reg,
 660    enum tgsi_opcode_type stype,
 661    unsigned swizzle)
 662 {
 663    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
 664    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
 665    LLVMBuilderRef builder = gallivm->builder;
 666    struct lp_build_context *uint_bld = &bld_base->uint_bld;
 667    LLVMValueRef indirect_index = NULL;
 668    LLVMValueRef res;
 669
 670    if (reg->Register.Indirect) {
 671       indirect_index = get_indirect_index(bld,
 672                                           reg->Register.File,
 673                                           reg->Register.Index,
 674                                           &reg->Indirect);
 675    }
 676
 677    if (reg->Register.Indirect) {
 678       LLVMValueRef swizzle_vec =
 679          lp_build_const_int_vec(gallivm, uint_bld->type, swizzle);
 680       LLVMValueRef length_vec =
 681          lp_build_const_int_vec(gallivm, uint_bld->type, bld->bld_base.base.type.length);
 682       LLVMValueRef index_vec;  /* index into the const buffer */
 683       LLVMValueRef inputs_array;
 684       LLVMTypeRef float4_ptr_type;
 685
 686       /* index_vec = (indirect_index * 4 + swizzle) * length */
 687       index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
 688       index_vec = lp_build_add(uint_bld, index_vec, swizzle_vec);
 689       index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
 690
 691       /* cast inputs_array pointer to float* */
 692       float4_ptr_type = LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
 693       inputs_array = LLVMBuildBitCast(builder, bld->inputs_array,
 694                                          float4_ptr_type, "");
 695
 696       /* Gather values from the temporary register array */
 697       res = build_gather(&bld_base->base, inputs_array, index_vec);
 698    } else {
 699       if (bld->indirect_files & (1 << TGSI_FILE_INPUT)) {
 700          LLVMValueRef lindex = lp_build_const_int32(gallivm,
 701                                         reg->Register.Index * 4 + swizzle);
 702          LLVMValueRef input_ptr =  LLVMBuildGEP(builder,
 703                                                 bld->inputs_array, &lindex, 1, "");
 704          res = LLVMBuildLoad(builder, input_ptr, "");
 705       }
 706       else {
 707          res = bld->inputs[reg->Register.Index][swizzle];
 708       }
 709    }
 710    assert(res);
 711    return res;
 712 }
 713
 714 static LLVMValueRef
 715 emit_fetch_temporary(
 716    struct lp_build_tgsi_context * bld_base,
 717    const struct tgsi_full_src_register * reg,
 718    enum tgsi_opcode_type stype,
 719    unsigned swizzle)
 720 {
 721    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
 722    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
 723    LLVMBuilderRef builder = gallivm->builder;
 724    struct lp_build_context *uint_bld = &bld_base->uint_bld;
 725    LLVMValueRef indirect_index = NULL;
 726    LLVMValueRef res;
 727
 728    if (reg->Register.Indirect) {
 729       indirect_index = get_indirect_index(bld,
 730                                           reg->Register.File,
 731                                           reg->Register.Index,
 732                                           &reg->Indirect);
 733    }
 734
 735    if (reg->Register.Indirect) {
 736       LLVMValueRef swizzle_vec =
 737          lp_build_const_int_vec(bld->bld_base.base.gallivm, uint_bld->type, swizzle);
 738       LLVMValueRef length_vec =
 739          lp_build_const_int_vec(bld->bld_base.base.gallivm, uint_bld->type,
 740                                 bld->bld_base.base.type.length);
 741       LLVMValueRef index_vec;  /* index into the const buffer */
 742       LLVMValueRef temps_array;
 743       LLVMTypeRef float4_ptr_type;
 744
 745       /* index_vec = (indirect_index * 4 + swizzle) * length */
 746       index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
 747       index_vec = lp_build_add(uint_bld, index_vec, swizzle_vec);
 748       index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
 749
 750       /* cast temps_array pointer to float* */
 751       float4_ptr_type = LLVMPointerType(LLVMFloatTypeInContext(bld->bld_base.base.gallivm->context), 0);
 752       temps_array = LLVMBuildBitCast(builder, bld->temps_array,
 753                                      float4_ptr_type, "");
 754
 755       /* Gather values from the temporary register array */
 756       res = build_gather(&bld_base->base, temps_array, index_vec);
 757    }
 758    else {
 759       LLVMValueRef temp_ptr;
 760       if (stype != TGSI_TYPE_FLOAT && stype != TGSI_TYPE_UNTYPED) {
 761          LLVMTypeRef itype = LLVMPointerType(LLVMVectorType(LLVMInt32TypeInContext(gallivm->context), 4), 0);
 762          LLVMValueRef tint_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index,
 763                                                      swizzle);
 764          temp_ptr = LLVMBuildBitCast(builder, tint_ptr, itype, "");
 765       } else
 766          temp_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index, swizzle);
 767       res = LLVMBuildLoad(builder, temp_ptr, "");
 768       if (!res)
 769          return bld->bld_base.base.undef;
 770    }
 771
 772    return res;
 773 }
 774
 775 static LLVMValueRef
 776 emit_fetch_system_value(
 777    struct lp_build_tgsi_context * bld_base,
 778    const struct tgsi_full_src_register * reg,
 779    enum tgsi_opcode_type stype,
 780    unsigned swizzle)
 781 {
 782    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
 783    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
 784    LLVMBuilderRef builder = gallivm->builder;
 785    LLVMValueRef index;  /* index into the system value array */
 786    LLVMValueRef scalar, scalar_ptr;
 787
 788    assert(!reg->Register.Indirect);
 789
 790    index = lp_build_const_int32(gallivm, reg->Register.Index * 4 + swizzle);
 791
 792    scalar_ptr = LLVMBuildGEP(builder, bld->system_values_array, &index, 1, "");
 793    scalar = LLVMBuildLoad(builder, scalar_ptr, "");
 794
 795    return lp_build_broadcast_scalar(&bld->bld_base.base, scalar);
 796 }
 797
 798 /**
 799  * Register fetch with derivatives.
 800  */
 801 static void
 802 emit_fetch_deriv(
 803    struct lp_build_tgsi_soa_context *bld,
 804    LLVMValueRef src,
 805    LLVMValueRef *res,
 806    LLVMValueRef *ddx,
 807    LLVMValueRef *ddy)
 808 {
 809    if(res)
 810       *res = src;
 811
 812    /* TODO: use interpolation coeffs for inputs */
 813
 814    if(ddx)
 815       *ddx = lp_build_ddx(&bld->bld_base.base, src);
 816
 817    if(ddy)
 818       *ddy = lp_build_ddy(&bld->bld_base.base, src);
 819 }
 820
 821
 822 /**
 823  * Predicate.
 824  */
 825 static void
 826 emit_fetch_predicate(
 827    struct lp_build_tgsi_soa_context *bld,
 828    const struct tgsi_full_instruction *inst,
 829    LLVMValueRef *pred)
 830 {
 831    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
 832    unsigned index;
 833    unsigned char swizzles[4];
 834    LLVMValueRef unswizzled[4] = {NULL, NULL, NULL, NULL};
 835    LLVMValueRef value;
 836    unsigned chan;
 837
 838    if (!inst->Instruction.Predicate) {
 839       TGSI_FOR_EACH_CHANNEL( chan ) {
 840          pred[chan] = NULL;
 841       }
 842       return;
 843    }
 844
 845    swizzles[0] = inst->Predicate.SwizzleX;
 846    swizzles[1] = inst->Predicate.SwizzleY;
 847    swizzles[2] = inst->Predicate.SwizzleZ;
 848    swizzles[3] = inst->Predicate.SwizzleW;
 849
 850    index = inst->Predicate.Index;
 851    assert(index < LP_MAX_TGSI_PREDS);
 852
 853    TGSI_FOR_EACH_CHANNEL( chan ) {
 854       unsigned swizzle = swizzles[chan];
 855
 856       /*
 857        * Only fetch the predicate register channels that are actually listed
 858        * in the swizzles
 859        */
 860       if (!unswizzled[swizzle]) {
 861          value = LLVMBuildLoad(builder,
 862                                bld->preds[index][swizzle], "");
 863
 864          /*
 865           * Convert the value to an integer mask.
 866           *
 867           * TODO: Short-circuit this comparison -- a D3D setp_xx instructions
 868           * is needlessly causing two comparisons due to storing the intermediate
 869           * result as float vector instead of an integer mask vector.
 870           */
 871          value = lp_build_compare(bld->bld_base.base.gallivm,
 872                                   bld->bld_base.base.type,
 873                                   PIPE_FUNC_NOTEQUAL,
 874                                   value,
 875                                   bld->bld_base.base.zero);
 876          if (inst->Predicate.Negate) {
 877             value = LLVMBuildNot(builder, value, "");
 878          }
 879
 880          unswizzled[swizzle] = value;
 881       } else {
 882          value = unswizzled[swizzle];
 883       }
 884
 885       pred[chan] = value;
 886    }
 887 }
 888
 889 /**
 890  * Register store.
 891  */
 892 static void
 893 emit_store_chan(
 894    struct lp_build_tgsi_context *bld_base,
 895    const struct tgsi_full_instruction *inst,
 896    unsigned index,
 897    unsigned chan_index,
 898    LLVMValueRef pred,
 899    LLVMValueRef value)
 900 {
 901    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
 902    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
 903    LLVMBuilderRef builder = gallivm->builder;
 904    const struct tgsi_full_dst_register *reg = &inst->Dst[index];
 905    struct lp_build_context *uint_bld = &bld_base->uint_bld;
 906    LLVMValueRef indirect_index = NULL;
 907    struct lp_build_context *bld_store;
 908    enum tgsi_opcode_type dtype = tgsi_opcode_infer_dst_type(inst->Instruction.Opcode);
 909
 910    switch (dtype) {
 911    default:
 912    case TGSI_TYPE_FLOAT:
 913    case TGSI_TYPE_UNTYPED:
 914       bld_store = &bld_base->base;
 915       break;
 916    case TGSI_TYPE_UNSIGNED:
 917       bld_store = &bld_base->uint_bld;
 918       break;
 919    case TGSI_TYPE_SIGNED:
 920       bld_store = &bld_base->int_bld;
 921       break;
 922    case TGSI_TYPE_DOUBLE:
 923    case TGSI_TYPE_VOID:
 924       assert(0);
 925       bld_store = NULL;
 926       break;
 927    }
 928
 929    switch( inst->Instruction.Saturate ) {
 930    case TGSI_SAT_NONE:
 931       break;
 932
 933    case TGSI_SAT_ZERO_ONE:
 934       value = lp_build_max(&bld->bld_base.base, value, bld->bld_base.base.zero);
 935       value = lp_build_min(&bld->bld_base.base, value, bld->bld_base.base.one);
 936       break;
 937
 938    case TGSI_SAT_MINUS_PLUS_ONE:
 939       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));
 940       value = lp_build_min(&bld->bld_base.base, value, bld->bld_base.base.one);
 941       break;
 942
 943    default:
 944       assert(0);
 945    }
 946
 947    if (reg->Register.Indirect) {
 948       indirect_index = get_indirect_index(bld,
 949                                           reg->Register.File,
 950                                           reg->Register.Index,
 951                                           &reg->Indirect);
 952    } else {
 953       assert(reg->Register.Index <=
 954                              bld->bld_base.info->file_max[reg->Register.File]);
 955    }
 956
 957    switch( reg->Register.File ) {
 958    case TGSI_FILE_OUTPUT:
 959       if (reg->Register.Indirect) {
 960          LLVMValueRef chan_vec =
 961             lp_build_const_int_vec(gallivm, uint_bld->type, chan_index);
 962          LLVMValueRef length_vec =
 963             lp_build_const_int_vec(gallivm, uint_bld->type, bld->bld_base.base.type.length);
 964          LLVMValueRef index_vec;  /* indexes into the temp registers */
 965          LLVMValueRef outputs_array;
 966          LLVMValueRef pixel_offsets;
 967          LLVMTypeRef float_ptr_type;
 968          int i;
 969
 970          /* build pixel offset vector: {0, 1, 2, 3, ...} */
 971          pixel_offsets = uint_bld->undef;
 972          for (i = 0; i < bld->bld_base.base.type.length; i++) {
 973             LLVMValueRef ii = lp_build_const_int32(gallivm, i);
 974             pixel_offsets = LLVMBuildInsertElement(builder, pixel_offsets,
 975                                                    ii, ii, "");
 976          }
 977
 978          /* index_vec = (indirect_index * 4 + chan_index) * length + offsets */
 979          index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
 980          index_vec = lp_build_add(uint_bld, index_vec, chan_vec);
 981          index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
 982          index_vec = lp_build_add(uint_bld, index_vec, pixel_offsets);
 983
 984          float_ptr_type =
 985             LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
 986          outputs_array = LLVMBuildBitCast(builder, bld->outputs_array,
 987                                           float_ptr_type, "");
 988
 989          /* Scatter store values into temp registers */
 990          emit_mask_scatter(bld, outputs_array, index_vec, value,
 991                            &bld->exec_mask, pred);
 992       }
 993       else {
 994          LLVMValueRef out_ptr = lp_get_output_ptr(bld, reg->Register.Index,
 995                                                chan_index);
 996          lp_exec_mask_store(&bld->exec_mask, bld_store, pred, value, out_ptr);
 997       }
 998       break;
 999
1000    case TGSI_FILE_TEMPORARY:
1001       if (reg->Register.Indirect) {
1002          LLVMValueRef chan_vec =
1003             lp_build_const_int_vec(gallivm, uint_bld->type, chan_index);
1004          LLVMValueRef length_vec =
1005             lp_build_const_int_vec(gallivm, uint_bld->type,
1006                                    bld->bld_base.base.type.length);
1007          LLVMValueRef index_vec;  /* indexes into the temp registers */
1008          LLVMValueRef temps_array;
1009          LLVMValueRef pixel_offsets;
1010          LLVMTypeRef float_ptr_type;
1011          int i;
1012
1013          /* build pixel offset vector: {0, 1, 2, 3, ...} */
1014          pixel_offsets = uint_bld->undef;
1015          for (i = 0; i < bld->bld_base.base.type.length; i++) {
1016             LLVMValueRef ii = lp_build_const_int32(gallivm, i);
1017             pixel_offsets = LLVMBuildInsertElement(builder, pixel_offsets,
1018                                                    ii, ii, "");
1019          }
1020
1021          /* index_vec = (indirect_index * 4 + chan_index) * length + offsets */
1022          index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
1023          index_vec = lp_build_add(uint_bld, index_vec, chan_vec);
1024          index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
1025          index_vec = lp_build_add(uint_bld, index_vec, pixel_offsets);
1026
1027          float_ptr_type =
1028             LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
1029          temps_array = LLVMBuildBitCast(builder, bld->temps_array,
1030                                         float_ptr_type, "");
1031
1032          /* Scatter store values into temp registers */
1033          emit_mask_scatter(bld, temps_array, index_vec, value,
1034                            &bld->exec_mask, pred);
1035       }
1036       else {
1037          LLVMValueRef temp_ptr;
1038
1039          switch (dtype) {
1040          case TGSI_TYPE_UNSIGNED:
1041          case TGSI_TYPE_SIGNED: {
1042             LLVMTypeRef itype = LLVMVectorType(LLVMInt32TypeInContext(gallivm->context), 4);
1043             LLVMTypeRef ivtype = LLVMPointerType(itype, 0);
1044             LLVMValueRef tint_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index,
1045                                                         chan_index);
1046             LLVMValueRef temp_value_ptr;
1047
1048             temp_ptr = LLVMBuildBitCast(builder, tint_ptr, ivtype, "");
1049             temp_value_ptr = LLVMBuildBitCast(builder, value, itype, "");
1050             value = temp_value_ptr;
1051             break;
1052          }
1053          default:
1054          case TGSI_TYPE_FLOAT:
1055          case TGSI_TYPE_UNTYPED:
1056             temp_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index,
1057                                            chan_index);
1058             break;
1059          }
1060
1061          lp_exec_mask_store(&bld->exec_mask, bld_store, pred, value, temp_ptr);
1062       }
1063       break;
1064
1065    case TGSI_FILE_ADDRESS:
1066       assert(dtype == TGSI_TYPE_SIGNED);
1067       assert(LLVMTypeOf(value) == bld_base->base.int_vec_type);
1068       lp_exec_mask_store(&bld->exec_mask, bld_store, pred, value,
1069                          bld->addr[reg->Register.Index][chan_index]);
1070       break;
1071
1072    case TGSI_FILE_PREDICATE:
1073       lp_exec_mask_store(&bld->exec_mask, bld_store, pred, value,
1074                          bld->preds[reg->Register.Index][chan_index]);
1075       break;
1076
1077    default:
1078       assert( 0 );
1079    }
1080 }
1081
1082 static void
1083 emit_store(
1084    struct lp_build_tgsi_context * bld_base,
1085    const struct tgsi_full_instruction * inst,
1086    const struct tgsi_opcode_info * info,
1087    LLVMValueRef dst[4])
1088
1089 {
1090    unsigned chan_index;
1091    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1092
1093    if(info->num_dst) {
1094       LLVMValueRef pred[TGSI_NUM_CHANNELS];
1095
1096       emit_fetch_predicate( bld, inst, pred );
1097
1098       TGSI_FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1099          emit_store_chan(bld_base, inst, 0, chan_index, pred[chan_index], dst[chan_index]);
1100       }
1101    }
1102 }
1103
1104 /**
1105  * High-level instruction translators.
1106  */
1107
1108 static void
1109 emit_tex( struct lp_build_tgsi_soa_context *bld,
1110           const struct tgsi_full_instruction *inst,
1111           enum lp_build_tex_modifier modifier,
1112           LLVMValueRef *texel)
1113 {
1114    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
1115    unsigned unit;
1116    LLVMValueRef lod_bias, explicit_lod;
1117    LLVMValueRef oow = NULL;
1118    LLVMValueRef coords[3];
1119    LLVMValueRef ddx[3];
1120    LLVMValueRef ddy[3];
1121    unsigned num_coords;
1122    unsigned i;
1123
1124    if (!bld->sampler) {
1125       _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
1126       for (i = 0; i < 4; i++) {
1127          texel[i] = bld->bld_base.base.undef;
1128       }
1129       return;
1130    }
1131
1132    switch (inst->Texture.Texture) {
1133    case TGSI_TEXTURE_1D:
1134       num_coords = 1;
1135       break;
1136    case TGSI_TEXTURE_1D_ARRAY:
1137    case TGSI_TEXTURE_2D:
1138    case TGSI_TEXTURE_RECT:
1139       num_coords = 2;
1140       break;
1141    case TGSI_TEXTURE_SHADOW1D:
1142    case TGSI_TEXTURE_SHADOW1D_ARRAY:
1143    case TGSI_TEXTURE_SHADOW2D:
1144    case TGSI_TEXTURE_SHADOWRECT:
1145    case TGSI_TEXTURE_2D_ARRAY:
1146    case TGSI_TEXTURE_3D:
1147    case TGSI_TEXTURE_CUBE:
1148       num_coords = 3;
1149       break;
1150    case TGSI_TEXTURE_SHADOW2D_ARRAY:
1151       num_coords = 4;
1152       break;
1153    default:
1154       assert(0);
1155       return;
1156    }
1157
1158    if (modifier == LP_BLD_TEX_MODIFIER_LOD_BIAS) {
1159       lod_bias = lp_build_emit_fetch( &bld->bld_base, inst, 0, 3 );
1160       explicit_lod = NULL;
1161    }
1162    else if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_LOD) {
1163       lod_bias = NULL;
1164       explicit_lod = lp_build_emit_fetch( &bld->bld_base, inst, 0, 3 );
1165    }
1166    else {
1167       lod_bias = NULL;
1168       explicit_lod = NULL;
1169    }
1170
1171    if (modifier == LP_BLD_TEX_MODIFIER_PROJECTED) {
1172       oow = lp_build_emit_fetch( &bld->bld_base, inst, 0, 3 );
1173       oow = lp_build_rcp(&bld->bld_base.base, oow);
1174    }
1175
1176    for (i = 0; i < num_coords; i++) {
1177       coords[i] = lp_build_emit_fetch( &bld->bld_base, inst, 0, i );
1178       if (modifier == LP_BLD_TEX_MODIFIER_PROJECTED)
1179          coords[i] = lp_build_mul(&bld->bld_base.base, coords[i], oow);
1180    }
1181    for (i = num_coords; i < 3; i++) {
1182       coords[i] = bld->bld_base.base.undef;
1183    }
1184
1185    if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV) {
1186       LLVMValueRef index0 = lp_build_const_int32(bld->bld_base.base.gallivm, 0);
1187       for (i = 0; i < num_coords; i++) {
1188          LLVMValueRef src1 = lp_build_emit_fetch( &bld->bld_base, inst, 1, i );
1189          LLVMValueRef src2 = lp_build_emit_fetch( &bld->bld_base, inst, 2, i );
1190          ddx[i] = LLVMBuildExtractElement(builder, src1, index0, "");
1191          ddy[i] = LLVMBuildExtractElement(builder, src2, index0, "");
1192       }
1193       unit = inst->Src[3].Register.Index;
1194    }  else {
1195       for (i = 0; i < num_coords; i++) {
1196          ddx[i] = lp_build_scalar_ddx( &bld->bld_base.base, coords[i] );
1197          ddy[i] = lp_build_scalar_ddy( &bld->bld_base.base, coords[i] );
1198       }
1199       unit = inst->Src[1].Register.Index;
1200    }
1201    for (i = num_coords; i < 3; i++) {
1202       ddx[i] = LLVMGetUndef(bld->bld_base.base.elem_type);
1203       ddy[i] = LLVMGetUndef(bld->bld_base.base.elem_type);
1204    }
1205
1206    bld->sampler->emit_fetch_texel(bld->sampler,
1207                                   bld->bld_base.base.gallivm,
1208                                   bld->bld_base.base.type,
1209                                   unit, num_coords, coords,
1210                                   ddx, ddy,
1211                                   lod_bias, explicit_lod,
1212                                   texel);
1213 }
1214
1215 static boolean
1216 near_end_of_shader(struct lp_build_tgsi_soa_context *bld,
1217                    int pc)
1218 {
1219    int i;
1220
1221    for (i = 0; i < 5; i++) {
1222       unsigned opcode;
1223
1224       if (pc + i >= bld->bld_base.info->num_instructions)
1225          return TRUE;
1226
1227       opcode = bld->bld_base.instructions[pc + i].Instruction.Opcode;
1228
1229       if (opcode == TGSI_OPCODE_END)
1230          return TRUE;
1231
1232       if (opcode == TGSI_OPCODE_TEX ||
1233           opcode == TGSI_OPCODE_TXP ||
1234           opcode == TGSI_OPCODE_TXD ||
1235           opcode == TGSI_OPCODE_TXB ||
1236           opcode == TGSI_OPCODE_TXL ||
1237           opcode == TGSI_OPCODE_TXF ||
1238           opcode == TGSI_OPCODE_TXQ ||
1239           opcode == TGSI_OPCODE_CAL ||
1240           opcode == TGSI_OPCODE_CALLNZ ||
1241           opcode == TGSI_OPCODE_IF ||
1242           opcode == TGSI_OPCODE_IFC ||
1243           opcode == TGSI_OPCODE_BGNLOOP ||
1244           opcode == TGSI_OPCODE_SWITCH)
1245          return FALSE;
1246    }
1247
1248    return TRUE;
1249 }
1250
1251
1252
1253 /**
1254  * Kill fragment if any of the src register values are negative.
1255  */
1256 static void
1257 emit_kil(
1258    struct lp_build_tgsi_soa_context *bld,
1259    const struct tgsi_full_instruction *inst,
1260    int pc)
1261 {
1262    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
1263    const struct tgsi_full_src_register *reg = &inst->Src[0];
1264    LLVMValueRef terms[TGSI_NUM_CHANNELS];
1265    LLVMValueRef mask;
1266    unsigned chan_index;
1267
1268    memset(&terms, 0, sizeof terms);
1269
1270    TGSI_FOR_EACH_CHANNEL( chan_index ) {
1271       unsigned swizzle;
1272
1273       /* Unswizzle channel */
1274       swizzle = tgsi_util_get_full_src_register_swizzle( reg, chan_index );
1275
1276       /* Check if the component has not been already tested. */
1277       assert(swizzle < TGSI_NUM_CHANNELS);
1278       if( !terms[swizzle] )
1279          /* TODO: change the comparison operator instead of setting the sign */
1280          terms[swizzle] =  lp_build_emit_fetch(&bld->bld_base, inst, 0, chan_index );
1281    }
1282
1283    mask = NULL;
1284    TGSI_FOR_EACH_CHANNEL( chan_index ) {
1285       if(terms[chan_index]) {
1286          LLVMValueRef chan_mask;
1287
1288          /*
1289           * If term < 0 then mask = 0 else mask = ~0.
1290           */
1291          chan_mask = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_GEQUAL, terms[chan_index], bld->bld_base.base.zero);
1292
1293          if(mask)
1294             mask = LLVMBuildAnd(builder, mask, chan_mask, "");
1295          else
1296             mask = chan_mask;
1297       }
1298    }
1299
1300    if(mask) {
1301       lp_build_mask_update(bld->mask, mask);
1302
1303       if (!near_end_of_shader(bld, pc))
1304          lp_build_mask_check(bld->mask);
1305    }
1306 }
1307
1308
1309 /**
1310  * Predicated fragment kill.
1311  * XXX Actually, we do an unconditional kill (as in tgsi_exec.c).
1312  * The only predication is the execution mask which will apply if
1313  * we're inside a loop or conditional.
1314  */
1315 static void
1316 emit_kilp(struct lp_build_tgsi_soa_context *bld,
1317           int pc)
1318 {
1319    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
1320    LLVMValueRef mask;
1321
1322    /* For those channels which are "alive", disable fragment shader
1323     * execution.
1324     */
1325    if (bld->exec_mask.has_mask) {
1326       mask = LLVMBuildNot(builder, bld->exec_mask.exec_mask, "kilp");
1327    }
1328    else {
1329       LLVMValueRef zero = LLVMConstNull(bld->bld_base.base.int_vec_type);
1330       mask = zero;
1331    }
1332
1333    lp_build_mask_update(bld->mask, mask);
1334
1335    if (!near_end_of_shader(bld, pc))
1336       lp_build_mask_check(bld->mask);
1337 }
1338
1339
1340 /**
1341  * Emit code which will dump the value of all the temporary registers
1342  * to stdout.
1343  */
1344 static void
1345 emit_dump_temps(struct lp_build_tgsi_soa_context *bld)
1346 {
1347    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
1348    LLVMBuilderRef builder = gallivm->builder;
1349    LLVMValueRef temp_ptr;
1350    LLVMValueRef i0 = lp_build_const_int32(gallivm, 0);
1351    LLVMValueRef i1 = lp_build_const_int32(gallivm, 1);
1352    LLVMValueRef i2 = lp_build_const_int32(gallivm, 2);
1353    LLVMValueRef i3 = lp_build_const_int32(gallivm, 3);
1354    int index;
1355    int n = bld->bld_base.info->file_max[TGSI_FILE_TEMPORARY];
1356
1357    for (index = 0; index < n; index++) {
1358       LLVMValueRef idx = lp_build_const_int32(gallivm, index);
1359       LLVMValueRef v[4][4], res;
1360       int chan;
1361
1362       lp_build_printf(gallivm, "TEMP[%d]:\n", idx);
1363
1364       for (chan = 0; chan < 4; chan++) {
1365          temp_ptr = lp_get_temp_ptr_soa(bld, index, chan);
1366          res = LLVMBuildLoad(builder, temp_ptr, "");
1367          v[chan][0] = LLVMBuildExtractElement(builder, res, i0, "");
1368          v[chan][1] = LLVMBuildExtractElement(builder, res, i1, "");
1369          v[chan][2] = LLVMBuildExtractElement(builder, res, i2, "");
1370          v[chan][3] = LLVMBuildExtractElement(builder, res, i3, "");
1371       }
1372
1373       lp_build_printf(gallivm, "  X: %f %f %f %f\n",
1374                       v[0][0], v[0][1], v[0][2], v[0][3]);
1375       lp_build_printf(gallivm, "  Y: %f %f %f %f\n",
1376                       v[1][0], v[1][1], v[1][2], v[1][3]);
1377       lp_build_printf(gallivm, "  Z: %f %f %f %f\n",
1378                       v[2][0], v[2][1], v[2][2], v[2][3]);
1379       lp_build_printf(gallivm, "  W: %f %f %f %f\n",
1380                       v[3][0], v[3][1], v[3][2], v[3][3]);
1381    }
1382 }
1383
1384
1385
1386 void
1387 lp_emit_declaration_soa(
1388    struct lp_build_tgsi_context *bld_base,
1389    const struct tgsi_full_declaration *decl)
1390 {
1391    struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
1392    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
1393    LLVMTypeRef vec_type = bld->bld_base.base.vec_type;
1394    const unsigned first = decl->Range.First;
1395    const unsigned last = decl->Range.Last;
1396    unsigned idx, i;
1397
1398    for (idx = first; idx <= last; ++idx) {
1399       assert(last <= bld->bld_base.info->file_max[decl->Declaration.File]);
1400       switch (decl->Declaration.File) {
1401       case TGSI_FILE_TEMPORARY:
1402          assert(idx < LP_MAX_TGSI_TEMPS);
1403          if (!(bld->indirect_files & (1 << TGSI_FILE_TEMPORARY))) {
1404             for (i = 0; i < TGSI_NUM_CHANNELS; i++)
1405                bld->temps[idx][i] = lp_build_alloca(gallivm, vec_type, "temp");
1406          }
1407          break;
1408
1409       case TGSI_FILE_OUTPUT:
1410          if (!(bld->indirect_files & (1 << TGSI_FILE_OUTPUT))) {
1411             for (i = 0; i < TGSI_NUM_CHANNELS; i++)
1412                bld->outputs[idx][i] = lp_build_alloca(gallivm,
1413                                                       vec_type, "output");
1414          }
1415          break;
1416
1417       case TGSI_FILE_ADDRESS:
1418          /* ADDR registers are the only allocated with an integer LLVM IR type,
1419           * as they are guaranteed to always have integers.
1420           * XXX: Not sure if this exception is worthwhile (or the whole idea of
1421           * an ADDR register for that matter).
1422           */
1423          assert(idx < LP_MAX_TGSI_ADDRS);
1424          for (i = 0; i < TGSI_NUM_CHANNELS; i++)
1425             bld->addr[idx][i] = lp_build_alloca(gallivm, bld_base->base.int_vec_type, "addr");
1426          break;
1427
1428       case TGSI_FILE_PREDICATE:
1429          assert(idx < LP_MAX_TGSI_PREDS);
1430          for (i = 0; i < TGSI_NUM_CHANNELS; i++)
1431             bld->preds[idx][i] = lp_build_alloca(gallivm, vec_type,
1432                                                  "predicate");
1433          break;
1434
1435       default:
1436          /* don't need to declare other vars */
1437          break;
1438       }
1439    }
1440 }
1441
1442
1443 void lp_emit_immediate_soa(
1444    struct lp_build_tgsi_context *bld_base,
1445    const struct tgsi_full_immediate *imm)
1446 {
1447    struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
1448    struct gallivm_state * gallivm = bld_base->base.gallivm;
1449
1450    /* simply copy the immediate values into the next immediates[] slot */
1451    unsigned i;
1452    const uint size = imm->Immediate.NrTokens - 1;
1453    assert(size <= 4);
1454    assert(bld->num_immediates < LP_MAX_TGSI_IMMEDIATES);
1455    switch (imm->Immediate.DataType) {
1456    case TGSI_IMM_FLOAT32:
1457       for( i = 0; i < size; ++i )
1458          bld->immediates[bld->num_immediates][i] =
1459             lp_build_const_vec(gallivm, bld_base->base.type, imm->u[i].Float);
1460
1461       break;
1462    case TGSI_IMM_UINT32:
1463       for( i = 0; i < size; ++i ) {
1464          LLVMValueRef tmp = lp_build_const_vec(gallivm, bld_base->uint_bld.type, imm->u[i].Uint);
1465          bld->immediates[bld->num_immediates][i] =
1466             LLVMConstBitCast(tmp, bld_base->base.vec_type);
1467       }
1468
1469       break;
1470    case TGSI_IMM_INT32:
1471       for( i = 0; i < size; ++i ) {
1472          LLVMValueRef tmp = lp_build_const_vec(gallivm, bld_base->int_bld.type, imm->u[i].Int);
1473          bld->immediates[bld->num_immediates][i] =
1474             LLVMConstBitCast(tmp, bld_base->base.vec_type);
1475       }
1476
1477       break;
1478    }
1479    for( i = size; i < 4; ++i )
1480       bld->immediates[bld->num_immediates][i] = bld_base->base.undef;
1481
1482    bld->num_immediates++;
1483 }
1484
1485 static void
1486 ddx_emit(
1487    const struct lp_build_tgsi_action * action,
1488    struct lp_build_tgsi_context * bld_base,
1489    struct lp_build_emit_data * emit_data)
1490 {
1491    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1492
1493    emit_fetch_deriv(bld, emit_data->args[0], NULL,
1494                     &emit_data->output[emit_data->chan], NULL);
1495 }
1496
1497 static void
1498 ddy_emit(
1499    const struct lp_build_tgsi_action * action,
1500    struct lp_build_tgsi_context * bld_base,
1501    struct lp_build_emit_data * emit_data)
1502 {
1503    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1504
1505    emit_fetch_deriv(bld, emit_data->args[0], NULL, NULL,
1506                     &emit_data->output[emit_data->chan]);
1507 }
1508
1509 static void
1510 kilp_emit(
1511    const struct lp_build_tgsi_action * action,
1512    struct lp_build_tgsi_context * bld_base,
1513    struct lp_build_emit_data * emit_data)
1514 {
1515    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1516
1517    emit_kilp(bld, bld_base->pc - 1);
1518 }
1519
1520 static void
1521 kil_emit(
1522    const struct lp_build_tgsi_action * action,
1523    struct lp_build_tgsi_context * bld_base,
1524    struct lp_build_emit_data * emit_data)
1525 {
1526    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1527
1528    emit_kil(bld, emit_data->inst, bld_base->pc - 1);
1529 }
1530
1531 static void
1532 tex_emit(
1533    const struct lp_build_tgsi_action * action,
1534    struct lp_build_tgsi_context * bld_base,
1535    struct lp_build_emit_data * emit_data)
1536 {
1537    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1538
1539    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_NONE, emit_data->output);
1540 }
1541
1542 static void
1543 txb_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_LOD_BIAS,
1551             emit_data->output);
1552 }
1553
1554 static void
1555 txd_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    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV,
1563             emit_data->output);
1564 }
1565
1566 static void
1567 txl_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    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD,
1575             emit_data->output);
1576 }
1577
1578 static void
1579 txp_emit(
1580    const struct lp_build_tgsi_action * action,
1581    struct lp_build_tgsi_context * bld_base,
1582    struct lp_build_emit_data * emit_data)
1583 {
1584    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1585
1586    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_PROJECTED,
1587             emit_data->output);
1588 }
1589
1590 static void
1591 cal_emit(
1592    const struct lp_build_tgsi_action * action,
1593    struct lp_build_tgsi_context * bld_base,
1594    struct lp_build_emit_data * emit_data)
1595 {
1596    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1597
1598    lp_exec_mask_call(&bld->exec_mask, emit_data->inst->Label.Label,
1599                      &bld_base->pc);
1600 }
1601
1602 static void
1603 ret_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_mask_ret(&bld->exec_mask, &bld_base->pc);
1611 }
1612
1613 static void
1614 brk_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_break(&bld->exec_mask);
1622 }
1623
1624 static void
1625 if_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    LLVMValueRef tmp;
1631    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1632
1633    tmp = lp_build_cmp(&bld_base->base, PIPE_FUNC_NOTEQUAL,
1634                       emit_data->args[0], bld->bld_base.base.zero);
1635    lp_exec_mask_cond_push(&bld->exec_mask, tmp);
1636 }
1637
1638 static void
1639 bgnloop_emit(
1640    const struct lp_build_tgsi_action * action,
1641    struct lp_build_tgsi_context * bld_base,
1642    struct lp_build_emit_data * emit_data)
1643 {
1644    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1645
1646    lp_exec_bgnloop(&bld->exec_mask);
1647 }
1648
1649 static void
1650 bgnsub_emit(
1651    const struct lp_build_tgsi_action * action,
1652    struct lp_build_tgsi_context * bld_base,
1653    struct lp_build_emit_data * emit_data)
1654 {
1655    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1656
1657    lp_exec_mask_bgnsub(&bld->exec_mask);
1658 }
1659
1660 static void
1661 else_emit(
1662    const struct lp_build_tgsi_action * action,
1663    struct lp_build_tgsi_context * bld_base,
1664    struct lp_build_emit_data * emit_data)
1665 {
1666    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1667
1668    lp_exec_mask_cond_invert(&bld->exec_mask);
1669 }
1670
1671 static void
1672 endif_emit(
1673    const struct lp_build_tgsi_action * action,
1674    struct lp_build_tgsi_context * bld_base,
1675    struct lp_build_emit_data * emit_data)
1676 {
1677    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1678
1679    lp_exec_mask_cond_pop(&bld->exec_mask);
1680 }
1681
1682 static void
1683 endloop_emit(
1684    const struct lp_build_tgsi_action * action,
1685    struct lp_build_tgsi_context * bld_base,
1686    struct lp_build_emit_data * emit_data)
1687 {
1688    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1689
1690    lp_exec_endloop(bld_base->base.gallivm, &bld->exec_mask);
1691 }
1692
1693 static void
1694 endsub_emit(
1695    const struct lp_build_tgsi_action * action,
1696    struct lp_build_tgsi_context * bld_base,
1697    struct lp_build_emit_data * emit_data)
1698 {
1699    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1700
1701    lp_exec_mask_endsub(&bld->exec_mask, &bld_base->pc);
1702 }
1703
1704 static void
1705 cont_emit(
1706    const struct lp_build_tgsi_action * action,
1707    struct lp_build_tgsi_context * bld_base,
1708    struct lp_build_emit_data * emit_data)
1709 {
1710    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1711
1712    lp_exec_continue(&bld->exec_mask);
1713 }
1714
1715 /* XXX: Refactor and move it to lp_bld_tgsi_action.c
1716  *
1717  * XXX: What do the comments about xmm registers mean?  Maybe they are left over
1718  * from old code, but there is no garauntee that LLVM will use those registers
1719  * for this code.
1720  *
1721  * XXX: There should be no calls to lp_build_emit_fetch in this function.  This
1722  * should be handled by the emit_data->fetch_args function. */
1723 static void
1724 nrm_emit(
1725    const struct lp_build_tgsi_action * action,
1726    struct lp_build_tgsi_context * bld_base,
1727    struct lp_build_emit_data * emit_data)
1728 {
1729    LLVMValueRef tmp0, tmp1;
1730    LLVMValueRef tmp4 = NULL;
1731    LLVMValueRef tmp5 = NULL;
1732    LLVMValueRef tmp6 = NULL;
1733    LLVMValueRef tmp7 = NULL;
1734    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1735
1736    uint dims = (emit_data->inst->Instruction.Opcode == TGSI_OPCODE_NRM) ? 3 : 4;
1737
1738   if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_X) ||
1739       TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Y) ||
1740       TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Z) ||
1741       (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_W) && dims == 4)) {
1742
1743       /* NOTE: Cannot use xmm regs 2/3 here (see emit_rsqrt() above). */
1744
1745       /* xmm4 = src.x */
1746       /* xmm0 = src.x * src.x */
1747       tmp0 = lp_build_emit_fetch(&bld->bld_base, emit_data->inst, 0, TGSI_CHAN_X);
1748       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_X)) {
1749          tmp4 = tmp0;
1750       }
1751       tmp0 = lp_build_mul( &bld->bld_base.base, tmp0, tmp0);
1752
1753       /* xmm5 = src.y */
1754       /* xmm0 = xmm0 + src.y * src.y */
1755       tmp1 = lp_build_emit_fetch(&bld->bld_base, emit_data->inst, 0, TGSI_CHAN_Y);
1756       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Y)) {
1757          tmp5 = tmp1;
1758       }
1759       tmp1 = lp_build_mul( &bld->bld_base.base, tmp1, tmp1);
1760       tmp0 = lp_build_add( &bld->bld_base.base, tmp0, tmp1);
1761
1762       /* xmm6 = src.z */
1763       /* xmm0 = xmm0 + src.z * src.z */
1764       tmp1 = lp_build_emit_fetch(&bld->bld_base, emit_data->inst, 0, TGSI_CHAN_Z);
1765       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Z)) {
1766          tmp6 = tmp1;
1767       }
1768       tmp1 = lp_build_mul( &bld->bld_base.base, tmp1, tmp1);
1769       tmp0 = lp_build_add( &bld->bld_base.base, tmp0, tmp1);
1770
1771       if (dims == 4) {
1772          /* xmm7 = src.w */
1773          /* xmm0 = xmm0 + src.w * src.w */
1774          tmp1 = lp_build_emit_fetch(&bld->bld_base, emit_data->inst, 0, TGSI_CHAN_W);
1775          if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_W)) {
1776             tmp7 = tmp1;
1777          }
1778          tmp1 = lp_build_mul( &bld->bld_base.base, tmp1, tmp1);
1779          tmp0 = lp_build_add( &bld->bld_base.base, tmp0, tmp1);
1780       }
1781       /* xmm1 = 1 / sqrt(xmm0) */
1782       tmp1 = lp_build_rsqrt( &bld->bld_base.base, tmp0);
1783        /* dst.x = xmm1 * src.x */
1784       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_X)) {
1785          emit_data->output[TGSI_CHAN_X] = lp_build_mul( &bld->bld_base.base, tmp4, tmp1);
1786       }
1787       /* dst.y = xmm1 * src.y */
1788       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Y)) {
1789          emit_data->output[TGSI_CHAN_Y] = lp_build_mul( &bld->bld_base.base, tmp5, tmp1);
1790       }
1791
1792       /* dst.z = xmm1 * src.z */
1793       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Z)) {
1794          emit_data->output[TGSI_CHAN_Z] = lp_build_mul( &bld->bld_base.base, tmp6, tmp1);
1795       }
1796       /* dst.w = xmm1 * src.w */
1797       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_X) && dims == 4) {
1798          emit_data->output[TGSI_CHAN_W] = lp_build_mul( &bld->bld_base.base, tmp7, tmp1);
1799       }
1800    }
1801
1802    /* dst.w = 1.0 */
1803    if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_W) && dims == 3) {
1804        emit_data->output[TGSI_CHAN_W] = bld->bld_base.base.one;
1805    }
1806 }
1807
1808 static void emit_prologue(struct lp_build_tgsi_context * bld_base)
1809 {
1810    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1811    struct gallivm_state * gallivm = bld_base->base.gallivm;
1812
1813    if (bld->indirect_files & (1 << TGSI_FILE_TEMPORARY)) {
1814       LLVMValueRef array_size =
1815          lp_build_const_int32(gallivm,
1816                          bld_base->info->file_max[TGSI_FILE_TEMPORARY] * 4 + 4);
1817       bld->temps_array = lp_build_array_alloca(gallivm,
1818                                               bld_base->base.vec_type, array_size,
1819                                               "temp_array");
1820    }
1821
1822    if (bld->indirect_files & (1 << TGSI_FILE_OUTPUT)) {
1823       LLVMValueRef array_size =
1824          lp_build_const_int32(gallivm,
1825                             bld_base->info->file_max[TGSI_FILE_OUTPUT] * 4 + 4);
1826       bld->outputs_array = lp_build_array_alloca(gallivm,
1827                                                 bld_base->base.vec_type, array_size,
1828                                                 "output_array");
1829    }
1830
1831    /* If we have indirect addressing in inputs we need to copy them into
1832     * our alloca array to be able to iterate over them */
1833    if (bld->indirect_files & (1 << TGSI_FILE_INPUT)) {
1834       unsigned index, chan;
1835       LLVMTypeRef vec_type = bld_base->base.vec_type;
1836       LLVMValueRef array_size = lp_build_const_int32(gallivm,
1837             bld_base->info->file_max[TGSI_FILE_INPUT]*4 + 4);
1838       bld->inputs_array = lp_build_array_alloca(gallivm,
1839                                                vec_type, array_size,
1840                                                "input_array");
1841
1842       assert(bld_base->info->num_inputs
1843                         <= bld_base->info->file_max[TGSI_FILE_INPUT] + 1);
1844
1845       for (index = 0; index < bld_base->info->num_inputs; ++index) {
1846          for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
1847             LLVMValueRef lindex =
1848                lp_build_const_int32(gallivm, index * 4 + chan);
1849             LLVMValueRef input_ptr =
1850                LLVMBuildGEP(gallivm->builder, bld->inputs_array,
1851                             &lindex, 1, "");
1852             LLVMValueRef value = bld->inputs[index][chan];
1853             if (value)
1854                LLVMBuildStore(gallivm->builder, value, input_ptr);
1855          }
1856       }
1857    }
1858 }
1859
1860 static void emit_epilogue(struct lp_build_tgsi_context * bld_base)
1861 {
1862    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1863
1864    if (0) {
1865       /* for debugging */
1866       emit_dump_temps(bld);
1867    }
1868
1869    /* If we have indirect addressing in outputs we need to copy our alloca array
1870     * to the outputs slots specified by the called */
1871    if (bld->indirect_files & (1 << TGSI_FILE_OUTPUT)) {
1872       unsigned index, chan;
1873       assert(bld_base->info->num_outputs <=
1874                         bld_base->info->file_max[TGSI_FILE_OUTPUT] + 1);
1875       for (index = 0; index < bld_base->info->num_outputs; ++index) {
1876          for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
1877             bld->outputs[index][chan] = lp_get_output_ptr(bld, index, chan);
1878          }
1879       }
1880    }
1881 }
1882
1883 void
1884 lp_build_tgsi_soa(struct gallivm_state *gallivm,
1885                   const struct tgsi_token *tokens,
1886                   struct lp_type type,
1887                   struct lp_build_mask_context *mask,
1888                   LLVMValueRef consts_ptr,
1889                   LLVMValueRef system_values_array,
1890                   const LLVMValueRef *pos,
1891                   const LLVMValueRef (*inputs)[TGSI_NUM_CHANNELS],
1892                   LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1893                   struct lp_build_sampler_soa *sampler,
1894                   const struct tgsi_shader_info *info)
1895 {
1896    struct lp_build_tgsi_soa_context bld;
1897
1898    struct lp_type res_type;
1899
1900    assert(type.length <= LP_MAX_VECTOR_LENGTH);
1901    memset(&res_type, 0, sizeof res_type);
1902    res_type.width = type.width;
1903    res_type.length = type.length;
1904    res_type.sign = 1;
1905
1906    /* Setup build context */
1907    memset(&bld, 0, sizeof bld);
1908    lp_build_context_init(&bld.bld_base.base, gallivm, type);
1909    lp_build_context_init(&bld.bld_base.uint_bld, gallivm, lp_uint_type(type));
1910    lp_build_context_init(&bld.bld_base.int_bld, gallivm, lp_int_type(type));
1911    lp_build_context_init(&bld.elem_bld, gallivm, lp_elem_type(type));
1912    bld.mask = mask;
1913    bld.pos = pos;
1914    bld.inputs = inputs;
1915    bld.outputs = outputs;
1916    bld.consts_ptr = consts_ptr;
1917    bld.sampler = sampler;
1918    bld.bld_base.info = info;
1919    bld.indirect_files = info->indirect_files;
1920
1921    bld.bld_base.soa = TRUE;
1922    bld.bld_base.emit_fetch_funcs[TGSI_FILE_CONSTANT] = emit_fetch_constant;
1923    bld.bld_base.emit_fetch_funcs[TGSI_FILE_IMMEDIATE] = emit_fetch_immediate;
1924    bld.bld_base.emit_fetch_funcs[TGSI_FILE_INPUT] = emit_fetch_input;
1925    bld.bld_base.emit_fetch_funcs[TGSI_FILE_TEMPORARY] = emit_fetch_temporary;
1926    bld.bld_base.emit_fetch_funcs[TGSI_FILE_SYSTEM_VALUE] = emit_fetch_system_value;
1927    bld.bld_base.emit_store = emit_store;
1928
1929    bld.bld_base.emit_declaration = lp_emit_declaration_soa;
1930    bld.bld_base.emit_immediate = lp_emit_immediate_soa;
1931
1932    bld.bld_base.emit_prologue = emit_prologue;
1933    bld.bld_base.emit_epilogue = emit_epilogue;
1934
1935    /* Set opcode actions */
1936    lp_set_default_actions_cpu(&bld.bld_base);
1937
1938    bld.bld_base.op_actions[TGSI_OPCODE_BGNLOOP].emit = bgnloop_emit;
1939    bld.bld_base.op_actions[TGSI_OPCODE_BGNSUB].emit = bgnsub_emit;
1940    bld.bld_base.op_actions[TGSI_OPCODE_BRK].emit = brk_emit;
1941    bld.bld_base.op_actions[TGSI_OPCODE_CAL].emit = cal_emit;
1942    bld.bld_base.op_actions[TGSI_OPCODE_CONT].emit = cont_emit;
1943    bld.bld_base.op_actions[TGSI_OPCODE_DDX].emit = ddx_emit;
1944    bld.bld_base.op_actions[TGSI_OPCODE_DDY].emit = ddy_emit;
1945    bld.bld_base.op_actions[TGSI_OPCODE_ELSE].emit = else_emit;
1946    bld.bld_base.op_actions[TGSI_OPCODE_ENDIF].emit = endif_emit;
1947    bld.bld_base.op_actions[TGSI_OPCODE_ENDLOOP].emit = endloop_emit;
1948    bld.bld_base.op_actions[TGSI_OPCODE_ENDSUB].emit = endsub_emit;
1949    bld.bld_base.op_actions[TGSI_OPCODE_IF].emit = if_emit;
1950    bld.bld_base.op_actions[TGSI_OPCODE_KIL].emit = kil_emit;
1951    bld.bld_base.op_actions[TGSI_OPCODE_KILP].emit = kilp_emit;
1952    bld.bld_base.op_actions[TGSI_OPCODE_NRM].emit = nrm_emit;
1953    bld.bld_base.op_actions[TGSI_OPCODE_NRM4].emit = nrm_emit;
1954    bld.bld_base.op_actions[TGSI_OPCODE_RET].emit = ret_emit;
1955    bld.bld_base.op_actions[TGSI_OPCODE_TEX].emit = tex_emit;
1956    bld.bld_base.op_actions[TGSI_OPCODE_TXB].emit = txb_emit;
1957    bld.bld_base.op_actions[TGSI_OPCODE_TXD].emit = txd_emit;
1958    bld.bld_base.op_actions[TGSI_OPCODE_TXL].emit = txl_emit;
1959    bld.bld_base.op_actions[TGSI_OPCODE_TXP].emit = txp_emit;
1960
1961    lp_exec_mask_init(&bld.exec_mask, &bld.bld_base.base);
1962
1963
1964    bld.system_values_array = system_values_array;
1965
1966    lp_build_tgsi_llvm(&bld.bld_base, tokens);
1967
1968    if (0) {
1969       LLVMBasicBlockRef block = LLVMGetInsertBlock(gallivm->builder);
1970       LLVMValueRef function = LLVMGetBasicBlockParent(block);
1971       debug_printf("11111111111111111111111111111 \n");
1972       tgsi_dump(tokens, 0);
1973       lp_debug_dump_value(function);
1974       debug_printf("2222222222222222222222222222 \n");
1975    }
1976
1977    if (0) {
1978       LLVMModuleRef module = LLVMGetGlobalParent(
1979          LLVMGetBasicBlockParent(LLVMGetInsertBlock(gallivm->builder)));
1980       LLVMDumpModule(module);
1981
1982    }
1983 }
1984
1985
1986 /**
1987  * Build up the system values array out of individual values such as
1988  * the instance ID, front-face, primitive ID, etc.  The shader info is
1989  * used to determine which system values are needed and where to put
1990  * them in the system values array.
1991  *
1992  * XXX only instance ID is implemented at this time.
1993  *
1994  * The system values register file is similar to the constants buffer.
1995  * Example declaration:
1996  *    DCL SV[0], INSTANCEID
1997  * Example instruction:
1998  *    MOVE foo, SV[0].xxxx;
1999  *
2000  * \return  LLVM float array (interpreted as float [][4])
2001  */
2002 LLVMValueRef
2003 lp_build_system_values_array(struct gallivm_state *gallivm,
2004                              const struct tgsi_shader_info *info,
2005                              LLVMValueRef instance_id,
2006                              LLVMValueRef facing)
2007 {
2008    LLVMValueRef size = lp_build_const_int32(gallivm, 4 * info->num_system_values);
2009    LLVMTypeRef float_t = LLVMFloatTypeInContext(gallivm->context);
2010    LLVMValueRef array = lp_build_array_alloca(gallivm, float_t,
2011                                               size, "sysvals_array");
2012    unsigned i;
2013
2014    for (i = 0; i < info->num_system_values; i++) {
2015       LLVMValueRef index = lp_build_const_int32(gallivm, i * 4);
2016       LLVMValueRef ptr, value = 0;
2017
2018       switch (info->system_value_semantic_name[i]) {
2019       case TGSI_SEMANTIC_INSTANCEID:
2020          /* convert instance ID from int to float */
2021          value = LLVMBuildSIToFP(gallivm->builder, instance_id, float_t,
2022                                  "sysval_instanceid");
2023          break;
2024       case TGSI_SEMANTIC_FACE:
2025          /* fall-through */
2026       default:
2027          assert(0 && "unexpected semantic in build_system_values_array()");
2028       }
2029
2030       ptr = LLVMBuildGEP(gallivm->builder, array, &index, 1, "");
2031       LLVMBuildStore(gallivm->builder, value, ptr);
2032    }
2033
2034    return array;
2035 }