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
 708    assert(res);
 709
 710    if (stype == TGSI_TYPE_UNSIGNED) {
 711       res = LLVMBuildBitCast(builder, res, bld_base->uint_bld.vec_type, "");
 712    } else if (stype == TGSI_TYPE_SIGNED) {
 713       res = LLVMBuildBitCast(builder, res, bld_base->int_bld.vec_type, "");
 714    }
 715
 716    return res;
 717 }
 718
 719 static LLVMValueRef
 720 emit_fetch_temporary(
 721    struct lp_build_tgsi_context * bld_base,
 722    const struct tgsi_full_src_register * reg,
 723    enum tgsi_opcode_type stype,
 724    unsigned swizzle)
 725 {
 726    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
 727    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
 728    LLVMBuilderRef builder = gallivm->builder;
 729    struct lp_build_context *uint_bld = &bld_base->uint_bld;
 730    LLVMValueRef indirect_index = NULL;
 731    LLVMValueRef res;
 732
 733    if (reg->Register.Indirect) {
 734       indirect_index = get_indirect_index(bld,
 735                                           reg->Register.File,
 736                                           reg->Register.Index,
 737                                           &reg->Indirect);
 738    }
 739
 740    if (reg->Register.Indirect) {
 741       LLVMValueRef swizzle_vec =
 742          lp_build_const_int_vec(bld->bld_base.base.gallivm, uint_bld->type, swizzle);
 743       LLVMValueRef length_vec =
 744          lp_build_const_int_vec(bld->bld_base.base.gallivm, uint_bld->type,
 745                                 bld->bld_base.base.type.length);
 746       LLVMValueRef index_vec;  /* index into the const buffer */
 747       LLVMValueRef temps_array;
 748       LLVMTypeRef float4_ptr_type;
 749
 750       /* index_vec = (indirect_index * 4 + swizzle) * length */
 751       index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
 752       index_vec = lp_build_add(uint_bld, index_vec, swizzle_vec);
 753       index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
 754
 755       /* cast temps_array pointer to float* */
 756       float4_ptr_type = LLVMPointerType(LLVMFloatTypeInContext(bld->bld_base.base.gallivm->context), 0);
 757       temps_array = LLVMBuildBitCast(builder, bld->temps_array,
 758                                      float4_ptr_type, "");
 759
 760       /* Gather values from the temporary register array */
 761       res = build_gather(&bld_base->base, temps_array, index_vec);
 762    }
 763    else {
 764       LLVMValueRef temp_ptr;
 765       if (stype != TGSI_TYPE_FLOAT && stype != TGSI_TYPE_UNTYPED) {
 766          LLVMTypeRef itype = LLVMPointerType(LLVMVectorType(LLVMInt32TypeInContext(gallivm->context), 4), 0);
 767          LLVMValueRef tint_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index,
 768                                                      swizzle);
 769          temp_ptr = LLVMBuildBitCast(builder, tint_ptr, itype, "");
 770       } else
 771          temp_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index, swizzle);
 772       res = LLVMBuildLoad(builder, temp_ptr, "");
 773       if (!res)
 774          return bld->bld_base.base.undef;
 775    }
 776
 777    return res;
 778 }
 779
 780 static LLVMValueRef
 781 emit_fetch_system_value(
 782    struct lp_build_tgsi_context * bld_base,
 783    const struct tgsi_full_src_register * reg,
 784    enum tgsi_opcode_type stype,
 785    unsigned swizzle)
 786 {
 787    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
 788    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
 789    LLVMBuilderRef builder = gallivm->builder;
 790    LLVMValueRef index;  /* index into the system value array */
 791    LLVMValueRef scalar, scalar_ptr;
 792
 793    assert(!reg->Register.Indirect);
 794
 795    index = lp_build_const_int32(gallivm, reg->Register.Index * 4 + swizzle);
 796
 797    scalar_ptr = LLVMBuildGEP(builder, bld->system_values_array, &index, 1, "");
 798    scalar = LLVMBuildLoad(builder, scalar_ptr, "");
 799
 800    return lp_build_broadcast_scalar(&bld->bld_base.base, scalar);
 801 }
 802
 803 /**
 804  * Register fetch with derivatives.
 805  */
 806 static void
 807 emit_fetch_deriv(
 808    struct lp_build_tgsi_soa_context *bld,
 809    LLVMValueRef src,
 810    LLVMValueRef *res,
 811    LLVMValueRef *ddx,
 812    LLVMValueRef *ddy)
 813 {
 814    if(res)
 815       *res = src;
 816
 817    /* TODO: use interpolation coeffs for inputs */
 818
 819    if(ddx)
 820       *ddx = lp_build_ddx(&bld->bld_base.base, src);
 821
 822    if(ddy)
 823       *ddy = lp_build_ddy(&bld->bld_base.base, src);
 824 }
 825
 826
 827 /**
 828  * Predicate.
 829  */
 830 static void
 831 emit_fetch_predicate(
 832    struct lp_build_tgsi_soa_context *bld,
 833    const struct tgsi_full_instruction *inst,
 834    LLVMValueRef *pred)
 835 {
 836    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
 837    unsigned index;
 838    unsigned char swizzles[4];
 839    LLVMValueRef unswizzled[4] = {NULL, NULL, NULL, NULL};
 840    LLVMValueRef value;
 841    unsigned chan;
 842
 843    if (!inst->Instruction.Predicate) {
 844       TGSI_FOR_EACH_CHANNEL( chan ) {
 845          pred[chan] = NULL;
 846       }
 847       return;
 848    }
 849
 850    swizzles[0] = inst->Predicate.SwizzleX;
 851    swizzles[1] = inst->Predicate.SwizzleY;
 852    swizzles[2] = inst->Predicate.SwizzleZ;
 853    swizzles[3] = inst->Predicate.SwizzleW;
 854
 855    index = inst->Predicate.Index;
 856    assert(index < LP_MAX_TGSI_PREDS);
 857
 858    TGSI_FOR_EACH_CHANNEL( chan ) {
 859       unsigned swizzle = swizzles[chan];
 860
 861       /*
 862        * Only fetch the predicate register channels that are actually listed
 863        * in the swizzles
 864        */
 865       if (!unswizzled[swizzle]) {
 866          value = LLVMBuildLoad(builder,
 867                                bld->preds[index][swizzle], "");
 868
 869          /*
 870           * Convert the value to an integer mask.
 871           *
 872           * TODO: Short-circuit this comparison -- a D3D setp_xx instructions
 873           * is needlessly causing two comparisons due to storing the intermediate
 874           * result as float vector instead of an integer mask vector.
 875           */
 876          value = lp_build_compare(bld->bld_base.base.gallivm,
 877                                   bld->bld_base.base.type,
 878                                   PIPE_FUNC_NOTEQUAL,
 879                                   value,
 880                                   bld->bld_base.base.zero);
 881          if (inst->Predicate.Negate) {
 882             value = LLVMBuildNot(builder, value, "");
 883          }
 884
 885          unswizzled[swizzle] = value;
 886       } else {
 887          value = unswizzled[swizzle];
 888       }
 889
 890       pred[chan] = value;
 891    }
 892 }
 893
 894 /**
 895  * Register store.
 896  */
 897 static void
 898 emit_store_chan(
 899    struct lp_build_tgsi_context *bld_base,
 900    const struct tgsi_full_instruction *inst,
 901    unsigned index,
 902    unsigned chan_index,
 903    LLVMValueRef pred,
 904    LLVMValueRef value)
 905 {
 906    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
 907    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
 908    LLVMBuilderRef builder = gallivm->builder;
 909    const struct tgsi_full_dst_register *reg = &inst->Dst[index];
 910    struct lp_build_context *uint_bld = &bld_base->uint_bld;
 911    LLVMValueRef indirect_index = NULL;
 912    struct lp_build_context *bld_store;
 913    enum tgsi_opcode_type dtype = tgsi_opcode_infer_dst_type(inst->Instruction.Opcode);
 914
 915    switch (dtype) {
 916    default:
 917    case TGSI_TYPE_FLOAT:
 918    case TGSI_TYPE_UNTYPED:
 919       bld_store = &bld_base->base;
 920       break;
 921    case TGSI_TYPE_UNSIGNED:
 922       bld_store = &bld_base->uint_bld;
 923       break;
 924    case TGSI_TYPE_SIGNED:
 925       bld_store = &bld_base->int_bld;
 926       break;
 927    case TGSI_TYPE_DOUBLE:
 928    case TGSI_TYPE_VOID:
 929       assert(0);
 930       bld_store = NULL;
 931       break;
 932    }
 933
 934    switch( inst->Instruction.Saturate ) {
 935    case TGSI_SAT_NONE:
 936       break;
 937
 938    case TGSI_SAT_ZERO_ONE:
 939       value = lp_build_max(&bld->bld_base.base, value, bld->bld_base.base.zero);
 940       value = lp_build_min(&bld->bld_base.base, value, bld->bld_base.base.one);
 941       break;
 942
 943    case TGSI_SAT_MINUS_PLUS_ONE:
 944       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));
 945       value = lp_build_min(&bld->bld_base.base, value, bld->bld_base.base.one);
 946       break;
 947
 948    default:
 949       assert(0);
 950    }
 951
 952    if (reg->Register.Indirect) {
 953       indirect_index = get_indirect_index(bld,
 954                                           reg->Register.File,
 955                                           reg->Register.Index,
 956                                           &reg->Indirect);
 957    } else {
 958       assert(reg->Register.Index <=
 959                              bld->bld_base.info->file_max[reg->Register.File]);
 960    }
 961
 962    switch( reg->Register.File ) {
 963    case TGSI_FILE_OUTPUT:
 964       if (reg->Register.Indirect) {
 965          LLVMValueRef chan_vec =
 966             lp_build_const_int_vec(gallivm, uint_bld->type, chan_index);
 967          LLVMValueRef length_vec =
 968             lp_build_const_int_vec(gallivm, uint_bld->type, bld->bld_base.base.type.length);
 969          LLVMValueRef index_vec;  /* indexes into the temp registers */
 970          LLVMValueRef outputs_array;
 971          LLVMValueRef pixel_offsets;
 972          LLVMTypeRef float_ptr_type;
 973          int i;
 974
 975          /* build pixel offset vector: {0, 1, 2, 3, ...} */
 976          pixel_offsets = uint_bld->undef;
 977          for (i = 0; i < bld->bld_base.base.type.length; i++) {
 978             LLVMValueRef ii = lp_build_const_int32(gallivm, i);
 979             pixel_offsets = LLVMBuildInsertElement(builder, pixel_offsets,
 980                                                    ii, ii, "");
 981          }
 982
 983          /* index_vec = (indirect_index * 4 + chan_index) * length + offsets */
 984          index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
 985          index_vec = lp_build_add(uint_bld, index_vec, chan_vec);
 986          index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
 987          index_vec = lp_build_add(uint_bld, index_vec, pixel_offsets);
 988
 989          float_ptr_type =
 990             LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
 991          outputs_array = LLVMBuildBitCast(builder, bld->outputs_array,
 992                                           float_ptr_type, "");
 993
 994          /* Scatter store values into temp registers */
 995          emit_mask_scatter(bld, outputs_array, index_vec, value,
 996                            &bld->exec_mask, pred);
 997       }
 998       else {
 999          LLVMValueRef out_ptr = lp_get_output_ptr(bld, reg->Register.Index,
1000                                                chan_index);
1001          lp_exec_mask_store(&bld->exec_mask, bld_store, pred, value, out_ptr);
1002       }
1003       break;
1004
1005    case TGSI_FILE_TEMPORARY:
1006       if (reg->Register.Indirect) {
1007          LLVMValueRef chan_vec =
1008             lp_build_const_int_vec(gallivm, uint_bld->type, chan_index);
1009          LLVMValueRef length_vec =
1010             lp_build_const_int_vec(gallivm, uint_bld->type,
1011                                    bld->bld_base.base.type.length);
1012          LLVMValueRef index_vec;  /* indexes into the temp registers */
1013          LLVMValueRef temps_array;
1014          LLVMValueRef pixel_offsets;
1015          LLVMTypeRef float_ptr_type;
1016          int i;
1017
1018          /* build pixel offset vector: {0, 1, 2, 3, ...} */
1019          pixel_offsets = uint_bld->undef;
1020          for (i = 0; i < bld->bld_base.base.type.length; i++) {
1021             LLVMValueRef ii = lp_build_const_int32(gallivm, i);
1022             pixel_offsets = LLVMBuildInsertElement(builder, pixel_offsets,
1023                                                    ii, ii, "");
1024          }
1025
1026          /* index_vec = (indirect_index * 4 + chan_index) * length + offsets */
1027          index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
1028          index_vec = lp_build_add(uint_bld, index_vec, chan_vec);
1029          index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
1030          index_vec = lp_build_add(uint_bld, index_vec, pixel_offsets);
1031
1032          float_ptr_type =
1033             LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
1034          temps_array = LLVMBuildBitCast(builder, bld->temps_array,
1035                                         float_ptr_type, "");
1036
1037          /* Scatter store values into temp registers */
1038          emit_mask_scatter(bld, temps_array, index_vec, value,
1039                            &bld->exec_mask, pred);
1040       }
1041       else {
1042          LLVMValueRef temp_ptr;
1043
1044          switch (dtype) {
1045          case TGSI_TYPE_UNSIGNED:
1046          case TGSI_TYPE_SIGNED: {
1047             LLVMTypeRef itype = LLVMVectorType(LLVMInt32TypeInContext(gallivm->context), 4);
1048             LLVMTypeRef ivtype = LLVMPointerType(itype, 0);
1049             LLVMValueRef tint_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index,
1050                                                         chan_index);
1051             LLVMValueRef temp_value_ptr;
1052
1053             temp_ptr = LLVMBuildBitCast(builder, tint_ptr, ivtype, "");
1054             temp_value_ptr = LLVMBuildBitCast(builder, value, itype, "");
1055             value = temp_value_ptr;
1056             break;
1057          }
1058          default:
1059          case TGSI_TYPE_FLOAT:
1060          case TGSI_TYPE_UNTYPED:
1061             temp_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index,
1062                                            chan_index);
1063             break;
1064          }
1065
1066          lp_exec_mask_store(&bld->exec_mask, bld_store, pred, value, temp_ptr);
1067       }
1068       break;
1069
1070    case TGSI_FILE_ADDRESS:
1071       assert(dtype == TGSI_TYPE_SIGNED);
1072       assert(LLVMTypeOf(value) == bld_base->base.int_vec_type);
1073       lp_exec_mask_store(&bld->exec_mask, bld_store, pred, value,
1074                          bld->addr[reg->Register.Index][chan_index]);
1075       break;
1076
1077    case TGSI_FILE_PREDICATE:
1078       lp_exec_mask_store(&bld->exec_mask, bld_store, pred, value,
1079                          bld->preds[reg->Register.Index][chan_index]);
1080       break;
1081
1082    default:
1083       assert( 0 );
1084    }
1085 }
1086
1087 static void
1088 emit_store(
1089    struct lp_build_tgsi_context * bld_base,
1090    const struct tgsi_full_instruction * inst,
1091    const struct tgsi_opcode_info * info,
1092    LLVMValueRef dst[4])
1093
1094 {
1095    unsigned chan_index;
1096    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1097
1098    if(info->num_dst) {
1099       LLVMValueRef pred[TGSI_NUM_CHANNELS];
1100
1101       emit_fetch_predicate( bld, inst, pred );
1102
1103       TGSI_FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1104          emit_store_chan(bld_base, inst, 0, chan_index, pred[chan_index], dst[chan_index]);
1105       }
1106    }
1107 }
1108
1109 /**
1110  * High-level instruction translators.
1111  */
1112
1113 static void
1114 emit_tex( struct lp_build_tgsi_soa_context *bld,
1115           const struct tgsi_full_instruction *inst,
1116           enum lp_build_tex_modifier modifier,
1117           LLVMValueRef *texel)
1118 {
1119    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
1120    unsigned unit;
1121    LLVMValueRef lod_bias, explicit_lod;
1122    LLVMValueRef oow = NULL;
1123    LLVMValueRef coords[3];
1124    LLVMValueRef ddx[3];
1125    LLVMValueRef ddy[3];
1126    unsigned num_coords;
1127    unsigned i;
1128
1129    if (!bld->sampler) {
1130       _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
1131       for (i = 0; i < 4; i++) {
1132          texel[i] = bld->bld_base.base.undef;
1133       }
1134       return;
1135    }
1136
1137    switch (inst->Texture.Texture) {
1138    case TGSI_TEXTURE_1D:
1139       num_coords = 1;
1140       break;
1141    case TGSI_TEXTURE_1D_ARRAY:
1142    case TGSI_TEXTURE_2D:
1143    case TGSI_TEXTURE_RECT:
1144       num_coords = 2;
1145       break;
1146    case TGSI_TEXTURE_SHADOW1D:
1147    case TGSI_TEXTURE_SHADOW1D_ARRAY:
1148    case TGSI_TEXTURE_SHADOW2D:
1149    case TGSI_TEXTURE_SHADOWRECT:
1150    case TGSI_TEXTURE_2D_ARRAY:
1151    case TGSI_TEXTURE_3D:
1152    case TGSI_TEXTURE_CUBE:
1153       num_coords = 3;
1154       break;
1155    case TGSI_TEXTURE_SHADOW2D_ARRAY:
1156       num_coords = 4;
1157       break;
1158    default:
1159       assert(0);
1160       return;
1161    }
1162
1163    if (modifier == LP_BLD_TEX_MODIFIER_LOD_BIAS) {
1164       lod_bias = lp_build_emit_fetch( &bld->bld_base, inst, 0, 3 );
1165       explicit_lod = NULL;
1166    }
1167    else if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_LOD) {
1168       lod_bias = NULL;
1169       explicit_lod = lp_build_emit_fetch( &bld->bld_base, inst, 0, 3 );
1170    }
1171    else {
1172       lod_bias = NULL;
1173       explicit_lod = NULL;
1174    }
1175
1176    if (modifier == LP_BLD_TEX_MODIFIER_PROJECTED) {
1177       oow = lp_build_emit_fetch( &bld->bld_base, inst, 0, 3 );
1178       oow = lp_build_rcp(&bld->bld_base.base, oow);
1179    }
1180
1181    for (i = 0; i < num_coords; i++) {
1182       coords[i] = lp_build_emit_fetch( &bld->bld_base, inst, 0, i );
1183       if (modifier == LP_BLD_TEX_MODIFIER_PROJECTED)
1184          coords[i] = lp_build_mul(&bld->bld_base.base, coords[i], oow);
1185    }
1186    for (i = num_coords; i < 3; i++) {
1187       coords[i] = bld->bld_base.base.undef;
1188    }
1189
1190    if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV) {
1191       LLVMValueRef index0 = lp_build_const_int32(bld->bld_base.base.gallivm, 0);
1192       for (i = 0; i < num_coords; i++) {
1193          LLVMValueRef src1 = lp_build_emit_fetch( &bld->bld_base, inst, 1, i );
1194          LLVMValueRef src2 = lp_build_emit_fetch( &bld->bld_base, inst, 2, i );
1195          ddx[i] = LLVMBuildExtractElement(builder, src1, index0, "");
1196          ddy[i] = LLVMBuildExtractElement(builder, src2, index0, "");
1197       }
1198       unit = inst->Src[3].Register.Index;
1199    }  else {
1200       for (i = 0; i < num_coords; i++) {
1201          ddx[i] = lp_build_scalar_ddx( &bld->bld_base.base, coords[i] );
1202          ddy[i] = lp_build_scalar_ddy( &bld->bld_base.base, coords[i] );
1203       }
1204       unit = inst->Src[1].Register.Index;
1205    }
1206    for (i = num_coords; i < 3; i++) {
1207       ddx[i] = LLVMGetUndef(bld->bld_base.base.elem_type);
1208       ddy[i] = LLVMGetUndef(bld->bld_base.base.elem_type);
1209    }
1210
1211    bld->sampler->emit_fetch_texel(bld->sampler,
1212                                   bld->bld_base.base.gallivm,
1213                                   bld->bld_base.base.type,
1214                                   unit, num_coords, coords,
1215                                   ddx, ddy,
1216                                   lod_bias, explicit_lod,
1217                                   texel);
1218 }
1219
1220 static void
1221 emit_txq( struct lp_build_tgsi_soa_context *bld,
1222           const struct tgsi_full_instruction *inst,
1223           LLVMValueRef *sizes_out)
1224 {
1225    LLVMValueRef explicit_lod;
1226    unsigned num_coords, has_lod;
1227    unsigned i;
1228
1229    switch (inst->Texture.Texture) {
1230    case TGSI_TEXTURE_1D:
1231    case TGSI_TEXTURE_SHADOW1D:
1232    case TGSI_TEXTURE_SHADOW2D:
1233    case TGSI_TEXTURE_SHADOWCUBE:
1234       num_coords = 1;
1235       has_lod = 1;
1236       break;
1237    case TGSI_TEXTURE_2D:
1238    case TGSI_TEXTURE_CUBE:
1239    case TGSI_TEXTURE_1D_ARRAY:
1240    case TGSI_TEXTURE_SHADOW1D_ARRAY:
1241       num_coords = 2;
1242       has_lod = 1;
1243       break;
1244    case TGSI_TEXTURE_3D:
1245 // case TGSI_TEXTURE_CUBE_ARRAY:
1246 // case TGSI_TEXTURE_SHADOWCUBE_ARRAY:
1247    case TGSI_TEXTURE_2D_ARRAY:
1248    case TGSI_TEXTURE_SHADOW2D_ARRAY:
1249       num_coords = 3;
1250       has_lod = 1;
1251       break;
1252
1253    case TGSI_TEXTURE_BUFFER:
1254       num_coords = 1;
1255       has_lod = 0;
1256       break;
1257
1258    case TGSI_TEXTURE_RECT:
1259    case TGSI_TEXTURE_SHADOWRECT:
1260 // case TGSI_TEXTURE_2D_MS:
1261       num_coords = 2;
1262       has_lod = 0;
1263       break;
1264
1265 // case TGSI_TEXTURE_2D_MS_ARRAY:
1266 //    num_coords = 3;
1267 //    has_lod = 0;
1268 //    break;
1269
1270    default:
1271       assert(0);
1272       return;
1273    }
1274
1275    if (!bld->sampler) {
1276       _debug_printf("warning: found texture query instruction but no sampler generator supplied\n");
1277       for (i = 0; i < num_coords; i++)
1278          sizes_out[i] = bld->bld_base.base.undef;
1279       return;
1280    }
1281
1282    if (has_lod)
1283       explicit_lod = lp_build_emit_fetch( &bld->bld_base, inst, 0, 2 );
1284    else
1285       explicit_lod = NULL;
1286
1287    bld->sampler->emit_size_query(bld->sampler,
1288                                  bld->bld_base.base.gallivm,
1289                                  inst->Src[1].Register.Index,
1290                                  explicit_lod,
1291                                  sizes_out);
1292 }
1293
1294 static boolean
1295 near_end_of_shader(struct lp_build_tgsi_soa_context *bld,
1296                    int pc)
1297 {
1298    int i;
1299
1300    for (i = 0; i < 5; i++) {
1301       unsigned opcode;
1302
1303       if (pc + i >= bld->bld_base.info->num_instructions)
1304          return TRUE;
1305
1306       opcode = bld->bld_base.instructions[pc + i].Instruction.Opcode;
1307
1308       if (opcode == TGSI_OPCODE_END)
1309          return TRUE;
1310
1311       if (opcode == TGSI_OPCODE_TEX ||
1312           opcode == TGSI_OPCODE_TXP ||
1313           opcode == TGSI_OPCODE_TXD ||
1314           opcode == TGSI_OPCODE_TXB ||
1315           opcode == TGSI_OPCODE_TXL ||
1316           opcode == TGSI_OPCODE_TXF ||
1317           opcode == TGSI_OPCODE_TXQ ||
1318           opcode == TGSI_OPCODE_CAL ||
1319           opcode == TGSI_OPCODE_CALLNZ ||
1320           opcode == TGSI_OPCODE_IF ||
1321           opcode == TGSI_OPCODE_IFC ||
1322           opcode == TGSI_OPCODE_BGNLOOP ||
1323           opcode == TGSI_OPCODE_SWITCH)
1324          return FALSE;
1325    }
1326
1327    return TRUE;
1328 }
1329
1330
1331
1332 /**
1333  * Kill fragment if any of the src register values are negative.
1334  */
1335 static void
1336 emit_kil(
1337    struct lp_build_tgsi_soa_context *bld,
1338    const struct tgsi_full_instruction *inst,
1339    int pc)
1340 {
1341    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
1342    const struct tgsi_full_src_register *reg = &inst->Src[0];
1343    LLVMValueRef terms[TGSI_NUM_CHANNELS];
1344    LLVMValueRef mask;
1345    unsigned chan_index;
1346
1347    memset(&terms, 0, sizeof terms);
1348
1349    TGSI_FOR_EACH_CHANNEL( chan_index ) {
1350       unsigned swizzle;
1351
1352       /* Unswizzle channel */
1353       swizzle = tgsi_util_get_full_src_register_swizzle( reg, chan_index );
1354
1355       /* Check if the component has not been already tested. */
1356       assert(swizzle < TGSI_NUM_CHANNELS);
1357       if( !terms[swizzle] )
1358          /* TODO: change the comparison operator instead of setting the sign */
1359          terms[swizzle] =  lp_build_emit_fetch(&bld->bld_base, inst, 0, chan_index );
1360    }
1361
1362    mask = NULL;
1363    TGSI_FOR_EACH_CHANNEL( chan_index ) {
1364       if(terms[chan_index]) {
1365          LLVMValueRef chan_mask;
1366
1367          /*
1368           * If term < 0 then mask = 0 else mask = ~0.
1369           */
1370          chan_mask = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_GEQUAL, terms[chan_index], bld->bld_base.base.zero);
1371
1372          if(mask)
1373             mask = LLVMBuildAnd(builder, mask, chan_mask, "");
1374          else
1375             mask = chan_mask;
1376       }
1377    }
1378
1379    if(mask) {
1380       lp_build_mask_update(bld->mask, mask);
1381
1382       if (!near_end_of_shader(bld, pc))
1383          lp_build_mask_check(bld->mask);
1384    }
1385 }
1386
1387
1388 /**
1389  * Predicated fragment kill.
1390  * XXX Actually, we do an unconditional kill (as in tgsi_exec.c).
1391  * The only predication is the execution mask which will apply if
1392  * we're inside a loop or conditional.
1393  */
1394 static void
1395 emit_kilp(struct lp_build_tgsi_soa_context *bld,
1396           int pc)
1397 {
1398    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
1399    LLVMValueRef mask;
1400
1401    /* For those channels which are "alive", disable fragment shader
1402     * execution.
1403     */
1404    if (bld->exec_mask.has_mask) {
1405       mask = LLVMBuildNot(builder, bld->exec_mask.exec_mask, "kilp");
1406    }
1407    else {
1408       LLVMValueRef zero = LLVMConstNull(bld->bld_base.base.int_vec_type);
1409       mask = zero;
1410    }
1411
1412    lp_build_mask_update(bld->mask, mask);
1413
1414    if (!near_end_of_shader(bld, pc))
1415       lp_build_mask_check(bld->mask);
1416 }
1417
1418
1419 /**
1420  * Emit code which will dump the value of all the temporary registers
1421  * to stdout.
1422  */
1423 static void
1424 emit_dump_temps(struct lp_build_tgsi_soa_context *bld)
1425 {
1426    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
1427    LLVMBuilderRef builder = gallivm->builder;
1428    LLVMValueRef temp_ptr;
1429    LLVMValueRef i0 = lp_build_const_int32(gallivm, 0);
1430    LLVMValueRef i1 = lp_build_const_int32(gallivm, 1);
1431    LLVMValueRef i2 = lp_build_const_int32(gallivm, 2);
1432    LLVMValueRef i3 = lp_build_const_int32(gallivm, 3);
1433    int index;
1434    int n = bld->bld_base.info->file_max[TGSI_FILE_TEMPORARY];
1435
1436    for (index = 0; index < n; index++) {
1437       LLVMValueRef idx = lp_build_const_int32(gallivm, index);
1438       LLVMValueRef v[4][4], res;
1439       int chan;
1440
1441       lp_build_printf(gallivm, "TEMP[%d]:\n", idx);
1442
1443       for (chan = 0; chan < 4; chan++) {
1444          temp_ptr = lp_get_temp_ptr_soa(bld, index, chan);
1445          res = LLVMBuildLoad(builder, temp_ptr, "");
1446          v[chan][0] = LLVMBuildExtractElement(builder, res, i0, "");
1447          v[chan][1] = LLVMBuildExtractElement(builder, res, i1, "");
1448          v[chan][2] = LLVMBuildExtractElement(builder, res, i2, "");
1449          v[chan][3] = LLVMBuildExtractElement(builder, res, i3, "");
1450       }
1451
1452       lp_build_printf(gallivm, "  X: %f %f %f %f\n",
1453                       v[0][0], v[0][1], v[0][2], v[0][3]);
1454       lp_build_printf(gallivm, "  Y: %f %f %f %f\n",
1455                       v[1][0], v[1][1], v[1][2], v[1][3]);
1456       lp_build_printf(gallivm, "  Z: %f %f %f %f\n",
1457                       v[2][0], v[2][1], v[2][2], v[2][3]);
1458       lp_build_printf(gallivm, "  W: %f %f %f %f\n",
1459                       v[3][0], v[3][1], v[3][2], v[3][3]);
1460    }
1461 }
1462
1463
1464
1465 void
1466 lp_emit_declaration_soa(
1467    struct lp_build_tgsi_context *bld_base,
1468    const struct tgsi_full_declaration *decl)
1469 {
1470    struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
1471    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
1472    LLVMTypeRef vec_type = bld->bld_base.base.vec_type;
1473    const unsigned first = decl->Range.First;
1474    const unsigned last = decl->Range.Last;
1475    unsigned idx, i;
1476
1477    for (idx = first; idx <= last; ++idx) {
1478       assert(last <= bld->bld_base.info->file_max[decl->Declaration.File]);
1479       switch (decl->Declaration.File) {
1480       case TGSI_FILE_TEMPORARY:
1481          assert(idx < LP_MAX_TGSI_TEMPS);
1482          if (!(bld->indirect_files & (1 << TGSI_FILE_TEMPORARY))) {
1483             for (i = 0; i < TGSI_NUM_CHANNELS; i++)
1484                bld->temps[idx][i] = lp_build_alloca(gallivm, vec_type, "temp");
1485          }
1486          break;
1487
1488       case TGSI_FILE_OUTPUT:
1489          if (!(bld->indirect_files & (1 << TGSI_FILE_OUTPUT))) {
1490             for (i = 0; i < TGSI_NUM_CHANNELS; i++)
1491                bld->outputs[idx][i] = lp_build_alloca(gallivm,
1492                                                       vec_type, "output");
1493          }
1494          break;
1495
1496       case TGSI_FILE_ADDRESS:
1497          /* ADDR registers are the only allocated with an integer LLVM IR type,
1498           * as they are guaranteed to always have integers.
1499           * XXX: Not sure if this exception is worthwhile (or the whole idea of
1500           * an ADDR register for that matter).
1501           */
1502          assert(idx < LP_MAX_TGSI_ADDRS);
1503          for (i = 0; i < TGSI_NUM_CHANNELS; i++)
1504             bld->addr[idx][i] = lp_build_alloca(gallivm, bld_base->base.int_vec_type, "addr");
1505          break;
1506
1507       case TGSI_FILE_PREDICATE:
1508          assert(idx < LP_MAX_TGSI_PREDS);
1509          for (i = 0; i < TGSI_NUM_CHANNELS; i++)
1510             bld->preds[idx][i] = lp_build_alloca(gallivm, vec_type,
1511                                                  "predicate");
1512          break;
1513
1514       default:
1515          /* don't need to declare other vars */
1516          break;
1517       }
1518    }
1519 }
1520
1521
1522 void lp_emit_immediate_soa(
1523    struct lp_build_tgsi_context *bld_base,
1524    const struct tgsi_full_immediate *imm)
1525 {
1526    struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
1527    struct gallivm_state * gallivm = bld_base->base.gallivm;
1528
1529    /* simply copy the immediate values into the next immediates[] slot */
1530    unsigned i;
1531    const uint size = imm->Immediate.NrTokens - 1;
1532    assert(size <= 4);
1533    assert(bld->num_immediates < LP_MAX_TGSI_IMMEDIATES);
1534    switch (imm->Immediate.DataType) {
1535    case TGSI_IMM_FLOAT32:
1536       for( i = 0; i < size; ++i )
1537          bld->immediates[bld->num_immediates][i] =
1538             lp_build_const_vec(gallivm, bld_base->base.type, imm->u[i].Float);
1539
1540       break;
1541    case TGSI_IMM_UINT32:
1542       for( i = 0; i < size; ++i ) {
1543          LLVMValueRef tmp = lp_build_const_vec(gallivm, bld_base->uint_bld.type, imm->u[i].Uint);
1544          bld->immediates[bld->num_immediates][i] =
1545             LLVMConstBitCast(tmp, bld_base->base.vec_type);
1546       }
1547
1548       break;
1549    case TGSI_IMM_INT32:
1550       for( i = 0; i < size; ++i ) {
1551          LLVMValueRef tmp = lp_build_const_vec(gallivm, bld_base->int_bld.type, imm->u[i].Int);
1552          bld->immediates[bld->num_immediates][i] =
1553             LLVMConstBitCast(tmp, bld_base->base.vec_type);
1554       }
1555
1556       break;
1557    }
1558    for( i = size; i < 4; ++i )
1559       bld->immediates[bld->num_immediates][i] = bld_base->base.undef;
1560
1561    bld->num_immediates++;
1562 }
1563
1564 static void
1565 ddx_emit(
1566    const struct lp_build_tgsi_action * action,
1567    struct lp_build_tgsi_context * bld_base,
1568    struct lp_build_emit_data * emit_data)
1569 {
1570    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1571
1572    emit_fetch_deriv(bld, emit_data->args[0], NULL,
1573                     &emit_data->output[emit_data->chan], NULL);
1574 }
1575
1576 static void
1577 ddy_emit(
1578    const struct lp_build_tgsi_action * action,
1579    struct lp_build_tgsi_context * bld_base,
1580    struct lp_build_emit_data * emit_data)
1581 {
1582    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1583
1584    emit_fetch_deriv(bld, emit_data->args[0], NULL, NULL,
1585                     &emit_data->output[emit_data->chan]);
1586 }
1587
1588 static void
1589 kilp_emit(
1590    const struct lp_build_tgsi_action * action,
1591    struct lp_build_tgsi_context * bld_base,
1592    struct lp_build_emit_data * emit_data)
1593 {
1594    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1595
1596    emit_kilp(bld, bld_base->pc - 1);
1597 }
1598
1599 static void
1600 kil_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    emit_kil(bld, emit_data->inst, bld_base->pc - 1);
1608 }
1609
1610 static void
1611 tex_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    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_NONE, emit_data->output);
1619 }
1620
1621 static void
1622 txb_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    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1628
1629    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_LOD_BIAS,
1630             emit_data->output);
1631 }
1632
1633 static void
1634 txd_emit(
1635    const struct lp_build_tgsi_action * action,
1636    struct lp_build_tgsi_context * bld_base,
1637    struct lp_build_emit_data * emit_data)
1638 {
1639    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1640
1641    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV,
1642             emit_data->output);
1643 }
1644
1645 static void
1646 txl_emit(
1647    const struct lp_build_tgsi_action * action,
1648    struct lp_build_tgsi_context * bld_base,
1649    struct lp_build_emit_data * emit_data)
1650 {
1651    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1652
1653    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD,
1654             emit_data->output);
1655 }
1656
1657 static void
1658 txp_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    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_PROJECTED,
1666             emit_data->output);
1667 }
1668
1669 static void
1670 txq_emit(
1671    const struct lp_build_tgsi_action * action,
1672    struct lp_build_tgsi_context * bld_base,
1673    struct lp_build_emit_data * emit_data)
1674 {
1675    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1676
1677    emit_txq(bld, emit_data->inst, emit_data->output);
1678 }
1679
1680 static void
1681 cal_emit(
1682    const struct lp_build_tgsi_action * action,
1683    struct lp_build_tgsi_context * bld_base,
1684    struct lp_build_emit_data * emit_data)
1685 {
1686    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1687
1688    lp_exec_mask_call(&bld->exec_mask, emit_data->inst->Label.Label,
1689                      &bld_base->pc);
1690 }
1691
1692 static void
1693 ret_emit(
1694    const struct lp_build_tgsi_action * action,
1695    struct lp_build_tgsi_context * bld_base,
1696    struct lp_build_emit_data * emit_data)
1697 {
1698    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1699
1700    lp_exec_mask_ret(&bld->exec_mask, &bld_base->pc);
1701 }
1702
1703 static void
1704 brk_emit(
1705    const struct lp_build_tgsi_action * action,
1706    struct lp_build_tgsi_context * bld_base,
1707    struct lp_build_emit_data * emit_data)
1708 {
1709    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1710
1711    lp_exec_break(&bld->exec_mask);
1712 }
1713
1714 static void
1715 if_emit(
1716    const struct lp_build_tgsi_action * action,
1717    struct lp_build_tgsi_context * bld_base,
1718    struct lp_build_emit_data * emit_data)
1719 {
1720    LLVMValueRef tmp;
1721    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1722
1723    tmp = lp_build_cmp(&bld_base->base, PIPE_FUNC_NOTEQUAL,
1724                       emit_data->args[0], bld->bld_base.base.zero);
1725    lp_exec_mask_cond_push(&bld->exec_mask, tmp);
1726 }
1727
1728 static void
1729 bgnloop_emit(
1730    const struct lp_build_tgsi_action * action,
1731    struct lp_build_tgsi_context * bld_base,
1732    struct lp_build_emit_data * emit_data)
1733 {
1734    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1735
1736    lp_exec_bgnloop(&bld->exec_mask);
1737 }
1738
1739 static void
1740 bgnsub_emit(
1741    const struct lp_build_tgsi_action * action,
1742    struct lp_build_tgsi_context * bld_base,
1743    struct lp_build_emit_data * emit_data)
1744 {
1745    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1746
1747    lp_exec_mask_bgnsub(&bld->exec_mask);
1748 }
1749
1750 static void
1751 else_emit(
1752    const struct lp_build_tgsi_action * action,
1753    struct lp_build_tgsi_context * bld_base,
1754    struct lp_build_emit_data * emit_data)
1755 {
1756    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1757
1758    lp_exec_mask_cond_invert(&bld->exec_mask);
1759 }
1760
1761 static void
1762 endif_emit(
1763    const struct lp_build_tgsi_action * action,
1764    struct lp_build_tgsi_context * bld_base,
1765    struct lp_build_emit_data * emit_data)
1766 {
1767    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1768
1769    lp_exec_mask_cond_pop(&bld->exec_mask);
1770 }
1771
1772 static void
1773 endloop_emit(
1774    const struct lp_build_tgsi_action * action,
1775    struct lp_build_tgsi_context * bld_base,
1776    struct lp_build_emit_data * emit_data)
1777 {
1778    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1779
1780    lp_exec_endloop(bld_base->base.gallivm, &bld->exec_mask);
1781 }
1782
1783 static void
1784 endsub_emit(
1785    const struct lp_build_tgsi_action * action,
1786    struct lp_build_tgsi_context * bld_base,
1787    struct lp_build_emit_data * emit_data)
1788 {
1789    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1790
1791    lp_exec_mask_endsub(&bld->exec_mask, &bld_base->pc);
1792 }
1793
1794 static void
1795 cont_emit(
1796    const struct lp_build_tgsi_action * action,
1797    struct lp_build_tgsi_context * bld_base,
1798    struct lp_build_emit_data * emit_data)
1799 {
1800    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1801
1802    lp_exec_continue(&bld->exec_mask);
1803 }
1804
1805 /* XXX: Refactor and move it to lp_bld_tgsi_action.c
1806  *
1807  * XXX: What do the comments about xmm registers mean?  Maybe they are left over
1808  * from old code, but there is no garauntee that LLVM will use those registers
1809  * for this code.
1810  *
1811  * XXX: There should be no calls to lp_build_emit_fetch in this function.  This
1812  * should be handled by the emit_data->fetch_args function. */
1813 static void
1814 nrm_emit(
1815    const struct lp_build_tgsi_action * action,
1816    struct lp_build_tgsi_context * bld_base,
1817    struct lp_build_emit_data * emit_data)
1818 {
1819    LLVMValueRef tmp0, tmp1;
1820    LLVMValueRef tmp4 = NULL;
1821    LLVMValueRef tmp5 = NULL;
1822    LLVMValueRef tmp6 = NULL;
1823    LLVMValueRef tmp7 = NULL;
1824    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1825
1826    uint dims = (emit_data->inst->Instruction.Opcode == TGSI_OPCODE_NRM) ? 3 : 4;
1827
1828   if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_X) ||
1829       TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Y) ||
1830       TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Z) ||
1831       (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_W) && dims == 4)) {
1832
1833       /* NOTE: Cannot use xmm regs 2/3 here (see emit_rsqrt() above). */
1834
1835       /* xmm4 = src.x */
1836       /* xmm0 = src.x * src.x */
1837       tmp0 = lp_build_emit_fetch(&bld->bld_base, emit_data->inst, 0, TGSI_CHAN_X);
1838       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_X)) {
1839          tmp4 = tmp0;
1840       }
1841       tmp0 = lp_build_mul( &bld->bld_base.base, tmp0, tmp0);
1842
1843       /* xmm5 = src.y */
1844       /* xmm0 = xmm0 + src.y * src.y */
1845       tmp1 = lp_build_emit_fetch(&bld->bld_base, emit_data->inst, 0, TGSI_CHAN_Y);
1846       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Y)) {
1847          tmp5 = tmp1;
1848       }
1849       tmp1 = lp_build_mul( &bld->bld_base.base, tmp1, tmp1);
1850       tmp0 = lp_build_add( &bld->bld_base.base, tmp0, tmp1);
1851
1852       /* xmm6 = src.z */
1853       /* xmm0 = xmm0 + src.z * src.z */
1854       tmp1 = lp_build_emit_fetch(&bld->bld_base, emit_data->inst, 0, TGSI_CHAN_Z);
1855       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Z)) {
1856          tmp6 = tmp1;
1857       }
1858       tmp1 = lp_build_mul( &bld->bld_base.base, tmp1, tmp1);
1859       tmp0 = lp_build_add( &bld->bld_base.base, tmp0, tmp1);
1860
1861       if (dims == 4) {
1862          /* xmm7 = src.w */
1863          /* xmm0 = xmm0 + src.w * src.w */
1864          tmp1 = lp_build_emit_fetch(&bld->bld_base, emit_data->inst, 0, TGSI_CHAN_W);
1865          if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_W)) {
1866             tmp7 = tmp1;
1867          }
1868          tmp1 = lp_build_mul( &bld->bld_base.base, tmp1, tmp1);
1869          tmp0 = lp_build_add( &bld->bld_base.base, tmp0, tmp1);
1870       }
1871       /* xmm1 = 1 / sqrt(xmm0) */
1872       tmp1 = lp_build_rsqrt( &bld->bld_base.base, tmp0);
1873        /* dst.x = xmm1 * src.x */
1874       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_X)) {
1875          emit_data->output[TGSI_CHAN_X] = lp_build_mul( &bld->bld_base.base, tmp4, tmp1);
1876       }
1877       /* dst.y = xmm1 * src.y */
1878       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Y)) {
1879          emit_data->output[TGSI_CHAN_Y] = lp_build_mul( &bld->bld_base.base, tmp5, tmp1);
1880       }
1881
1882       /* dst.z = xmm1 * src.z */
1883       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Z)) {
1884          emit_data->output[TGSI_CHAN_Z] = lp_build_mul( &bld->bld_base.base, tmp6, tmp1);
1885       }
1886       /* dst.w = xmm1 * src.w */
1887       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_X) && dims == 4) {
1888          emit_data->output[TGSI_CHAN_W] = lp_build_mul( &bld->bld_base.base, tmp7, tmp1);
1889       }
1890    }
1891
1892    /* dst.w = 1.0 */
1893    if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_W) && dims == 3) {
1894        emit_data->output[TGSI_CHAN_W] = bld->bld_base.base.one;
1895    }
1896 }
1897
1898 static void emit_prologue(struct lp_build_tgsi_context * bld_base)
1899 {
1900    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1901    struct gallivm_state * gallivm = bld_base->base.gallivm;
1902
1903    if (bld->indirect_files & (1 << TGSI_FILE_TEMPORARY)) {
1904       LLVMValueRef array_size =
1905          lp_build_const_int32(gallivm,
1906                          bld_base->info->file_max[TGSI_FILE_TEMPORARY] * 4 + 4);
1907       bld->temps_array = lp_build_array_alloca(gallivm,
1908                                               bld_base->base.vec_type, array_size,
1909                                               "temp_array");
1910    }
1911
1912    if (bld->indirect_files & (1 << TGSI_FILE_OUTPUT)) {
1913       LLVMValueRef array_size =
1914          lp_build_const_int32(gallivm,
1915                             bld_base->info->file_max[TGSI_FILE_OUTPUT] * 4 + 4);
1916       bld->outputs_array = lp_build_array_alloca(gallivm,
1917                                                 bld_base->base.vec_type, array_size,
1918                                                 "output_array");
1919    }
1920
1921    /* If we have indirect addressing in inputs we need to copy them into
1922     * our alloca array to be able to iterate over them */
1923    if (bld->indirect_files & (1 << TGSI_FILE_INPUT)) {
1924       unsigned index, chan;
1925       LLVMTypeRef vec_type = bld_base->base.vec_type;
1926       LLVMValueRef array_size = lp_build_const_int32(gallivm,
1927             bld_base->info->file_max[TGSI_FILE_INPUT]*4 + 4);
1928       bld->inputs_array = lp_build_array_alloca(gallivm,
1929                                                vec_type, array_size,
1930                                                "input_array");
1931
1932       assert(bld_base->info->num_inputs
1933                         <= bld_base->info->file_max[TGSI_FILE_INPUT] + 1);
1934
1935       for (index = 0; index < bld_base->info->num_inputs; ++index) {
1936          for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
1937             LLVMValueRef lindex =
1938                lp_build_const_int32(gallivm, index * 4 + chan);
1939             LLVMValueRef input_ptr =
1940                LLVMBuildGEP(gallivm->builder, bld->inputs_array,
1941                             &lindex, 1, "");
1942             LLVMValueRef value = bld->inputs[index][chan];
1943             if (value)
1944                LLVMBuildStore(gallivm->builder, value, input_ptr);
1945          }
1946       }
1947    }
1948 }
1949
1950 static void emit_epilogue(struct lp_build_tgsi_context * bld_base)
1951 {
1952    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1953
1954    if (0) {
1955       /* for debugging */
1956       emit_dump_temps(bld);
1957    }
1958
1959    /* If we have indirect addressing in outputs we need to copy our alloca array
1960     * to the outputs slots specified by the called */
1961    if (bld->indirect_files & (1 << TGSI_FILE_OUTPUT)) {
1962       unsigned index, chan;
1963       assert(bld_base->info->num_outputs <=
1964                         bld_base->info->file_max[TGSI_FILE_OUTPUT] + 1);
1965       for (index = 0; index < bld_base->info->num_outputs; ++index) {
1966          for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
1967             bld->outputs[index][chan] = lp_get_output_ptr(bld, index, chan);
1968          }
1969       }
1970    }
1971 }
1972
1973 void
1974 lp_build_tgsi_soa(struct gallivm_state *gallivm,
1975                   const struct tgsi_token *tokens,
1976                   struct lp_type type,
1977                   struct lp_build_mask_context *mask,
1978                   LLVMValueRef consts_ptr,
1979                   LLVMValueRef system_values_array,
1980                   const LLVMValueRef *pos,
1981                   const LLVMValueRef (*inputs)[TGSI_NUM_CHANNELS],
1982                   LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1983                   struct lp_build_sampler_soa *sampler,
1984                   const struct tgsi_shader_info *info)
1985 {
1986    struct lp_build_tgsi_soa_context bld;
1987
1988    struct lp_type res_type;
1989
1990    assert(type.length <= LP_MAX_VECTOR_LENGTH);
1991    memset(&res_type, 0, sizeof res_type);
1992    res_type.width = type.width;
1993    res_type.length = type.length;
1994    res_type.sign = 1;
1995
1996    /* Setup build context */
1997    memset(&bld, 0, sizeof bld);
1998    lp_build_context_init(&bld.bld_base.base, gallivm, type);
1999    lp_build_context_init(&bld.bld_base.uint_bld, gallivm, lp_uint_type(type));
2000    lp_build_context_init(&bld.bld_base.int_bld, gallivm, lp_int_type(type));
2001    lp_build_context_init(&bld.elem_bld, gallivm, lp_elem_type(type));
2002    bld.mask = mask;
2003    bld.pos = pos;
2004    bld.inputs = inputs;
2005    bld.outputs = outputs;
2006    bld.consts_ptr = consts_ptr;
2007    bld.sampler = sampler;
2008    bld.bld_base.info = info;
2009    bld.indirect_files = info->indirect_files;
2010
2011    bld.bld_base.soa = TRUE;
2012    bld.bld_base.emit_fetch_funcs[TGSI_FILE_CONSTANT] = emit_fetch_constant;
2013    bld.bld_base.emit_fetch_funcs[TGSI_FILE_IMMEDIATE] = emit_fetch_immediate;
2014    bld.bld_base.emit_fetch_funcs[TGSI_FILE_INPUT] = emit_fetch_input;
2015    bld.bld_base.emit_fetch_funcs[TGSI_FILE_TEMPORARY] = emit_fetch_temporary;
2016    bld.bld_base.emit_fetch_funcs[TGSI_FILE_SYSTEM_VALUE] = emit_fetch_system_value;
2017    bld.bld_base.emit_store = emit_store;
2018
2019    bld.bld_base.emit_declaration = lp_emit_declaration_soa;
2020    bld.bld_base.emit_immediate = lp_emit_immediate_soa;
2021
2022    bld.bld_base.emit_prologue = emit_prologue;
2023    bld.bld_base.emit_epilogue = emit_epilogue;
2024
2025    /* Set opcode actions */
2026    lp_set_default_actions_cpu(&bld.bld_base);
2027
2028    bld.bld_base.op_actions[TGSI_OPCODE_BGNLOOP].emit = bgnloop_emit;
2029    bld.bld_base.op_actions[TGSI_OPCODE_BGNSUB].emit = bgnsub_emit;
2030    bld.bld_base.op_actions[TGSI_OPCODE_BRK].emit = brk_emit;
2031    bld.bld_base.op_actions[TGSI_OPCODE_CAL].emit = cal_emit;
2032    bld.bld_base.op_actions[TGSI_OPCODE_CONT].emit = cont_emit;
2033    bld.bld_base.op_actions[TGSI_OPCODE_DDX].emit = ddx_emit;
2034    bld.bld_base.op_actions[TGSI_OPCODE_DDY].emit = ddy_emit;
2035    bld.bld_base.op_actions[TGSI_OPCODE_ELSE].emit = else_emit;
2036    bld.bld_base.op_actions[TGSI_OPCODE_ENDIF].emit = endif_emit;
2037    bld.bld_base.op_actions[TGSI_OPCODE_ENDLOOP].emit = endloop_emit;
2038    bld.bld_base.op_actions[TGSI_OPCODE_ENDSUB].emit = endsub_emit;
2039    bld.bld_base.op_actions[TGSI_OPCODE_IF].emit = if_emit;
2040    bld.bld_base.op_actions[TGSI_OPCODE_KIL].emit = kil_emit;
2041    bld.bld_base.op_actions[TGSI_OPCODE_KILP].emit = kilp_emit;
2042    bld.bld_base.op_actions[TGSI_OPCODE_NRM].emit = nrm_emit;
2043    bld.bld_base.op_actions[TGSI_OPCODE_NRM4].emit = nrm_emit;
2044    bld.bld_base.op_actions[TGSI_OPCODE_RET].emit = ret_emit;
2045    bld.bld_base.op_actions[TGSI_OPCODE_TEX].emit = tex_emit;
2046    bld.bld_base.op_actions[TGSI_OPCODE_TXB].emit = txb_emit;
2047    bld.bld_base.op_actions[TGSI_OPCODE_TXD].emit = txd_emit;
2048    bld.bld_base.op_actions[TGSI_OPCODE_TXL].emit = txl_emit;
2049    bld.bld_base.op_actions[TGSI_OPCODE_TXP].emit = txp_emit;
2050    bld.bld_base.op_actions[TGSI_OPCODE_TXQ].emit = txq_emit;
2051
2052    lp_exec_mask_init(&bld.exec_mask, &bld.bld_base.base);
2053
2054
2055    bld.system_values_array = system_values_array;
2056
2057    lp_build_tgsi_llvm(&bld.bld_base, tokens);
2058
2059    if (0) {
2060       LLVMBasicBlockRef block = LLVMGetInsertBlock(gallivm->builder);
2061       LLVMValueRef function = LLVMGetBasicBlockParent(block);
2062       debug_printf("11111111111111111111111111111 \n");
2063       tgsi_dump(tokens, 0);
2064       lp_debug_dump_value(function);
2065       debug_printf("2222222222222222222222222222 \n");
2066    }
2067
2068    if (0) {
2069       LLVMModuleRef module = LLVMGetGlobalParent(
2070          LLVMGetBasicBlockParent(LLVMGetInsertBlock(gallivm->builder)));
2071       LLVMDumpModule(module);
2072
2073    }
2074 }
2075
2076
2077 /**
2078  * Build up the system values array out of individual values such as
2079  * the instance ID, front-face, primitive ID, etc.  The shader info is
2080  * used to determine which system values are needed and where to put
2081  * them in the system values array.
2082  *
2083  * XXX only instance ID is implemented at this time.
2084  *
2085  * The system values register file is similar to the constants buffer.
2086  * Example declaration:
2087  *    DCL SV[0], INSTANCEID
2088  * Example instruction:
2089  *    MOVE foo, SV[0].xxxx;
2090  *
2091  * \return  LLVM float array (interpreted as float [][4])
2092  */
2093 LLVMValueRef
2094 lp_build_system_values_array(struct gallivm_state *gallivm,
2095                              const struct tgsi_shader_info *info,
2096                              LLVMValueRef instance_id,
2097                              LLVMValueRef facing)
2098 {
2099    LLVMValueRef size = lp_build_const_int32(gallivm, 4 * info->num_system_values);
2100    LLVMTypeRef float_t = LLVMFloatTypeInContext(gallivm->context);
2101    LLVMValueRef array = lp_build_array_alloca(gallivm, float_t,
2102                                               size, "sysvals_array");
2103    unsigned i;
2104
2105    for (i = 0; i < info->num_system_values; i++) {
2106       LLVMValueRef index = lp_build_const_int32(gallivm, i * 4);
2107       LLVMValueRef ptr, value = 0;
2108
2109       switch (info->system_value_semantic_name[i]) {
2110       case TGSI_SEMANTIC_INSTANCEID:
2111          /* convert instance ID from int to float */
2112          value = LLVMBuildSIToFP(gallivm->builder, instance_id, float_t,
2113                                  "sysval_instanceid");
2114          break;
2115       case TGSI_SEMANTIC_FACE:
2116          /* fall-through */
2117       default:
2118          assert(0 && "unexpected semantic in build_system_values_array()");
2119       }
2120
2121       ptr = LLVMBuildGEP(gallivm->builder, array, &index, 1, "");
2122       LLVMBuildStore(gallivm->builder, value, ptr);
2123    }
2124
2125    return array;
2126 }