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    const struct tgsi_shader_info *info = bld->bld_base.info;
 790    LLVMBuilderRef builder = gallivm->builder;
 791    LLVMValueRef res;
 792    enum tgsi_opcode_type atype; // Actual type of the value
 793
 794    assert(!reg->Register.Indirect);
 795
 796    switch (info->system_value_semantic_name[reg->Register.Index]) {
 797    case TGSI_SEMANTIC_INSTANCEID:
 798       res = lp_build_broadcast_scalar(&bld_base->uint_bld, bld->system_values.instance_id);
 799       atype = TGSI_TYPE_UNSIGNED;
 800       break;
 801
 802    case TGSI_SEMANTIC_VERTEXID:
 803       res = bld->system_values.vertex_id;
 804       atype = TGSI_TYPE_UNSIGNED;
 805       break;
 806
 807    default:
 808       assert(!"unexpected semantic in emit_fetch_system_value");
 809       res = bld_base->base.zero;
 810       atype = TGSI_TYPE_FLOAT;
 811       break;
 812    }
 813
 814    if (atype != stype) {
 815       if (stype == TGSI_TYPE_FLOAT) {
 816          res = LLVMBuildBitCast(builder, res, bld_base->base.vec_type, "");
 817       } else if (stype == TGSI_TYPE_UNSIGNED) {
 818          res = LLVMBuildBitCast(builder, res, bld_base->uint_bld.vec_type, "");
 819       } else if (stype == TGSI_TYPE_SIGNED) {
 820          res = LLVMBuildBitCast(builder, res, bld_base->int_bld.vec_type, "");
 821       }
 822    }
 823
 824    return res;
 825 }
 826
 827 /**
 828  * Register fetch with derivatives.
 829  */
 830 static void
 831 emit_fetch_deriv(
 832    struct lp_build_tgsi_soa_context *bld,
 833    LLVMValueRef src,
 834    LLVMValueRef *res,
 835    LLVMValueRef *ddx,
 836    LLVMValueRef *ddy)
 837 {
 838    if(res)
 839       *res = src;
 840
 841    /* TODO: use interpolation coeffs for inputs */
 842
 843    if(ddx)
 844       *ddx = lp_build_ddx(&bld->bld_base.base, src);
 845
 846    if(ddy)
 847       *ddy = lp_build_ddy(&bld->bld_base.base, src);
 848 }
 849
 850
 851 /**
 852  * Predicate.
 853  */
 854 static void
 855 emit_fetch_predicate(
 856    struct lp_build_tgsi_soa_context *bld,
 857    const struct tgsi_full_instruction *inst,
 858    LLVMValueRef *pred)
 859 {
 860    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
 861    unsigned index;
 862    unsigned char swizzles[4];
 863    LLVMValueRef unswizzled[4] = {NULL, NULL, NULL, NULL};
 864    LLVMValueRef value;
 865    unsigned chan;
 866
 867    if (!inst->Instruction.Predicate) {
 868       TGSI_FOR_EACH_CHANNEL( chan ) {
 869          pred[chan] = NULL;
 870       }
 871       return;
 872    }
 873
 874    swizzles[0] = inst->Predicate.SwizzleX;
 875    swizzles[1] = inst->Predicate.SwizzleY;
 876    swizzles[2] = inst->Predicate.SwizzleZ;
 877    swizzles[3] = inst->Predicate.SwizzleW;
 878
 879    index = inst->Predicate.Index;
 880    assert(index < LP_MAX_TGSI_PREDS);
 881
 882    TGSI_FOR_EACH_CHANNEL( chan ) {
 883       unsigned swizzle = swizzles[chan];
 884
 885       /*
 886        * Only fetch the predicate register channels that are actually listed
 887        * in the swizzles
 888        */
 889       if (!unswizzled[swizzle]) {
 890          value = LLVMBuildLoad(builder,
 891                                bld->preds[index][swizzle], "");
 892
 893          /*
 894           * Convert the value to an integer mask.
 895           *
 896           * TODO: Short-circuit this comparison -- a D3D setp_xx instructions
 897           * is needlessly causing two comparisons due to storing the intermediate
 898           * result as float vector instead of an integer mask vector.
 899           */
 900          value = lp_build_compare(bld->bld_base.base.gallivm,
 901                                   bld->bld_base.base.type,
 902                                   PIPE_FUNC_NOTEQUAL,
 903                                   value,
 904                                   bld->bld_base.base.zero);
 905          if (inst->Predicate.Negate) {
 906             value = LLVMBuildNot(builder, value, "");
 907          }
 908
 909          unswizzled[swizzle] = value;
 910       } else {
 911          value = unswizzled[swizzle];
 912       }
 913
 914       pred[chan] = value;
 915    }
 916 }
 917
 918 /**
 919  * Register store.
 920  */
 921 static void
 922 emit_store_chan(
 923    struct lp_build_tgsi_context *bld_base,
 924    const struct tgsi_full_instruction *inst,
 925    unsigned index,
 926    unsigned chan_index,
 927    LLVMValueRef pred,
 928    LLVMValueRef value)
 929 {
 930    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
 931    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
 932    LLVMBuilderRef builder = gallivm->builder;
 933    const struct tgsi_full_dst_register *reg = &inst->Dst[index];
 934    struct lp_build_context *uint_bld = &bld_base->uint_bld;
 935    LLVMValueRef indirect_index = NULL;
 936    struct lp_build_context *bld_store;
 937    enum tgsi_opcode_type dtype = tgsi_opcode_infer_dst_type(inst->Instruction.Opcode);
 938
 939    switch (dtype) {
 940    default:
 941    case TGSI_TYPE_FLOAT:
 942    case TGSI_TYPE_UNTYPED:
 943       bld_store = &bld_base->base;
 944       break;
 945    case TGSI_TYPE_UNSIGNED:
 946       bld_store = &bld_base->uint_bld;
 947       break;
 948    case TGSI_TYPE_SIGNED:
 949       bld_store = &bld_base->int_bld;
 950       break;
 951    case TGSI_TYPE_DOUBLE:
 952    case TGSI_TYPE_VOID:
 953       assert(0);
 954       bld_store = NULL;
 955       break;
 956    }
 957
 958    switch( inst->Instruction.Saturate ) {
 959    case TGSI_SAT_NONE:
 960       break;
 961
 962    case TGSI_SAT_ZERO_ONE:
 963       value = lp_build_max(&bld->bld_base.base, value, bld->bld_base.base.zero);
 964       value = lp_build_min(&bld->bld_base.base, value, bld->bld_base.base.one);
 965       break;
 966
 967    case TGSI_SAT_MINUS_PLUS_ONE:
 968       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));
 969       value = lp_build_min(&bld->bld_base.base, value, bld->bld_base.base.one);
 970       break;
 971
 972    default:
 973       assert(0);
 974    }
 975
 976    if (reg->Register.Indirect) {
 977       indirect_index = get_indirect_index(bld,
 978                                           reg->Register.File,
 979                                           reg->Register.Index,
 980                                           &reg->Indirect);
 981    } else {
 982       assert(reg->Register.Index <=
 983                              bld->bld_base.info->file_max[reg->Register.File]);
 984    }
 985
 986    switch( reg->Register.File ) {
 987    case TGSI_FILE_OUTPUT:
 988       if (reg->Register.Indirect) {
 989          LLVMValueRef chan_vec =
 990             lp_build_const_int_vec(gallivm, uint_bld->type, chan_index);
 991          LLVMValueRef length_vec =
 992             lp_build_const_int_vec(gallivm, uint_bld->type, bld->bld_base.base.type.length);
 993          LLVMValueRef index_vec;  /* indexes into the temp registers */
 994          LLVMValueRef outputs_array;
 995          LLVMValueRef pixel_offsets;
 996          LLVMTypeRef float_ptr_type;
 997          int i;
 998
 999          /* build pixel offset vector: {0, 1, 2, 3, ...} */
1000          pixel_offsets = uint_bld->undef;
1001          for (i = 0; i < bld->bld_base.base.type.length; i++) {
1002             LLVMValueRef ii = lp_build_const_int32(gallivm, i);
1003             pixel_offsets = LLVMBuildInsertElement(builder, pixel_offsets,
1004                                                    ii, ii, "");
1005          }
1006
1007          /* index_vec = (indirect_index * 4 + chan_index) * length + offsets */
1008          index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
1009          index_vec = lp_build_add(uint_bld, index_vec, chan_vec);
1010          index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
1011          index_vec = lp_build_add(uint_bld, index_vec, pixel_offsets);
1012
1013          float_ptr_type =
1014             LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
1015          outputs_array = LLVMBuildBitCast(builder, bld->outputs_array,
1016                                           float_ptr_type, "");
1017
1018          /* Scatter store values into temp registers */
1019          emit_mask_scatter(bld, outputs_array, index_vec, value,
1020                            &bld->exec_mask, pred);
1021       }
1022       else {
1023          LLVMValueRef out_ptr = lp_get_output_ptr(bld, reg->Register.Index,
1024                                                chan_index);
1025          lp_exec_mask_store(&bld->exec_mask, bld_store, pred, value, out_ptr);
1026       }
1027       break;
1028
1029    case TGSI_FILE_TEMPORARY:
1030       if (reg->Register.Indirect) {
1031          LLVMValueRef chan_vec =
1032             lp_build_const_int_vec(gallivm, uint_bld->type, chan_index);
1033          LLVMValueRef length_vec =
1034             lp_build_const_int_vec(gallivm, uint_bld->type,
1035                                    bld->bld_base.base.type.length);
1036          LLVMValueRef index_vec;  /* indexes into the temp registers */
1037          LLVMValueRef temps_array;
1038          LLVMValueRef pixel_offsets;
1039          LLVMTypeRef float_ptr_type;
1040          int i;
1041
1042          /* build pixel offset vector: {0, 1, 2, 3, ...} */
1043          pixel_offsets = uint_bld->undef;
1044          for (i = 0; i < bld->bld_base.base.type.length; i++) {
1045             LLVMValueRef ii = lp_build_const_int32(gallivm, i);
1046             pixel_offsets = LLVMBuildInsertElement(builder, pixel_offsets,
1047                                                    ii, ii, "");
1048          }
1049
1050          /* index_vec = (indirect_index * 4 + chan_index) * length + offsets */
1051          index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
1052          index_vec = lp_build_add(uint_bld, index_vec, chan_vec);
1053          index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
1054          index_vec = lp_build_add(uint_bld, index_vec, pixel_offsets);
1055
1056          float_ptr_type =
1057             LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
1058          temps_array = LLVMBuildBitCast(builder, bld->temps_array,
1059                                         float_ptr_type, "");
1060
1061          /* Scatter store values into temp registers */
1062          emit_mask_scatter(bld, temps_array, index_vec, value,
1063                            &bld->exec_mask, pred);
1064       }
1065       else {
1066          LLVMValueRef temp_ptr;
1067
1068          switch (dtype) {
1069          case TGSI_TYPE_UNSIGNED:
1070          case TGSI_TYPE_SIGNED: {
1071             LLVMTypeRef itype = LLVMVectorType(LLVMInt32TypeInContext(gallivm->context), 4);
1072             LLVMTypeRef ivtype = LLVMPointerType(itype, 0);
1073             LLVMValueRef tint_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index,
1074                                                         chan_index);
1075             LLVMValueRef temp_value_ptr;
1076
1077             temp_ptr = LLVMBuildBitCast(builder, tint_ptr, ivtype, "");
1078             temp_value_ptr = LLVMBuildBitCast(builder, value, itype, "");
1079             value = temp_value_ptr;
1080             break;
1081          }
1082          default:
1083          case TGSI_TYPE_FLOAT:
1084          case TGSI_TYPE_UNTYPED:
1085             temp_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index,
1086                                            chan_index);
1087             break;
1088          }
1089
1090          lp_exec_mask_store(&bld->exec_mask, bld_store, pred, value, temp_ptr);
1091       }
1092       break;
1093
1094    case TGSI_FILE_ADDRESS:
1095       assert(dtype == TGSI_TYPE_SIGNED);
1096       assert(LLVMTypeOf(value) == bld_base->base.int_vec_type);
1097       lp_exec_mask_store(&bld->exec_mask, bld_store, pred, value,
1098                          bld->addr[reg->Register.Index][chan_index]);
1099       break;
1100
1101    case TGSI_FILE_PREDICATE:
1102       lp_exec_mask_store(&bld->exec_mask, bld_store, pred, value,
1103                          bld->preds[reg->Register.Index][chan_index]);
1104       break;
1105
1106    default:
1107       assert( 0 );
1108    }
1109 }
1110
1111 static void
1112 emit_store(
1113    struct lp_build_tgsi_context * bld_base,
1114    const struct tgsi_full_instruction * inst,
1115    const struct tgsi_opcode_info * info,
1116    LLVMValueRef dst[4])
1117
1118 {
1119    unsigned chan_index;
1120    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1121
1122    if(info->num_dst) {
1123       LLVMValueRef pred[TGSI_NUM_CHANNELS];
1124
1125       emit_fetch_predicate( bld, inst, pred );
1126
1127       TGSI_FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1128          emit_store_chan(bld_base, inst, 0, chan_index, pred[chan_index], dst[chan_index]);
1129       }
1130    }
1131 }
1132
1133 /**
1134  * High-level instruction translators.
1135  */
1136
1137 static void
1138 emit_tex( struct lp_build_tgsi_soa_context *bld,
1139           const struct tgsi_full_instruction *inst,
1140           enum lp_build_tex_modifier modifier,
1141           LLVMValueRef *texel)
1142 {
1143    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
1144    unsigned unit;
1145    LLVMValueRef lod_bias, explicit_lod;
1146    LLVMValueRef oow = NULL;
1147    LLVMValueRef coords[3];
1148    LLVMValueRef ddx[3];
1149    LLVMValueRef ddy[3];
1150    unsigned num_coords;
1151    unsigned i;
1152
1153    if (!bld->sampler) {
1154       _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
1155       for (i = 0; i < 4; i++) {
1156          texel[i] = bld->bld_base.base.undef;
1157       }
1158       return;
1159    }
1160
1161    switch (inst->Texture.Texture) {
1162    case TGSI_TEXTURE_1D:
1163       num_coords = 1;
1164       break;
1165    case TGSI_TEXTURE_1D_ARRAY:
1166    case TGSI_TEXTURE_2D:
1167    case TGSI_TEXTURE_RECT:
1168       num_coords = 2;
1169       break;
1170    case TGSI_TEXTURE_SHADOW1D:
1171    case TGSI_TEXTURE_SHADOW1D_ARRAY:
1172    case TGSI_TEXTURE_SHADOW2D:
1173    case TGSI_TEXTURE_SHADOWRECT:
1174    case TGSI_TEXTURE_2D_ARRAY:
1175    case TGSI_TEXTURE_3D:
1176    case TGSI_TEXTURE_CUBE:
1177       num_coords = 3;
1178       break;
1179    case TGSI_TEXTURE_SHADOW2D_ARRAY:
1180       num_coords = 4;
1181       break;
1182    default:
1183       assert(0);
1184       return;
1185    }
1186
1187    if (modifier == LP_BLD_TEX_MODIFIER_LOD_BIAS) {
1188       lod_bias = lp_build_emit_fetch( &bld->bld_base, inst, 0, 3 );
1189       explicit_lod = NULL;
1190    }
1191    else if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_LOD) {
1192       lod_bias = NULL;
1193       explicit_lod = lp_build_emit_fetch( &bld->bld_base, inst, 0, 3 );
1194    }
1195    else {
1196       lod_bias = NULL;
1197       explicit_lod = NULL;
1198    }
1199
1200    if (modifier == LP_BLD_TEX_MODIFIER_PROJECTED) {
1201       oow = lp_build_emit_fetch( &bld->bld_base, inst, 0, 3 );
1202       oow = lp_build_rcp(&bld->bld_base.base, oow);
1203    }
1204
1205    for (i = 0; i < num_coords; i++) {
1206       coords[i] = lp_build_emit_fetch( &bld->bld_base, inst, 0, i );
1207       if (modifier == LP_BLD_TEX_MODIFIER_PROJECTED)
1208          coords[i] = lp_build_mul(&bld->bld_base.base, coords[i], oow);
1209    }
1210    for (i = num_coords; i < 3; i++) {
1211       coords[i] = bld->bld_base.base.undef;
1212    }
1213
1214    if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV) {
1215       LLVMValueRef index0 = lp_build_const_int32(bld->bld_base.base.gallivm, 0);
1216       for (i = 0; i < num_coords; i++) {
1217          LLVMValueRef src1 = lp_build_emit_fetch( &bld->bld_base, inst, 1, i );
1218          LLVMValueRef src2 = lp_build_emit_fetch( &bld->bld_base, inst, 2, i );
1219          ddx[i] = LLVMBuildExtractElement(builder, src1, index0, "");
1220          ddy[i] = LLVMBuildExtractElement(builder, src2, index0, "");
1221       }
1222       unit = inst->Src[3].Register.Index;
1223    }  else {
1224       for (i = 0; i < num_coords; i++) {
1225          ddx[i] = lp_build_scalar_ddx( &bld->bld_base.base, coords[i] );
1226          ddy[i] = lp_build_scalar_ddy( &bld->bld_base.base, coords[i] );
1227       }
1228       unit = inst->Src[1].Register.Index;
1229    }
1230    for (i = num_coords; i < 3; i++) {
1231       ddx[i] = LLVMGetUndef(bld->bld_base.base.elem_type);
1232       ddy[i] = LLVMGetUndef(bld->bld_base.base.elem_type);
1233    }
1234
1235    bld->sampler->emit_fetch_texel(bld->sampler,
1236                                   bld->bld_base.base.gallivm,
1237                                   bld->bld_base.base.type,
1238                                   unit, num_coords, coords,
1239                                   ddx, ddy,
1240                                   lod_bias, explicit_lod,
1241                                   texel);
1242 }
1243
1244 static void
1245 emit_txq( struct lp_build_tgsi_soa_context *bld,
1246           const struct tgsi_full_instruction *inst,
1247           LLVMValueRef *sizes_out)
1248 {
1249    LLVMValueRef explicit_lod;
1250    unsigned num_coords, has_lod;
1251    unsigned i;
1252
1253    switch (inst->Texture.Texture) {
1254    case TGSI_TEXTURE_1D:
1255    case TGSI_TEXTURE_SHADOW1D:
1256    case TGSI_TEXTURE_SHADOW2D:
1257    case TGSI_TEXTURE_SHADOWCUBE:
1258       num_coords = 1;
1259       has_lod = 1;
1260       break;
1261    case TGSI_TEXTURE_2D:
1262    case TGSI_TEXTURE_CUBE:
1263    case TGSI_TEXTURE_1D_ARRAY:
1264    case TGSI_TEXTURE_SHADOW1D_ARRAY:
1265       num_coords = 2;
1266       has_lod = 1;
1267       break;
1268    case TGSI_TEXTURE_3D:
1269 // case TGSI_TEXTURE_CUBE_ARRAY:
1270 // case TGSI_TEXTURE_SHADOWCUBE_ARRAY:
1271    case TGSI_TEXTURE_2D_ARRAY:
1272    case TGSI_TEXTURE_SHADOW2D_ARRAY:
1273       num_coords = 3;
1274       has_lod = 1;
1275       break;
1276
1277    case TGSI_TEXTURE_BUFFER:
1278       num_coords = 1;
1279       has_lod = 0;
1280       break;
1281
1282    case TGSI_TEXTURE_RECT:
1283    case TGSI_TEXTURE_SHADOWRECT:
1284 // case TGSI_TEXTURE_2D_MS:
1285       num_coords = 2;
1286       has_lod = 0;
1287       break;
1288
1289 // case TGSI_TEXTURE_2D_MS_ARRAY:
1290 //    num_coords = 3;
1291 //    has_lod = 0;
1292 //    break;
1293
1294    default:
1295       assert(0);
1296       return;
1297    }
1298
1299    if (!bld->sampler) {
1300       _debug_printf("warning: found texture query instruction but no sampler generator supplied\n");
1301       for (i = 0; i < num_coords; i++)
1302          sizes_out[i] = bld->bld_base.base.undef;
1303       return;
1304    }
1305
1306    if (has_lod)
1307       explicit_lod = lp_build_emit_fetch( &bld->bld_base, inst, 0, 2 );
1308    else
1309       explicit_lod = NULL;
1310
1311    bld->sampler->emit_size_query(bld->sampler,
1312                                  bld->bld_base.base.gallivm,
1313                                  inst->Src[1].Register.Index,
1314                                  explicit_lod,
1315                                  sizes_out);
1316 }
1317
1318 static boolean
1319 near_end_of_shader(struct lp_build_tgsi_soa_context *bld,
1320                    int pc)
1321 {
1322    int i;
1323
1324    for (i = 0; i < 5; i++) {
1325       unsigned opcode;
1326
1327       if (pc + i >= bld->bld_base.info->num_instructions)
1328          return TRUE;
1329
1330       opcode = bld->bld_base.instructions[pc + i].Instruction.Opcode;
1331
1332       if (opcode == TGSI_OPCODE_END)
1333          return TRUE;
1334
1335       if (opcode == TGSI_OPCODE_TEX ||
1336           opcode == TGSI_OPCODE_TXP ||
1337           opcode == TGSI_OPCODE_TXD ||
1338           opcode == TGSI_OPCODE_TXB ||
1339           opcode == TGSI_OPCODE_TXL ||
1340           opcode == TGSI_OPCODE_TXF ||
1341           opcode == TGSI_OPCODE_TXQ ||
1342           opcode == TGSI_OPCODE_CAL ||
1343           opcode == TGSI_OPCODE_CALLNZ ||
1344           opcode == TGSI_OPCODE_IF ||
1345           opcode == TGSI_OPCODE_IFC ||
1346           opcode == TGSI_OPCODE_BGNLOOP ||
1347           opcode == TGSI_OPCODE_SWITCH)
1348          return FALSE;
1349    }
1350
1351    return TRUE;
1352 }
1353
1354
1355
1356 /**
1357  * Kill fragment if any of the src register values are negative.
1358  */
1359 static void
1360 emit_kil(
1361    struct lp_build_tgsi_soa_context *bld,
1362    const struct tgsi_full_instruction *inst,
1363    int pc)
1364 {
1365    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
1366    const struct tgsi_full_src_register *reg = &inst->Src[0];
1367    LLVMValueRef terms[TGSI_NUM_CHANNELS];
1368    LLVMValueRef mask;
1369    unsigned chan_index;
1370
1371    memset(&terms, 0, sizeof terms);
1372
1373    TGSI_FOR_EACH_CHANNEL( chan_index ) {
1374       unsigned swizzle;
1375
1376       /* Unswizzle channel */
1377       swizzle = tgsi_util_get_full_src_register_swizzle( reg, chan_index );
1378
1379       /* Check if the component has not been already tested. */
1380       assert(swizzle < TGSI_NUM_CHANNELS);
1381       if( !terms[swizzle] )
1382          /* TODO: change the comparison operator instead of setting the sign */
1383          terms[swizzle] =  lp_build_emit_fetch(&bld->bld_base, inst, 0, chan_index );
1384    }
1385
1386    mask = NULL;
1387    TGSI_FOR_EACH_CHANNEL( chan_index ) {
1388       if(terms[chan_index]) {
1389          LLVMValueRef chan_mask;
1390
1391          /*
1392           * If term < 0 then mask = 0 else mask = ~0.
1393           */
1394          chan_mask = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_GEQUAL, terms[chan_index], bld->bld_base.base.zero);
1395
1396          if(mask)
1397             mask = LLVMBuildAnd(builder, mask, chan_mask, "");
1398          else
1399             mask = chan_mask;
1400       }
1401    }
1402
1403    if(mask) {
1404       lp_build_mask_update(bld->mask, mask);
1405
1406       if (!near_end_of_shader(bld, pc))
1407          lp_build_mask_check(bld->mask);
1408    }
1409 }
1410
1411
1412 /**
1413  * Predicated fragment kill.
1414  * XXX Actually, we do an unconditional kill (as in tgsi_exec.c).
1415  * The only predication is the execution mask which will apply if
1416  * we're inside a loop or conditional.
1417  */
1418 static void
1419 emit_kilp(struct lp_build_tgsi_soa_context *bld,
1420           int pc)
1421 {
1422    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
1423    LLVMValueRef mask;
1424
1425    /* For those channels which are "alive", disable fragment shader
1426     * execution.
1427     */
1428    if (bld->exec_mask.has_mask) {
1429       mask = LLVMBuildNot(builder, bld->exec_mask.exec_mask, "kilp");
1430    }
1431    else {
1432       LLVMValueRef zero = LLVMConstNull(bld->bld_base.base.int_vec_type);
1433       mask = zero;
1434    }
1435
1436    lp_build_mask_update(bld->mask, mask);
1437
1438    if (!near_end_of_shader(bld, pc))
1439       lp_build_mask_check(bld->mask);
1440 }
1441
1442
1443 /**
1444  * Emit code which will dump the value of all the temporary registers
1445  * to stdout.
1446  */
1447 static void
1448 emit_dump_temps(struct lp_build_tgsi_soa_context *bld)
1449 {
1450    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
1451    LLVMBuilderRef builder = gallivm->builder;
1452    LLVMValueRef temp_ptr;
1453    LLVMValueRef i0 = lp_build_const_int32(gallivm, 0);
1454    LLVMValueRef i1 = lp_build_const_int32(gallivm, 1);
1455    LLVMValueRef i2 = lp_build_const_int32(gallivm, 2);
1456    LLVMValueRef i3 = lp_build_const_int32(gallivm, 3);
1457    int index;
1458    int n = bld->bld_base.info->file_max[TGSI_FILE_TEMPORARY];
1459
1460    for (index = 0; index < n; index++) {
1461       LLVMValueRef idx = lp_build_const_int32(gallivm, index);
1462       LLVMValueRef v[4][4], res;
1463       int chan;
1464
1465       lp_build_printf(gallivm, "TEMP[%d]:\n", idx);
1466
1467       for (chan = 0; chan < 4; chan++) {
1468          temp_ptr = lp_get_temp_ptr_soa(bld, index, chan);
1469          res = LLVMBuildLoad(builder, temp_ptr, "");
1470          v[chan][0] = LLVMBuildExtractElement(builder, res, i0, "");
1471          v[chan][1] = LLVMBuildExtractElement(builder, res, i1, "");
1472          v[chan][2] = LLVMBuildExtractElement(builder, res, i2, "");
1473          v[chan][3] = LLVMBuildExtractElement(builder, res, i3, "");
1474       }
1475
1476       lp_build_printf(gallivm, "  X: %f %f %f %f\n",
1477                       v[0][0], v[0][1], v[0][2], v[0][3]);
1478       lp_build_printf(gallivm, "  Y: %f %f %f %f\n",
1479                       v[1][0], v[1][1], v[1][2], v[1][3]);
1480       lp_build_printf(gallivm, "  Z: %f %f %f %f\n",
1481                       v[2][0], v[2][1], v[2][2], v[2][3]);
1482       lp_build_printf(gallivm, "  W: %f %f %f %f\n",
1483                       v[3][0], v[3][1], v[3][2], v[3][3]);
1484    }
1485 }
1486
1487
1488
1489 void
1490 lp_emit_declaration_soa(
1491    struct lp_build_tgsi_context *bld_base,
1492    const struct tgsi_full_declaration *decl)
1493 {
1494    struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
1495    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
1496    LLVMTypeRef vec_type = bld->bld_base.base.vec_type;
1497    const unsigned first = decl->Range.First;
1498    const unsigned last = decl->Range.Last;
1499    unsigned idx, i;
1500
1501    for (idx = first; idx <= last; ++idx) {
1502       assert(last <= bld->bld_base.info->file_max[decl->Declaration.File]);
1503       switch (decl->Declaration.File) {
1504       case TGSI_FILE_TEMPORARY:
1505          assert(idx < LP_MAX_TGSI_TEMPS);
1506          if (!(bld->indirect_files & (1 << TGSI_FILE_TEMPORARY))) {
1507             for (i = 0; i < TGSI_NUM_CHANNELS; i++)
1508                bld->temps[idx][i] = lp_build_alloca(gallivm, vec_type, "temp");
1509          }
1510          break;
1511
1512       case TGSI_FILE_OUTPUT:
1513          if (!(bld->indirect_files & (1 << TGSI_FILE_OUTPUT))) {
1514             for (i = 0; i < TGSI_NUM_CHANNELS; i++)
1515                bld->outputs[idx][i] = lp_build_alloca(gallivm,
1516                                                       vec_type, "output");
1517          }
1518          break;
1519
1520       case TGSI_FILE_ADDRESS:
1521          /* ADDR registers are the only allocated with an integer LLVM IR type,
1522           * as they are guaranteed to always have integers.
1523           * XXX: Not sure if this exception is worthwhile (or the whole idea of
1524           * an ADDR register for that matter).
1525           */
1526          assert(idx < LP_MAX_TGSI_ADDRS);
1527          for (i = 0; i < TGSI_NUM_CHANNELS; i++)
1528             bld->addr[idx][i] = lp_build_alloca(gallivm, bld_base->base.int_vec_type, "addr");
1529          break;
1530
1531       case TGSI_FILE_PREDICATE:
1532          assert(idx < LP_MAX_TGSI_PREDS);
1533          for (i = 0; i < TGSI_NUM_CHANNELS; i++)
1534             bld->preds[idx][i] = lp_build_alloca(gallivm, vec_type,
1535                                                  "predicate");
1536          break;
1537
1538       default:
1539          /* don't need to declare other vars */
1540          break;
1541       }
1542    }
1543 }
1544
1545
1546 void lp_emit_immediate_soa(
1547    struct lp_build_tgsi_context *bld_base,
1548    const struct tgsi_full_immediate *imm)
1549 {
1550    struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
1551    struct gallivm_state * gallivm = bld_base->base.gallivm;
1552
1553    /* simply copy the immediate values into the next immediates[] slot */
1554    unsigned i;
1555    const uint size = imm->Immediate.NrTokens - 1;
1556    assert(size <= 4);
1557    assert(bld->num_immediates < LP_MAX_TGSI_IMMEDIATES);
1558    switch (imm->Immediate.DataType) {
1559    case TGSI_IMM_FLOAT32:
1560       for( i = 0; i < size; ++i )
1561          bld->immediates[bld->num_immediates][i] =
1562             lp_build_const_vec(gallivm, bld_base->base.type, imm->u[i].Float);
1563
1564       break;
1565    case TGSI_IMM_UINT32:
1566       for( i = 0; i < size; ++i ) {
1567          LLVMValueRef tmp = lp_build_const_vec(gallivm, bld_base->uint_bld.type, imm->u[i].Uint);
1568          bld->immediates[bld->num_immediates][i] =
1569             LLVMConstBitCast(tmp, bld_base->base.vec_type);
1570       }
1571
1572       break;
1573    case TGSI_IMM_INT32:
1574       for( i = 0; i < size; ++i ) {
1575          LLVMValueRef tmp = lp_build_const_vec(gallivm, bld_base->int_bld.type, imm->u[i].Int);
1576          bld->immediates[bld->num_immediates][i] =
1577             LLVMConstBitCast(tmp, bld_base->base.vec_type);
1578       }
1579
1580       break;
1581    }
1582    for( i = size; i < 4; ++i )
1583       bld->immediates[bld->num_immediates][i] = bld_base->base.undef;
1584
1585    bld->num_immediates++;
1586 }
1587
1588 static void
1589 ddx_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_fetch_deriv(bld, emit_data->args[0], NULL,
1597                     &emit_data->output[emit_data->chan], NULL);
1598 }
1599
1600 static void
1601 ddy_emit(
1602    const struct lp_build_tgsi_action * action,
1603    struct lp_build_tgsi_context * bld_base,
1604    struct lp_build_emit_data * emit_data)
1605 {
1606    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1607
1608    emit_fetch_deriv(bld, emit_data->args[0], NULL, NULL,
1609                     &emit_data->output[emit_data->chan]);
1610 }
1611
1612 static void
1613 kilp_emit(
1614    const struct lp_build_tgsi_action * action,
1615    struct lp_build_tgsi_context * bld_base,
1616    struct lp_build_emit_data * emit_data)
1617 {
1618    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1619
1620    emit_kilp(bld, bld_base->pc - 1);
1621 }
1622
1623 static void
1624 kil_emit(
1625    const struct lp_build_tgsi_action * action,
1626    struct lp_build_tgsi_context * bld_base,
1627    struct lp_build_emit_data * emit_data)
1628 {
1629    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1630
1631    emit_kil(bld, emit_data->inst, bld_base->pc - 1);
1632 }
1633
1634 static void
1635 tex_emit(
1636    const struct lp_build_tgsi_action * action,
1637    struct lp_build_tgsi_context * bld_base,
1638    struct lp_build_emit_data * emit_data)
1639 {
1640    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1641
1642    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_NONE, emit_data->output);
1643 }
1644
1645 static void
1646 txb_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_LOD_BIAS,
1654             emit_data->output);
1655 }
1656
1657 static void
1658 txd_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_EXPLICIT_DERIV,
1666             emit_data->output);
1667 }
1668
1669 static void
1670 txl_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_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD,
1678             emit_data->output);
1679 }
1680
1681 static void
1682 txp_emit(
1683    const struct lp_build_tgsi_action * action,
1684    struct lp_build_tgsi_context * bld_base,
1685    struct lp_build_emit_data * emit_data)
1686 {
1687    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1688
1689    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_PROJECTED,
1690             emit_data->output);
1691 }
1692
1693 static void
1694 txq_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    emit_txq(bld, emit_data->inst, emit_data->output);
1702 }
1703
1704 static void
1705 cal_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_mask_call(&bld->exec_mask, emit_data->inst->Label.Label,
1713                      &bld_base->pc);
1714 }
1715
1716 static void
1717 ret_emit(
1718    const struct lp_build_tgsi_action * action,
1719    struct lp_build_tgsi_context * bld_base,
1720    struct lp_build_emit_data * emit_data)
1721 {
1722    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1723
1724    lp_exec_mask_ret(&bld->exec_mask, &bld_base->pc);
1725 }
1726
1727 static void
1728 brk_emit(
1729    const struct lp_build_tgsi_action * action,
1730    struct lp_build_tgsi_context * bld_base,
1731    struct lp_build_emit_data * emit_data)
1732 {
1733    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1734
1735    lp_exec_break(&bld->exec_mask);
1736 }
1737
1738 static void
1739 if_emit(
1740    const struct lp_build_tgsi_action * action,
1741    struct lp_build_tgsi_context * bld_base,
1742    struct lp_build_emit_data * emit_data)
1743 {
1744    LLVMValueRef tmp;
1745    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1746
1747    tmp = lp_build_cmp(&bld_base->base, PIPE_FUNC_NOTEQUAL,
1748                       emit_data->args[0], bld->bld_base.base.zero);
1749    lp_exec_mask_cond_push(&bld->exec_mask, tmp);
1750 }
1751
1752 static void
1753 bgnloop_emit(
1754    const struct lp_build_tgsi_action * action,
1755    struct lp_build_tgsi_context * bld_base,
1756    struct lp_build_emit_data * emit_data)
1757 {
1758    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1759
1760    lp_exec_bgnloop(&bld->exec_mask);
1761 }
1762
1763 static void
1764 bgnsub_emit(
1765    const struct lp_build_tgsi_action * action,
1766    struct lp_build_tgsi_context * bld_base,
1767    struct lp_build_emit_data * emit_data)
1768 {
1769    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1770
1771    lp_exec_mask_bgnsub(&bld->exec_mask);
1772 }
1773
1774 static void
1775 else_emit(
1776    const struct lp_build_tgsi_action * action,
1777    struct lp_build_tgsi_context * bld_base,
1778    struct lp_build_emit_data * emit_data)
1779 {
1780    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1781
1782    lp_exec_mask_cond_invert(&bld->exec_mask);
1783 }
1784
1785 static void
1786 endif_emit(
1787    const struct lp_build_tgsi_action * action,
1788    struct lp_build_tgsi_context * bld_base,
1789    struct lp_build_emit_data * emit_data)
1790 {
1791    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1792
1793    lp_exec_mask_cond_pop(&bld->exec_mask);
1794 }
1795
1796 static void
1797 endloop_emit(
1798    const struct lp_build_tgsi_action * action,
1799    struct lp_build_tgsi_context * bld_base,
1800    struct lp_build_emit_data * emit_data)
1801 {
1802    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1803
1804    lp_exec_endloop(bld_base->base.gallivm, &bld->exec_mask);
1805 }
1806
1807 static void
1808 endsub_emit(
1809    const struct lp_build_tgsi_action * action,
1810    struct lp_build_tgsi_context * bld_base,
1811    struct lp_build_emit_data * emit_data)
1812 {
1813    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1814
1815    lp_exec_mask_endsub(&bld->exec_mask, &bld_base->pc);
1816 }
1817
1818 static void
1819 cont_emit(
1820    const struct lp_build_tgsi_action * action,
1821    struct lp_build_tgsi_context * bld_base,
1822    struct lp_build_emit_data * emit_data)
1823 {
1824    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1825
1826    lp_exec_continue(&bld->exec_mask);
1827 }
1828
1829 /* XXX: Refactor and move it to lp_bld_tgsi_action.c
1830  *
1831  * XXX: What do the comments about xmm registers mean?  Maybe they are left over
1832  * from old code, but there is no garauntee that LLVM will use those registers
1833  * for this code.
1834  *
1835  * XXX: There should be no calls to lp_build_emit_fetch in this function.  This
1836  * should be handled by the emit_data->fetch_args function. */
1837 static void
1838 nrm_emit(
1839    const struct lp_build_tgsi_action * action,
1840    struct lp_build_tgsi_context * bld_base,
1841    struct lp_build_emit_data * emit_data)
1842 {
1843    LLVMValueRef tmp0, tmp1;
1844    LLVMValueRef tmp4 = NULL;
1845    LLVMValueRef tmp5 = NULL;
1846    LLVMValueRef tmp6 = NULL;
1847    LLVMValueRef tmp7 = NULL;
1848    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1849
1850    uint dims = (emit_data->inst->Instruction.Opcode == TGSI_OPCODE_NRM) ? 3 : 4;
1851
1852   if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_X) ||
1853       TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Y) ||
1854       TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Z) ||
1855       (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_W) && dims == 4)) {
1856
1857       /* NOTE: Cannot use xmm regs 2/3 here (see emit_rsqrt() above). */
1858
1859       /* xmm4 = src.x */
1860       /* xmm0 = src.x * src.x */
1861       tmp0 = lp_build_emit_fetch(&bld->bld_base, emit_data->inst, 0, TGSI_CHAN_X);
1862       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_X)) {
1863          tmp4 = tmp0;
1864       }
1865       tmp0 = lp_build_mul( &bld->bld_base.base, tmp0, tmp0);
1866
1867       /* xmm5 = src.y */
1868       /* xmm0 = xmm0 + src.y * src.y */
1869       tmp1 = lp_build_emit_fetch(&bld->bld_base, emit_data->inst, 0, TGSI_CHAN_Y);
1870       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Y)) {
1871          tmp5 = tmp1;
1872       }
1873       tmp1 = lp_build_mul( &bld->bld_base.base, tmp1, tmp1);
1874       tmp0 = lp_build_add( &bld->bld_base.base, tmp0, tmp1);
1875
1876       /* xmm6 = src.z */
1877       /* xmm0 = xmm0 + src.z * src.z */
1878       tmp1 = lp_build_emit_fetch(&bld->bld_base, emit_data->inst, 0, TGSI_CHAN_Z);
1879       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Z)) {
1880          tmp6 = tmp1;
1881       }
1882       tmp1 = lp_build_mul( &bld->bld_base.base, tmp1, tmp1);
1883       tmp0 = lp_build_add( &bld->bld_base.base, tmp0, tmp1);
1884
1885       if (dims == 4) {
1886          /* xmm7 = src.w */
1887          /* xmm0 = xmm0 + src.w * src.w */
1888          tmp1 = lp_build_emit_fetch(&bld->bld_base, emit_data->inst, 0, TGSI_CHAN_W);
1889          if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_W)) {
1890             tmp7 = tmp1;
1891          }
1892          tmp1 = lp_build_mul( &bld->bld_base.base, tmp1, tmp1);
1893          tmp0 = lp_build_add( &bld->bld_base.base, tmp0, tmp1);
1894       }
1895       /* xmm1 = 1 / sqrt(xmm0) */
1896       tmp1 = lp_build_rsqrt( &bld->bld_base.base, tmp0);
1897        /* dst.x = xmm1 * src.x */
1898       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_X)) {
1899          emit_data->output[TGSI_CHAN_X] = lp_build_mul( &bld->bld_base.base, tmp4, tmp1);
1900       }
1901       /* dst.y = xmm1 * src.y */
1902       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Y)) {
1903          emit_data->output[TGSI_CHAN_Y] = lp_build_mul( &bld->bld_base.base, tmp5, tmp1);
1904       }
1905
1906       /* dst.z = xmm1 * src.z */
1907       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Z)) {
1908          emit_data->output[TGSI_CHAN_Z] = lp_build_mul( &bld->bld_base.base, tmp6, tmp1);
1909       }
1910       /* dst.w = xmm1 * src.w */
1911       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_X) && dims == 4) {
1912          emit_data->output[TGSI_CHAN_W] = lp_build_mul( &bld->bld_base.base, tmp7, tmp1);
1913       }
1914    }
1915
1916    /* dst.w = 1.0 */
1917    if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_W) && dims == 3) {
1918        emit_data->output[TGSI_CHAN_W] = bld->bld_base.base.one;
1919    }
1920 }
1921
1922 static void emit_prologue(struct lp_build_tgsi_context * bld_base)
1923 {
1924    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1925    struct gallivm_state * gallivm = bld_base->base.gallivm;
1926
1927    if (bld->indirect_files & (1 << TGSI_FILE_TEMPORARY)) {
1928       LLVMValueRef array_size =
1929          lp_build_const_int32(gallivm,
1930                          bld_base->info->file_max[TGSI_FILE_TEMPORARY] * 4 + 4);
1931       bld->temps_array = lp_build_array_alloca(gallivm,
1932                                               bld_base->base.vec_type, array_size,
1933                                               "temp_array");
1934    }
1935
1936    if (bld->indirect_files & (1 << TGSI_FILE_OUTPUT)) {
1937       LLVMValueRef array_size =
1938          lp_build_const_int32(gallivm,
1939                             bld_base->info->file_max[TGSI_FILE_OUTPUT] * 4 + 4);
1940       bld->outputs_array = lp_build_array_alloca(gallivm,
1941                                                 bld_base->base.vec_type, array_size,
1942                                                 "output_array");
1943    }
1944
1945    /* If we have indirect addressing in inputs we need to copy them into
1946     * our alloca array to be able to iterate over them */
1947    if (bld->indirect_files & (1 << TGSI_FILE_INPUT)) {
1948       unsigned index, chan;
1949       LLVMTypeRef vec_type = bld_base->base.vec_type;
1950       LLVMValueRef array_size = lp_build_const_int32(gallivm,
1951             bld_base->info->file_max[TGSI_FILE_INPUT]*4 + 4);
1952       bld->inputs_array = lp_build_array_alloca(gallivm,
1953                                                vec_type, array_size,
1954                                                "input_array");
1955
1956       assert(bld_base->info->num_inputs
1957                         <= bld_base->info->file_max[TGSI_FILE_INPUT] + 1);
1958
1959       for (index = 0; index < bld_base->info->num_inputs; ++index) {
1960          for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
1961             LLVMValueRef lindex =
1962                lp_build_const_int32(gallivm, index * 4 + chan);
1963             LLVMValueRef input_ptr =
1964                LLVMBuildGEP(gallivm->builder, bld->inputs_array,
1965                             &lindex, 1, "");
1966             LLVMValueRef value = bld->inputs[index][chan];
1967             if (value)
1968                LLVMBuildStore(gallivm->builder, value, input_ptr);
1969          }
1970       }
1971    }
1972 }
1973
1974 static void emit_epilogue(struct lp_build_tgsi_context * bld_base)
1975 {
1976    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1977
1978    if (0) {
1979       /* for debugging */
1980       emit_dump_temps(bld);
1981    }
1982
1983    /* If we have indirect addressing in outputs we need to copy our alloca array
1984     * to the outputs slots specified by the called */
1985    if (bld->indirect_files & (1 << TGSI_FILE_OUTPUT)) {
1986       unsigned index, chan;
1987       assert(bld_base->info->num_outputs <=
1988                         bld_base->info->file_max[TGSI_FILE_OUTPUT] + 1);
1989       for (index = 0; index < bld_base->info->num_outputs; ++index) {
1990          for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
1991             bld->outputs[index][chan] = lp_get_output_ptr(bld, index, chan);
1992          }
1993       }
1994    }
1995 }
1996
1997 void
1998 lp_build_tgsi_soa(struct gallivm_state *gallivm,
1999                   const struct tgsi_token *tokens,
2000                   struct lp_type type,
2001                   struct lp_build_mask_context *mask,
2002                   LLVMValueRef consts_ptr,
2003                   const struct lp_bld_tgsi_system_values *system_values,
2004                   const LLVMValueRef *pos,
2005                   const LLVMValueRef (*inputs)[TGSI_NUM_CHANNELS],
2006                   LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
2007                   struct lp_build_sampler_soa *sampler,
2008                   const struct tgsi_shader_info *info)
2009 {
2010    struct lp_build_tgsi_soa_context bld;
2011
2012    struct lp_type res_type;
2013
2014    assert(type.length <= LP_MAX_VECTOR_LENGTH);
2015    memset(&res_type, 0, sizeof res_type);
2016    res_type.width = type.width;
2017    res_type.length = type.length;
2018    res_type.sign = 1;
2019
2020    /* Setup build context */
2021    memset(&bld, 0, sizeof bld);
2022    lp_build_context_init(&bld.bld_base.base, gallivm, type);
2023    lp_build_context_init(&bld.bld_base.uint_bld, gallivm, lp_uint_type(type));
2024    lp_build_context_init(&bld.bld_base.int_bld, gallivm, lp_int_type(type));
2025    lp_build_context_init(&bld.elem_bld, gallivm, lp_elem_type(type));
2026    bld.mask = mask;
2027    bld.pos = pos;
2028    bld.inputs = inputs;
2029    bld.outputs = outputs;
2030    bld.consts_ptr = consts_ptr;
2031    bld.sampler = sampler;
2032    bld.bld_base.info = info;
2033    bld.indirect_files = info->indirect_files;
2034
2035    bld.bld_base.soa = TRUE;
2036    bld.bld_base.emit_fetch_funcs[TGSI_FILE_CONSTANT] = emit_fetch_constant;
2037    bld.bld_base.emit_fetch_funcs[TGSI_FILE_IMMEDIATE] = emit_fetch_immediate;
2038    bld.bld_base.emit_fetch_funcs[TGSI_FILE_INPUT] = emit_fetch_input;
2039    bld.bld_base.emit_fetch_funcs[TGSI_FILE_TEMPORARY] = emit_fetch_temporary;
2040    bld.bld_base.emit_fetch_funcs[TGSI_FILE_SYSTEM_VALUE] = emit_fetch_system_value;
2041    bld.bld_base.emit_store = emit_store;
2042
2043    bld.bld_base.emit_declaration = lp_emit_declaration_soa;
2044    bld.bld_base.emit_immediate = lp_emit_immediate_soa;
2045
2046    bld.bld_base.emit_prologue = emit_prologue;
2047    bld.bld_base.emit_epilogue = emit_epilogue;
2048
2049    /* Set opcode actions */
2050    lp_set_default_actions_cpu(&bld.bld_base);
2051
2052    bld.bld_base.op_actions[TGSI_OPCODE_BGNLOOP].emit = bgnloop_emit;
2053    bld.bld_base.op_actions[TGSI_OPCODE_BGNSUB].emit = bgnsub_emit;
2054    bld.bld_base.op_actions[TGSI_OPCODE_BRK].emit = brk_emit;
2055    bld.bld_base.op_actions[TGSI_OPCODE_CAL].emit = cal_emit;
2056    bld.bld_base.op_actions[TGSI_OPCODE_CONT].emit = cont_emit;
2057    bld.bld_base.op_actions[TGSI_OPCODE_DDX].emit = ddx_emit;
2058    bld.bld_base.op_actions[TGSI_OPCODE_DDY].emit = ddy_emit;
2059    bld.bld_base.op_actions[TGSI_OPCODE_ELSE].emit = else_emit;
2060    bld.bld_base.op_actions[TGSI_OPCODE_ENDIF].emit = endif_emit;
2061    bld.bld_base.op_actions[TGSI_OPCODE_ENDLOOP].emit = endloop_emit;
2062    bld.bld_base.op_actions[TGSI_OPCODE_ENDSUB].emit = endsub_emit;
2063    bld.bld_base.op_actions[TGSI_OPCODE_IF].emit = if_emit;
2064    bld.bld_base.op_actions[TGSI_OPCODE_KIL].emit = kil_emit;
2065    bld.bld_base.op_actions[TGSI_OPCODE_KILP].emit = kilp_emit;
2066    bld.bld_base.op_actions[TGSI_OPCODE_NRM].emit = nrm_emit;
2067    bld.bld_base.op_actions[TGSI_OPCODE_NRM4].emit = nrm_emit;
2068    bld.bld_base.op_actions[TGSI_OPCODE_RET].emit = ret_emit;
2069    bld.bld_base.op_actions[TGSI_OPCODE_TEX].emit = tex_emit;
2070    bld.bld_base.op_actions[TGSI_OPCODE_TXB].emit = txb_emit;
2071    bld.bld_base.op_actions[TGSI_OPCODE_TXD].emit = txd_emit;
2072    bld.bld_base.op_actions[TGSI_OPCODE_TXL].emit = txl_emit;
2073    bld.bld_base.op_actions[TGSI_OPCODE_TXP].emit = txp_emit;
2074    bld.bld_base.op_actions[TGSI_OPCODE_TXQ].emit = txq_emit;
2075
2076    lp_exec_mask_init(&bld.exec_mask, &bld.bld_base.base);
2077
2078    bld.system_values = *system_values;
2079
2080    lp_build_tgsi_llvm(&bld.bld_base, tokens);
2081
2082    if (0) {
2083       LLVMBasicBlockRef block = LLVMGetInsertBlock(gallivm->builder);
2084       LLVMValueRef function = LLVMGetBasicBlockParent(block);
2085       debug_printf("11111111111111111111111111111 \n");
2086       tgsi_dump(tokens, 0);
2087       lp_debug_dump_value(function);
2088       debug_printf("2222222222222222222222222222 \n");
2089    }
2090
2091    if (0) {
2092       LLVMModuleRef module = LLVMGetGlobalParent(
2093          LLVMGetBasicBlockParent(LLVMGetInsertBlock(gallivm->builder)));
2094       LLVMDumpModule(module);
2095
2096    }
2097 }