src/gallium/auxiliary/gallivm/lp_bld_tgsi_soa.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2009 VMware, Inc.
   4  * Copyright 2007-2008 Tungsten Graphics, Inc., Cedar Park, Texas.
   5  * All Rights Reserved.
   6  *
   7  * Permission is hereby granted, free of charge, to any person obtaining a
   8  * copy of this software and associated documentation files (the
   9  * "Software"), to deal in the Software without restriction, including
  10  * without limitation the rights to use, copy, modify, merge, publish,
  11  * distribute, sub license, and/or sell copies of the Software, and to
  12  * permit persons to whom the Software is furnished to do so, subject to
  13  * the following conditions:
  14  *
  15  * The above copyright notice and this permission notice (including the
  16  * next paragraph) shall be included in all copies or substantial portions
  17  * of the Software.
  18  *
  19  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  20  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  21  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  22  * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
  23  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  24  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  25  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  26  *
  27  **************************************************************************/
  28
  29 /**
  30  * @file
  31  * TGSI to LLVM IR translation -- SoA.
  32  *
  33  * @author Jose Fonseca <jfonseca@vmware.com>
  34  *
  35  * Based on tgsi_sse2.c code written by Michal Krol, Keith Whitwell,
  36  * Brian Paul, and others.
  37  */
  38
  39 #include "pipe/p_config.h"
  40 #include "pipe/p_shader_tokens.h"
  41 #include "util/u_debug.h"
  42 #include "util/u_math.h"
  43 #include "util/u_memory.h"
  44 #include "tgsi/tgsi_dump.h"
  45 #include "tgsi/tgsi_exec.h"
  46 #include "tgsi/tgsi_info.h"
  47 #include "tgsi/tgsi_parse.h"
  48 #include "tgsi/tgsi_util.h"
  49 #include "tgsi/tgsi_scan.h"
  50 #include "lp_bld_tgsi_action.h"
  51 #include "lp_bld_type.h"
  52 #include "lp_bld_const.h"
  53 #include "lp_bld_arit.h"
  54 #include "lp_bld_bitarit.h"
  55 #include "lp_bld_gather.h"
  56 #include "lp_bld_init.h"
  57 #include "lp_bld_logic.h"
  58 #include "lp_bld_swizzle.h"
  59 #include "lp_bld_flow.h"
  60 #include "lp_bld_quad.h"
  61 #include "lp_bld_tgsi.h"
  62 #include "lp_bld_limits.h"
  63 #include "lp_bld_debug.h"
  64 #include "lp_bld_printf.h"
  65
  66
  67 static void lp_exec_mask_init(struct lp_exec_mask *mask, struct lp_build_context *bld)
  68 {
  69    LLVMTypeRef int_type = LLVMInt32TypeInContext(bld->gallivm->context);
  70    LLVMBuilderRef builder = bld->gallivm->builder;
  71
  72    mask->bld = bld;
  73    mask->has_mask = FALSE;
  74    mask->cond_stack_size = 0;
  75    mask->loop_stack_size = 0;
  76    mask->call_stack_size = 0;
  77
  78    mask->int_vec_type = lp_build_int_vec_type(bld->gallivm, mask->bld->type);
  79    mask->exec_mask = mask->ret_mask = mask->break_mask = mask->cont_mask = mask->cond_mask =
  80          LLVMConstAllOnes(mask->int_vec_type);
  81
  82    mask->loop_limiter = lp_build_alloca(bld->gallivm, int_type, "looplimiter");
  83
  84    LLVMBuildStore(
  85       builder,
  86       LLVMConstInt(int_type, LP_MAX_TGSI_LOOP_ITERATIONS, false),
  87       mask->loop_limiter);
  88 }
  89
  90 static void lp_exec_mask_update(struct lp_exec_mask *mask)
  91 {
  92    LLVMBuilderRef builder = mask->bld->gallivm->builder;
  93
  94    if (mask->loop_stack_size) {
  95       /*for loops we need to update the entire mask at runtime */
  96       LLVMValueRef tmp;
  97       assert(mask->break_mask);
  98       tmp = LLVMBuildAnd(builder,
  99                          mask->cont_mask,
 100                          mask->break_mask,
 101                          "maskcb");
 102       mask->exec_mask = LLVMBuildAnd(builder,
 103                                      mask->cond_mask,
 104                                      tmp,
 105                                      "maskfull");
 106    } else
 107       mask->exec_mask = mask->cond_mask;
 108
 109    if (mask->call_stack_size) {
 110       mask->exec_mask = LLVMBuildAnd(builder,
 111                                      mask->exec_mask,
 112                                      mask->ret_mask,
 113                                      "callmask");
 114    }
 115
 116    mask->has_mask = (mask->cond_stack_size > 0 ||
 117                      mask->loop_stack_size > 0 ||
 118                      mask->call_stack_size > 0);
 119 }
 120
 121 static void lp_exec_mask_cond_push(struct lp_exec_mask *mask,
 122                                    LLVMValueRef val)
 123 {
 124    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 125
 126    assert(mask->cond_stack_size < LP_MAX_TGSI_NESTING);
 127    if (mask->cond_stack_size == 0) {
 128       assert(mask->cond_mask == LLVMConstAllOnes(mask->int_vec_type));
 129    }
 130    mask->cond_stack[mask->cond_stack_size++] = mask->cond_mask;
 131    assert(LLVMTypeOf(val) == mask->int_vec_type);
 132    mask->cond_mask = LLVMBuildAnd(builder,
 133                                   mask->cond_mask,
 134                                   val,
 135                                   "");
 136    lp_exec_mask_update(mask);
 137 }
 138
 139 static void lp_exec_mask_cond_invert(struct lp_exec_mask *mask)
 140 {
 141    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 142    LLVMValueRef prev_mask;
 143    LLVMValueRef inv_mask;
 144
 145    assert(mask->cond_stack_size);
 146    prev_mask = mask->cond_stack[mask->cond_stack_size - 1];
 147    if (mask->cond_stack_size == 1) {
 148       assert(prev_mask == LLVMConstAllOnes(mask->int_vec_type));
 149    }
 150
 151    inv_mask = LLVMBuildNot(builder, mask->cond_mask, "");
 152
 153    mask->cond_mask = LLVMBuildAnd(builder,
 154                                   inv_mask,
 155                                   prev_mask, "");
 156    lp_exec_mask_update(mask);
 157 }
 158
 159 static void lp_exec_mask_cond_pop(struct lp_exec_mask *mask)
 160 {
 161    assert(mask->cond_stack_size);
 162    mask->cond_mask = mask->cond_stack[--mask->cond_stack_size];
 163    lp_exec_mask_update(mask);
 164 }
 165
 166 static void lp_exec_bgnloop(struct lp_exec_mask *mask)
 167 {
 168    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 169
 170    if (mask->loop_stack_size == 0) {
 171       assert(mask->loop_block == NULL);
 172       assert(mask->cont_mask == LLVMConstAllOnes(mask->int_vec_type));
 173       assert(mask->break_mask == LLVMConstAllOnes(mask->int_vec_type));
 174       assert(mask->break_var == NULL);
 175    }
 176
 177    assert(mask->loop_stack_size < LP_MAX_TGSI_NESTING);
 178
 179    mask->loop_stack[mask->loop_stack_size].loop_block = mask->loop_block;
 180    mask->loop_stack[mask->loop_stack_size].cont_mask = mask->cont_mask;
 181    mask->loop_stack[mask->loop_stack_size].break_mask = mask->break_mask;
 182    mask->loop_stack[mask->loop_stack_size].break_var = mask->break_var;
 183    ++mask->loop_stack_size;
 184
 185    mask->break_var = lp_build_alloca(mask->bld->gallivm, mask->int_vec_type, "");
 186    LLVMBuildStore(builder, mask->break_mask, mask->break_var);
 187
 188    mask->loop_block = lp_build_insert_new_block(mask->bld->gallivm, "bgnloop");
 189
 190    LLVMBuildBr(builder, mask->loop_block);
 191    LLVMPositionBuilderAtEnd(builder, mask->loop_block);
 192
 193    mask->break_mask = LLVMBuildLoad(builder, mask->break_var, "");
 194
 195    lp_exec_mask_update(mask);
 196 }
 197
 198 static void lp_exec_break(struct lp_exec_mask *mask)
 199 {
 200    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 201    LLVMValueRef exec_mask = LLVMBuildNot(builder,
 202                                          mask->exec_mask,
 203                                          "break");
 204
 205    mask->break_mask = LLVMBuildAnd(builder,
 206                                    mask->break_mask,
 207                                    exec_mask, "break_full");
 208
 209    lp_exec_mask_update(mask);
 210 }
 211
 212 static void lp_exec_continue(struct lp_exec_mask *mask)
 213 {
 214    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 215    LLVMValueRef exec_mask = LLVMBuildNot(builder,
 216                                          mask->exec_mask,
 217                                          "");
 218
 219    mask->cont_mask = LLVMBuildAnd(builder,
 220                                   mask->cont_mask,
 221                                   exec_mask, "");
 222
 223    lp_exec_mask_update(mask);
 224 }
 225
 226
 227 static void lp_exec_endloop(struct gallivm_state *gallivm,
 228                             struct lp_exec_mask *mask)
 229 {
 230    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 231    LLVMBasicBlockRef endloop;
 232    LLVMTypeRef int_type = LLVMInt32TypeInContext(mask->bld->gallivm->context);
 233    LLVMTypeRef reg_type = LLVMIntTypeInContext(gallivm->context,
 234                                                mask->bld->type.width *
 235                                                mask->bld->type.length);
 236    LLVMValueRef i1cond, i2cond, icond, limiter;
 237
 238    assert(mask->break_mask);
 239
 240    /*
 241     * Restore the cont_mask, but don't pop
 242     */
 243    assert(mask->loop_stack_size);
 244    mask->cont_mask = mask->loop_stack[mask->loop_stack_size - 1].cont_mask;
 245    lp_exec_mask_update(mask);
 246
 247    /*
 248     * Unlike the continue mask, the break_mask must be preserved across loop
 249     * iterations
 250     */
 251    LLVMBuildStore(builder, mask->break_mask, mask->break_var);
 252
 253    /* Decrement the loop limiter */
 254    limiter = LLVMBuildLoad(builder, mask->loop_limiter, "");
 255
 256    limiter = LLVMBuildSub(
 257       builder,
 258       limiter,
 259       LLVMConstInt(int_type, 1, false),
 260       "");
 261
 262    LLVMBuildStore(builder, limiter, mask->loop_limiter);
 263
 264    /* i1cond = (mask != 0) */
 265    i1cond = LLVMBuildICmp(
 266       builder,
 267       LLVMIntNE,
 268       LLVMBuildBitCast(builder, mask->exec_mask, reg_type, ""),
 269       LLVMConstNull(reg_type), "");
 270
 271    /* i2cond = (looplimiter > 0) */
 272    i2cond = LLVMBuildICmp(
 273       builder,
 274       LLVMIntSGT,
 275       limiter,
 276       LLVMConstNull(int_type), "");
 277
 278    /* if( i1cond && i2cond ) */
 279    icond = LLVMBuildAnd(builder, i1cond, i2cond, "");
 280
 281    endloop = lp_build_insert_new_block(mask->bld->gallivm, "endloop");
 282
 283    LLVMBuildCondBr(builder,
 284                    icond, mask->loop_block, endloop);
 285
 286    LLVMPositionBuilderAtEnd(builder, endloop);
 287
 288    assert(mask->loop_stack_size);
 289    --mask->loop_stack_size;
 290    mask->loop_block = mask->loop_stack[mask->loop_stack_size].loop_block;
 291    mask->cont_mask = mask->loop_stack[mask->loop_stack_size].cont_mask;
 292    mask->break_mask = mask->loop_stack[mask->loop_stack_size].break_mask;
 293    mask->break_var = mask->loop_stack[mask->loop_stack_size].break_var;
 294
 295    lp_exec_mask_update(mask);
 296 }
 297
 298 /* stores val into an address pointed to by dst.
 299  * mask->exec_mask is used to figure out which bits of val
 300  * should be stored into the address
 301  * (0 means don't store this bit, 1 means do store).
 302  */
 303 static void lp_exec_mask_store(struct lp_exec_mask *mask,
 304                                struct lp_build_context *bld_store,
 305                                LLVMValueRef pred,
 306                                LLVMValueRef val,
 307                                LLVMValueRef dst)
 308 {
 309    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 310
 311    /* Mix the predicate and execution mask */
 312    if (mask->has_mask) {
 313       if (pred) {
 314          pred = LLVMBuildAnd(builder, pred, mask->exec_mask, "");
 315       } else {
 316          pred = mask->exec_mask;
 317       }
 318    }
 319
 320    if (pred) {
 321       LLVMValueRef real_val, dst_val;
 322
 323       dst_val = LLVMBuildLoad(builder, dst, "");
 324       real_val = lp_build_select(bld_store,
 325                                  pred,
 326                                  val, dst_val);
 327
 328       LLVMBuildStore(builder, real_val, dst);
 329    } else
 330       LLVMBuildStore(builder, val, dst);
 331 }
 332
 333 static void lp_exec_mask_call(struct lp_exec_mask *mask,
 334                               int func,
 335                               int *pc)
 336 {
 337    assert(mask->call_stack_size < LP_MAX_TGSI_NESTING);
 338    mask->call_stack[mask->call_stack_size].pc = *pc;
 339    mask->call_stack[mask->call_stack_size].ret_mask = mask->ret_mask;
 340    mask->call_stack_size++;
 341    *pc = func;
 342 }
 343
 344 static void lp_exec_mask_ret(struct lp_exec_mask *mask, int *pc)
 345 {
 346    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 347    LLVMValueRef exec_mask;
 348
 349    if (mask->call_stack_size == 0) {
 350       /* returning from main() */
 351       *pc = -1;
 352       return;
 353    }
 354    exec_mask = LLVMBuildNot(builder,
 355                             mask->exec_mask,
 356                             "ret");
 357
 358    mask->ret_mask = LLVMBuildAnd(builder,
 359                                  mask->ret_mask,
 360                                  exec_mask, "ret_full");
 361
 362    lp_exec_mask_update(mask);
 363 }
 364
 365 static void lp_exec_mask_bgnsub(struct lp_exec_mask *mask)
 366 {
 367 }
 368
 369 static void lp_exec_mask_endsub(struct lp_exec_mask *mask, int *pc)
 370 {
 371    assert(mask->call_stack_size);
 372    mask->call_stack_size--;
 373    *pc = mask->call_stack[mask->call_stack_size].pc;
 374    mask->ret_mask = mask->call_stack[mask->call_stack_size].ret_mask;
 375    lp_exec_mask_update(mask);
 376 }
 377
 378
 379 /**
 380  * Return pointer to a temporary register channel (src or dest).
 381  * Note that indirect addressing cannot be handled here.
 382  * \param index  which temporary register
 383  * \param chan  which channel of the temp register.
 384  */
 385 LLVMValueRef
 386 lp_get_temp_ptr_soa(struct lp_build_tgsi_soa_context *bld,
 387              unsigned index,
 388              unsigned chan)
 389 {
 390    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
 391    assert(chan < 4);
 392    if (bld->indirect_files & (1 << TGSI_FILE_TEMPORARY)) {
 393       LLVMValueRef lindex = lp_build_const_int32(bld->bld_base.base.gallivm, index * 4 + chan);
 394       return LLVMBuildGEP(builder, bld->temps_array, &lindex, 1, "");
 395    }
 396    else {
 397       return bld->temps[index][chan];
 398    }
 399 }
 400
 401 /**
 402  * Return pointer to a output register channel (src or dest).
 403  * Note that indirect addressing cannot be handled here.
 404  * \param index  which output register
 405  * \param chan  which channel of the output register.
 406  */
 407 LLVMValueRef
 408 lp_get_output_ptr(struct lp_build_tgsi_soa_context *bld,
 409                unsigned index,
 410                unsigned chan)
 411 {
 412    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
 413    assert(chan < 4);
 414    if (bld->indirect_files & (1 << TGSI_FILE_OUTPUT)) {
 415       LLVMValueRef lindex = lp_build_const_int32(bld->bld_base.base.gallivm,
 416                                                  index * 4 + chan);
 417       return LLVMBuildGEP(builder, bld->outputs_array, &lindex, 1, "");
 418    }
 419    else {
 420       return bld->outputs[index][chan];
 421    }
 422 }
 423
 424 /**
 425  * Gather vector.
 426  * XXX the lp_build_gather() function should be capable of doing this
 427  * with a little work.
 428  */
 429 static LLVMValueRef
 430 build_gather(struct lp_build_context *bld,
 431              LLVMValueRef base_ptr,
 432              LLVMValueRef indexes)
 433 {
 434    LLVMBuilderRef builder = bld->gallivm->builder;
 435    LLVMValueRef res = bld->undef;
 436    unsigned i;
 437
 438    /*
 439     * Loop over elements of index_vec, load scalar value, insert it into 'res'.
 440     */
 441    for (i = 0; i < bld->type.length; i++) {
 442       LLVMValueRef ii = lp_build_const_int32(bld->gallivm, i);
 443       LLVMValueRef index = LLVMBuildExtractElement(builder,
 444                                                    indexes, ii, "");
 445       LLVMValueRef scalar_ptr = LLVMBuildGEP(builder, base_ptr,
 446                                              &index, 1, "gather_ptr");
 447       LLVMValueRef scalar = LLVMBuildLoad(builder, scalar_ptr, "");
 448
 449       res = LLVMBuildInsertElement(builder, res, scalar, ii, "");
 450    }
 451
 452    return res;
 453 }
 454
 455
 456 /**
 457  * Scatter/store vector.
 458  */
 459 static void
 460 emit_mask_scatter(struct lp_build_tgsi_soa_context *bld,
 461                   LLVMValueRef base_ptr,
 462                   LLVMValueRef indexes,
 463                   LLVMValueRef values,
 464                   struct lp_exec_mask *mask,
 465                   LLVMValueRef pred)
 466 {
 467    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
 468    LLVMBuilderRef builder = gallivm->builder;
 469    unsigned i;
 470
 471    /* Mix the predicate and execution mask */
 472    if (mask->has_mask) {
 473       if (pred) {
 474          pred = LLVMBuildAnd(builder, pred, mask->exec_mask, "");
 475       }
 476       else {
 477          pred = mask->exec_mask;
 478       }
 479    }
 480
 481    /*
 482     * Loop over elements of index_vec, store scalar value.
 483     */
 484    for (i = 0; i < bld->bld_base.base.type.length; i++) {
 485       LLVMValueRef ii = lp_build_const_int32(gallivm, i);
 486       LLVMValueRef index = LLVMBuildExtractElement(builder, indexes, ii, "");
 487       LLVMValueRef scalar_ptr = LLVMBuildGEP(builder, base_ptr, &index, 1, "scatter_ptr");
 488       LLVMValueRef val = LLVMBuildExtractElement(builder, values, ii, "scatter_val");
 489       LLVMValueRef scalar_pred = pred ?
 490          LLVMBuildExtractElement(builder, pred, ii, "scatter_pred") : NULL;
 491
 492       if (0)
 493          lp_build_printf(gallivm, "scatter %d: val %f at %d %p\n",
 494                          ii, val, index, scalar_ptr);
 495
 496       if (scalar_pred) {
 497          LLVMValueRef real_val, dst_val;
 498          dst_val = LLVMBuildLoad(builder, scalar_ptr, "");
 499          real_val = lp_build_select(&bld->elem_bld, scalar_pred, val, dst_val);
 500          LLVMBuildStore(builder, real_val, scalar_ptr);
 501       }
 502       else {
 503          LLVMBuildStore(builder, val, scalar_ptr);
 504       }
 505    }
 506 }
 507
 508
 509 /**
 510  * Read the current value of the ADDR register, convert the floats to
 511  * ints, add the base index and return the vector of offsets.
 512  * The offsets will be used to index into the constant buffer or
 513  * temporary register file.
 514  */
 515 static LLVMValueRef
 516 get_indirect_index(struct lp_build_tgsi_soa_context *bld,
 517                    unsigned reg_file, unsigned reg_index,
 518                    const struct tgsi_src_register *indirect_reg)
 519 {
 520    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
 521    struct lp_build_context *uint_bld = &bld->bld_base.uint_bld;
 522    /* always use X component of address register */
 523    unsigned swizzle = indirect_reg->SwizzleX;
 524    LLVMValueRef base;
 525    LLVMValueRef rel;
 526    LLVMValueRef max_index;
 527    LLVMValueRef index;
 528
 529    assert(bld->indirect_files & (1 << reg_file));
 530
 531    base = lp_build_const_int_vec(bld->bld_base.base.gallivm, uint_bld->type, reg_index);
 532
 533    assert(swizzle < 4);
 534    rel = LLVMBuildLoad(builder,
 535                         bld->addr[indirect_reg->Index][swizzle],
 536                         "load addr reg");
 537
 538    index = lp_build_add(uint_bld, base, rel);
 539
 540    max_index = lp_build_const_int_vec(bld->bld_base.base.gallivm,
 541                                       uint_bld->type,
 542                                       bld->bld_base.info->file_max[reg_file]);
 543
 544    assert(!uint_bld->type.sign);
 545    index = lp_build_min(uint_bld, index, max_index);
 546
 547    return index;
 548 }
 549
 550 static struct lp_build_context *
 551 stype_to_fetch(struct lp_build_tgsi_context * bld_base,
 552                enum tgsi_opcode_type stype)
 553 {
 554    struct lp_build_context *bld_fetch;
 555
 556    switch (stype) {
 557    case TGSI_TYPE_FLOAT:
 558    case TGSI_TYPE_UNTYPED:
 559       bld_fetch = &bld_base->base;
 560       break;
 561    case TGSI_TYPE_UNSIGNED:
 562       bld_fetch = &bld_base->uint_bld;
 563       break;
 564    case TGSI_TYPE_SIGNED:
 565       bld_fetch = &bld_base->int_bld;
 566       break;
 567    case TGSI_TYPE_VOID:
 568    case TGSI_TYPE_DOUBLE:
 569    default:
 570       assert(0);
 571       bld_fetch = NULL;
 572       break;
 573    }
 574    return bld_fetch;
 575 }
 576
 577 static LLVMValueRef
 578 emit_fetch_constant(
 579    struct lp_build_tgsi_context * bld_base,
 580    const struct tgsi_full_src_register * reg,
 581    enum tgsi_opcode_type stype,
 582    unsigned swizzle)
 583 {
 584    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
 585    struct gallivm_state *gallivm = bld_base->base.gallivm;
 586    LLVMBuilderRef builder = gallivm->builder;
 587    struct lp_build_context *uint_bld = &bld_base->uint_bld;
 588    LLVMValueRef indirect_index = NULL;
 589    struct lp_build_context *bld_fetch = stype_to_fetch(bld_base, stype);
 590
 591    /* XXX: Handle fetching xyzw components as a vector */
 592    assert(swizzle != ~0);
 593
 594    if (reg->Register.Indirect) {
 595       indirect_index = get_indirect_index(bld,
 596                                           reg->Register.File,
 597                                           reg->Register.Index,
 598                                           &reg->Indirect);
 599    }
 600
 601    if (reg->Register.Indirect) {
 602       LLVMValueRef swizzle_vec =
 603          lp_build_const_int_vec(bld->bld_base.base.gallivm, uint_bld->type, swizzle);
 604       LLVMValueRef index_vec;  /* index into the const buffer */
 605
 606       /* index_vec = indirect_index * 4 + swizzle */
 607       index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
 608       index_vec = lp_build_add(uint_bld, index_vec, swizzle_vec);
 609
 610       /* Gather values from the constant buffer */
 611       return build_gather(bld_fetch, bld->consts_ptr, index_vec);
 612    }
 613    else {
 614       LLVMValueRef index;  /* index into the const buffer */
 615       LLVMValueRef scalar, scalar_ptr;
 616
 617       index = lp_build_const_int32(gallivm, reg->Register.Index*4 + swizzle);
 618
 619       scalar_ptr = LLVMBuildGEP(builder, bld->consts_ptr,
 620                                    &index, 1, "");
 621
 622       if (stype != TGSI_TYPE_FLOAT && stype != TGSI_TYPE_UNTYPED) {
 623          LLVMTypeRef ivtype = LLVMPointerType(LLVMInt32TypeInContext(gallivm->context), 0);
 624          LLVMValueRef temp_ptr;
 625          temp_ptr = LLVMBuildBitCast(builder, scalar_ptr, ivtype, "");
 626          scalar = LLVMBuildLoad(builder, temp_ptr, "");
 627       } else
 628          scalar = LLVMBuildLoad(builder, scalar_ptr, "");
 629
 630       return lp_build_broadcast_scalar(bld_fetch, scalar);
 631    }
 632 }
 633
 634 static LLVMValueRef
 635 emit_fetch_immediate(
 636    struct lp_build_tgsi_context * bld_base,
 637    const struct tgsi_full_src_register * reg,
 638    enum tgsi_opcode_type stype,
 639    unsigned swizzle)
 640 {
 641    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
 642    LLVMValueRef res = bld->immediates[reg->Register.Index][swizzle];
 643    assert(res);
 644
 645    if (stype == TGSI_TYPE_UNSIGNED) {
 646       res = LLVMConstBitCast(res, bld_base->uint_bld.vec_type);
 647    } else if (stype == TGSI_TYPE_SIGNED) {
 648       res = LLVMConstBitCast(res, bld_base->int_bld.vec_type);
 649    }
 650    return res;
 651 }
 652
 653 static LLVMValueRef
 654 emit_fetch_input(
 655    struct lp_build_tgsi_context * bld_base,
 656    const struct tgsi_full_src_register * reg,
 657    enum tgsi_opcode_type stype,
 658    unsigned swizzle)
 659 {
 660    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
 661    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
 662    LLVMBuilderRef builder = gallivm->builder;
 663    struct lp_build_context *uint_bld = &bld_base->uint_bld;
 664    LLVMValueRef indirect_index = NULL;
 665    LLVMValueRef res;
 666
 667    if (reg->Register.Indirect) {
 668       indirect_index = get_indirect_index(bld,
 669                                           reg->Register.File,
 670                                           reg->Register.Index,
 671                                           &reg->Indirect);
 672    }
 673
 674    if (reg->Register.Indirect) {
 675       LLVMValueRef swizzle_vec =
 676          lp_build_const_int_vec(gallivm, uint_bld->type, swizzle);
 677       LLVMValueRef length_vec =
 678          lp_build_const_int_vec(gallivm, uint_bld->type, bld->bld_base.base.type.length);
 679       LLVMValueRef index_vec;  /* index into the const buffer */
 680       LLVMValueRef inputs_array;
 681       LLVMTypeRef float4_ptr_type;
 682
 683       /* index_vec = (indirect_index * 4 + swizzle) * length */
 684       index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
 685       index_vec = lp_build_add(uint_bld, index_vec, swizzle_vec);
 686       index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
 687
 688       /* cast inputs_array pointer to float* */
 689       float4_ptr_type = LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
 690       inputs_array = LLVMBuildBitCast(builder, bld->inputs_array,
 691                                          float4_ptr_type, "");
 692
 693       /* Gather values from the temporary register array */
 694       res = build_gather(&bld_base->base, inputs_array, index_vec);
 695    } else {
 696       if (bld->indirect_files & (1 << TGSI_FILE_INPUT)) {
 697          LLVMValueRef lindex = lp_build_const_int32(gallivm,
 698                                         reg->Register.Index * 4 + swizzle);
 699          LLVMValueRef input_ptr =  LLVMBuildGEP(builder,
 700                                                 bld->inputs_array, &lindex, 1, "");
 701          res = LLVMBuildLoad(builder, input_ptr, "");
 702       }
 703       else {
 704          res = bld->inputs[reg->Register.Index][swizzle];
 705       }
 706    }
 707    assert(res);
 708    return res;
 709 }
 710
 711 static LLVMValueRef
 712 emit_fetch_temporary(
 713    struct lp_build_tgsi_context * bld_base,
 714    const struct tgsi_full_src_register * reg,
 715    enum tgsi_opcode_type stype,
 716    unsigned swizzle)
 717 {
 718    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
 719    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
 720    LLVMBuilderRef builder = gallivm->builder;
 721    struct lp_build_context *uint_bld = &bld_base->uint_bld;
 722    LLVMValueRef indirect_index = NULL;
 723    LLVMValueRef res;
 724
 725    if (reg->Register.Indirect) {
 726       indirect_index = get_indirect_index(bld,
 727                                           reg->Register.File,
 728                                           reg->Register.Index,
 729                                           &reg->Indirect);
 730    }
 731
 732    if (reg->Register.Indirect) {
 733       LLVMValueRef swizzle_vec =
 734          lp_build_const_int_vec(bld->bld_base.base.gallivm, uint_bld->type, swizzle);
 735       LLVMValueRef length_vec =
 736          lp_build_const_int_vec(bld->bld_base.base.gallivm, uint_bld->type,
 737                                 bld->bld_base.base.type.length);
 738       LLVMValueRef index_vec;  /* index into the const buffer */
 739       LLVMValueRef temps_array;
 740       LLVMTypeRef float4_ptr_type;
 741
 742       /* index_vec = (indirect_index * 4 + swizzle) * length */
 743       index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
 744       index_vec = lp_build_add(uint_bld, index_vec, swizzle_vec);
 745       index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
 746
 747       /* cast temps_array pointer to float* */
 748       float4_ptr_type = LLVMPointerType(LLVMFloatTypeInContext(bld->bld_base.base.gallivm->context), 0);
 749       temps_array = LLVMBuildBitCast(builder, bld->temps_array,
 750                                      float4_ptr_type, "");
 751
 752       /* Gather values from the temporary register array */
 753       res = build_gather(&bld_base->base, temps_array, index_vec);
 754    }
 755    else {
 756       LLVMValueRef temp_ptr;
 757       if (stype != TGSI_TYPE_FLOAT && stype != TGSI_TYPE_UNTYPED) {
 758          LLVMTypeRef itype = LLVMPointerType(LLVMVectorType(LLVMInt32TypeInContext(gallivm->context), 4), 0);
 759          LLVMValueRef tint_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index,
 760                                                      swizzle);
 761          temp_ptr = LLVMBuildBitCast(builder, tint_ptr, itype, "");
 762       } else
 763          temp_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index, swizzle);
 764       res = LLVMBuildLoad(builder, temp_ptr, "");
 765       if (!res)
 766          return bld->bld_base.base.undef;
 767    }
 768
 769    return res;
 770 }
 771
 772 static LLVMValueRef
 773 emit_fetch_system_value(
 774    struct lp_build_tgsi_context * bld_base,
 775    const struct tgsi_full_src_register * reg,
 776    enum tgsi_opcode_type stype,
 777    unsigned swizzle)
 778 {
 779    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
 780    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
 781    LLVMBuilderRef builder = gallivm->builder;
 782    LLVMValueRef index;  /* index into the system value array */
 783    LLVMValueRef scalar, scalar_ptr;
 784
 785    assert(!reg->Register.Indirect);
 786
 787    index = lp_build_const_int32(gallivm, reg->Register.Index * 4 + swizzle);
 788
 789    scalar_ptr = LLVMBuildGEP(builder, bld->system_values_array, &index, 1, "");
 790    scalar = LLVMBuildLoad(builder, scalar_ptr, "");
 791
 792    return lp_build_broadcast_scalar(&bld->bld_base.base, scalar);
 793 }
 794
 795 /**
 796  * Register fetch with derivatives.
 797  */
 798 static void
 799 emit_fetch_deriv(
 800    struct lp_build_tgsi_soa_context *bld,
 801    LLVMValueRef src,
 802    LLVMValueRef *res,
 803    LLVMValueRef *ddx,
 804    LLVMValueRef *ddy)
 805 {
 806    if(res)
 807       *res = src;
 808
 809    /* TODO: use interpolation coeffs for inputs */
 810
 811    if(ddx)
 812       *ddx = lp_build_ddx(&bld->bld_base.base, src);
 813
 814    if(ddy)
 815       *ddy = lp_build_ddy(&bld->bld_base.base, src);
 816 }
 817
 818
 819 /**
 820  * Predicate.
 821  */
 822 static void
 823 emit_fetch_predicate(
 824    struct lp_build_tgsi_soa_context *bld,
 825    const struct tgsi_full_instruction *inst,
 826    LLVMValueRef *pred)
 827 {
 828    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
 829    unsigned index;
 830    unsigned char swizzles[4];
 831    LLVMValueRef unswizzled[4] = {NULL, NULL, NULL, NULL};
 832    LLVMValueRef value;
 833    unsigned chan;
 834
 835    if (!inst->Instruction.Predicate) {
 836       TGSI_FOR_EACH_CHANNEL( chan ) {
 837          pred[chan] = NULL;
 838       }
 839       return;
 840    }
 841
 842    swizzles[0] = inst->Predicate.SwizzleX;
 843    swizzles[1] = inst->Predicate.SwizzleY;
 844    swizzles[2] = inst->Predicate.SwizzleZ;
 845    swizzles[3] = inst->Predicate.SwizzleW;
 846
 847    index = inst->Predicate.Index;
 848    assert(index < LP_MAX_TGSI_PREDS);
 849
 850    TGSI_FOR_EACH_CHANNEL( chan ) {
 851       unsigned swizzle = swizzles[chan];
 852
 853       /*
 854        * Only fetch the predicate register channels that are actually listed
 855        * in the swizzles
 856        */
 857       if (!unswizzled[swizzle]) {
 858          value = LLVMBuildLoad(builder,
 859                                bld->preds[index][swizzle], "");
 860
 861          /*
 862           * Convert the value to an integer mask.
 863           *
 864           * TODO: Short-circuit this comparison -- a D3D setp_xx instructions
 865           * is needlessly causing two comparisons due to storing the intermediate
 866           * result as float vector instead of an integer mask vector.
 867           */
 868          value = lp_build_compare(bld->bld_base.base.gallivm,
 869                                   bld->bld_base.base.type,
 870                                   PIPE_FUNC_NOTEQUAL,
 871                                   value,
 872                                   bld->bld_base.base.zero);
 873          if (inst->Predicate.Negate) {
 874             value = LLVMBuildNot(builder, value, "");
 875          }
 876
 877          unswizzled[swizzle] = value;
 878       } else {
 879          value = unswizzled[swizzle];
 880       }
 881
 882       pred[chan] = value;
 883    }
 884 }
 885
 886 /**
 887  * Register store.
 888  */
 889 static void
 890 emit_store_chan(
 891    struct lp_build_tgsi_context *bld_base,
 892    const struct tgsi_full_instruction *inst,
 893    unsigned index,
 894    unsigned chan_index,
 895    LLVMValueRef pred,
 896    LLVMValueRef value)
 897 {
 898    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
 899    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
 900    LLVMBuilderRef builder = gallivm->builder;
 901    const struct tgsi_full_dst_register *reg = &inst->Dst[index];
 902    struct lp_build_context *uint_bld = &bld_base->uint_bld;
 903    LLVMValueRef indirect_index = NULL;
 904    struct lp_build_context *bld_store;
 905    enum tgsi_opcode_type dtype = tgsi_opcode_infer_dst_type(inst->Instruction.Opcode);
 906
 907    switch (dtype) {
 908    default:
 909    case TGSI_TYPE_FLOAT:
 910    case TGSI_TYPE_UNTYPED:
 911       bld_store = &bld_base->base;
 912       break;
 913    case TGSI_TYPE_UNSIGNED:
 914       bld_store = &bld_base->uint_bld;
 915       break;
 916    case TGSI_TYPE_SIGNED:
 917       bld_store = &bld_base->int_bld;
 918       break;
 919    case TGSI_TYPE_DOUBLE:
 920    case TGSI_TYPE_VOID:
 921       assert(0);
 922       bld_store = NULL;
 923       break;
 924    }
 925
 926    switch( inst->Instruction.Saturate ) {
 927    case TGSI_SAT_NONE:
 928       break;
 929
 930    case TGSI_SAT_ZERO_ONE:
 931       value = lp_build_max(&bld->bld_base.base, value, bld->bld_base.base.zero);
 932       value = lp_build_min(&bld->bld_base.base, value, bld->bld_base.base.one);
 933       break;
 934
 935    case TGSI_SAT_MINUS_PLUS_ONE:
 936       value = lp_build_max(&bld->bld_base.base, value, lp_build_const_vec(bld->bld_base.base.gallivm, bld->bld_base.base.type, -1.0));
 937       value = lp_build_min(&bld->bld_base.base, value, bld->bld_base.base.one);
 938       break;
 939
 940    default:
 941       assert(0);
 942    }
 943
 944    if (reg->Register.Indirect) {
 945       indirect_index = get_indirect_index(bld,
 946                                           reg->Register.File,
 947                                           reg->Register.Index,
 948                                           &reg->Indirect);
 949    } else {
 950       assert(reg->Register.Index <=
 951                              bld->bld_base.info->file_max[reg->Register.File]);
 952    }
 953
 954    switch( reg->Register.File ) {
 955    case TGSI_FILE_OUTPUT:
 956       if (reg->Register.Indirect) {
 957          LLVMValueRef chan_vec =
 958             lp_build_const_int_vec(gallivm, uint_bld->type, chan_index);
 959          LLVMValueRef length_vec =
 960             lp_build_const_int_vec(gallivm, uint_bld->type, bld->bld_base.base.type.length);
 961          LLVMValueRef index_vec;  /* indexes into the temp registers */
 962          LLVMValueRef outputs_array;
 963          LLVMValueRef pixel_offsets;
 964          LLVMTypeRef float_ptr_type;
 965          int i;
 966
 967          /* build pixel offset vector: {0, 1, 2, 3, ...} */
 968          pixel_offsets = uint_bld->undef;
 969          for (i = 0; i < bld->bld_base.base.type.length; i++) {
 970             LLVMValueRef ii = lp_build_const_int32(gallivm, i);
 971             pixel_offsets = LLVMBuildInsertElement(builder, pixel_offsets,
 972                                                    ii, ii, "");
 973          }
 974
 975          /* index_vec = (indirect_index * 4 + chan_index) * length + offsets */
 976          index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
 977          index_vec = lp_build_add(uint_bld, index_vec, chan_vec);
 978          index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
 979          index_vec = lp_build_add(uint_bld, index_vec, pixel_offsets);
 980
 981          float_ptr_type =
 982             LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
 983          outputs_array = LLVMBuildBitCast(builder, bld->outputs_array,
 984                                           float_ptr_type, "");
 985
 986          /* Scatter store values into temp registers */
 987          emit_mask_scatter(bld, outputs_array, index_vec, value,
 988                            &bld->exec_mask, pred);
 989       }
 990       else {
 991          LLVMValueRef out_ptr = lp_get_output_ptr(bld, reg->Register.Index,
 992                                                chan_index);
 993          lp_exec_mask_store(&bld->exec_mask, bld_store, pred, value, out_ptr);
 994       }
 995       break;
 996
 997    case TGSI_FILE_TEMPORARY:
 998       if (reg->Register.Indirect) {
 999          LLVMValueRef chan_vec =
1000             lp_build_const_int_vec(gallivm, uint_bld->type, chan_index);
1001          LLVMValueRef length_vec =
1002             lp_build_const_int_vec(gallivm, uint_bld->type,
1003                                    bld->bld_base.base.type.length);
1004          LLVMValueRef index_vec;  /* indexes into the temp registers */
1005          LLVMValueRef temps_array;
1006          LLVMValueRef pixel_offsets;
1007          LLVMTypeRef float_ptr_type;
1008          int i;
1009
1010          /* build pixel offset vector: {0, 1, 2, 3, ...} */
1011          pixel_offsets = uint_bld->undef;
1012          for (i = 0; i < bld->bld_base.base.type.length; i++) {
1013             LLVMValueRef ii = lp_build_const_int32(gallivm, i);
1014             pixel_offsets = LLVMBuildInsertElement(builder, pixel_offsets,
1015                                                    ii, ii, "");
1016          }
1017
1018          /* index_vec = (indirect_index * 4 + chan_index) * length + offsets */
1019          index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
1020          index_vec = lp_build_add(uint_bld, index_vec, chan_vec);
1021          index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
1022          index_vec = lp_build_add(uint_bld, index_vec, pixel_offsets);
1023
1024          float_ptr_type =
1025             LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
1026          temps_array = LLVMBuildBitCast(builder, bld->temps_array,
1027                                         float_ptr_type, "");
1028
1029          /* Scatter store values into temp registers */
1030          emit_mask_scatter(bld, temps_array, index_vec, value,
1031                            &bld->exec_mask, pred);
1032       }
1033       else {
1034          LLVMValueRef temp_ptr;
1035
1036          switch (dtype) {
1037          case TGSI_TYPE_UNSIGNED:
1038          case TGSI_TYPE_SIGNED: {
1039             LLVMTypeRef itype = LLVMVectorType(LLVMInt32TypeInContext(gallivm->context), 4);
1040             LLVMTypeRef ivtype = LLVMPointerType(itype, 0);
1041             LLVMValueRef tint_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index,
1042                                                         chan_index);
1043             LLVMValueRef temp_value_ptr;
1044
1045             temp_ptr = LLVMBuildBitCast(builder, tint_ptr, ivtype, "");
1046             temp_value_ptr = LLVMBuildBitCast(builder, value, itype, "");
1047             value = temp_value_ptr;
1048             break;
1049          }
1050          default:
1051          case TGSI_TYPE_FLOAT:
1052          case TGSI_TYPE_UNTYPED:
1053             temp_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index,
1054                                            chan_index);
1055             break;
1056          }
1057
1058          lp_exec_mask_store(&bld->exec_mask, bld_store, pred, value, temp_ptr);
1059       }
1060       break;
1061
1062    case TGSI_FILE_ADDRESS:
1063       assert(dtype == TGSI_TYPE_SIGNED);
1064       assert(LLVMTypeOf(value) == bld_base->base.int_vec_type);
1065       lp_exec_mask_store(&bld->exec_mask, bld_store, pred, value,
1066                          bld->addr[reg->Register.Index][chan_index]);
1067       break;
1068
1069    case TGSI_FILE_PREDICATE:
1070       lp_exec_mask_store(&bld->exec_mask, bld_store, pred, value,
1071                          bld->preds[reg->Register.Index][chan_index]);
1072       break;
1073
1074    default:
1075       assert( 0 );
1076    }
1077 }
1078
1079 static void
1080 emit_store(
1081    struct lp_build_tgsi_context * bld_base,
1082    const struct tgsi_full_instruction * inst,
1083    const struct tgsi_opcode_info * info,
1084    LLVMValueRef dst[4])
1085
1086 {
1087    unsigned chan_index;
1088    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1089
1090    if(info->num_dst) {
1091       LLVMValueRef pred[TGSI_NUM_CHANNELS];
1092
1093       emit_fetch_predicate( bld, inst, pred );
1094
1095       TGSI_FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1096          emit_store_chan(bld_base, inst, 0, chan_index, pred[chan_index], dst[chan_index]);
1097       }
1098    }
1099 }
1100
1101 /**
1102  * High-level instruction translators.
1103  */
1104
1105 static void
1106 emit_tex( struct lp_build_tgsi_soa_context *bld,
1107           const struct tgsi_full_instruction *inst,
1108           enum lp_build_tex_modifier modifier,
1109           LLVMValueRef *texel)
1110 {
1111    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
1112    unsigned unit;
1113    LLVMValueRef lod_bias, explicit_lod;
1114    LLVMValueRef oow = NULL;
1115    LLVMValueRef coords[3];
1116    LLVMValueRef ddx[3];
1117    LLVMValueRef ddy[3];
1118    unsigned num_coords;
1119    unsigned i;
1120
1121    if (!bld->sampler) {
1122       _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
1123       for (i = 0; i < 4; i++) {
1124          texel[i] = bld->bld_base.base.undef;
1125       }
1126       return;
1127    }
1128
1129    switch (inst->Texture.Texture) {
1130    case TGSI_TEXTURE_1D:
1131       num_coords = 1;
1132       break;
1133    case TGSI_TEXTURE_1D_ARRAY:
1134    case TGSI_TEXTURE_2D:
1135    case TGSI_TEXTURE_RECT:
1136       num_coords = 2;
1137       break;
1138    case TGSI_TEXTURE_SHADOW1D:
1139    case TGSI_TEXTURE_SHADOW1D_ARRAY:
1140    case TGSI_TEXTURE_SHADOW2D:
1141    case TGSI_TEXTURE_SHADOWRECT:
1142    case TGSI_TEXTURE_2D_ARRAY:
1143    case TGSI_TEXTURE_3D:
1144    case TGSI_TEXTURE_CUBE:
1145       num_coords = 3;
1146       break;
1147    case TGSI_TEXTURE_SHADOW2D_ARRAY:
1148       num_coords = 4;
1149       break;
1150    default:
1151       assert(0);
1152       return;
1153    }
1154
1155    if (modifier == LP_BLD_TEX_MODIFIER_LOD_BIAS) {
1156       lod_bias = lp_build_emit_fetch( &bld->bld_base, inst, 0, 3 );
1157       explicit_lod = NULL;
1158    }
1159    else if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_LOD) {
1160       lod_bias = NULL;
1161       explicit_lod = lp_build_emit_fetch( &bld->bld_base, inst, 0, 3 );
1162    }
1163    else {
1164       lod_bias = NULL;
1165       explicit_lod = NULL;
1166    }
1167
1168    if (modifier == LP_BLD_TEX_MODIFIER_PROJECTED) {
1169       oow = lp_build_emit_fetch( &bld->bld_base, inst, 0, 3 );
1170       oow = lp_build_rcp(&bld->bld_base.base, oow);
1171    }
1172
1173    for (i = 0; i < num_coords; i++) {
1174       coords[i] = lp_build_emit_fetch( &bld->bld_base, inst, 0, i );
1175       if (modifier == LP_BLD_TEX_MODIFIER_PROJECTED)
1176          coords[i] = lp_build_mul(&bld->bld_base.base, coords[i], oow);
1177    }
1178    for (i = num_coords; i < 3; i++) {
1179       coords[i] = bld->bld_base.base.undef;
1180    }
1181
1182    if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV) {
1183       LLVMValueRef index0 = lp_build_const_int32(bld->bld_base.base.gallivm, 0);
1184       for (i = 0; i < num_coords; i++) {
1185          LLVMValueRef src1 = lp_build_emit_fetch( &bld->bld_base, inst, 1, i );
1186          LLVMValueRef src2 = lp_build_emit_fetch( &bld->bld_base, inst, 2, i );
1187          ddx[i] = LLVMBuildExtractElement(builder, src1, index0, "");
1188          ddy[i] = LLVMBuildExtractElement(builder, src2, index0, "");
1189       }
1190       unit = inst->Src[3].Register.Index;
1191    }  else {
1192       for (i = 0; i < num_coords; i++) {
1193          ddx[i] = lp_build_scalar_ddx( &bld->bld_base.base, coords[i] );
1194          ddy[i] = lp_build_scalar_ddy( &bld->bld_base.base, coords[i] );
1195       }
1196       unit = inst->Src[1].Register.Index;
1197    }
1198    for (i = num_coords; i < 3; i++) {
1199       ddx[i] = LLVMGetUndef(bld->bld_base.base.elem_type);
1200       ddy[i] = LLVMGetUndef(bld->bld_base.base.elem_type);
1201    }
1202
1203    bld->sampler->emit_fetch_texel(bld->sampler,
1204                                   bld->bld_base.base.gallivm,
1205                                   bld->bld_base.base.type,
1206                                   unit, num_coords, coords,
1207                                   ddx, ddy,
1208                                   lod_bias, explicit_lod,
1209                                   texel);
1210 }
1211
1212 static void
1213 emit_txq( struct lp_build_tgsi_soa_context *bld,
1214           const struct tgsi_full_instruction *inst,
1215           LLVMValueRef *sizes_out)
1216 {
1217    LLVMValueRef explicit_lod;
1218    unsigned num_coords, has_lod;
1219    unsigned i;
1220
1221    switch (inst->Texture.Texture) {
1222    case TGSI_TEXTURE_1D:
1223    case TGSI_TEXTURE_SHADOW1D:
1224    case TGSI_TEXTURE_SHADOW2D:
1225    case TGSI_TEXTURE_SHADOWCUBE:
1226       num_coords = 1;
1227       has_lod = 1;
1228       break;
1229    case TGSI_TEXTURE_2D:
1230    case TGSI_TEXTURE_CUBE:
1231    case TGSI_TEXTURE_1D_ARRAY:
1232    case TGSI_TEXTURE_SHADOW1D_ARRAY:
1233       num_coords = 2;
1234       has_lod = 1;
1235       break;
1236    case TGSI_TEXTURE_3D:
1237 // case TGSI_TEXTURE_CUBE_ARRAY:
1238 // case TGSI_TEXTURE_SHADOWCUBE_ARRAY:
1239    case TGSI_TEXTURE_2D_ARRAY:
1240    case TGSI_TEXTURE_SHADOW2D_ARRAY:
1241       num_coords = 3;
1242       has_lod = 1;
1243       break;
1244
1245    case TGSI_TEXTURE_BUFFER:
1246       num_coords = 1;
1247       has_lod = 0;
1248       break;
1249
1250    case TGSI_TEXTURE_RECT:
1251    case TGSI_TEXTURE_SHADOWRECT:
1252 // case TGSI_TEXTURE_2D_MS:
1253       num_coords = 2;
1254       has_lod = 0;
1255       break;
1256
1257 // case TGSI_TEXTURE_2D_MS_ARRAY:
1258 //    num_coords = 3;
1259 //    has_lod = 0;
1260 //    break;
1261
1262    default:
1263       assert(0);
1264       return;
1265    }
1266
1267    if (!bld->sampler) {
1268       _debug_printf("warning: found texture query instruction but no sampler generator supplied\n");
1269       for (i = 0; i < num_coords; i++)
1270          sizes_out[i] = bld->bld_base.base.undef;
1271       return;
1272    }
1273
1274    if (has_lod)
1275       explicit_lod = lp_build_emit_fetch( &bld->bld_base, inst, 0, 2 );
1276    else
1277       explicit_lod = NULL;
1278
1279    bld->sampler->emit_size_query(bld->sampler,
1280                                  bld->bld_base.base.gallivm,
1281                                  inst->Src[1].Register.Index,
1282                                  explicit_lod,
1283                                  sizes_out);
1284 }
1285
1286 static boolean
1287 near_end_of_shader(struct lp_build_tgsi_soa_context *bld,
1288                    int pc)
1289 {
1290    int i;
1291
1292    for (i = 0; i < 5; i++) {
1293       unsigned opcode;
1294
1295       if (pc + i >= bld->bld_base.info->num_instructions)
1296          return TRUE;
1297
1298       opcode = bld->bld_base.instructions[pc + i].Instruction.Opcode;
1299
1300       if (opcode == TGSI_OPCODE_END)
1301          return TRUE;
1302
1303       if (opcode == TGSI_OPCODE_TEX ||
1304           opcode == TGSI_OPCODE_TXP ||
1305           opcode == TGSI_OPCODE_TXD ||
1306           opcode == TGSI_OPCODE_TXB ||
1307           opcode == TGSI_OPCODE_TXL ||
1308           opcode == TGSI_OPCODE_TXF ||
1309           opcode == TGSI_OPCODE_TXQ ||
1310           opcode == TGSI_OPCODE_CAL ||
1311           opcode == TGSI_OPCODE_CALLNZ ||
1312           opcode == TGSI_OPCODE_IF ||
1313           opcode == TGSI_OPCODE_IFC ||
1314           opcode == TGSI_OPCODE_BGNLOOP ||
1315           opcode == TGSI_OPCODE_SWITCH)
1316          return FALSE;
1317    }
1318
1319    return TRUE;
1320 }
1321
1322
1323
1324 /**
1325  * Kill fragment if any of the src register values are negative.
1326  */
1327 static void
1328 emit_kil(
1329    struct lp_build_tgsi_soa_context *bld,
1330    const struct tgsi_full_instruction *inst,
1331    int pc)
1332 {
1333    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
1334    const struct tgsi_full_src_register *reg = &inst->Src[0];
1335    LLVMValueRef terms[TGSI_NUM_CHANNELS];
1336    LLVMValueRef mask;
1337    unsigned chan_index;
1338
1339    memset(&terms, 0, sizeof terms);
1340
1341    TGSI_FOR_EACH_CHANNEL( chan_index ) {
1342       unsigned swizzle;
1343
1344       /* Unswizzle channel */
1345       swizzle = tgsi_util_get_full_src_register_swizzle( reg, chan_index );
1346
1347       /* Check if the component has not been already tested. */
1348       assert(swizzle < TGSI_NUM_CHANNELS);
1349       if( !terms[swizzle] )
1350          /* TODO: change the comparison operator instead of setting the sign */
1351          terms[swizzle] =  lp_build_emit_fetch(&bld->bld_base, inst, 0, chan_index );
1352    }
1353
1354    mask = NULL;
1355    TGSI_FOR_EACH_CHANNEL( chan_index ) {
1356       if(terms[chan_index]) {
1357          LLVMValueRef chan_mask;
1358
1359          /*
1360           * If term < 0 then mask = 0 else mask = ~0.
1361           */
1362          chan_mask = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_GEQUAL, terms[chan_index], bld->bld_base.base.zero);
1363
1364          if(mask)
1365             mask = LLVMBuildAnd(builder, mask, chan_mask, "");
1366          else
1367             mask = chan_mask;
1368       }
1369    }
1370
1371    if(mask) {
1372       lp_build_mask_update(bld->mask, mask);
1373
1374       if (!near_end_of_shader(bld, pc))
1375          lp_build_mask_check(bld->mask);
1376    }
1377 }
1378
1379
1380 /**
1381  * Predicated fragment kill.
1382  * XXX Actually, we do an unconditional kill (as in tgsi_exec.c).
1383  * The only predication is the execution mask which will apply if
1384  * we're inside a loop or conditional.
1385  */
1386 static void
1387 emit_kilp(struct lp_build_tgsi_soa_context *bld,
1388           int pc)
1389 {
1390    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
1391    LLVMValueRef mask;
1392
1393    /* For those channels which are "alive", disable fragment shader
1394     * execution.
1395     */
1396    if (bld->exec_mask.has_mask) {
1397       mask = LLVMBuildNot(builder, bld->exec_mask.exec_mask, "kilp");
1398    }
1399    else {
1400       LLVMValueRef zero = LLVMConstNull(bld->bld_base.base.int_vec_type);
1401       mask = zero;
1402    }
1403
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  * Emit code which will dump the value of all the temporary registers
1413  * to stdout.
1414  */
1415 static void
1416 emit_dump_temps(struct lp_build_tgsi_soa_context *bld)
1417 {
1418    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
1419    LLVMBuilderRef builder = gallivm->builder;
1420    LLVMValueRef temp_ptr;
1421    LLVMValueRef i0 = lp_build_const_int32(gallivm, 0);
1422    LLVMValueRef i1 = lp_build_const_int32(gallivm, 1);
1423    LLVMValueRef i2 = lp_build_const_int32(gallivm, 2);
1424    LLVMValueRef i3 = lp_build_const_int32(gallivm, 3);
1425    int index;
1426    int n = bld->bld_base.info->file_max[TGSI_FILE_TEMPORARY];
1427
1428    for (index = 0; index < n; index++) {
1429       LLVMValueRef idx = lp_build_const_int32(gallivm, index);
1430       LLVMValueRef v[4][4], res;
1431       int chan;
1432
1433       lp_build_printf(gallivm, "TEMP[%d]:\n", idx);
1434
1435       for (chan = 0; chan < 4; chan++) {
1436          temp_ptr = lp_get_temp_ptr_soa(bld, index, chan);
1437          res = LLVMBuildLoad(builder, temp_ptr, "");
1438          v[chan][0] = LLVMBuildExtractElement(builder, res, i0, "");
1439          v[chan][1] = LLVMBuildExtractElement(builder, res, i1, "");
1440          v[chan][2] = LLVMBuildExtractElement(builder, res, i2, "");
1441          v[chan][3] = LLVMBuildExtractElement(builder, res, i3, "");
1442       }
1443
1444       lp_build_printf(gallivm, "  X: %f %f %f %f\n",
1445                       v[0][0], v[0][1], v[0][2], v[0][3]);
1446       lp_build_printf(gallivm, "  Y: %f %f %f %f\n",
1447                       v[1][0], v[1][1], v[1][2], v[1][3]);
1448       lp_build_printf(gallivm, "  Z: %f %f %f %f\n",
1449                       v[2][0], v[2][1], v[2][2], v[2][3]);
1450       lp_build_printf(gallivm, "  W: %f %f %f %f\n",
1451                       v[3][0], v[3][1], v[3][2], v[3][3]);
1452    }
1453 }
1454
1455
1456
1457 void
1458 lp_emit_declaration_soa(
1459    struct lp_build_tgsi_context *bld_base,
1460    const struct tgsi_full_declaration *decl)
1461 {
1462    struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
1463    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
1464    LLVMTypeRef vec_type = bld->bld_base.base.vec_type;
1465    const unsigned first = decl->Range.First;
1466    const unsigned last = decl->Range.Last;
1467    unsigned idx, i;
1468
1469    for (idx = first; idx <= last; ++idx) {
1470       assert(last <= bld->bld_base.info->file_max[decl->Declaration.File]);
1471       switch (decl->Declaration.File) {
1472       case TGSI_FILE_TEMPORARY:
1473          assert(idx < LP_MAX_TGSI_TEMPS);
1474          if (!(bld->indirect_files & (1 << TGSI_FILE_TEMPORARY))) {
1475             for (i = 0; i < TGSI_NUM_CHANNELS; i++)
1476                bld->temps[idx][i] = lp_build_alloca(gallivm, vec_type, "temp");
1477          }
1478          break;
1479
1480       case TGSI_FILE_OUTPUT:
1481          if (!(bld->indirect_files & (1 << TGSI_FILE_OUTPUT))) {
1482             for (i = 0; i < TGSI_NUM_CHANNELS; i++)
1483                bld->outputs[idx][i] = lp_build_alloca(gallivm,
1484                                                       vec_type, "output");
1485          }
1486          break;
1487
1488       case TGSI_FILE_ADDRESS:
1489          /* ADDR registers are the only allocated with an integer LLVM IR type,
1490           * as they are guaranteed to always have integers.
1491           * XXX: Not sure if this exception is worthwhile (or the whole idea of
1492           * an ADDR register for that matter).
1493           */
1494          assert(idx < LP_MAX_TGSI_ADDRS);
1495          for (i = 0; i < TGSI_NUM_CHANNELS; i++)
1496             bld->addr[idx][i] = lp_build_alloca(gallivm, bld_base->base.int_vec_type, "addr");
1497          break;
1498
1499       case TGSI_FILE_PREDICATE:
1500          assert(idx < LP_MAX_TGSI_PREDS);
1501          for (i = 0; i < TGSI_NUM_CHANNELS; i++)
1502             bld->preds[idx][i] = lp_build_alloca(gallivm, vec_type,
1503                                                  "predicate");
1504          break;
1505
1506       default:
1507          /* don't need to declare other vars */
1508          break;
1509       }
1510    }
1511 }
1512
1513
1514 void lp_emit_immediate_soa(
1515    struct lp_build_tgsi_context *bld_base,
1516    const struct tgsi_full_immediate *imm)
1517 {
1518    struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
1519    struct gallivm_state * gallivm = bld_base->base.gallivm;
1520
1521    /* simply copy the immediate values into the next immediates[] slot */
1522    unsigned i;
1523    const uint size = imm->Immediate.NrTokens - 1;
1524    assert(size <= 4);
1525    assert(bld->num_immediates < LP_MAX_TGSI_IMMEDIATES);
1526    switch (imm->Immediate.DataType) {
1527    case TGSI_IMM_FLOAT32:
1528       for( i = 0; i < size; ++i )
1529          bld->immediates[bld->num_immediates][i] =
1530             lp_build_const_vec(gallivm, bld_base->base.type, imm->u[i].Float);
1531
1532       break;
1533    case TGSI_IMM_UINT32:
1534       for( i = 0; i < size; ++i ) {
1535          LLVMValueRef tmp = lp_build_const_vec(gallivm, bld_base->uint_bld.type, imm->u[i].Uint);
1536          bld->immediates[bld->num_immediates][i] =
1537             LLVMConstBitCast(tmp, bld_base->base.vec_type);
1538       }
1539
1540       break;
1541    case TGSI_IMM_INT32:
1542       for( i = 0; i < size; ++i ) {
1543          LLVMValueRef tmp = lp_build_const_vec(gallivm, bld_base->int_bld.type, imm->u[i].Int);
1544          bld->immediates[bld->num_immediates][i] =
1545             LLVMConstBitCast(tmp, bld_base->base.vec_type);
1546       }
1547
1548       break;
1549    }
1550    for( i = size; i < 4; ++i )
1551       bld->immediates[bld->num_immediates][i] = bld_base->base.undef;
1552
1553    bld->num_immediates++;
1554 }
1555
1556 static void
1557 ddx_emit(
1558    const struct lp_build_tgsi_action * action,
1559    struct lp_build_tgsi_context * bld_base,
1560    struct lp_build_emit_data * emit_data)
1561 {
1562    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1563
1564    emit_fetch_deriv(bld, emit_data->args[0], NULL,
1565                     &emit_data->output[emit_data->chan], NULL);
1566 }
1567
1568 static void
1569 ddy_emit(
1570    const struct lp_build_tgsi_action * action,
1571    struct lp_build_tgsi_context * bld_base,
1572    struct lp_build_emit_data * emit_data)
1573 {
1574    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1575
1576    emit_fetch_deriv(bld, emit_data->args[0], NULL, NULL,
1577                     &emit_data->output[emit_data->chan]);
1578 }
1579
1580 static void
1581 kilp_emit(
1582    const struct lp_build_tgsi_action * action,
1583    struct lp_build_tgsi_context * bld_base,
1584    struct lp_build_emit_data * emit_data)
1585 {
1586    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1587
1588    emit_kilp(bld, bld_base->pc - 1);
1589 }
1590
1591 static void
1592 kil_emit(
1593    const struct lp_build_tgsi_action * action,
1594    struct lp_build_tgsi_context * bld_base,
1595    struct lp_build_emit_data * emit_data)
1596 {
1597    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1598
1599    emit_kil(bld, emit_data->inst, bld_base->pc - 1);
1600 }
1601
1602 static void
1603 tex_emit(
1604    const struct lp_build_tgsi_action * action,
1605    struct lp_build_tgsi_context * bld_base,
1606    struct lp_build_emit_data * emit_data)
1607 {
1608    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1609
1610    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_NONE, emit_data->output);
1611 }
1612
1613 static void
1614 txb_emit(
1615    const struct lp_build_tgsi_action * action,
1616    struct lp_build_tgsi_context * bld_base,
1617    struct lp_build_emit_data * emit_data)
1618 {
1619    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1620
1621    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_LOD_BIAS,
1622             emit_data->output);
1623 }
1624
1625 static void
1626 txd_emit(
1627    const struct lp_build_tgsi_action * action,
1628    struct lp_build_tgsi_context * bld_base,
1629    struct lp_build_emit_data * emit_data)
1630 {
1631    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1632
1633    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV,
1634             emit_data->output);
1635 }
1636
1637 static void
1638 txl_emit(
1639    const struct lp_build_tgsi_action * action,
1640    struct lp_build_tgsi_context * bld_base,
1641    struct lp_build_emit_data * emit_data)
1642 {
1643    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1644
1645    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD,
1646             emit_data->output);
1647 }
1648
1649 static void
1650 txp_emit(
1651    const struct lp_build_tgsi_action * action,
1652    struct lp_build_tgsi_context * bld_base,
1653    struct lp_build_emit_data * emit_data)
1654 {
1655    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1656
1657    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_PROJECTED,
1658             emit_data->output);
1659 }
1660
1661 static void
1662 txq_emit(
1663    const struct lp_build_tgsi_action * action,
1664    struct lp_build_tgsi_context * bld_base,
1665    struct lp_build_emit_data * emit_data)
1666 {
1667    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1668
1669    emit_txq(bld, emit_data->inst, emit_data->output);
1670 }
1671
1672 static void
1673 cal_emit(
1674    const struct lp_build_tgsi_action * action,
1675    struct lp_build_tgsi_context * bld_base,
1676    struct lp_build_emit_data * emit_data)
1677 {
1678    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1679
1680    lp_exec_mask_call(&bld->exec_mask, emit_data->inst->Label.Label,
1681                      &bld_base->pc);
1682 }
1683
1684 static void
1685 ret_emit(
1686    const struct lp_build_tgsi_action * action,
1687    struct lp_build_tgsi_context * bld_base,
1688    struct lp_build_emit_data * emit_data)
1689 {
1690    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1691
1692    lp_exec_mask_ret(&bld->exec_mask, &bld_base->pc);
1693 }
1694
1695 static void
1696 brk_emit(
1697    const struct lp_build_tgsi_action * action,
1698    struct lp_build_tgsi_context * bld_base,
1699    struct lp_build_emit_data * emit_data)
1700 {
1701    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1702
1703    lp_exec_break(&bld->exec_mask);
1704 }
1705
1706 static void
1707 if_emit(
1708    const struct lp_build_tgsi_action * action,
1709    struct lp_build_tgsi_context * bld_base,
1710    struct lp_build_emit_data * emit_data)
1711 {
1712    LLVMValueRef tmp;
1713    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1714
1715    tmp = lp_build_cmp(&bld_base->base, PIPE_FUNC_NOTEQUAL,
1716                       emit_data->args[0], bld->bld_base.base.zero);
1717    lp_exec_mask_cond_push(&bld->exec_mask, tmp);
1718 }
1719
1720 static void
1721 bgnloop_emit(
1722    const struct lp_build_tgsi_action * action,
1723    struct lp_build_tgsi_context * bld_base,
1724    struct lp_build_emit_data * emit_data)
1725 {
1726    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1727
1728    lp_exec_bgnloop(&bld->exec_mask);
1729 }
1730
1731 static void
1732 bgnsub_emit(
1733    const struct lp_build_tgsi_action * action,
1734    struct lp_build_tgsi_context * bld_base,
1735    struct lp_build_emit_data * emit_data)
1736 {
1737    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1738
1739    lp_exec_mask_bgnsub(&bld->exec_mask);
1740 }
1741
1742 static void
1743 else_emit(
1744    const struct lp_build_tgsi_action * action,
1745    struct lp_build_tgsi_context * bld_base,
1746    struct lp_build_emit_data * emit_data)
1747 {
1748    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1749
1750    lp_exec_mask_cond_invert(&bld->exec_mask);
1751 }
1752
1753 static void
1754 endif_emit(
1755    const struct lp_build_tgsi_action * action,
1756    struct lp_build_tgsi_context * bld_base,
1757    struct lp_build_emit_data * emit_data)
1758 {
1759    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1760
1761    lp_exec_mask_cond_pop(&bld->exec_mask);
1762 }
1763
1764 static void
1765 endloop_emit(
1766    const struct lp_build_tgsi_action * action,
1767    struct lp_build_tgsi_context * bld_base,
1768    struct lp_build_emit_data * emit_data)
1769 {
1770    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1771
1772    lp_exec_endloop(bld_base->base.gallivm, &bld->exec_mask);
1773 }
1774
1775 static void
1776 endsub_emit(
1777    const struct lp_build_tgsi_action * action,
1778    struct lp_build_tgsi_context * bld_base,
1779    struct lp_build_emit_data * emit_data)
1780 {
1781    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1782
1783    lp_exec_mask_endsub(&bld->exec_mask, &bld_base->pc);
1784 }
1785
1786 static void
1787 cont_emit(
1788    const struct lp_build_tgsi_action * action,
1789    struct lp_build_tgsi_context * bld_base,
1790    struct lp_build_emit_data * emit_data)
1791 {
1792    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1793
1794    lp_exec_continue(&bld->exec_mask);
1795 }
1796
1797 /* XXX: Refactor and move it to lp_bld_tgsi_action.c
1798  *
1799  * XXX: What do the comments about xmm registers mean?  Maybe they are left over
1800  * from old code, but there is no garauntee that LLVM will use those registers
1801  * for this code.
1802  *
1803  * XXX: There should be no calls to lp_build_emit_fetch in this function.  This
1804  * should be handled by the emit_data->fetch_args function. */
1805 static void
1806 nrm_emit(
1807    const struct lp_build_tgsi_action * action,
1808    struct lp_build_tgsi_context * bld_base,
1809    struct lp_build_emit_data * emit_data)
1810 {
1811    LLVMValueRef tmp0, tmp1;
1812    LLVMValueRef tmp4 = NULL;
1813    LLVMValueRef tmp5 = NULL;
1814    LLVMValueRef tmp6 = NULL;
1815    LLVMValueRef tmp7 = NULL;
1816    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1817
1818    uint dims = (emit_data->inst->Instruction.Opcode == TGSI_OPCODE_NRM) ? 3 : 4;
1819
1820   if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_X) ||
1821       TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Y) ||
1822       TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Z) ||
1823       (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_W) && dims == 4)) {
1824
1825       /* NOTE: Cannot use xmm regs 2/3 here (see emit_rsqrt() above). */
1826
1827       /* xmm4 = src.x */
1828       /* xmm0 = src.x * src.x */
1829       tmp0 = lp_build_emit_fetch(&bld->bld_base, emit_data->inst, 0, TGSI_CHAN_X);
1830       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_X)) {
1831          tmp4 = tmp0;
1832       }
1833       tmp0 = lp_build_mul( &bld->bld_base.base, tmp0, tmp0);
1834
1835       /* xmm5 = src.y */
1836       /* xmm0 = xmm0 + src.y * src.y */
1837       tmp1 = lp_build_emit_fetch(&bld->bld_base, emit_data->inst, 0, TGSI_CHAN_Y);
1838       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Y)) {
1839          tmp5 = tmp1;
1840       }
1841       tmp1 = lp_build_mul( &bld->bld_base.base, tmp1, tmp1);
1842       tmp0 = lp_build_add( &bld->bld_base.base, tmp0, tmp1);
1843
1844       /* xmm6 = src.z */
1845       /* xmm0 = xmm0 + src.z * src.z */
1846       tmp1 = lp_build_emit_fetch(&bld->bld_base, emit_data->inst, 0, TGSI_CHAN_Z);
1847       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Z)) {
1848          tmp6 = tmp1;
1849       }
1850       tmp1 = lp_build_mul( &bld->bld_base.base, tmp1, tmp1);
1851       tmp0 = lp_build_add( &bld->bld_base.base, tmp0, tmp1);
1852
1853       if (dims == 4) {
1854          /* xmm7 = src.w */
1855          /* xmm0 = xmm0 + src.w * src.w */
1856          tmp1 = lp_build_emit_fetch(&bld->bld_base, emit_data->inst, 0, TGSI_CHAN_W);
1857          if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_W)) {
1858             tmp7 = tmp1;
1859          }
1860          tmp1 = lp_build_mul( &bld->bld_base.base, tmp1, tmp1);
1861          tmp0 = lp_build_add( &bld->bld_base.base, tmp0, tmp1);
1862       }
1863       /* xmm1 = 1 / sqrt(xmm0) */
1864       tmp1 = lp_build_rsqrt( &bld->bld_base.base, tmp0);
1865        /* dst.x = xmm1 * src.x */
1866       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_X)) {
1867          emit_data->output[TGSI_CHAN_X] = lp_build_mul( &bld->bld_base.base, tmp4, tmp1);
1868       }
1869       /* dst.y = xmm1 * src.y */
1870       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Y)) {
1871          emit_data->output[TGSI_CHAN_Y] = lp_build_mul( &bld->bld_base.base, tmp5, tmp1);
1872       }
1873
1874       /* dst.z = xmm1 * src.z */
1875       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Z)) {
1876          emit_data->output[TGSI_CHAN_Z] = lp_build_mul( &bld->bld_base.base, tmp6, tmp1);
1877       }
1878       /* dst.w = xmm1 * src.w */
1879       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_X) && dims == 4) {
1880          emit_data->output[TGSI_CHAN_W] = lp_build_mul( &bld->bld_base.base, tmp7, tmp1);
1881       }
1882    }
1883
1884    /* dst.w = 1.0 */
1885    if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_W) && dims == 3) {
1886        emit_data->output[TGSI_CHAN_W] = bld->bld_base.base.one;
1887    }
1888 }
1889
1890 static void emit_prologue(struct lp_build_tgsi_context * bld_base)
1891 {
1892    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1893    struct gallivm_state * gallivm = bld_base->base.gallivm;
1894
1895    if (bld->indirect_files & (1 << TGSI_FILE_TEMPORARY)) {
1896       LLVMValueRef array_size =
1897          lp_build_const_int32(gallivm,
1898                          bld_base->info->file_max[TGSI_FILE_TEMPORARY] * 4 + 4);
1899       bld->temps_array = lp_build_array_alloca(gallivm,
1900                                               bld_base->base.vec_type, array_size,
1901                                               "temp_array");
1902    }
1903
1904    if (bld->indirect_files & (1 << TGSI_FILE_OUTPUT)) {
1905       LLVMValueRef array_size =
1906          lp_build_const_int32(gallivm,
1907                             bld_base->info->file_max[TGSI_FILE_OUTPUT] * 4 + 4);
1908       bld->outputs_array = lp_build_array_alloca(gallivm,
1909                                                 bld_base->base.vec_type, array_size,
1910                                                 "output_array");
1911    }
1912
1913    /* If we have indirect addressing in inputs we need to copy them into
1914     * our alloca array to be able to iterate over them */
1915    if (bld->indirect_files & (1 << TGSI_FILE_INPUT)) {
1916       unsigned index, chan;
1917       LLVMTypeRef vec_type = bld_base->base.vec_type;
1918       LLVMValueRef array_size = lp_build_const_int32(gallivm,
1919             bld_base->info->file_max[TGSI_FILE_INPUT]*4 + 4);
1920       bld->inputs_array = lp_build_array_alloca(gallivm,
1921                                                vec_type, array_size,
1922                                                "input_array");
1923
1924       assert(bld_base->info->num_inputs
1925                         <= bld_base->info->file_max[TGSI_FILE_INPUT] + 1);
1926
1927       for (index = 0; index < bld_base->info->num_inputs; ++index) {
1928          for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
1929             LLVMValueRef lindex =
1930                lp_build_const_int32(gallivm, index * 4 + chan);
1931             LLVMValueRef input_ptr =
1932                LLVMBuildGEP(gallivm->builder, bld->inputs_array,
1933                             &lindex, 1, "");
1934             LLVMValueRef value = bld->inputs[index][chan];
1935             if (value)
1936                LLVMBuildStore(gallivm->builder, value, input_ptr);
1937          }
1938       }
1939    }
1940 }
1941
1942 static void emit_epilogue(struct lp_build_tgsi_context * bld_base)
1943 {
1944    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1945
1946    if (0) {
1947       /* for debugging */
1948       emit_dump_temps(bld);
1949    }
1950
1951    /* If we have indirect addressing in outputs we need to copy our alloca array
1952     * to the outputs slots specified by the called */
1953    if (bld->indirect_files & (1 << TGSI_FILE_OUTPUT)) {
1954       unsigned index, chan;
1955       assert(bld_base->info->num_outputs <=
1956                         bld_base->info->file_max[TGSI_FILE_OUTPUT] + 1);
1957       for (index = 0; index < bld_base->info->num_outputs; ++index) {
1958          for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
1959             bld->outputs[index][chan] = lp_get_output_ptr(bld, index, chan);
1960          }
1961       }
1962    }
1963 }
1964
1965 void
1966 lp_build_tgsi_soa(struct gallivm_state *gallivm,
1967                   const struct tgsi_token *tokens,
1968                   struct lp_type type,
1969                   struct lp_build_mask_context *mask,
1970                   LLVMValueRef consts_ptr,
1971                   LLVMValueRef system_values_array,
1972                   const LLVMValueRef *pos,
1973                   const LLVMValueRef (*inputs)[TGSI_NUM_CHANNELS],
1974                   LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1975                   struct lp_build_sampler_soa *sampler,
1976                   const struct tgsi_shader_info *info)
1977 {
1978    struct lp_build_tgsi_soa_context bld;
1979
1980    struct lp_type res_type;
1981
1982    assert(type.length <= LP_MAX_VECTOR_LENGTH);
1983    memset(&res_type, 0, sizeof res_type);
1984    res_type.width = type.width;
1985    res_type.length = type.length;
1986    res_type.sign = 1;
1987
1988    /* Setup build context */
1989    memset(&bld, 0, sizeof bld);
1990    lp_build_context_init(&bld.bld_base.base, gallivm, type);
1991    lp_build_context_init(&bld.bld_base.uint_bld, gallivm, lp_uint_type(type));
1992    lp_build_context_init(&bld.bld_base.int_bld, gallivm, lp_int_type(type));
1993    lp_build_context_init(&bld.elem_bld, gallivm, lp_elem_type(type));
1994    bld.mask = mask;
1995    bld.pos = pos;
1996    bld.inputs = inputs;
1997    bld.outputs = outputs;
1998    bld.consts_ptr = consts_ptr;
1999    bld.sampler = sampler;
2000    bld.bld_base.info = info;
2001    bld.indirect_files = info->indirect_files;
2002
2003    bld.bld_base.soa = TRUE;
2004    bld.bld_base.emit_fetch_funcs[TGSI_FILE_CONSTANT] = emit_fetch_constant;
2005    bld.bld_base.emit_fetch_funcs[TGSI_FILE_IMMEDIATE] = emit_fetch_immediate;
2006    bld.bld_base.emit_fetch_funcs[TGSI_FILE_INPUT] = emit_fetch_input;
2007    bld.bld_base.emit_fetch_funcs[TGSI_FILE_TEMPORARY] = emit_fetch_temporary;
2008    bld.bld_base.emit_fetch_funcs[TGSI_FILE_SYSTEM_VALUE] = emit_fetch_system_value;
2009    bld.bld_base.emit_store = emit_store;
2010
2011    bld.bld_base.emit_declaration = lp_emit_declaration_soa;
2012    bld.bld_base.emit_immediate = lp_emit_immediate_soa;
2013
2014    bld.bld_base.emit_prologue = emit_prologue;
2015    bld.bld_base.emit_epilogue = emit_epilogue;
2016
2017    /* Set opcode actions */
2018    lp_set_default_actions_cpu(&bld.bld_base);
2019
2020    bld.bld_base.op_actions[TGSI_OPCODE_BGNLOOP].emit = bgnloop_emit;
2021    bld.bld_base.op_actions[TGSI_OPCODE_BGNSUB].emit = bgnsub_emit;
2022    bld.bld_base.op_actions[TGSI_OPCODE_BRK].emit = brk_emit;
2023    bld.bld_base.op_actions[TGSI_OPCODE_CAL].emit = cal_emit;
2024    bld.bld_base.op_actions[TGSI_OPCODE_CONT].emit = cont_emit;
2025    bld.bld_base.op_actions[TGSI_OPCODE_DDX].emit = ddx_emit;
2026    bld.bld_base.op_actions[TGSI_OPCODE_DDY].emit = ddy_emit;
2027    bld.bld_base.op_actions[TGSI_OPCODE_ELSE].emit = else_emit;
2028    bld.bld_base.op_actions[TGSI_OPCODE_ENDIF].emit = endif_emit;
2029    bld.bld_base.op_actions[TGSI_OPCODE_ENDLOOP].emit = endloop_emit;
2030    bld.bld_base.op_actions[TGSI_OPCODE_ENDSUB].emit = endsub_emit;
2031    bld.bld_base.op_actions[TGSI_OPCODE_IF].emit = if_emit;
2032    bld.bld_base.op_actions[TGSI_OPCODE_KIL].emit = kil_emit;
2033    bld.bld_base.op_actions[TGSI_OPCODE_KILP].emit = kilp_emit;
2034    bld.bld_base.op_actions[TGSI_OPCODE_NRM].emit = nrm_emit;
2035    bld.bld_base.op_actions[TGSI_OPCODE_NRM4].emit = nrm_emit;
2036    bld.bld_base.op_actions[TGSI_OPCODE_RET].emit = ret_emit;
2037    bld.bld_base.op_actions[TGSI_OPCODE_TEX].emit = tex_emit;
2038    bld.bld_base.op_actions[TGSI_OPCODE_TXB].emit = txb_emit;
2039    bld.bld_base.op_actions[TGSI_OPCODE_TXD].emit = txd_emit;
2040    bld.bld_base.op_actions[TGSI_OPCODE_TXL].emit = txl_emit;
2041    bld.bld_base.op_actions[TGSI_OPCODE_TXP].emit = txp_emit;
2042    bld.bld_base.op_actions[TGSI_OPCODE_TXQ].emit = txq_emit;
2043
2044    lp_exec_mask_init(&bld.exec_mask, &bld.bld_base.base);
2045
2046
2047    bld.system_values_array = system_values_array;
2048
2049    lp_build_tgsi_llvm(&bld.bld_base, tokens);
2050
2051    if (0) {
2052       LLVMBasicBlockRef block = LLVMGetInsertBlock(gallivm->builder);
2053       LLVMValueRef function = LLVMGetBasicBlockParent(block);
2054       debug_printf("11111111111111111111111111111 \n");
2055       tgsi_dump(tokens, 0);
2056       lp_debug_dump_value(function);
2057       debug_printf("2222222222222222222222222222 \n");
2058    }
2059
2060    if (0) {
2061       LLVMModuleRef module = LLVMGetGlobalParent(
2062          LLVMGetBasicBlockParent(LLVMGetInsertBlock(gallivm->builder)));
2063       LLVMDumpModule(module);
2064
2065    }
2066 }
2067
2068
2069 /**
2070  * Build up the system values array out of individual values such as
2071  * the instance ID, front-face, primitive ID, etc.  The shader info is
2072  * used to determine which system values are needed and where to put
2073  * them in the system values array.
2074  *
2075  * XXX only instance ID is implemented at this time.
2076  *
2077  * The system values register file is similar to the constants buffer.
2078  * Example declaration:
2079  *    DCL SV[0], INSTANCEID
2080  * Example instruction:
2081  *    MOVE foo, SV[0].xxxx;
2082  *
2083  * \return  LLVM float array (interpreted as float [][4])
2084  */
2085 LLVMValueRef
2086 lp_build_system_values_array(struct gallivm_state *gallivm,
2087                              const struct tgsi_shader_info *info,
2088                              LLVMValueRef instance_id,
2089                              LLVMValueRef facing)
2090 {
2091    LLVMValueRef size = lp_build_const_int32(gallivm, 4 * info->num_system_values);
2092    LLVMTypeRef float_t = LLVMFloatTypeInContext(gallivm->context);
2093    LLVMValueRef array = lp_build_array_alloca(gallivm, float_t,
2094                                               size, "sysvals_array");
2095    unsigned i;
2096
2097    for (i = 0; i < info->num_system_values; i++) {
2098       LLVMValueRef index = lp_build_const_int32(gallivm, i * 4);
2099       LLVMValueRef ptr, value = 0;
2100
2101       switch (info->system_value_semantic_name[i]) {
2102       case TGSI_SEMANTIC_INSTANCEID:
2103          /* convert instance ID from int to float */
2104          value = LLVMBuildSIToFP(gallivm->builder, instance_id, float_t,
2105                                  "sysval_instanceid");
2106          break;
2107       case TGSI_SEMANTIC_FACE:
2108          /* fall-through */
2109       default:
2110          assert(0 && "unexpected semantic in build_system_values_array()");
2111       }
2112
2113       ptr = LLVMBuildGEP(gallivm->builder, array, &index, 1, "");
2114       LLVMBuildStore(gallivm->builder, value, ptr);
2115    }
2116
2117    return array;
2118 }