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 #include "lp_bld_sample.h"
  66 #include "lp_bld_struct.h"
  67
  68
  69 static void lp_exec_mask_init(struct lp_exec_mask *mask, struct lp_build_context *bld)
  70 {
  71    LLVMTypeRef int_type = LLVMInt32TypeInContext(bld->gallivm->context);
  72    LLVMBuilderRef builder = bld->gallivm->builder;
  73
  74    mask->bld = bld;
  75    mask->has_mask = FALSE;
  76    mask->cond_stack_size = 0;
  77    mask->loop_stack_size = 0;
  78    mask->call_stack_size = 0;
  79
  80    mask->int_vec_type = lp_build_int_vec_type(bld->gallivm, mask->bld->type);
  81    mask->exec_mask = mask->ret_mask = mask->break_mask = mask->cont_mask = mask->cond_mask =
  82          LLVMConstAllOnes(mask->int_vec_type);
  83
  84    mask->loop_limiter = lp_build_alloca(bld->gallivm, int_type, "looplimiter");
  85
  86    LLVMBuildStore(
  87       builder,
  88       LLVMConstInt(int_type, LP_MAX_TGSI_LOOP_ITERATIONS, false),
  89       mask->loop_limiter);
  90 }
  91
  92 static void lp_exec_mask_update(struct lp_exec_mask *mask)
  93 {
  94    LLVMBuilderRef builder = mask->bld->gallivm->builder;
  95
  96    if (mask->loop_stack_size) {
  97       /*for loops we need to update the entire mask at runtime */
  98       LLVMValueRef tmp;
  99       assert(mask->break_mask);
 100       tmp = LLVMBuildAnd(builder,
 101                          mask->cont_mask,
 102                          mask->break_mask,
 103                          "maskcb");
 104       mask->exec_mask = LLVMBuildAnd(builder,
 105                                      mask->cond_mask,
 106                                      tmp,
 107                                      "maskfull");
 108    } else
 109       mask->exec_mask = mask->cond_mask;
 110
 111    if (mask->call_stack_size) {
 112       mask->exec_mask = LLVMBuildAnd(builder,
 113                                      mask->exec_mask,
 114                                      mask->ret_mask,
 115                                      "callmask");
 116    }
 117
 118    mask->has_mask = (mask->cond_stack_size > 0 ||
 119                      mask->loop_stack_size > 0 ||
 120                      mask->call_stack_size > 0);
 121 }
 122
 123 static void lp_exec_mask_cond_push(struct lp_exec_mask *mask,
 124                                    LLVMValueRef val)
 125 {
 126    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 127
 128    assert(mask->cond_stack_size < LP_MAX_TGSI_NESTING);
 129    if (mask->cond_stack_size == 0) {
 130       assert(mask->cond_mask == LLVMConstAllOnes(mask->int_vec_type));
 131    }
 132    mask->cond_stack[mask->cond_stack_size++] = mask->cond_mask;
 133    assert(LLVMTypeOf(val) == mask->int_vec_type);
 134    mask->cond_mask = LLVMBuildAnd(builder,
 135                                   mask->cond_mask,
 136                                   val,
 137                                   "");
 138    lp_exec_mask_update(mask);
 139 }
 140
 141 static void lp_exec_mask_cond_invert(struct lp_exec_mask *mask)
 142 {
 143    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 144    LLVMValueRef prev_mask;
 145    LLVMValueRef inv_mask;
 146
 147    assert(mask->cond_stack_size);
 148    prev_mask = mask->cond_stack[mask->cond_stack_size - 1];
 149    if (mask->cond_stack_size == 1) {
 150       assert(prev_mask == LLVMConstAllOnes(mask->int_vec_type));
 151    }
 152
 153    inv_mask = LLVMBuildNot(builder, mask->cond_mask, "");
 154
 155    mask->cond_mask = LLVMBuildAnd(builder,
 156                                   inv_mask,
 157                                   prev_mask, "");
 158    lp_exec_mask_update(mask);
 159 }
 160
 161 static void lp_exec_mask_cond_pop(struct lp_exec_mask *mask)
 162 {
 163    assert(mask->cond_stack_size);
 164    mask->cond_mask = mask->cond_stack[--mask->cond_stack_size];
 165    lp_exec_mask_update(mask);
 166 }
 167
 168 static void lp_exec_bgnloop(struct lp_exec_mask *mask)
 169 {
 170    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 171
 172    if (mask->loop_stack_size == 0) {
 173       assert(mask->loop_block == NULL);
 174       assert(mask->cont_mask == LLVMConstAllOnes(mask->int_vec_type));
 175       assert(mask->break_mask == LLVMConstAllOnes(mask->int_vec_type));
 176       assert(mask->break_var == NULL);
 177    }
 178
 179    assert(mask->loop_stack_size < LP_MAX_TGSI_NESTING);
 180
 181    mask->loop_stack[mask->loop_stack_size].loop_block = mask->loop_block;
 182    mask->loop_stack[mask->loop_stack_size].cont_mask = mask->cont_mask;
 183    mask->loop_stack[mask->loop_stack_size].break_mask = mask->break_mask;
 184    mask->loop_stack[mask->loop_stack_size].break_var = mask->break_var;
 185    ++mask->loop_stack_size;
 186
 187    mask->break_var = lp_build_alloca(mask->bld->gallivm, mask->int_vec_type, "");
 188    LLVMBuildStore(builder, mask->break_mask, mask->break_var);
 189
 190    mask->loop_block = lp_build_insert_new_block(mask->bld->gallivm, "bgnloop");
 191
 192    LLVMBuildBr(builder, mask->loop_block);
 193    LLVMPositionBuilderAtEnd(builder, mask->loop_block);
 194
 195    mask->break_mask = LLVMBuildLoad(builder, mask->break_var, "");
 196
 197    lp_exec_mask_update(mask);
 198 }
 199
 200 static void lp_exec_break(struct lp_exec_mask *mask)
 201 {
 202    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 203    LLVMValueRef exec_mask = LLVMBuildNot(builder,
 204                                          mask->exec_mask,
 205                                          "break");
 206
 207    mask->break_mask = LLVMBuildAnd(builder,
 208                                    mask->break_mask,
 209                                    exec_mask, "break_full");
 210
 211    lp_exec_mask_update(mask);
 212 }
 213
 214 static void lp_exec_continue(struct lp_exec_mask *mask)
 215 {
 216    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 217    LLVMValueRef exec_mask = LLVMBuildNot(builder,
 218                                          mask->exec_mask,
 219                                          "");
 220
 221    mask->cont_mask = LLVMBuildAnd(builder,
 222                                   mask->cont_mask,
 223                                   exec_mask, "");
 224
 225    lp_exec_mask_update(mask);
 226 }
 227
 228
 229 static void lp_exec_endloop(struct gallivm_state *gallivm,
 230                             struct lp_exec_mask *mask)
 231 {
 232    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 233    LLVMBasicBlockRef endloop;
 234    LLVMTypeRef int_type = LLVMInt32TypeInContext(mask->bld->gallivm->context);
 235    LLVMTypeRef reg_type = LLVMIntTypeInContext(gallivm->context,
 236                                                mask->bld->type.width *
 237                                                mask->bld->type.length);
 238    LLVMValueRef i1cond, i2cond, icond, limiter;
 239
 240    assert(mask->break_mask);
 241
 242    /*
 243     * Restore the cont_mask, but don't pop
 244     */
 245    assert(mask->loop_stack_size);
 246    mask->cont_mask = mask->loop_stack[mask->loop_stack_size - 1].cont_mask;
 247    lp_exec_mask_update(mask);
 248
 249    /*
 250     * Unlike the continue mask, the break_mask must be preserved across loop
 251     * iterations
 252     */
 253    LLVMBuildStore(builder, mask->break_mask, mask->break_var);
 254
 255    /* Decrement the loop limiter */
 256    limiter = LLVMBuildLoad(builder, mask->loop_limiter, "");
 257
 258    limiter = LLVMBuildSub(
 259       builder,
 260       limiter,
 261       LLVMConstInt(int_type, 1, false),
 262       "");
 263
 264    LLVMBuildStore(builder, limiter, mask->loop_limiter);
 265
 266    /* i1cond = (mask != 0) */
 267    i1cond = LLVMBuildICmp(
 268       builder,
 269       LLVMIntNE,
 270       LLVMBuildBitCast(builder, mask->exec_mask, reg_type, ""),
 271       LLVMConstNull(reg_type), "");
 272
 273    /* i2cond = (looplimiter > 0) */
 274    i2cond = LLVMBuildICmp(
 275       builder,
 276       LLVMIntSGT,
 277       limiter,
 278       LLVMConstNull(int_type), "");
 279
 280    /* if( i1cond && i2cond ) */
 281    icond = LLVMBuildAnd(builder, i1cond, i2cond, "");
 282
 283    endloop = lp_build_insert_new_block(mask->bld->gallivm, "endloop");
 284
 285    LLVMBuildCondBr(builder,
 286                    icond, mask->loop_block, endloop);
 287
 288    LLVMPositionBuilderAtEnd(builder, endloop);
 289
 290    assert(mask->loop_stack_size);
 291    --mask->loop_stack_size;
 292    mask->loop_block = mask->loop_stack[mask->loop_stack_size].loop_block;
 293    mask->cont_mask = mask->loop_stack[mask->loop_stack_size].cont_mask;
 294    mask->break_mask = mask->loop_stack[mask->loop_stack_size].break_mask;
 295    mask->break_var = mask->loop_stack[mask->loop_stack_size].break_var;
 296
 297    lp_exec_mask_update(mask);
 298 }
 299
 300 /* stores val into an address pointed to by dst.
 301  * mask->exec_mask is used to figure out which bits of val
 302  * should be stored into the address
 303  * (0 means don't store this bit, 1 means do store).
 304  */
 305 static void lp_exec_mask_store(struct lp_exec_mask *mask,
 306                                struct lp_build_context *bld_store,
 307                                LLVMValueRef pred,
 308                                LLVMValueRef val,
 309                                LLVMValueRef dst)
 310 {
 311    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 312
 313    /* Mix the predicate and execution mask */
 314    if (mask->has_mask) {
 315       if (pred) {
 316          pred = LLVMBuildAnd(builder, pred, mask->exec_mask, "");
 317       } else {
 318          pred = mask->exec_mask;
 319       }
 320    }
 321
 322    if (pred) {
 323       LLVMValueRef real_val, dst_val;
 324
 325       dst_val = LLVMBuildLoad(builder, dst, "");
 326       real_val = lp_build_select(bld_store,
 327                                  pred,
 328                                  val, dst_val);
 329
 330       LLVMBuildStore(builder, real_val, dst);
 331    } else
 332       LLVMBuildStore(builder, val, dst);
 333 }
 334
 335 static void lp_exec_mask_call(struct lp_exec_mask *mask,
 336                               int func,
 337                               int *pc)
 338 {
 339    assert(mask->call_stack_size < LP_MAX_TGSI_NESTING);
 340    mask->call_stack[mask->call_stack_size].pc = *pc;
 341    mask->call_stack[mask->call_stack_size].ret_mask = mask->ret_mask;
 342    mask->call_stack_size++;
 343    *pc = func;
 344 }
 345
 346 static void lp_exec_mask_ret(struct lp_exec_mask *mask, int *pc)
 347 {
 348    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 349    LLVMValueRef exec_mask;
 350
 351    if (mask->call_stack_size == 0) {
 352       /* returning from main() */
 353       *pc = -1;
 354       return;
 355    }
 356    exec_mask = LLVMBuildNot(builder,
 357                             mask->exec_mask,
 358                             "ret");
 359
 360    mask->ret_mask = LLVMBuildAnd(builder,
 361                                  mask->ret_mask,
 362                                  exec_mask, "ret_full");
 363
 364    lp_exec_mask_update(mask);
 365 }
 366
 367 static void lp_exec_mask_bgnsub(struct lp_exec_mask *mask)
 368 {
 369 }
 370
 371 static void lp_exec_mask_endsub(struct lp_exec_mask *mask, int *pc)
 372 {
 373    assert(mask->call_stack_size);
 374    mask->call_stack_size--;
 375    *pc = mask->call_stack[mask->call_stack_size].pc;
 376    mask->ret_mask = mask->call_stack[mask->call_stack_size].ret_mask;
 377    lp_exec_mask_update(mask);
 378 }
 379
 380
 381 /**
 382  * Return pointer to a temporary register channel (src or dest).
 383  * Note that indirect addressing cannot be handled here.
 384  * \param index  which temporary register
 385  * \param chan  which channel of the temp register.
 386  */
 387 LLVMValueRef
 388 lp_get_temp_ptr_soa(struct lp_build_tgsi_soa_context *bld,
 389              unsigned index,
 390              unsigned chan)
 391 {
 392    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
 393    assert(chan < 4);
 394    if (bld->indirect_files & (1 << TGSI_FILE_TEMPORARY)) {
 395       LLVMValueRef lindex = lp_build_const_int32(bld->bld_base.base.gallivm, index * 4 + chan);
 396       return LLVMBuildGEP(builder, bld->temps_array, &lindex, 1, "");
 397    }
 398    else {
 399       return bld->temps[index][chan];
 400    }
 401 }
 402
 403 /**
 404  * Return pointer to a output register channel (src or dest).
 405  * Note that indirect addressing cannot be handled here.
 406  * \param index  which output register
 407  * \param chan  which channel of the output register.
 408  */
 409 LLVMValueRef
 410 lp_get_output_ptr(struct lp_build_tgsi_soa_context *bld,
 411                unsigned index,
 412                unsigned chan)
 413 {
 414    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
 415    assert(chan < 4);
 416    if (bld->indirect_files & (1 << TGSI_FILE_OUTPUT)) {
 417       LLVMValueRef lindex = lp_build_const_int32(bld->bld_base.base.gallivm,
 418                                                  index * 4 + chan);
 419       return LLVMBuildGEP(builder, bld->outputs_array, &lindex, 1, "");
 420    }
 421    else {
 422       return bld->outputs[index][chan];
 423    }
 424 }
 425
 426 /**
 427  * Gather vector.
 428  * XXX the lp_build_gather() function should be capable of doing this
 429  * with a little work.
 430  */
 431 static LLVMValueRef
 432 build_gather(struct lp_build_context *bld,
 433              LLVMValueRef base_ptr,
 434              LLVMValueRef indexes)
 435 {
 436    LLVMBuilderRef builder = bld->gallivm->builder;
 437    LLVMValueRef res = bld->undef;
 438    unsigned i;
 439
 440    /*
 441     * Loop over elements of index_vec, load scalar value, insert it into 'res'.
 442     */
 443    for (i = 0; i < bld->type.length; i++) {
 444       LLVMValueRef ii = lp_build_const_int32(bld->gallivm, i);
 445       LLVMValueRef index = LLVMBuildExtractElement(builder,
 446                                                    indexes, ii, "");
 447       LLVMValueRef scalar_ptr = LLVMBuildGEP(builder, base_ptr,
 448                                              &index, 1, "gather_ptr");
 449       LLVMValueRef scalar = LLVMBuildLoad(builder, scalar_ptr, "");
 450
 451       res = LLVMBuildInsertElement(builder, res, scalar, ii, "");
 452    }
 453
 454    return res;
 455 }
 456
 457
 458 /**
 459  * Scatter/store vector.
 460  */
 461 static void
 462 emit_mask_scatter(struct lp_build_tgsi_soa_context *bld,
 463                   LLVMValueRef base_ptr,
 464                   LLVMValueRef indexes,
 465                   LLVMValueRef values,
 466                   struct lp_exec_mask *mask,
 467                   LLVMValueRef pred)
 468 {
 469    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
 470    LLVMBuilderRef builder = gallivm->builder;
 471    unsigned i;
 472
 473    /* Mix the predicate and execution mask */
 474    if (mask->has_mask) {
 475       if (pred) {
 476          pred = LLVMBuildAnd(builder, pred, mask->exec_mask, "");
 477       }
 478       else {
 479          pred = mask->exec_mask;
 480       }
 481    }
 482
 483    /*
 484     * Loop over elements of index_vec, store scalar value.
 485     */
 486    for (i = 0; i < bld->bld_base.base.type.length; i++) {
 487       LLVMValueRef ii = lp_build_const_int32(gallivm, i);
 488       LLVMValueRef index = LLVMBuildExtractElement(builder, indexes, ii, "");
 489       LLVMValueRef scalar_ptr = LLVMBuildGEP(builder, base_ptr, &index, 1, "scatter_ptr");
 490       LLVMValueRef val = LLVMBuildExtractElement(builder, values, ii, "scatter_val");
 491       LLVMValueRef scalar_pred = pred ?
 492          LLVMBuildExtractElement(builder, pred, ii, "scatter_pred") : NULL;
 493
 494       if (0)
 495          lp_build_printf(gallivm, "scatter %d: val %f at %d %p\n",
 496                          ii, val, index, scalar_ptr);
 497
 498       if (scalar_pred) {
 499          LLVMValueRef real_val, dst_val;
 500          dst_val = LLVMBuildLoad(builder, scalar_ptr, "");
 501          real_val = lp_build_select(&bld->elem_bld, scalar_pred, val, dst_val);
 502          LLVMBuildStore(builder, real_val, scalar_ptr);
 503       }
 504       else {
 505          LLVMBuildStore(builder, val, scalar_ptr);
 506       }
 507    }
 508 }
 509
 510
 511 /**
 512  * Read the current value of the ADDR register, convert the floats to
 513  * ints, add the base index and return the vector of offsets.
 514  * The offsets will be used to index into the constant buffer or
 515  * temporary register file.
 516  */
 517 static LLVMValueRef
 518 get_indirect_index(struct lp_build_tgsi_soa_context *bld,
 519                    unsigned reg_file, unsigned reg_index,
 520                    const struct tgsi_src_register *indirect_reg)
 521 {
 522    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
 523    struct lp_build_context *uint_bld = &bld->bld_base.uint_bld;
 524    /* always use X component of address register */
 525    unsigned swizzle = indirect_reg->SwizzleX;
 526    LLVMValueRef base;
 527    LLVMValueRef rel;
 528    LLVMValueRef max_index;
 529    LLVMValueRef index;
 530
 531    assert(bld->indirect_files & (1 << reg_file));
 532
 533    base = lp_build_const_int_vec(bld->bld_base.base.gallivm, uint_bld->type, reg_index);
 534
 535    assert(swizzle < 4);
 536    rel = LLVMBuildLoad(builder,
 537                         bld->addr[indirect_reg->Index][swizzle],
 538                         "load addr reg");
 539
 540    index = lp_build_add(uint_bld, base, rel);
 541
 542    max_index = lp_build_const_int_vec(bld->bld_base.base.gallivm,
 543                                       uint_bld->type,
 544                                       bld->bld_base.info->file_max[reg_file]);
 545
 546    assert(!uint_bld->type.sign);
 547    index = lp_build_min(uint_bld, index, max_index);
 548
 549    return index;
 550 }
 551
 552 static struct lp_build_context *
 553 stype_to_fetch(struct lp_build_tgsi_context * bld_base,
 554                enum tgsi_opcode_type stype)
 555 {
 556    struct lp_build_context *bld_fetch;
 557
 558    switch (stype) {
 559    case TGSI_TYPE_FLOAT:
 560    case TGSI_TYPE_UNTYPED:
 561       bld_fetch = &bld_base->base;
 562       break;
 563    case TGSI_TYPE_UNSIGNED:
 564       bld_fetch = &bld_base->uint_bld;
 565       break;
 566    case TGSI_TYPE_SIGNED:
 567       bld_fetch = &bld_base->int_bld;
 568       break;
 569    case TGSI_TYPE_VOID:
 570    case TGSI_TYPE_DOUBLE:
 571    default:
 572       assert(0);
 573       bld_fetch = NULL;
 574       break;
 575    }
 576    return bld_fetch;
 577 }
 578
 579 static LLVMValueRef
 580 emit_fetch_constant(
 581    struct lp_build_tgsi_context * bld_base,
 582    const struct tgsi_full_src_register * reg,
 583    enum tgsi_opcode_type stype,
 584    unsigned swizzle)
 585 {
 586    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
 587    struct gallivm_state *gallivm = bld_base->base.gallivm;
 588    LLVMBuilderRef builder = gallivm->builder;
 589    struct lp_build_context *uint_bld = &bld_base->uint_bld;
 590    LLVMValueRef indirect_index = NULL;
 591    struct lp_build_context *bld_fetch = stype_to_fetch(bld_base, stype);
 592    unsigned dimension = 0;
 593    LLVMValueRef dimension_index;
 594    LLVMValueRef consts_ptr;
 595
 596    /* XXX: Handle fetching xyzw components as a vector */
 597    assert(swizzle != ~0);
 598
 599    if (reg->Register.Dimension) {
 600       assert(!reg->Dimension.Indirect);
 601       dimension = reg->Dimension.Index;
 602       assert(dimension < LP_MAX_TGSI_CONST_BUFFERS);
 603    }
 604
 605    dimension_index = lp_build_const_int32(gallivm, dimension);
 606    consts_ptr = lp_build_array_get(gallivm, bld->consts_ptr, dimension_index);
 607
 608    if (reg->Register.Indirect) {
 609       indirect_index = get_indirect_index(bld,
 610                                           reg->Register.File,
 611                                           reg->Register.Index,
 612                                           &reg->Indirect);
 613    }
 614
 615    if (reg->Register.Indirect) {
 616       LLVMValueRef swizzle_vec =
 617          lp_build_const_int_vec(bld->bld_base.base.gallivm, uint_bld->type, swizzle);
 618       LLVMValueRef index_vec;  /* index into the const buffer */
 619
 620       /* index_vec = indirect_index * 4 + swizzle */
 621       index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
 622       index_vec = lp_build_add(uint_bld, index_vec, swizzle_vec);
 623
 624       /* Gather values from the constant buffer */
 625       return build_gather(bld_fetch, consts_ptr, index_vec);
 626    }
 627    else {
 628       LLVMValueRef index;  /* index into the const buffer */
 629       LLVMValueRef scalar, scalar_ptr;
 630
 631       index = lp_build_const_int32(gallivm, reg->Register.Index*4 + swizzle);
 632
 633       scalar_ptr = LLVMBuildGEP(builder, consts_ptr,
 634                                    &index, 1, "");
 635
 636       if (stype != TGSI_TYPE_FLOAT && stype != TGSI_TYPE_UNTYPED) {
 637          LLVMTypeRef ivtype = LLVMPointerType(LLVMInt32TypeInContext(gallivm->context), 0);
 638          LLVMValueRef temp_ptr;
 639          temp_ptr = LLVMBuildBitCast(builder, scalar_ptr, ivtype, "");
 640          scalar = LLVMBuildLoad(builder, temp_ptr, "");
 641       } else
 642          scalar = LLVMBuildLoad(builder, scalar_ptr, "");
 643
 644       return lp_build_broadcast_scalar(bld_fetch, scalar);
 645    }
 646 }
 647
 648 static LLVMValueRef
 649 emit_fetch_immediate(
 650    struct lp_build_tgsi_context * bld_base,
 651    const struct tgsi_full_src_register * reg,
 652    enum tgsi_opcode_type stype,
 653    unsigned swizzle)
 654 {
 655    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
 656    LLVMValueRef res = bld->immediates[reg->Register.Index][swizzle];
 657    assert(res);
 658
 659    if (stype == TGSI_TYPE_UNSIGNED) {
 660       res = LLVMConstBitCast(res, bld_base->uint_bld.vec_type);
 661    } else if (stype == TGSI_TYPE_SIGNED) {
 662       res = LLVMConstBitCast(res, bld_base->int_bld.vec_type);
 663    }
 664    return res;
 665 }
 666
 667 static LLVMValueRef
 668 emit_fetch_input(
 669    struct lp_build_tgsi_context * bld_base,
 670    const struct tgsi_full_src_register * reg,
 671    enum tgsi_opcode_type stype,
 672    unsigned swizzle)
 673 {
 674    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
 675    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
 676    LLVMBuilderRef builder = gallivm->builder;
 677    struct lp_build_context *uint_bld = &bld_base->uint_bld;
 678    LLVMValueRef indirect_index = NULL;
 679    LLVMValueRef res;
 680
 681    if (reg->Register.Indirect) {
 682       indirect_index = get_indirect_index(bld,
 683                                           reg->Register.File,
 684                                           reg->Register.Index,
 685                                           &reg->Indirect);
 686    }
 687
 688    if (reg->Register.Indirect) {
 689       LLVMValueRef swizzle_vec =
 690          lp_build_const_int_vec(gallivm, uint_bld->type, swizzle);
 691       LLVMValueRef length_vec =
 692          lp_build_const_int_vec(gallivm, uint_bld->type, bld->bld_base.base.type.length);
 693       LLVMValueRef index_vec;  /* index into the const buffer */
 694       LLVMValueRef inputs_array;
 695       LLVMTypeRef float4_ptr_type;
 696
 697       /* index_vec = (indirect_index * 4 + swizzle) * length */
 698       index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
 699       index_vec = lp_build_add(uint_bld, index_vec, swizzle_vec);
 700       index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
 701
 702       /* cast inputs_array pointer to float* */
 703       float4_ptr_type = LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
 704       inputs_array = LLVMBuildBitCast(builder, bld->inputs_array,
 705                                          float4_ptr_type, "");
 706
 707       /* Gather values from the temporary register array */
 708       res = build_gather(&bld_base->base, inputs_array, index_vec);
 709    } else {
 710       if (bld->indirect_files & (1 << TGSI_FILE_INPUT)) {
 711          LLVMValueRef lindex = lp_build_const_int32(gallivm,
 712                                         reg->Register.Index * 4 + swizzle);
 713          LLVMValueRef input_ptr =  LLVMBuildGEP(builder,
 714                                                 bld->inputs_array, &lindex, 1, "");
 715          res = LLVMBuildLoad(builder, input_ptr, "");
 716       }
 717       else {
 718          res = bld->inputs[reg->Register.Index][swizzle];
 719       }
 720    }
 721
 722    assert(res);
 723
 724    if (stype == TGSI_TYPE_UNSIGNED) {
 725       res = LLVMBuildBitCast(builder, res, bld_base->uint_bld.vec_type, "");
 726    } else if (stype == TGSI_TYPE_SIGNED) {
 727       res = LLVMBuildBitCast(builder, res, bld_base->int_bld.vec_type, "");
 728    }
 729
 730    return res;
 731 }
 732
 733 static LLVMValueRef
 734 emit_fetch_temporary(
 735    struct lp_build_tgsi_context * bld_base,
 736    const struct tgsi_full_src_register * reg,
 737    enum tgsi_opcode_type stype,
 738    unsigned swizzle)
 739 {
 740    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
 741    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
 742    LLVMBuilderRef builder = gallivm->builder;
 743    struct lp_build_context *uint_bld = &bld_base->uint_bld;
 744    LLVMValueRef indirect_index = NULL;
 745    LLVMValueRef res;
 746
 747    if (reg->Register.Indirect) {
 748       indirect_index = get_indirect_index(bld,
 749                                           reg->Register.File,
 750                                           reg->Register.Index,
 751                                           &reg->Indirect);
 752    }
 753
 754    if (reg->Register.Indirect) {
 755       LLVMValueRef swizzle_vec =
 756          lp_build_const_int_vec(bld->bld_base.base.gallivm, uint_bld->type, swizzle);
 757       LLVMValueRef length_vec =
 758          lp_build_const_int_vec(bld->bld_base.base.gallivm, uint_bld->type,
 759                                 bld->bld_base.base.type.length);
 760       LLVMValueRef index_vec;  /* index into the const buffer */
 761       LLVMValueRef temps_array;
 762       LLVMTypeRef float4_ptr_type;
 763
 764       /* index_vec = (indirect_index * 4 + swizzle) * length */
 765       index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
 766       index_vec = lp_build_add(uint_bld, index_vec, swizzle_vec);
 767       index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
 768
 769       /* cast temps_array pointer to float* */
 770       float4_ptr_type = LLVMPointerType(LLVMFloatTypeInContext(bld->bld_base.base.gallivm->context), 0);
 771       temps_array = LLVMBuildBitCast(builder, bld->temps_array,
 772                                      float4_ptr_type, "");
 773
 774       /* Gather values from the temporary register array */
 775       res = build_gather(&bld_base->base, temps_array, index_vec);
 776    }
 777    else {
 778       LLVMValueRef temp_ptr;
 779       if (stype != TGSI_TYPE_FLOAT && stype != TGSI_TYPE_UNTYPED) {
 780          LLVMTypeRef itype = LLVMPointerType(bld->bld_base.int_bld.vec_type, 0);
 781          LLVMValueRef tint_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index,
 782                                                      swizzle);
 783          temp_ptr = LLVMBuildBitCast(builder, tint_ptr, itype, "");
 784       } else
 785          temp_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index, swizzle);
 786       res = LLVMBuildLoad(builder, temp_ptr, "");
 787       if (!res)
 788          return bld->bld_base.base.undef;
 789    }
 790
 791    return res;
 792 }
 793
 794 static LLVMValueRef
 795 emit_fetch_system_value(
 796    struct lp_build_tgsi_context * bld_base,
 797    const struct tgsi_full_src_register * reg,
 798    enum tgsi_opcode_type stype,
 799    unsigned swizzle)
 800 {
 801    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
 802    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
 803    const struct tgsi_shader_info *info = bld->bld_base.info;
 804    LLVMBuilderRef builder = gallivm->builder;
 805    LLVMValueRef res;
 806    enum tgsi_opcode_type atype; // Actual type of the value
 807
 808    assert(!reg->Register.Indirect);
 809
 810    switch (info->system_value_semantic_name[reg->Register.Index]) {
 811    case TGSI_SEMANTIC_INSTANCEID:
 812       res = lp_build_broadcast_scalar(&bld_base->uint_bld, bld->system_values.instance_id);
 813       atype = TGSI_TYPE_UNSIGNED;
 814       break;
 815
 816    case TGSI_SEMANTIC_VERTEXID:
 817       res = bld->system_values.vertex_id;
 818       atype = TGSI_TYPE_UNSIGNED;
 819       break;
 820
 821    default:
 822       assert(!"unexpected semantic in emit_fetch_system_value");
 823       res = bld_base->base.zero;
 824       atype = TGSI_TYPE_FLOAT;
 825       break;
 826    }
 827
 828    if (atype != stype) {
 829       if (stype == TGSI_TYPE_FLOAT) {
 830          res = LLVMBuildBitCast(builder, res, bld_base->base.vec_type, "");
 831       } else if (stype == TGSI_TYPE_UNSIGNED) {
 832          res = LLVMBuildBitCast(builder, res, bld_base->uint_bld.vec_type, "");
 833       } else if (stype == TGSI_TYPE_SIGNED) {
 834          res = LLVMBuildBitCast(builder, res, bld_base->int_bld.vec_type, "");
 835       }
 836    }
 837
 838    return res;
 839 }
 840
 841 /**
 842  * Register fetch with derivatives.
 843  */
 844 static void
 845 emit_fetch_deriv(
 846    struct lp_build_tgsi_soa_context *bld,
 847    LLVMValueRef src,
 848    LLVMValueRef *res,
 849    LLVMValueRef *ddx,
 850    LLVMValueRef *ddy)
 851 {
 852    if(res)
 853       *res = src;
 854
 855    /* TODO: use interpolation coeffs for inputs */
 856
 857    if(ddx)
 858       *ddx = lp_build_ddx(&bld->bld_base.base, src);
 859
 860    if(ddy)
 861       *ddy = lp_build_ddy(&bld->bld_base.base, src);
 862 }
 863
 864
 865 /**
 866  * Predicate.
 867  */
 868 static void
 869 emit_fetch_predicate(
 870    struct lp_build_tgsi_soa_context *bld,
 871    const struct tgsi_full_instruction *inst,
 872    LLVMValueRef *pred)
 873 {
 874    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
 875    unsigned index;
 876    unsigned char swizzles[4];
 877    LLVMValueRef unswizzled[4] = {NULL, NULL, NULL, NULL};
 878    LLVMValueRef value;
 879    unsigned chan;
 880
 881    if (!inst->Instruction.Predicate) {
 882       TGSI_FOR_EACH_CHANNEL( chan ) {
 883          pred[chan] = NULL;
 884       }
 885       return;
 886    }
 887
 888    swizzles[0] = inst->Predicate.SwizzleX;
 889    swizzles[1] = inst->Predicate.SwizzleY;
 890    swizzles[2] = inst->Predicate.SwizzleZ;
 891    swizzles[3] = inst->Predicate.SwizzleW;
 892
 893    index = inst->Predicate.Index;
 894    assert(index < LP_MAX_TGSI_PREDS);
 895
 896    TGSI_FOR_EACH_CHANNEL( chan ) {
 897       unsigned swizzle = swizzles[chan];
 898
 899       /*
 900        * Only fetch the predicate register channels that are actually listed
 901        * in the swizzles
 902        */
 903       if (!unswizzled[swizzle]) {
 904          value = LLVMBuildLoad(builder,
 905                                bld->preds[index][swizzle], "");
 906
 907          /*
 908           * Convert the value to an integer mask.
 909           *
 910           * TODO: Short-circuit this comparison -- a D3D setp_xx instructions
 911           * is needlessly causing two comparisons due to storing the intermediate
 912           * result as float vector instead of an integer mask vector.
 913           */
 914          value = lp_build_compare(bld->bld_base.base.gallivm,
 915                                   bld->bld_base.base.type,
 916                                   PIPE_FUNC_NOTEQUAL,
 917                                   value,
 918                                   bld->bld_base.base.zero);
 919          if (inst->Predicate.Negate) {
 920             value = LLVMBuildNot(builder, value, "");
 921          }
 922
 923          unswizzled[swizzle] = value;
 924       } else {
 925          value = unswizzled[swizzle];
 926       }
 927
 928       pred[chan] = value;
 929    }
 930 }
 931
 932 /**
 933  * Register store.
 934  */
 935 static void
 936 emit_store_chan(
 937    struct lp_build_tgsi_context *bld_base,
 938    const struct tgsi_full_instruction *inst,
 939    unsigned index,
 940    unsigned chan_index,
 941    LLVMValueRef pred,
 942    LLVMValueRef value)
 943 {
 944    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
 945    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
 946    LLVMBuilderRef builder = gallivm->builder;
 947    const struct tgsi_full_dst_register *reg = &inst->Dst[index];
 948    struct lp_build_context *uint_bld = &bld_base->uint_bld;
 949    LLVMValueRef indirect_index = NULL;
 950    struct lp_build_context *bld_store;
 951    enum tgsi_opcode_type dtype = tgsi_opcode_infer_dst_type(inst->Instruction.Opcode);
 952
 953    switch (dtype) {
 954    default:
 955    case TGSI_TYPE_FLOAT:
 956    case TGSI_TYPE_UNTYPED:
 957       bld_store = &bld_base->base;
 958       break;
 959    case TGSI_TYPE_UNSIGNED:
 960       bld_store = &bld_base->uint_bld;
 961       break;
 962    case TGSI_TYPE_SIGNED:
 963       bld_store = &bld_base->int_bld;
 964       break;
 965    case TGSI_TYPE_DOUBLE:
 966    case TGSI_TYPE_VOID:
 967       assert(0);
 968       bld_store = NULL;
 969       break;
 970    }
 971
 972    switch( inst->Instruction.Saturate ) {
 973    case TGSI_SAT_NONE:
 974       break;
 975
 976    case TGSI_SAT_ZERO_ONE:
 977       value = lp_build_max(&bld->bld_base.base, value, bld->bld_base.base.zero);
 978       value = lp_build_min(&bld->bld_base.base, value, bld->bld_base.base.one);
 979       break;
 980
 981    case TGSI_SAT_MINUS_PLUS_ONE:
 982       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));
 983       value = lp_build_min(&bld->bld_base.base, value, bld->bld_base.base.one);
 984       break;
 985
 986    default:
 987       assert(0);
 988    }
 989
 990    if (reg->Register.Indirect) {
 991       indirect_index = get_indirect_index(bld,
 992                                           reg->Register.File,
 993                                           reg->Register.Index,
 994                                           &reg->Indirect);
 995    } else {
 996       assert(reg->Register.Index <=
 997                              bld->bld_base.info->file_max[reg->Register.File]);
 998    }
 999
1000    switch( reg->Register.File ) {
1001    case TGSI_FILE_OUTPUT:
1002       if (reg->Register.Indirect) {
1003          LLVMValueRef chan_vec =
1004             lp_build_const_int_vec(gallivm, uint_bld->type, chan_index);
1005          LLVMValueRef length_vec =
1006             lp_build_const_int_vec(gallivm, uint_bld->type, bld->bld_base.base.type.length);
1007          LLVMValueRef index_vec;  /* indexes into the temp registers */
1008          LLVMValueRef outputs_array;
1009          LLVMValueRef pixel_offsets;
1010          LLVMTypeRef float_ptr_type;
1011          int i;
1012
1013          /* build pixel offset vector: {0, 1, 2, 3, ...} */
1014          pixel_offsets = uint_bld->undef;
1015          for (i = 0; i < bld->bld_base.base.type.length; i++) {
1016             LLVMValueRef ii = lp_build_const_int32(gallivm, i);
1017             pixel_offsets = LLVMBuildInsertElement(builder, pixel_offsets,
1018                                                    ii, ii, "");
1019          }
1020
1021          /* index_vec = (indirect_index * 4 + chan_index) * length + offsets */
1022          index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
1023          index_vec = lp_build_add(uint_bld, index_vec, chan_vec);
1024          index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
1025          index_vec = lp_build_add(uint_bld, index_vec, pixel_offsets);
1026
1027          float_ptr_type =
1028             LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
1029          outputs_array = LLVMBuildBitCast(builder, bld->outputs_array,
1030                                           float_ptr_type, "");
1031
1032          /* Scatter store values into temp registers */
1033          emit_mask_scatter(bld, outputs_array, index_vec, value,
1034                            &bld->exec_mask, pred);
1035       }
1036       else {
1037          LLVMValueRef out_ptr = lp_get_output_ptr(bld, reg->Register.Index,
1038                                                chan_index);
1039          lp_exec_mask_store(&bld->exec_mask, bld_store, pred, value, out_ptr);
1040       }
1041       break;
1042
1043    case TGSI_FILE_TEMPORARY:
1044       if (reg->Register.Indirect) {
1045          LLVMValueRef chan_vec =
1046             lp_build_const_int_vec(gallivm, uint_bld->type, chan_index);
1047          LLVMValueRef length_vec =
1048             lp_build_const_int_vec(gallivm, uint_bld->type,
1049                                    bld->bld_base.base.type.length);
1050          LLVMValueRef index_vec;  /* indexes into the temp registers */
1051          LLVMValueRef temps_array;
1052          LLVMValueRef pixel_offsets;
1053          LLVMTypeRef float_ptr_type;
1054          int i;
1055
1056          /* build pixel offset vector: {0, 1, 2, 3, ...} */
1057          pixel_offsets = uint_bld->undef;
1058          for (i = 0; i < bld->bld_base.base.type.length; i++) {
1059             LLVMValueRef ii = lp_build_const_int32(gallivm, i);
1060             pixel_offsets = LLVMBuildInsertElement(builder, pixel_offsets,
1061                                                    ii, ii, "");
1062          }
1063
1064          /* index_vec = (indirect_index * 4 + chan_index) * length + offsets */
1065          index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
1066          index_vec = lp_build_add(uint_bld, index_vec, chan_vec);
1067          index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
1068          index_vec = lp_build_add(uint_bld, index_vec, pixel_offsets);
1069
1070          float_ptr_type =
1071             LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
1072          temps_array = LLVMBuildBitCast(builder, bld->temps_array,
1073                                         float_ptr_type, "");
1074
1075          /* Scatter store values into temp registers */
1076          emit_mask_scatter(bld, temps_array, index_vec, value,
1077                            &bld->exec_mask, pred);
1078       }
1079       else {
1080          LLVMValueRef temp_ptr;
1081
1082          switch (dtype) {
1083          case TGSI_TYPE_UNSIGNED:
1084          case TGSI_TYPE_SIGNED: {
1085             LLVMTypeRef itype = bld_base->int_bld.vec_type;
1086             LLVMTypeRef ivtype = LLVMPointerType(itype, 0);
1087             LLVMValueRef tint_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index,
1088                                                         chan_index);
1089             LLVMValueRef temp_value_ptr;
1090
1091             temp_ptr = LLVMBuildBitCast(builder, tint_ptr, ivtype, "");
1092             temp_value_ptr = LLVMBuildBitCast(builder, value, itype, "");
1093             value = temp_value_ptr;
1094             break;
1095          }
1096          default:
1097          case TGSI_TYPE_FLOAT:
1098          case TGSI_TYPE_UNTYPED:
1099             temp_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index,
1100                                            chan_index);
1101             break;
1102          }
1103
1104          lp_exec_mask_store(&bld->exec_mask, bld_store, pred, value, temp_ptr);
1105       }
1106       break;
1107
1108    case TGSI_FILE_ADDRESS:
1109       assert(dtype == TGSI_TYPE_SIGNED);
1110       assert(LLVMTypeOf(value) == bld_base->base.int_vec_type);
1111       lp_exec_mask_store(&bld->exec_mask, bld_store, pred, value,
1112                          bld->addr[reg->Register.Index][chan_index]);
1113       break;
1114
1115    case TGSI_FILE_PREDICATE:
1116       lp_exec_mask_store(&bld->exec_mask, bld_store, pred, value,
1117                          bld->preds[reg->Register.Index][chan_index]);
1118       break;
1119
1120    default:
1121       assert( 0 );
1122    }
1123 }
1124
1125 static void
1126 emit_store(
1127    struct lp_build_tgsi_context * bld_base,
1128    const struct tgsi_full_instruction * inst,
1129    const struct tgsi_opcode_info * info,
1130    LLVMValueRef dst[4])
1131
1132 {
1133    unsigned chan_index;
1134    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1135
1136    if(info->num_dst) {
1137       LLVMValueRef pred[TGSI_NUM_CHANNELS];
1138
1139       emit_fetch_predicate( bld, inst, pred );
1140
1141       TGSI_FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1142          emit_store_chan(bld_base, inst, 0, chan_index, pred[chan_index], dst[chan_index]);
1143       }
1144    }
1145 }
1146
1147 /**
1148  * High-level instruction translators.
1149  */
1150
1151 static void
1152 emit_tex( struct lp_build_tgsi_soa_context *bld,
1153           const struct tgsi_full_instruction *inst,
1154           enum lp_build_tex_modifier modifier,
1155           LLVMValueRef *texel)
1156 {
1157    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
1158    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
1159    unsigned unit;
1160    LLVMValueRef lod_bias, explicit_lod;
1161    LLVMValueRef oow = NULL;
1162    LLVMValueRef coords[4];
1163    LLVMValueRef offsets[3] = { NULL };
1164    struct lp_derivatives derivs;
1165    unsigned num_coords;
1166    unsigned dims;
1167    unsigned i;
1168
1169    if (!bld->sampler) {
1170       _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
1171       for (i = 0; i < 4; i++) {
1172          texel[i] = bld->bld_base.base.undef;
1173       }
1174       return;
1175    }
1176
1177    derivs.ddx_ddy[0] = bld->bld_base.base.undef;
1178    derivs.ddx_ddy[1] = bld->bld_base.base.undef;
1179
1180    switch (inst->Texture.Texture) {
1181    case TGSI_TEXTURE_1D:
1182       num_coords = 1;
1183       dims = 1;
1184       break;
1185    case TGSI_TEXTURE_1D_ARRAY:
1186       num_coords = 2;
1187       dims = 1;
1188       break;
1189    case TGSI_TEXTURE_2D:
1190    case TGSI_TEXTURE_RECT:
1191       num_coords = 2;
1192       dims = 2;
1193       break;
1194    case TGSI_TEXTURE_SHADOW1D:
1195    case TGSI_TEXTURE_SHADOW1D_ARRAY:
1196       num_coords = 3;
1197       dims = 1;
1198       break;
1199    case TGSI_TEXTURE_SHADOW2D:
1200    case TGSI_TEXTURE_SHADOWRECT:
1201    case TGSI_TEXTURE_2D_ARRAY:
1202    case TGSI_TEXTURE_CUBE:
1203       num_coords = 3;
1204       dims = 2;
1205       break;
1206    case TGSI_TEXTURE_3D:
1207       num_coords = 3;
1208       dims = 3;
1209       break;
1210    case TGSI_TEXTURE_SHADOW2D_ARRAY:
1211    case TGSI_TEXTURE_SHADOWCUBE:
1212       num_coords = 4;
1213       dims = 2;
1214       break;
1215    default:
1216       assert(0);
1217       return;
1218    }
1219
1220    /* Note lod and especially projected are illegal in a LOT of cases */
1221    if (modifier == LP_BLD_TEX_MODIFIER_LOD_BIAS) {
1222       assert(num_coords < 4);
1223       lod_bias = lp_build_emit_fetch( &bld->bld_base, inst, 0, 3 );
1224       explicit_lod = NULL;
1225    }
1226    else if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_LOD) {
1227       assert(num_coords < 4);
1228       lod_bias = NULL;
1229       explicit_lod = lp_build_emit_fetch( &bld->bld_base, inst, 0, 3 );
1230    }
1231    else {
1232       lod_bias = NULL;
1233       explicit_lod = NULL;
1234    }
1235
1236    if (modifier == LP_BLD_TEX_MODIFIER_PROJECTED) {
1237       assert(num_coords < 4);
1238       oow = lp_build_emit_fetch( &bld->bld_base, inst, 0, 3 );
1239       oow = lp_build_rcp(&bld->bld_base.base, oow);
1240    }
1241
1242    for (i = 0; i < num_coords; i++) {
1243       coords[i] = lp_build_emit_fetch( &bld->bld_base, inst, 0, i );
1244       if (modifier == LP_BLD_TEX_MODIFIER_PROJECTED)
1245          coords[i] = lp_build_mul(&bld->bld_base.base, coords[i], oow);
1246    }
1247    for (i = num_coords; i < 4; i++) {
1248       coords[i] = bld->bld_base.base.undef;
1249    }
1250
1251    if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV) {
1252       LLVMValueRef i32undef = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
1253       LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
1254       LLVMValueRef ddxdyonec[3];
1255       unsigned length = bld->bld_base.base.type.length;
1256       unsigned num_quads = length / 4;
1257       unsigned dim;
1258       unsigned quad;
1259
1260       for (dim = 0; dim < dims; ++dim) {
1261          LLVMValueRef srcx = lp_build_emit_fetch( &bld->bld_base, inst, 1, dim );
1262          LLVMValueRef srcy = lp_build_emit_fetch( &bld->bld_base, inst, 2, dim );
1263          for (quad = 0; quad < num_quads; ++quad) {
1264             unsigned s1 = 4*quad;
1265             unsigned s2 = 4*quad + length;
1266             shuffles[4*quad + 0] = lp_build_const_int32(gallivm, s1);
1267             shuffles[4*quad + 1] = lp_build_const_int32(gallivm, s2);
1268             shuffles[4*quad + 2] = i32undef;
1269             shuffles[4*quad + 3] = i32undef;
1270          }
1271          ddxdyonec[dim] = LLVMBuildShuffleVector(builder, srcx, srcy,
1272                                                LLVMConstVector(shuffles, length), "");
1273       }
1274       if (dims == 1) {
1275          derivs.ddx_ddy[0] = ddxdyonec[0];
1276       }
1277       else if (dims >= 2) {
1278          for (quad = 0; quad < num_quads; ++quad) {
1279             unsigned s1 = 4*quad;
1280             unsigned s2 = 4*quad + length;
1281             shuffles[4*quad + 0] = lp_build_const_int32(gallivm, s1);
1282             shuffles[4*quad + 1] = lp_build_const_int32(gallivm, s1 + 1);
1283             shuffles[4*quad + 2] = lp_build_const_int32(gallivm, s2);
1284             shuffles[4*quad + 3] = lp_build_const_int32(gallivm, s2 + 1);
1285          }
1286          derivs.ddx_ddy[0] = LLVMBuildShuffleVector(builder, ddxdyonec[0], ddxdyonec[1],
1287                                                   LLVMConstVector(shuffles, length), "");
1288          if (dims == 3) {
1289             derivs.ddx_ddy[1] = ddxdyonec[2];
1290          }
1291       }
1292       unit = inst->Src[3].Register.Index;
1293    }  else {
1294       if (dims == 1) {
1295          derivs.ddx_ddy[0] = lp_build_packed_ddx_ddy_onecoord(&bld->bld_base.base, coords[0]);
1296       }
1297       else if (dims >= 2) {
1298          derivs.ddx_ddy[0] = lp_build_packed_ddx_ddy_twocoord(&bld->bld_base.base,
1299                                                             coords[0], coords[1]);
1300          if (dims == 3) {
1301             derivs.ddx_ddy[1] = lp_build_packed_ddx_ddy_onecoord(&bld->bld_base.base, coords[2]);
1302          }
1303       }
1304       unit = inst->Src[1].Register.Index;
1305    }
1306
1307    /* some advanced gather instructions (txgo) would require 4 offsets */
1308    if (inst->Texture.NumOffsets == 1) {
1309       unsigned dim;
1310       for (dim = 0; dim < dims; dim++) {
1311          offsets[dim] = lp_build_emit_fetch_texoffset(&bld->bld_base, inst, 0, dim );
1312       }
1313    }
1314
1315    bld->sampler->emit_fetch_texel(bld->sampler,
1316                                   bld->bld_base.base.gallivm,
1317                                   bld->bld_base.base.type,
1318                                   FALSE,
1319                                   unit, coords,
1320                                   offsets,
1321                                   &derivs,
1322                                   lod_bias, explicit_lod,
1323                                   texel);
1324 }
1325
1326 static void
1327 emit_txf( struct lp_build_tgsi_soa_context *bld,
1328           const struct tgsi_full_instruction *inst,
1329           LLVMValueRef *texel)
1330 {
1331    unsigned unit;
1332    LLVMValueRef coord_undef = LLVMGetUndef(bld->bld_base.base.int_vec_type);
1333    LLVMValueRef explicit_lod = NULL;
1334    LLVMValueRef coords[3];
1335    LLVMValueRef offsets[3] = { NULL };
1336    struct lp_derivatives derivs;
1337    unsigned num_coords;
1338    unsigned dims;
1339    unsigned i;
1340
1341    if (!bld->sampler) {
1342       _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
1343       for (i = 0; i < 4; i++) {
1344          texel[i] = coord_undef;
1345       }
1346       return;
1347    }
1348
1349    derivs.ddx_ddy[0] = coord_undef;
1350    derivs.ddx_ddy[1] = coord_undef;
1351
1352    switch (inst->Texture.Texture) {
1353    case TGSI_TEXTURE_1D:
1354    case TGSI_TEXTURE_BUFFER:
1355       num_coords = 1;
1356       dims = 1;
1357       break;
1358    case TGSI_TEXTURE_1D_ARRAY:
1359       num_coords = 2;
1360       dims = 1;
1361       break;
1362    case TGSI_TEXTURE_2D:
1363    case TGSI_TEXTURE_RECT:
1364       num_coords = 2;
1365       dims = 2;
1366       break;
1367    case TGSI_TEXTURE_2D_ARRAY:
1368       num_coords = 3;
1369       dims = 2;
1370       break;
1371    case TGSI_TEXTURE_3D:
1372       num_coords = 3;
1373       dims = 3;
1374       break;
1375    default:
1376       assert(0);
1377       return;
1378    }
1379
1380    /* always have lod except for buffers ? */
1381    if (inst->Texture.Texture != TGSI_TEXTURE_BUFFER) {
1382       explicit_lod = lp_build_emit_fetch( &bld->bld_base, inst, 0, 3 );
1383    }
1384
1385    for (i = 0; i < num_coords; i++) {
1386       coords[i] = lp_build_emit_fetch( &bld->bld_base, inst, 0, i );
1387    }
1388    for (i = num_coords; i < 3; i++) {
1389       coords[i] = coord_undef;
1390    }
1391
1392    unit = inst->Src[1].Register.Index;
1393
1394    if (inst->Texture.NumOffsets == 1) {
1395       unsigned dim;
1396       for (dim = 0; dim < dims; dim++) {
1397          offsets[dim] = lp_build_emit_fetch_texoffset(&bld->bld_base, inst, 0, dim );
1398       }
1399    }
1400
1401    bld->sampler->emit_fetch_texel(bld->sampler,
1402                                   bld->bld_base.base.gallivm,
1403                                   bld->bld_base.base.type,
1404                                   TRUE,
1405                                   unit, coords,
1406                                   offsets,
1407                                   &derivs,
1408                                   NULL, explicit_lod,
1409                                   texel);
1410 }
1411
1412 static void
1413 emit_txq( struct lp_build_tgsi_soa_context *bld,
1414           const struct tgsi_full_instruction *inst,
1415           LLVMValueRef *sizes_out)
1416 {
1417    LLVMValueRef explicit_lod;
1418    unsigned num_coords, has_lod;
1419    unsigned i;
1420
1421    switch (inst->Texture.Texture) {
1422    case TGSI_TEXTURE_1D:
1423    case TGSI_TEXTURE_SHADOW1D:
1424       num_coords = 1;
1425       has_lod = 1;
1426       break;
1427    case TGSI_TEXTURE_2D:
1428    case TGSI_TEXTURE_SHADOW2D:
1429    case TGSI_TEXTURE_CUBE:
1430    case TGSI_TEXTURE_SHADOWCUBE:
1431    case TGSI_TEXTURE_1D_ARRAY:
1432    case TGSI_TEXTURE_SHADOW1D_ARRAY:
1433       num_coords = 2;
1434       has_lod = 1;
1435       break;
1436    case TGSI_TEXTURE_3D:
1437 // case TGSI_TEXTURE_CUBE_ARRAY:
1438 // case TGSI_TEXTURE_SHADOWCUBE_ARRAY:
1439    case TGSI_TEXTURE_2D_ARRAY:
1440    case TGSI_TEXTURE_SHADOW2D_ARRAY:
1441       num_coords = 3;
1442       has_lod = 1;
1443       break;
1444
1445    case TGSI_TEXTURE_BUFFER:
1446       num_coords = 1;
1447       has_lod = 0;
1448       break;
1449
1450    case TGSI_TEXTURE_RECT:
1451    case TGSI_TEXTURE_SHADOWRECT:
1452 // case TGSI_TEXTURE_2D_MS:
1453       num_coords = 2;
1454       has_lod = 0;
1455       break;
1456
1457 // case TGSI_TEXTURE_2D_MS_ARRAY:
1458 //    num_coords = 3;
1459 //    has_lod = 0;
1460 //    break;
1461
1462    default:
1463       assert(0);
1464       return;
1465    }
1466
1467    if (!bld->sampler) {
1468       _debug_printf("warning: found texture query instruction but no sampler generator supplied\n");
1469       for (i = 0; i < num_coords; i++)
1470          sizes_out[i] = bld->bld_base.base.undef;
1471       return;
1472    }
1473
1474    if (has_lod)
1475       explicit_lod = lp_build_emit_fetch( &bld->bld_base, inst, 0, 2 );
1476    else
1477       explicit_lod = NULL;
1478
1479    bld->sampler->emit_size_query(bld->sampler,
1480                                  bld->bld_base.base.gallivm,
1481                                  bld->bld_base.int_bld.type,
1482                                  inst->Src[1].Register.Index,
1483                                  explicit_lod,
1484                                  sizes_out);
1485 }
1486
1487 static boolean
1488 near_end_of_shader(struct lp_build_tgsi_soa_context *bld,
1489                    int pc)
1490 {
1491    int i;
1492
1493    for (i = 0; i < 5; i++) {
1494       unsigned opcode;
1495
1496       if (pc + i >= bld->bld_base.info->num_instructions)
1497          return TRUE;
1498
1499       opcode = bld->bld_base.instructions[pc + i].Instruction.Opcode;
1500
1501       if (opcode == TGSI_OPCODE_END)
1502          return TRUE;
1503
1504       if (opcode == TGSI_OPCODE_TEX ||
1505           opcode == TGSI_OPCODE_TXP ||
1506           opcode == TGSI_OPCODE_TXD ||
1507           opcode == TGSI_OPCODE_TXB ||
1508           opcode == TGSI_OPCODE_TXL ||
1509           opcode == TGSI_OPCODE_TXF ||
1510           opcode == TGSI_OPCODE_TXQ ||
1511           opcode == TGSI_OPCODE_CAL ||
1512           opcode == TGSI_OPCODE_CALLNZ ||
1513           opcode == TGSI_OPCODE_IF ||
1514           opcode == TGSI_OPCODE_IFC ||
1515           opcode == TGSI_OPCODE_BGNLOOP ||
1516           opcode == TGSI_OPCODE_SWITCH)
1517          return FALSE;
1518    }
1519
1520    return TRUE;
1521 }
1522
1523
1524
1525 /**
1526  * Kill fragment if any of the src register values are negative.
1527  */
1528 static void
1529 emit_kil(
1530    struct lp_build_tgsi_soa_context *bld,
1531    const struct tgsi_full_instruction *inst,
1532    int pc)
1533 {
1534    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
1535    const struct tgsi_full_src_register *reg = &inst->Src[0];
1536    LLVMValueRef terms[TGSI_NUM_CHANNELS];
1537    LLVMValueRef mask;
1538    unsigned chan_index;
1539
1540    memset(&terms, 0, sizeof terms);
1541
1542    TGSI_FOR_EACH_CHANNEL( chan_index ) {
1543       unsigned swizzle;
1544
1545       /* Unswizzle channel */
1546       swizzle = tgsi_util_get_full_src_register_swizzle( reg, chan_index );
1547
1548       /* Check if the component has not been already tested. */
1549       assert(swizzle < TGSI_NUM_CHANNELS);
1550       if( !terms[swizzle] )
1551          /* TODO: change the comparison operator instead of setting the sign */
1552          terms[swizzle] =  lp_build_emit_fetch(&bld->bld_base, inst, 0, chan_index );
1553    }
1554
1555    mask = NULL;
1556    TGSI_FOR_EACH_CHANNEL( chan_index ) {
1557       if(terms[chan_index]) {
1558          LLVMValueRef chan_mask;
1559
1560          /*
1561           * If term < 0 then mask = 0 else mask = ~0.
1562           */
1563          chan_mask = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_GEQUAL, terms[chan_index], bld->bld_base.base.zero);
1564
1565          if(mask)
1566             mask = LLVMBuildAnd(builder, mask, chan_mask, "");
1567          else
1568             mask = chan_mask;
1569       }
1570    }
1571
1572    if(mask) {
1573       lp_build_mask_update(bld->mask, mask);
1574
1575       if (!near_end_of_shader(bld, pc))
1576          lp_build_mask_check(bld->mask);
1577    }
1578 }
1579
1580
1581 /**
1582  * Predicated fragment kill.
1583  * XXX Actually, we do an unconditional kill (as in tgsi_exec.c).
1584  * The only predication is the execution mask which will apply if
1585  * we're inside a loop or conditional.
1586  */
1587 static void
1588 emit_kilp(struct lp_build_tgsi_soa_context *bld,
1589           int pc)
1590 {
1591    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
1592    LLVMValueRef mask;
1593
1594    /* For those channels which are "alive", disable fragment shader
1595     * execution.
1596     */
1597    if (bld->exec_mask.has_mask) {
1598       mask = LLVMBuildNot(builder, bld->exec_mask.exec_mask, "kilp");
1599    }
1600    else {
1601       LLVMValueRef zero = LLVMConstNull(bld->bld_base.base.int_vec_type);
1602       mask = zero;
1603    }
1604
1605    lp_build_mask_update(bld->mask, mask);
1606
1607    if (!near_end_of_shader(bld, pc))
1608       lp_build_mask_check(bld->mask);
1609 }
1610
1611
1612 /**
1613  * Emit code which will dump the value of all the temporary registers
1614  * to stdout.
1615  */
1616 static void
1617 emit_dump_temps(struct lp_build_tgsi_soa_context *bld)
1618 {
1619    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
1620    LLVMBuilderRef builder = gallivm->builder;
1621    LLVMValueRef temp_ptr;
1622    LLVMValueRef i0 = lp_build_const_int32(gallivm, 0);
1623    LLVMValueRef i1 = lp_build_const_int32(gallivm, 1);
1624    LLVMValueRef i2 = lp_build_const_int32(gallivm, 2);
1625    LLVMValueRef i3 = lp_build_const_int32(gallivm, 3);
1626    int index;
1627    int n = bld->bld_base.info->file_max[TGSI_FILE_TEMPORARY];
1628
1629    for (index = 0; index < n; index++) {
1630       LLVMValueRef idx = lp_build_const_int32(gallivm, index);
1631       LLVMValueRef v[4][4], res;
1632       int chan;
1633
1634       lp_build_printf(gallivm, "TEMP[%d]:\n", idx);
1635
1636       for (chan = 0; chan < 4; chan++) {
1637          temp_ptr = lp_get_temp_ptr_soa(bld, index, chan);
1638          res = LLVMBuildLoad(builder, temp_ptr, "");
1639          v[chan][0] = LLVMBuildExtractElement(builder, res, i0, "");
1640          v[chan][1] = LLVMBuildExtractElement(builder, res, i1, "");
1641          v[chan][2] = LLVMBuildExtractElement(builder, res, i2, "");
1642          v[chan][3] = LLVMBuildExtractElement(builder, res, i3, "");
1643       }
1644
1645       lp_build_printf(gallivm, "  X: %f %f %f %f\n",
1646                       v[0][0], v[0][1], v[0][2], v[0][3]);
1647       lp_build_printf(gallivm, "  Y: %f %f %f %f\n",
1648                       v[1][0], v[1][1], v[1][2], v[1][3]);
1649       lp_build_printf(gallivm, "  Z: %f %f %f %f\n",
1650                       v[2][0], v[2][1], v[2][2], v[2][3]);
1651       lp_build_printf(gallivm, "  W: %f %f %f %f\n",
1652                       v[3][0], v[3][1], v[3][2], v[3][3]);
1653    }
1654 }
1655
1656
1657
1658 void
1659 lp_emit_declaration_soa(
1660    struct lp_build_tgsi_context *bld_base,
1661    const struct tgsi_full_declaration *decl)
1662 {
1663    struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
1664    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
1665    LLVMTypeRef vec_type = bld->bld_base.base.vec_type;
1666    const unsigned first = decl->Range.First;
1667    const unsigned last = decl->Range.Last;
1668    unsigned idx, i;
1669
1670    for (idx = first; idx <= last; ++idx) {
1671       assert(last <= bld->bld_base.info->file_max[decl->Declaration.File]);
1672       switch (decl->Declaration.File) {
1673       case TGSI_FILE_TEMPORARY:
1674          assert(idx < LP_MAX_TGSI_TEMPS);
1675          if (!(bld->indirect_files & (1 << TGSI_FILE_TEMPORARY))) {
1676             for (i = 0; i < TGSI_NUM_CHANNELS; i++)
1677                bld->temps[idx][i] = lp_build_alloca(gallivm, vec_type, "temp");
1678          }
1679          break;
1680
1681       case TGSI_FILE_OUTPUT:
1682          if (!(bld->indirect_files & (1 << TGSI_FILE_OUTPUT))) {
1683             for (i = 0; i < TGSI_NUM_CHANNELS; i++)
1684                bld->outputs[idx][i] = lp_build_alloca(gallivm,
1685                                                       vec_type, "output");
1686          }
1687          break;
1688
1689       case TGSI_FILE_ADDRESS:
1690          /* ADDR registers are the only allocated with an integer LLVM IR type,
1691           * as they are guaranteed to always have integers.
1692           * XXX: Not sure if this exception is worthwhile (or the whole idea of
1693           * an ADDR register for that matter).
1694           */
1695          assert(idx < LP_MAX_TGSI_ADDRS);
1696          for (i = 0; i < TGSI_NUM_CHANNELS; i++)
1697             bld->addr[idx][i] = lp_build_alloca(gallivm, bld_base->base.int_vec_type, "addr");
1698          break;
1699
1700       case TGSI_FILE_PREDICATE:
1701          assert(idx < LP_MAX_TGSI_PREDS);
1702          for (i = 0; i < TGSI_NUM_CHANNELS; i++)
1703             bld->preds[idx][i] = lp_build_alloca(gallivm, vec_type,
1704                                                  "predicate");
1705          break;
1706
1707       default:
1708          /* don't need to declare other vars */
1709          break;
1710       }
1711    }
1712 }
1713
1714
1715 void lp_emit_immediate_soa(
1716    struct lp_build_tgsi_context *bld_base,
1717    const struct tgsi_full_immediate *imm)
1718 {
1719    struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
1720    struct gallivm_state * gallivm = bld_base->base.gallivm;
1721
1722    /* simply copy the immediate values into the next immediates[] slot */
1723    unsigned i;
1724    const uint size = imm->Immediate.NrTokens - 1;
1725    assert(size <= 4);
1726    assert(bld->num_immediates < LP_MAX_TGSI_IMMEDIATES);
1727    switch (imm->Immediate.DataType) {
1728    case TGSI_IMM_FLOAT32:
1729       for( i = 0; i < size; ++i )
1730          bld->immediates[bld->num_immediates][i] =
1731             lp_build_const_vec(gallivm, bld_base->base.type, imm->u[i].Float);
1732
1733       break;
1734    case TGSI_IMM_UINT32:
1735       for( i = 0; i < size; ++i ) {
1736          LLVMValueRef tmp = lp_build_const_vec(gallivm, bld_base->uint_bld.type, imm->u[i].Uint);
1737          bld->immediates[bld->num_immediates][i] =
1738             LLVMConstBitCast(tmp, bld_base->base.vec_type);
1739       }
1740
1741       break;
1742    case TGSI_IMM_INT32:
1743       for( i = 0; i < size; ++i ) {
1744          LLVMValueRef tmp = lp_build_const_vec(gallivm, bld_base->int_bld.type, imm->u[i].Int);
1745          bld->immediates[bld->num_immediates][i] =
1746             LLVMConstBitCast(tmp, bld_base->base.vec_type);
1747       }
1748
1749       break;
1750    }
1751    for( i = size; i < 4; ++i )
1752       bld->immediates[bld->num_immediates][i] = bld_base->base.undef;
1753
1754    bld->num_immediates++;
1755 }
1756
1757 static void
1758 ddx_emit(
1759    const struct lp_build_tgsi_action * action,
1760    struct lp_build_tgsi_context * bld_base,
1761    struct lp_build_emit_data * emit_data)
1762 {
1763    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1764
1765    emit_fetch_deriv(bld, emit_data->args[0], NULL,
1766                     &emit_data->output[emit_data->chan], NULL);
1767 }
1768
1769 static void
1770 ddy_emit(
1771    const struct lp_build_tgsi_action * action,
1772    struct lp_build_tgsi_context * bld_base,
1773    struct lp_build_emit_data * emit_data)
1774 {
1775    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1776
1777    emit_fetch_deriv(bld, emit_data->args[0], NULL, NULL,
1778                     &emit_data->output[emit_data->chan]);
1779 }
1780
1781 static void
1782 kilp_emit(
1783    const struct lp_build_tgsi_action * action,
1784    struct lp_build_tgsi_context * bld_base,
1785    struct lp_build_emit_data * emit_data)
1786 {
1787    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1788
1789    emit_kilp(bld, bld_base->pc - 1);
1790 }
1791
1792 static void
1793 kil_emit(
1794    const struct lp_build_tgsi_action * action,
1795    struct lp_build_tgsi_context * bld_base,
1796    struct lp_build_emit_data * emit_data)
1797 {
1798    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1799
1800    emit_kil(bld, emit_data->inst, bld_base->pc - 1);
1801 }
1802
1803 static void
1804 tex_emit(
1805    const struct lp_build_tgsi_action * action,
1806    struct lp_build_tgsi_context * bld_base,
1807    struct lp_build_emit_data * emit_data)
1808 {
1809    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1810
1811    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_NONE, emit_data->output);
1812 }
1813
1814 static void
1815 txb_emit(
1816    const struct lp_build_tgsi_action * action,
1817    struct lp_build_tgsi_context * bld_base,
1818    struct lp_build_emit_data * emit_data)
1819 {
1820    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1821
1822    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_LOD_BIAS,
1823             emit_data->output);
1824 }
1825
1826 static void
1827 txd_emit(
1828    const struct lp_build_tgsi_action * action,
1829    struct lp_build_tgsi_context * bld_base,
1830    struct lp_build_emit_data * emit_data)
1831 {
1832    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1833
1834    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV,
1835             emit_data->output);
1836 }
1837
1838 static void
1839 txl_emit(
1840    const struct lp_build_tgsi_action * action,
1841    struct lp_build_tgsi_context * bld_base,
1842    struct lp_build_emit_data * emit_data)
1843 {
1844    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1845
1846    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD,
1847             emit_data->output);
1848 }
1849
1850 static void
1851 txp_emit(
1852    const struct lp_build_tgsi_action * action,
1853    struct lp_build_tgsi_context * bld_base,
1854    struct lp_build_emit_data * emit_data)
1855 {
1856    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1857
1858    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_PROJECTED,
1859             emit_data->output);
1860 }
1861
1862 static void
1863 txq_emit(
1864    const struct lp_build_tgsi_action * action,
1865    struct lp_build_tgsi_context * bld_base,
1866    struct lp_build_emit_data * emit_data)
1867 {
1868    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1869
1870    emit_txq(bld, emit_data->inst, emit_data->output);
1871 }
1872
1873 static void
1874 txf_emit(
1875    const struct lp_build_tgsi_action * action,
1876    struct lp_build_tgsi_context * bld_base,
1877    struct lp_build_emit_data * emit_data)
1878 {
1879    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1880
1881    emit_txf(bld, emit_data->inst, emit_data->output);
1882 }
1883
1884 static void
1885 cal_emit(
1886    const struct lp_build_tgsi_action * action,
1887    struct lp_build_tgsi_context * bld_base,
1888    struct lp_build_emit_data * emit_data)
1889 {
1890    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1891
1892    lp_exec_mask_call(&bld->exec_mask, emit_data->inst->Label.Label,
1893                      &bld_base->pc);
1894 }
1895
1896 static void
1897 ret_emit(
1898    const struct lp_build_tgsi_action * action,
1899    struct lp_build_tgsi_context * bld_base,
1900    struct lp_build_emit_data * emit_data)
1901 {
1902    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1903
1904    lp_exec_mask_ret(&bld->exec_mask, &bld_base->pc);
1905 }
1906
1907 static void
1908 brk_emit(
1909    const struct lp_build_tgsi_action * action,
1910    struct lp_build_tgsi_context * bld_base,
1911    struct lp_build_emit_data * emit_data)
1912 {
1913    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1914
1915    lp_exec_break(&bld->exec_mask);
1916 }
1917
1918 static void
1919 if_emit(
1920    const struct lp_build_tgsi_action * action,
1921    struct lp_build_tgsi_context * bld_base,
1922    struct lp_build_emit_data * emit_data)
1923 {
1924    LLVMValueRef tmp;
1925    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1926
1927    tmp = lp_build_cmp(&bld_base->base, PIPE_FUNC_NOTEQUAL,
1928                       emit_data->args[0], bld->bld_base.base.zero);
1929    lp_exec_mask_cond_push(&bld->exec_mask, tmp);
1930 }
1931
1932 static void
1933 bgnloop_emit(
1934    const struct lp_build_tgsi_action * action,
1935    struct lp_build_tgsi_context * bld_base,
1936    struct lp_build_emit_data * emit_data)
1937 {
1938    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1939
1940    lp_exec_bgnloop(&bld->exec_mask);
1941 }
1942
1943 static void
1944 bgnsub_emit(
1945    const struct lp_build_tgsi_action * action,
1946    struct lp_build_tgsi_context * bld_base,
1947    struct lp_build_emit_data * emit_data)
1948 {
1949    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1950
1951    lp_exec_mask_bgnsub(&bld->exec_mask);
1952 }
1953
1954 static void
1955 else_emit(
1956    const struct lp_build_tgsi_action * action,
1957    struct lp_build_tgsi_context * bld_base,
1958    struct lp_build_emit_data * emit_data)
1959 {
1960    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1961
1962    lp_exec_mask_cond_invert(&bld->exec_mask);
1963 }
1964
1965 static void
1966 endif_emit(
1967    const struct lp_build_tgsi_action * action,
1968    struct lp_build_tgsi_context * bld_base,
1969    struct lp_build_emit_data * emit_data)
1970 {
1971    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1972
1973    lp_exec_mask_cond_pop(&bld->exec_mask);
1974 }
1975
1976 static void
1977 endloop_emit(
1978    const struct lp_build_tgsi_action * action,
1979    struct lp_build_tgsi_context * bld_base,
1980    struct lp_build_emit_data * emit_data)
1981 {
1982    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1983
1984    lp_exec_endloop(bld_base->base.gallivm, &bld->exec_mask);
1985 }
1986
1987 static void
1988 endsub_emit(
1989    const struct lp_build_tgsi_action * action,
1990    struct lp_build_tgsi_context * bld_base,
1991    struct lp_build_emit_data * emit_data)
1992 {
1993    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1994
1995    lp_exec_mask_endsub(&bld->exec_mask, &bld_base->pc);
1996 }
1997
1998 static void
1999 cont_emit(
2000    const struct lp_build_tgsi_action * action,
2001    struct lp_build_tgsi_context * bld_base,
2002    struct lp_build_emit_data * emit_data)
2003 {
2004    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2005
2006    lp_exec_continue(&bld->exec_mask);
2007 }
2008
2009 /* XXX: Refactor and move it to lp_bld_tgsi_action.c
2010  *
2011  * XXX: What do the comments about xmm registers mean?  Maybe they are left over
2012  * from old code, but there is no garauntee that LLVM will use those registers
2013  * for this code.
2014  *
2015  * XXX: There should be no calls to lp_build_emit_fetch in this function.  This
2016  * should be handled by the emit_data->fetch_args function. */
2017 static void
2018 nrm_emit(
2019    const struct lp_build_tgsi_action * action,
2020    struct lp_build_tgsi_context * bld_base,
2021    struct lp_build_emit_data * emit_data)
2022 {
2023    LLVMValueRef tmp0, tmp1;
2024    LLVMValueRef tmp4 = NULL;
2025    LLVMValueRef tmp5 = NULL;
2026    LLVMValueRef tmp6 = NULL;
2027    LLVMValueRef tmp7 = NULL;
2028    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2029
2030    uint dims = (emit_data->inst->Instruction.Opcode == TGSI_OPCODE_NRM) ? 3 : 4;
2031
2032   if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_X) ||
2033       TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Y) ||
2034       TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Z) ||
2035       (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_W) && dims == 4)) {
2036
2037       /* NOTE: Cannot use xmm regs 2/3 here (see emit_rsqrt() above). */
2038
2039       /* xmm4 = src.x */
2040       /* xmm0 = src.x * src.x */
2041       tmp0 = lp_build_emit_fetch(&bld->bld_base, emit_data->inst, 0, TGSI_CHAN_X);
2042       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_X)) {
2043          tmp4 = tmp0;
2044       }
2045       tmp0 = lp_build_mul( &bld->bld_base.base, tmp0, tmp0);
2046
2047       /* xmm5 = src.y */
2048       /* xmm0 = xmm0 + src.y * src.y */
2049       tmp1 = lp_build_emit_fetch(&bld->bld_base, emit_data->inst, 0, TGSI_CHAN_Y);
2050       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Y)) {
2051          tmp5 = tmp1;
2052       }
2053       tmp1 = lp_build_mul( &bld->bld_base.base, tmp1, tmp1);
2054       tmp0 = lp_build_add( &bld->bld_base.base, tmp0, tmp1);
2055
2056       /* xmm6 = src.z */
2057       /* xmm0 = xmm0 + src.z * src.z */
2058       tmp1 = lp_build_emit_fetch(&bld->bld_base, emit_data->inst, 0, TGSI_CHAN_Z);
2059       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Z)) {
2060          tmp6 = tmp1;
2061       }
2062       tmp1 = lp_build_mul( &bld->bld_base.base, tmp1, tmp1);
2063       tmp0 = lp_build_add( &bld->bld_base.base, tmp0, tmp1);
2064
2065       if (dims == 4) {
2066          /* xmm7 = src.w */
2067          /* xmm0 = xmm0 + src.w * src.w */
2068          tmp1 = lp_build_emit_fetch(&bld->bld_base, emit_data->inst, 0, TGSI_CHAN_W);
2069          if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_W)) {
2070             tmp7 = tmp1;
2071          }
2072          tmp1 = lp_build_mul( &bld->bld_base.base, tmp1, tmp1);
2073          tmp0 = lp_build_add( &bld->bld_base.base, tmp0, tmp1);
2074       }
2075       /* xmm1 = 1 / sqrt(xmm0) */
2076       tmp1 = lp_build_rsqrt( &bld->bld_base.base, tmp0);
2077        /* dst.x = xmm1 * src.x */
2078       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_X)) {
2079          emit_data->output[TGSI_CHAN_X] = lp_build_mul( &bld->bld_base.base, tmp4, tmp1);
2080       }
2081       /* dst.y = xmm1 * src.y */
2082       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Y)) {
2083          emit_data->output[TGSI_CHAN_Y] = lp_build_mul( &bld->bld_base.base, tmp5, tmp1);
2084       }
2085
2086       /* dst.z = xmm1 * src.z */
2087       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Z)) {
2088          emit_data->output[TGSI_CHAN_Z] = lp_build_mul( &bld->bld_base.base, tmp6, tmp1);
2089       }
2090       /* dst.w = xmm1 * src.w */
2091       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_X) && dims == 4) {
2092          emit_data->output[TGSI_CHAN_W] = lp_build_mul( &bld->bld_base.base, tmp7, tmp1);
2093       }
2094    }
2095
2096    /* dst.w = 1.0 */
2097    if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_W) && dims == 3) {
2098        emit_data->output[TGSI_CHAN_W] = bld->bld_base.base.one;
2099    }
2100 }
2101
2102 static void emit_prologue(struct lp_build_tgsi_context * bld_base)
2103 {
2104    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2105    struct gallivm_state * gallivm = bld_base->base.gallivm;
2106
2107    if (bld->indirect_files & (1 << TGSI_FILE_TEMPORARY)) {
2108       LLVMValueRef array_size =
2109          lp_build_const_int32(gallivm,
2110                          bld_base->info->file_max[TGSI_FILE_TEMPORARY] * 4 + 4);
2111       bld->temps_array = lp_build_array_alloca(gallivm,
2112                                               bld_base->base.vec_type, array_size,
2113                                               "temp_array");
2114    }
2115
2116    if (bld->indirect_files & (1 << TGSI_FILE_OUTPUT)) {
2117       LLVMValueRef array_size =
2118          lp_build_const_int32(gallivm,
2119                             bld_base->info->file_max[TGSI_FILE_OUTPUT] * 4 + 4);
2120       bld->outputs_array = lp_build_array_alloca(gallivm,
2121                                                 bld_base->base.vec_type, array_size,
2122                                                 "output_array");
2123    }
2124
2125    /* If we have indirect addressing in inputs we need to copy them into
2126     * our alloca array to be able to iterate over them */
2127    if (bld->indirect_files & (1 << TGSI_FILE_INPUT)) {
2128       unsigned index, chan;
2129       LLVMTypeRef vec_type = bld_base->base.vec_type;
2130       LLVMValueRef array_size = lp_build_const_int32(gallivm,
2131             bld_base->info->file_max[TGSI_FILE_INPUT]*4 + 4);
2132       bld->inputs_array = lp_build_array_alloca(gallivm,
2133                                                vec_type, array_size,
2134                                                "input_array");
2135
2136       assert(bld_base->info->num_inputs
2137                         <= bld_base->info->file_max[TGSI_FILE_INPUT] + 1);
2138
2139       for (index = 0; index < bld_base->info->num_inputs; ++index) {
2140          for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
2141             LLVMValueRef lindex =
2142                lp_build_const_int32(gallivm, index * 4 + chan);
2143             LLVMValueRef input_ptr =
2144                LLVMBuildGEP(gallivm->builder, bld->inputs_array,
2145                             &lindex, 1, "");
2146             LLVMValueRef value = bld->inputs[index][chan];
2147             if (value)
2148                LLVMBuildStore(gallivm->builder, value, input_ptr);
2149          }
2150       }
2151    }
2152 }
2153
2154 static void emit_epilogue(struct lp_build_tgsi_context * bld_base)
2155 {
2156    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2157
2158    if (0) {
2159       /* for debugging */
2160       emit_dump_temps(bld);
2161    }
2162
2163    /* If we have indirect addressing in outputs we need to copy our alloca array
2164     * to the outputs slots specified by the called */
2165    if (bld->indirect_files & (1 << TGSI_FILE_OUTPUT)) {
2166       unsigned index, chan;
2167       assert(bld_base->info->num_outputs <=
2168                         bld_base->info->file_max[TGSI_FILE_OUTPUT] + 1);
2169       for (index = 0; index < bld_base->info->num_outputs; ++index) {
2170          for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
2171             bld->outputs[index][chan] = lp_get_output_ptr(bld, index, chan);
2172          }
2173       }
2174    }
2175 }
2176
2177 void
2178 lp_build_tgsi_soa(struct gallivm_state *gallivm,
2179                   const struct tgsi_token *tokens,
2180                   struct lp_type type,
2181                   struct lp_build_mask_context *mask,
2182                   LLVMValueRef consts_ptr,
2183                   const struct lp_bld_tgsi_system_values *system_values,
2184                   const LLVMValueRef *pos,
2185                   const LLVMValueRef (*inputs)[TGSI_NUM_CHANNELS],
2186                   LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
2187                   struct lp_build_sampler_soa *sampler,
2188                   const struct tgsi_shader_info *info)
2189 {
2190    struct lp_build_tgsi_soa_context bld;
2191
2192    struct lp_type res_type;
2193
2194    assert(type.length <= LP_MAX_VECTOR_LENGTH);
2195    memset(&res_type, 0, sizeof res_type);
2196    res_type.width = type.width;
2197    res_type.length = type.length;
2198    res_type.sign = 1;
2199
2200    /* Setup build context */
2201    memset(&bld, 0, sizeof bld);
2202    lp_build_context_init(&bld.bld_base.base, gallivm, type);
2203    lp_build_context_init(&bld.bld_base.uint_bld, gallivm, lp_uint_type(type));
2204    lp_build_context_init(&bld.bld_base.int_bld, gallivm, lp_int_type(type));
2205    lp_build_context_init(&bld.elem_bld, gallivm, lp_elem_type(type));
2206    bld.mask = mask;
2207    bld.pos = pos;
2208    bld.inputs = inputs;
2209    bld.outputs = outputs;
2210    bld.consts_ptr = consts_ptr;
2211    bld.sampler = sampler;
2212    bld.bld_base.info = info;
2213    bld.indirect_files = info->indirect_files;
2214
2215    bld.bld_base.soa = TRUE;
2216    bld.bld_base.emit_fetch_funcs[TGSI_FILE_CONSTANT] = emit_fetch_constant;
2217    bld.bld_base.emit_fetch_funcs[TGSI_FILE_IMMEDIATE] = emit_fetch_immediate;
2218    bld.bld_base.emit_fetch_funcs[TGSI_FILE_INPUT] = emit_fetch_input;
2219    bld.bld_base.emit_fetch_funcs[TGSI_FILE_TEMPORARY] = emit_fetch_temporary;
2220    bld.bld_base.emit_fetch_funcs[TGSI_FILE_SYSTEM_VALUE] = emit_fetch_system_value;
2221    bld.bld_base.emit_store = emit_store;
2222
2223    bld.bld_base.emit_declaration = lp_emit_declaration_soa;
2224    bld.bld_base.emit_immediate = lp_emit_immediate_soa;
2225
2226    bld.bld_base.emit_prologue = emit_prologue;
2227    bld.bld_base.emit_epilogue = emit_epilogue;
2228
2229    /* Set opcode actions */
2230    lp_set_default_actions_cpu(&bld.bld_base);
2231
2232    bld.bld_base.op_actions[TGSI_OPCODE_BGNLOOP].emit = bgnloop_emit;
2233    bld.bld_base.op_actions[TGSI_OPCODE_BGNSUB].emit = bgnsub_emit;
2234    bld.bld_base.op_actions[TGSI_OPCODE_BRK].emit = brk_emit;
2235    bld.bld_base.op_actions[TGSI_OPCODE_CAL].emit = cal_emit;
2236    bld.bld_base.op_actions[TGSI_OPCODE_CONT].emit = cont_emit;
2237    bld.bld_base.op_actions[TGSI_OPCODE_DDX].emit = ddx_emit;
2238    bld.bld_base.op_actions[TGSI_OPCODE_DDY].emit = ddy_emit;
2239    bld.bld_base.op_actions[TGSI_OPCODE_ELSE].emit = else_emit;
2240    bld.bld_base.op_actions[TGSI_OPCODE_ENDIF].emit = endif_emit;
2241    bld.bld_base.op_actions[TGSI_OPCODE_ENDLOOP].emit = endloop_emit;
2242    bld.bld_base.op_actions[TGSI_OPCODE_ENDSUB].emit = endsub_emit;
2243    bld.bld_base.op_actions[TGSI_OPCODE_IF].emit = if_emit;
2244    bld.bld_base.op_actions[TGSI_OPCODE_KIL].emit = kil_emit;
2245    bld.bld_base.op_actions[TGSI_OPCODE_KILP].emit = kilp_emit;
2246    bld.bld_base.op_actions[TGSI_OPCODE_NRM].emit = nrm_emit;
2247    bld.bld_base.op_actions[TGSI_OPCODE_NRM4].emit = nrm_emit;
2248    bld.bld_base.op_actions[TGSI_OPCODE_RET].emit = ret_emit;
2249    bld.bld_base.op_actions[TGSI_OPCODE_TEX].emit = tex_emit;
2250    bld.bld_base.op_actions[TGSI_OPCODE_TXB].emit = txb_emit;
2251    bld.bld_base.op_actions[TGSI_OPCODE_TXD].emit = txd_emit;
2252    bld.bld_base.op_actions[TGSI_OPCODE_TXL].emit = txl_emit;
2253    bld.bld_base.op_actions[TGSI_OPCODE_TXP].emit = txp_emit;
2254    bld.bld_base.op_actions[TGSI_OPCODE_TXQ].emit = txq_emit;
2255    bld.bld_base.op_actions[TGSI_OPCODE_TXF].emit = txf_emit;
2256
2257    lp_exec_mask_init(&bld.exec_mask, &bld.bld_base.base);
2258
2259    bld.system_values = *system_values;
2260
2261    lp_build_tgsi_llvm(&bld.bld_base, tokens);
2262
2263    if (0) {
2264       LLVMBasicBlockRef block = LLVMGetInsertBlock(gallivm->builder);
2265       LLVMValueRef function = LLVMGetBasicBlockParent(block);
2266       debug_printf("11111111111111111111111111111 \n");
2267       tgsi_dump(tokens, 0);
2268       lp_debug_dump_value(function);
2269       debug_printf("2222222222222222222222222222 \n");
2270    }
2271
2272    if (0) {
2273       LLVMModuleRef module = LLVMGetGlobalParent(
2274          LLVMGetBasicBlockParent(LLVMGetInsertBlock(gallivm->builder)));
2275       LLVMDumpModule(module);
2276
2277    }
2278 }