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 #define DUMP_GS_EMITS 0
  69
  70 static void lp_exec_mask_init(struct lp_exec_mask *mask, struct lp_build_context *bld)
  71 {
  72    LLVMTypeRef int_type = LLVMInt32TypeInContext(bld->gallivm->context);
  73    LLVMBuilderRef builder = bld->gallivm->builder;
  74
  75    mask->bld = bld;
  76    mask->has_mask = FALSE;
  77    mask->ret_in_main = FALSE;
  78    mask->cond_stack_size = 0;
  79    mask->loop_stack_size = 0;
  80    mask->call_stack_size = 0;
  81
  82    mask->int_vec_type = lp_build_int_vec_type(bld->gallivm, mask->bld->type);
  83    mask->exec_mask = mask->ret_mask = mask->break_mask = mask->cont_mask = mask->cond_mask =
  84          LLVMConstAllOnes(mask->int_vec_type);
  85
  86    mask->loop_limiter = lp_build_alloca(bld->gallivm, int_type, "looplimiter");
  87
  88    LLVMBuildStore(
  89       builder,
  90       LLVMConstInt(int_type, LP_MAX_TGSI_LOOP_ITERATIONS, false),
  91       mask->loop_limiter);
  92 }
  93
  94 static void lp_exec_mask_update(struct lp_exec_mask *mask)
  95 {
  96    LLVMBuilderRef builder = mask->bld->gallivm->builder;
  97
  98    if (mask->loop_stack_size) {
  99       /*for loops we need to update the entire mask at runtime */
 100       LLVMValueRef tmp;
 101       assert(mask->break_mask);
 102       tmp = LLVMBuildAnd(builder,
 103                          mask->cont_mask,
 104                          mask->break_mask,
 105                          "maskcb");
 106       mask->exec_mask = LLVMBuildAnd(builder,
 107                                      mask->cond_mask,
 108                                      tmp,
 109                                      "maskfull");
 110    } else
 111       mask->exec_mask = mask->cond_mask;
 112
 113    if (mask->call_stack_size || mask->ret_in_main) {
 114       mask->exec_mask = LLVMBuildAnd(builder,
 115                                      mask->exec_mask,
 116                                      mask->ret_mask,
 117                                      "callmask");
 118    }
 119
 120    mask->has_mask = (mask->cond_stack_size > 0 ||
 121                      mask->loop_stack_size > 0 ||
 122                      mask->call_stack_size > 0 ||
 123                      mask->ret_in_main);
 124 }
 125
 126 static void lp_exec_mask_cond_push(struct lp_exec_mask *mask,
 127                                    LLVMValueRef val)
 128 {
 129    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 130
 131    assert(mask->cond_stack_size < LP_MAX_TGSI_NESTING);
 132    if (mask->cond_stack_size == 0) {
 133       assert(mask->cond_mask == LLVMConstAllOnes(mask->int_vec_type));
 134    }
 135    mask->cond_stack[mask->cond_stack_size++] = mask->cond_mask;
 136    assert(LLVMTypeOf(val) == mask->int_vec_type);
 137    mask->cond_mask = LLVMBuildAnd(builder,
 138                                   mask->cond_mask,
 139                                   val,
 140                                   "");
 141    lp_exec_mask_update(mask);
 142 }
 143
 144 static void lp_exec_mask_cond_invert(struct lp_exec_mask *mask)
 145 {
 146    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 147    LLVMValueRef prev_mask;
 148    LLVMValueRef inv_mask;
 149
 150    assert(mask->cond_stack_size);
 151    prev_mask = mask->cond_stack[mask->cond_stack_size - 1];
 152    if (mask->cond_stack_size == 1) {
 153       assert(prev_mask == LLVMConstAllOnes(mask->int_vec_type));
 154    }
 155
 156    inv_mask = LLVMBuildNot(builder, mask->cond_mask, "");
 157
 158    mask->cond_mask = LLVMBuildAnd(builder,
 159                                   inv_mask,
 160                                   prev_mask, "");
 161    lp_exec_mask_update(mask);
 162 }
 163
 164 static void lp_exec_mask_cond_pop(struct lp_exec_mask *mask)
 165 {
 166    assert(mask->cond_stack_size);
 167    mask->cond_mask = mask->cond_stack[--mask->cond_stack_size];
 168    lp_exec_mask_update(mask);
 169 }
 170
 171 static void lp_exec_bgnloop(struct lp_exec_mask *mask)
 172 {
 173    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 174
 175    if (mask->loop_stack_size == 0) {
 176       assert(mask->loop_block == NULL);
 177       assert(mask->cont_mask == LLVMConstAllOnes(mask->int_vec_type));
 178       assert(mask->break_mask == LLVMConstAllOnes(mask->int_vec_type));
 179       assert(mask->break_var == NULL);
 180    }
 181
 182    assert(mask->loop_stack_size < LP_MAX_TGSI_NESTING);
 183
 184    mask->loop_stack[mask->loop_stack_size].loop_block = mask->loop_block;
 185    mask->loop_stack[mask->loop_stack_size].cont_mask = mask->cont_mask;
 186    mask->loop_stack[mask->loop_stack_size].break_mask = mask->break_mask;
 187    mask->loop_stack[mask->loop_stack_size].break_var = mask->break_var;
 188    ++mask->loop_stack_size;
 189
 190    mask->break_var = lp_build_alloca(mask->bld->gallivm, mask->int_vec_type, "");
 191    LLVMBuildStore(builder, mask->break_mask, mask->break_var);
 192
 193    mask->loop_block = lp_build_insert_new_block(mask->bld->gallivm, "bgnloop");
 194
 195    LLVMBuildBr(builder, mask->loop_block);
 196    LLVMPositionBuilderAtEnd(builder, mask->loop_block);
 197
 198    mask->break_mask = LLVMBuildLoad(builder, mask->break_var, "");
 199
 200    lp_exec_mask_update(mask);
 201 }
 202
 203 static void lp_exec_break(struct lp_exec_mask *mask)
 204 {
 205    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 206    LLVMValueRef exec_mask = LLVMBuildNot(builder,
 207                                          mask->exec_mask,
 208                                          "break");
 209
 210    mask->break_mask = LLVMBuildAnd(builder,
 211                                    mask->break_mask,
 212                                    exec_mask, "break_full");
 213
 214    lp_exec_mask_update(mask);
 215 }
 216
 217 static void lp_exec_break_condition(struct lp_exec_mask *mask,
 218                                     LLVMValueRef cond)
 219 {
 220    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 221    LLVMValueRef cond_mask = LLVMBuildAnd(builder,
 222                                          mask->exec_mask,
 223                                          cond, "cond_mask");
 224    cond_mask = LLVMBuildNot(builder, cond_mask, "break_cond");
 225
 226    mask->break_mask = LLVMBuildAnd(builder,
 227                                    mask->break_mask,
 228                                    cond_mask, "breakc_full");
 229
 230    lp_exec_mask_update(mask);
 231 }
 232
 233 static void lp_exec_continue(struct lp_exec_mask *mask)
 234 {
 235    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 236    LLVMValueRef exec_mask = LLVMBuildNot(builder,
 237                                          mask->exec_mask,
 238                                          "");
 239
 240    mask->cont_mask = LLVMBuildAnd(builder,
 241                                   mask->cont_mask,
 242                                   exec_mask, "");
 243
 244    lp_exec_mask_update(mask);
 245 }
 246
 247
 248 static void lp_exec_endloop(struct gallivm_state *gallivm,
 249                             struct lp_exec_mask *mask)
 250 {
 251    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 252    LLVMBasicBlockRef endloop;
 253    LLVMTypeRef int_type = LLVMInt32TypeInContext(mask->bld->gallivm->context);
 254    LLVMTypeRef reg_type = LLVMIntTypeInContext(gallivm->context,
 255                                                mask->bld->type.width *
 256                                                mask->bld->type.length);
 257    LLVMValueRef i1cond, i2cond, icond, limiter;
 258
 259    assert(mask->break_mask);
 260
 261    /*
 262     * Restore the cont_mask, but don't pop
 263     */
 264    assert(mask->loop_stack_size);
 265    mask->cont_mask = mask->loop_stack[mask->loop_stack_size - 1].cont_mask;
 266    lp_exec_mask_update(mask);
 267
 268    /*
 269     * Unlike the continue mask, the break_mask must be preserved across loop
 270     * iterations
 271     */
 272    LLVMBuildStore(builder, mask->break_mask, mask->break_var);
 273
 274    /* Decrement the loop limiter */
 275    limiter = LLVMBuildLoad(builder, mask->loop_limiter, "");
 276
 277    limiter = LLVMBuildSub(
 278       builder,
 279       limiter,
 280       LLVMConstInt(int_type, 1, false),
 281       "");
 282
 283    LLVMBuildStore(builder, limiter, mask->loop_limiter);
 284
 285    /* i1cond = (mask != 0) */
 286    i1cond = LLVMBuildICmp(
 287       builder,
 288       LLVMIntNE,
 289       LLVMBuildBitCast(builder, mask->exec_mask, reg_type, ""),
 290       LLVMConstNull(reg_type), "i1cond");
 291
 292    /* i2cond = (looplimiter > 0) */
 293    i2cond = LLVMBuildICmp(
 294       builder,
 295       LLVMIntSGT,
 296       limiter,
 297       LLVMConstNull(int_type), "i2cond");
 298
 299    /* if( i1cond && i2cond ) */
 300    icond = LLVMBuildAnd(builder, i1cond, i2cond, "");
 301
 302    endloop = lp_build_insert_new_block(mask->bld->gallivm, "endloop");
 303
 304    LLVMBuildCondBr(builder,
 305                    icond, mask->loop_block, endloop);
 306
 307    LLVMPositionBuilderAtEnd(builder, endloop);
 308
 309    assert(mask->loop_stack_size);
 310    --mask->loop_stack_size;
 311    mask->loop_block = mask->loop_stack[mask->loop_stack_size].loop_block;
 312    mask->cont_mask = mask->loop_stack[mask->loop_stack_size].cont_mask;
 313    mask->break_mask = mask->loop_stack[mask->loop_stack_size].break_mask;
 314    mask->break_var = mask->loop_stack[mask->loop_stack_size].break_var;
 315
 316    lp_exec_mask_update(mask);
 317 }
 318
 319 /* stores val into an address pointed to by dst.
 320  * mask->exec_mask is used to figure out which bits of val
 321  * should be stored into the address
 322  * (0 means don't store this bit, 1 means do store).
 323  */
 324 static void lp_exec_mask_store(struct lp_exec_mask *mask,
 325                                struct lp_build_context *bld_store,
 326                                LLVMValueRef pred,
 327                                LLVMValueRef val,
 328                                LLVMValueRef dst)
 329 {
 330    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 331
 332    /* Mix the predicate and execution mask */
 333    if (mask->has_mask) {
 334       if (pred) {
 335          pred = LLVMBuildAnd(builder, pred, mask->exec_mask, "");
 336       } else {
 337          pred = mask->exec_mask;
 338       }
 339    }
 340
 341    if (pred) {
 342       LLVMValueRef real_val, dst_val;
 343
 344       dst_val = LLVMBuildLoad(builder, dst, "");
 345       real_val = lp_build_select(bld_store,
 346                                  pred,
 347                                  val, dst_val);
 348
 349       LLVMBuildStore(builder, real_val, dst);
 350    } else
 351       LLVMBuildStore(builder, val, dst);
 352 }
 353
 354 static void lp_exec_mask_call(struct lp_exec_mask *mask,
 355                               int func,
 356                               int *pc)
 357 {
 358    assert(mask->call_stack_size < LP_MAX_TGSI_NESTING);
 359    mask->call_stack[mask->call_stack_size].pc = *pc;
 360    mask->call_stack[mask->call_stack_size].ret_mask = mask->ret_mask;
 361    mask->call_stack_size++;
 362    *pc = func;
 363 }
 364
 365 static void lp_exec_mask_ret(struct lp_exec_mask *mask, int *pc)
 366 {
 367    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 368    LLVMValueRef exec_mask;
 369
 370    if (mask->cond_stack_size == 0 &&
 371        mask->loop_stack_size == 0 &&
 372        mask->call_stack_size == 0) {
 373       /* returning from main() */
 374       *pc = -1;
 375       return;
 376    }
 377
 378    if (mask->call_stack_size == 0) {
 379       /*
 380        * This requires special handling since we need to ensure
 381        * we don't drop the mask even if we have no call stack
 382        * (e.g. after a ret in a if clause after the endif)
 383        */
 384       mask->ret_in_main = TRUE;
 385    }
 386
 387    exec_mask = LLVMBuildNot(builder,
 388                             mask->exec_mask,
 389                             "ret");
 390
 391    mask->ret_mask = LLVMBuildAnd(builder,
 392                                  mask->ret_mask,
 393                                  exec_mask, "ret_full");
 394
 395    lp_exec_mask_update(mask);
 396 }
 397
 398 static void lp_exec_mask_bgnsub(struct lp_exec_mask *mask)
 399 {
 400 }
 401
 402 static void lp_exec_mask_endsub(struct lp_exec_mask *mask, int *pc)
 403 {
 404    assert(mask->call_stack_size);
 405    mask->call_stack_size--;
 406    *pc = mask->call_stack[mask->call_stack_size].pc;
 407    mask->ret_mask = mask->call_stack[mask->call_stack_size].ret_mask;
 408    lp_exec_mask_update(mask);
 409 }
 410
 411
 412 /**
 413  * Return pointer to a temporary register channel (src or dest).
 414  * Note that indirect addressing cannot be handled here.
 415  * \param index  which temporary register
 416  * \param chan  which channel of the temp register.
 417  */
 418 LLVMValueRef
 419 lp_get_temp_ptr_soa(struct lp_build_tgsi_soa_context *bld,
 420              unsigned index,
 421              unsigned chan)
 422 {
 423    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
 424    assert(chan < 4);
 425    if (bld->indirect_files & (1 << TGSI_FILE_TEMPORARY)) {
 426       LLVMValueRef lindex = lp_build_const_int32(bld->bld_base.base.gallivm, index * 4 + chan);
 427       return LLVMBuildGEP(builder, bld->temps_array, &lindex, 1, "");
 428    }
 429    else {
 430       return bld->temps[index][chan];
 431    }
 432 }
 433
 434 /**
 435  * Return pointer to a output register channel (src or dest).
 436  * Note that indirect addressing cannot be handled here.
 437  * \param index  which output register
 438  * \param chan  which channel of the output register.
 439  */
 440 LLVMValueRef
 441 lp_get_output_ptr(struct lp_build_tgsi_soa_context *bld,
 442                unsigned index,
 443                unsigned chan)
 444 {
 445    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
 446    assert(chan < 4);
 447    if (bld->indirect_files & (1 << TGSI_FILE_OUTPUT)) {
 448       LLVMValueRef lindex = lp_build_const_int32(bld->bld_base.base.gallivm,
 449                                                  index * 4 + chan);
 450       return LLVMBuildGEP(builder, bld->outputs_array, &lindex, 1, "");
 451    }
 452    else {
 453       return bld->outputs[index][chan];
 454    }
 455 }
 456
 457 /*
 458  * If we have indirect addressing in outputs copy our alloca array
 459  * to the outputs slots specified by the caller to make sure
 460  * our outputs are delivered consistently via the same interface.
 461  */
 462 static void
 463 gather_outputs(struct lp_build_tgsi_soa_context * bld)
 464 {
 465    if ((bld->indirect_files & (1 << TGSI_FILE_OUTPUT))) {
 466       unsigned index, chan;
 467       assert(bld->bld_base.info->num_outputs <=
 468              bld->bld_base.info->file_max[TGSI_FILE_OUTPUT] + 1);
 469       for (index = 0; index < bld->bld_base.info->num_outputs; ++index) {
 470          for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
 471             bld->outputs[index][chan] = lp_get_output_ptr(bld, index, chan);
 472          }
 473       }
 474    }
 475 }
 476
 477 /**
 478  * Gather vector.
 479  * XXX the lp_build_gather() function should be capable of doing this
 480  * with a little work.
 481  */
 482 static LLVMValueRef
 483 build_gather(struct lp_build_context *bld,
 484              LLVMValueRef base_ptr,
 485              LLVMValueRef indexes)
 486 {
 487    LLVMBuilderRef builder = bld->gallivm->builder;
 488    LLVMValueRef res = bld->undef;
 489    unsigned i;
 490
 491    /*
 492     * Loop over elements of index_vec, load scalar value, insert it into 'res'.
 493     */
 494    for (i = 0; i < bld->type.length; i++) {
 495       LLVMValueRef ii = lp_build_const_int32(bld->gallivm, i);
 496       LLVMValueRef index = LLVMBuildExtractElement(builder,
 497                                                    indexes, ii, "");
 498       LLVMValueRef scalar_ptr = LLVMBuildGEP(builder, base_ptr,
 499                                              &index, 1, "gather_ptr");
 500       LLVMValueRef scalar = LLVMBuildLoad(builder, scalar_ptr, "");
 501
 502       res = LLVMBuildInsertElement(builder, res, scalar, ii, "");
 503    }
 504
 505    return res;
 506 }
 507
 508
 509 /**
 510  * Scatter/store vector.
 511  */
 512 static void
 513 emit_mask_scatter(struct lp_build_tgsi_soa_context *bld,
 514                   LLVMValueRef base_ptr,
 515                   LLVMValueRef indexes,
 516                   LLVMValueRef values,
 517                   struct lp_exec_mask *mask,
 518                   LLVMValueRef pred)
 519 {
 520    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
 521    LLVMBuilderRef builder = gallivm->builder;
 522    unsigned i;
 523
 524    /* Mix the predicate and execution mask */
 525    if (mask->has_mask) {
 526       if (pred) {
 527          pred = LLVMBuildAnd(builder, pred, mask->exec_mask, "");
 528       }
 529       else {
 530          pred = mask->exec_mask;
 531       }
 532    }
 533
 534    /*
 535     * Loop over elements of index_vec, store scalar value.
 536     */
 537    for (i = 0; i < bld->bld_base.base.type.length; i++) {
 538       LLVMValueRef ii = lp_build_const_int32(gallivm, i);
 539       LLVMValueRef index = LLVMBuildExtractElement(builder, indexes, ii, "");
 540       LLVMValueRef scalar_ptr = LLVMBuildGEP(builder, base_ptr, &index, 1, "scatter_ptr");
 541       LLVMValueRef val = LLVMBuildExtractElement(builder, values, ii, "scatter_val");
 542       LLVMValueRef scalar_pred = pred ?
 543          LLVMBuildExtractElement(builder, pred, ii, "scatter_pred") : NULL;
 544
 545       if (0)
 546          lp_build_printf(gallivm, "scatter %d: val %f at %d %p\n",
 547                          ii, val, index, scalar_ptr);
 548
 549       if (scalar_pred) {
 550          LLVMValueRef real_val, dst_val;
 551          dst_val = LLVMBuildLoad(builder, scalar_ptr, "");
 552          real_val = lp_build_select(&bld->elem_bld, scalar_pred, val, dst_val);
 553          LLVMBuildStore(builder, real_val, scalar_ptr);
 554       }
 555       else {
 556          LLVMBuildStore(builder, val, scalar_ptr);
 557       }
 558    }
 559 }
 560
 561
 562 /**
 563  * Read the current value of the ADDR register, convert the floats to
 564  * ints, add the base index and return the vector of offsets.
 565  * The offsets will be used to index into the constant buffer or
 566  * temporary register file.
 567  */
 568 static LLVMValueRef
 569 get_indirect_index(struct lp_build_tgsi_soa_context *bld,
 570                    unsigned reg_file, unsigned reg_index,
 571                    const struct tgsi_ind_register *indirect_reg)
 572 {
 573    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
 574    struct lp_build_context *uint_bld = &bld->bld_base.uint_bld;
 575    /* always use X component of address register */
 576    unsigned swizzle = indirect_reg->Swizzle;
 577    LLVMValueRef base;
 578    LLVMValueRef rel;
 579    LLVMValueRef max_index;
 580    LLVMValueRef index;
 581
 582    assert(bld->indirect_files & (1 << reg_file));
 583
 584    base = lp_build_const_int_vec(bld->bld_base.base.gallivm, uint_bld->type, reg_index);
 585
 586    assert(swizzle < 4);
 587    switch (indirect_reg->File) {
 588    case TGSI_FILE_ADDRESS:
 589       rel = LLVMBuildLoad(builder,
 590                           bld->addr[indirect_reg->Index][swizzle],
 591                           "load addr reg");
 592       /* ADDR LLVM values already have LLVM integer type. */
 593       break;
 594    case TGSI_FILE_TEMPORARY:
 595       rel = lp_get_temp_ptr_soa(bld, indirect_reg->Index, swizzle);
 596       rel = LLVMBuildLoad(builder, rel, "load temp reg");
 597       /* TEMP LLVM values always have LLVM float type, but for indirection, the
 598        * value actually stored is expected to be an integer */
 599       rel = LLVMBuildBitCast(builder, rel, uint_bld->vec_type, "");
 600       break;
 601    default:
 602       assert(0);
 603       rel = uint_bld->zero;
 604    }
 605
 606    index = lp_build_add(uint_bld, base, rel);
 607
 608    max_index = lp_build_const_int_vec(bld->bld_base.base.gallivm,
 609                                       uint_bld->type,
 610                                       bld->bld_base.info->file_max[reg_file]);
 611
 612    assert(!uint_bld->type.sign);
 613    index = lp_build_min(uint_bld, index, max_index);
 614
 615    return index;
 616 }
 617
 618 static struct lp_build_context *
 619 stype_to_fetch(struct lp_build_tgsi_context * bld_base,
 620                enum tgsi_opcode_type stype)
 621 {
 622    struct lp_build_context *bld_fetch;
 623
 624    switch (stype) {
 625    case TGSI_TYPE_FLOAT:
 626    case TGSI_TYPE_UNTYPED:
 627       bld_fetch = &bld_base->base;
 628       break;
 629    case TGSI_TYPE_UNSIGNED:
 630       bld_fetch = &bld_base->uint_bld;
 631       break;
 632    case TGSI_TYPE_SIGNED:
 633       bld_fetch = &bld_base->int_bld;
 634       break;
 635    case TGSI_TYPE_VOID:
 636    case TGSI_TYPE_DOUBLE:
 637    default:
 638       assert(0);
 639       bld_fetch = NULL;
 640       break;
 641    }
 642    return bld_fetch;
 643 }
 644
 645 static LLVMValueRef
 646 emit_fetch_constant(
 647    struct lp_build_tgsi_context * bld_base,
 648    const struct tgsi_full_src_register * reg,
 649    enum tgsi_opcode_type stype,
 650    unsigned swizzle)
 651 {
 652    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
 653    struct gallivm_state *gallivm = bld_base->base.gallivm;
 654    LLVMBuilderRef builder = gallivm->builder;
 655    struct lp_build_context *uint_bld = &bld_base->uint_bld;
 656    LLVMValueRef indirect_index = NULL;
 657    unsigned dimension = 0;
 658    LLVMValueRef dimension_index;
 659    LLVMValueRef consts_ptr;
 660    LLVMValueRef res;
 661
 662    /* XXX: Handle fetching xyzw components as a vector */
 663    assert(swizzle != ~0);
 664
 665    if (reg->Register.Dimension) {
 666       assert(!reg->Dimension.Indirect);
 667       dimension = reg->Dimension.Index;
 668       assert(dimension < LP_MAX_TGSI_CONST_BUFFERS);
 669    }
 670
 671    dimension_index = lp_build_const_int32(gallivm, dimension);
 672    consts_ptr = lp_build_array_get(gallivm, bld->consts_ptr, dimension_index);
 673
 674    if (reg->Register.Indirect) {
 675       indirect_index = get_indirect_index(bld,
 676                                           reg->Register.File,
 677                                           reg->Register.Index,
 678                                           &reg->Indirect);
 679    }
 680
 681    if (reg->Register.Indirect) {
 682       LLVMValueRef swizzle_vec =
 683          lp_build_const_int_vec(bld->bld_base.base.gallivm, uint_bld->type, swizzle);
 684       LLVMValueRef index_vec;  /* index into the const buffer */
 685
 686       /* index_vec = indirect_index * 4 + swizzle */
 687       index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
 688       index_vec = lp_build_add(uint_bld, index_vec, swizzle_vec);
 689
 690       /* Gather values from the constant buffer */
 691       res = build_gather(&bld_base->base, consts_ptr, index_vec);
 692    }
 693    else {
 694       LLVMValueRef index;  /* index into the const buffer */
 695       LLVMValueRef scalar, scalar_ptr;
 696
 697       index = lp_build_const_int32(gallivm, reg->Register.Index*4 + swizzle);
 698
 699       scalar_ptr = LLVMBuildGEP(builder, consts_ptr,
 700                                 &index, 1, "");
 701       scalar = LLVMBuildLoad(builder, scalar_ptr, "");
 702       res = lp_build_broadcast_scalar(&bld_base->base, scalar);
 703    }
 704
 705    if (stype == TGSI_TYPE_SIGNED || stype == TGSI_TYPE_UNSIGNED) {
 706       struct lp_build_context *bld_fetch = stype_to_fetch(bld_base, stype);
 707       res = LLVMBuildBitCast(builder, res, bld_fetch->vec_type, "");
 708    }
 709    return res;
 710 }
 711
 712 static LLVMValueRef
 713 emit_fetch_immediate(
 714    struct lp_build_tgsi_context * bld_base,
 715    const struct tgsi_full_src_register * reg,
 716    enum tgsi_opcode_type stype,
 717    unsigned swizzle)
 718 {
 719    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
 720    LLVMValueRef res = bld->immediates[reg->Register.Index][swizzle];
 721    assert(res);
 722
 723    if (stype == TGSI_TYPE_UNSIGNED) {
 724       res = LLVMConstBitCast(res, bld_base->uint_bld.vec_type);
 725    } else if (stype == TGSI_TYPE_SIGNED) {
 726       res = LLVMConstBitCast(res, bld_base->int_bld.vec_type);
 727    }
 728    return res;
 729 }
 730
 731 static LLVMValueRef
 732 emit_fetch_input(
 733    struct lp_build_tgsi_context * bld_base,
 734    const struct tgsi_full_src_register * reg,
 735    enum tgsi_opcode_type stype,
 736    unsigned swizzle)
 737 {
 738    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
 739    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
 740    LLVMBuilderRef builder = gallivm->builder;
 741    struct lp_build_context *uint_bld = &bld_base->uint_bld;
 742    LLVMValueRef indirect_index = NULL;
 743    LLVMValueRef res;
 744
 745    if (reg->Register.Indirect) {
 746       indirect_index = get_indirect_index(bld,
 747                                           reg->Register.File,
 748                                           reg->Register.Index,
 749                                           &reg->Indirect);
 750    }
 751
 752    if (reg->Register.Indirect) {
 753       LLVMValueRef swizzle_vec =
 754          lp_build_const_int_vec(gallivm, uint_bld->type, swizzle);
 755       LLVMValueRef length_vec =
 756          lp_build_const_int_vec(gallivm, uint_bld->type, bld->bld_base.base.type.length);
 757       LLVMValueRef index_vec;  /* index into the const buffer */
 758       LLVMValueRef inputs_array;
 759       LLVMTypeRef float4_ptr_type;
 760
 761       /* index_vec = (indirect_index * 4 + swizzle) * length */
 762       index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
 763       index_vec = lp_build_add(uint_bld, index_vec, swizzle_vec);
 764       index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
 765
 766       /* cast inputs_array pointer to float* */
 767       float4_ptr_type = LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
 768       inputs_array = LLVMBuildBitCast(builder, bld->inputs_array,
 769                                          float4_ptr_type, "");
 770
 771       /* Gather values from the temporary register array */
 772       res = build_gather(&bld_base->base, inputs_array, index_vec);
 773    } else {
 774       if (bld->indirect_files & (1 << TGSI_FILE_INPUT)) {
 775          LLVMValueRef lindex = lp_build_const_int32(gallivm,
 776                                         reg->Register.Index * 4 + swizzle);
 777          LLVMValueRef input_ptr =  LLVMBuildGEP(builder,
 778                                                 bld->inputs_array, &lindex, 1, "");
 779          res = LLVMBuildLoad(builder, input_ptr, "");
 780       }
 781       else {
 782          res = bld->inputs[reg->Register.Index][swizzle];
 783       }
 784    }
 785
 786    assert(res);
 787
 788    if (stype == TGSI_TYPE_UNSIGNED) {
 789       res = LLVMBuildBitCast(builder, res, bld_base->uint_bld.vec_type, "");
 790    } else if (stype == TGSI_TYPE_SIGNED) {
 791       res = LLVMBuildBitCast(builder, res, bld_base->int_bld.vec_type, "");
 792    }
 793
 794    return res;
 795 }
 796
 797
 798 static LLVMValueRef
 799 emit_fetch_gs_input(
 800    struct lp_build_tgsi_context * bld_base,
 801    const struct tgsi_full_src_register * reg,
 802    enum tgsi_opcode_type stype,
 803    unsigned swizzle)
 804 {
 805    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
 806    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
 807    LLVMBuilderRef builder = gallivm->builder;
 808    LLVMValueRef attrib_index = NULL;
 809    LLVMValueRef vertex_index = NULL;
 810    LLVMValueRef swizzle_index = lp_build_const_int32(gallivm, swizzle);
 811    LLVMValueRef res;
 812
 813    if (reg->Register.Indirect) {
 814       attrib_index = get_indirect_index(bld,
 815                                           reg->Register.File,
 816                                           reg->Register.Index,
 817                                           &reg->Indirect);
 818    } else {
 819       attrib_index = lp_build_const_int32(gallivm, reg->Register.Index);
 820    }
 821
 822    if (reg->Dimension.Indirect) {
 823       vertex_index = get_indirect_index(bld,
 824                                         reg->Register.File,
 825                                         reg->Dimension.Index,
 826                                         &reg->DimIndirect);
 827    } else {
 828       vertex_index = lp_build_const_int32(gallivm, reg->Dimension.Index);
 829    }
 830
 831    res = bld->gs_iface->fetch_input(bld->gs_iface, bld_base,
 832                                     reg->Dimension.Indirect,
 833                                     vertex_index, attrib_index,
 834                                     swizzle_index);
 835
 836    assert(res);
 837
 838    if (stype == TGSI_TYPE_UNSIGNED) {
 839       res = LLVMBuildBitCast(builder, res, bld_base->uint_bld.vec_type, "");
 840    } else if (stype == TGSI_TYPE_SIGNED) {
 841       res = LLVMBuildBitCast(builder, res, bld_base->int_bld.vec_type, "");
 842    }
 843
 844    return res;
 845 }
 846
 847 static LLVMValueRef
 848 emit_fetch_temporary(
 849    struct lp_build_tgsi_context * bld_base,
 850    const struct tgsi_full_src_register * reg,
 851    enum tgsi_opcode_type stype,
 852    unsigned swizzle)
 853 {
 854    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
 855    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
 856    LLVMBuilderRef builder = gallivm->builder;
 857    struct lp_build_context *uint_bld = &bld_base->uint_bld;
 858    LLVMValueRef indirect_index = NULL;
 859    LLVMValueRef res;
 860
 861    if (reg->Register.Indirect) {
 862       indirect_index = get_indirect_index(bld,
 863                                           reg->Register.File,
 864                                           reg->Register.Index,
 865                                           &reg->Indirect);
 866    }
 867
 868    if (reg->Register.Indirect) {
 869       LLVMValueRef swizzle_vec =
 870          lp_build_const_int_vec(bld->bld_base.base.gallivm, uint_bld->type, swizzle);
 871       LLVMValueRef length_vec =
 872          lp_build_const_int_vec(bld->bld_base.base.gallivm, uint_bld->type,
 873                                 bld->bld_base.base.type.length);
 874       LLVMValueRef index_vec;  /* index into the const buffer */
 875       LLVMValueRef temps_array;
 876       LLVMTypeRef float4_ptr_type;
 877
 878       /* index_vec = (indirect_index * 4 + swizzle) * length */
 879       index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
 880       index_vec = lp_build_add(uint_bld, index_vec, swizzle_vec);
 881       index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
 882
 883       /* cast temps_array pointer to float* */
 884       float4_ptr_type = LLVMPointerType(LLVMFloatTypeInContext(bld->bld_base.base.gallivm->context), 0);
 885       temps_array = LLVMBuildBitCast(builder, bld->temps_array,
 886                                      float4_ptr_type, "");
 887
 888       /* Gather values from the temporary register array */
 889       res = build_gather(&bld_base->base, temps_array, index_vec);
 890    }
 891    else {
 892       LLVMValueRef temp_ptr;
 893       temp_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index, swizzle);
 894       res = LLVMBuildLoad(builder, temp_ptr, "");
 895    }
 896
 897    if (stype == TGSI_TYPE_SIGNED || stype == TGSI_TYPE_UNSIGNED) {
 898       struct lp_build_context *bld_fetch = stype_to_fetch(bld_base, stype);
 899       res = LLVMBuildBitCast(builder, res, bld_fetch->vec_type, "");
 900    }
 901
 902    return res;
 903 }
 904
 905 static LLVMValueRef
 906 emit_fetch_system_value(
 907    struct lp_build_tgsi_context * bld_base,
 908    const struct tgsi_full_src_register * reg,
 909    enum tgsi_opcode_type stype,
 910    unsigned swizzle)
 911 {
 912    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
 913    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
 914    const struct tgsi_shader_info *info = bld->bld_base.info;
 915    LLVMBuilderRef builder = gallivm->builder;
 916    LLVMValueRef res;
 917    enum tgsi_opcode_type atype; // Actual type of the value
 918
 919    assert(!reg->Register.Indirect);
 920
 921    switch (info->system_value_semantic_name[reg->Register.Index]) {
 922    case TGSI_SEMANTIC_INSTANCEID:
 923       res = lp_build_broadcast_scalar(&bld_base->uint_bld, bld->system_values.instance_id);
 924       atype = TGSI_TYPE_UNSIGNED;
 925       break;
 926
 927    case TGSI_SEMANTIC_VERTEXID:
 928       res = bld->system_values.vertex_id;
 929       atype = TGSI_TYPE_UNSIGNED;
 930       break;
 931
 932    case TGSI_SEMANTIC_PRIMID:
 933       res = bld->system_values.prim_id;
 934       atype = TGSI_TYPE_UNSIGNED;
 935       break;
 936
 937    default:
 938       assert(!"unexpected semantic in emit_fetch_system_value");
 939       res = bld_base->base.zero;
 940       atype = TGSI_TYPE_FLOAT;
 941       break;
 942    }
 943
 944    if (atype != stype) {
 945       if (stype == TGSI_TYPE_FLOAT) {
 946          res = LLVMBuildBitCast(builder, res, bld_base->base.vec_type, "");
 947       } else if (stype == TGSI_TYPE_UNSIGNED) {
 948          res = LLVMBuildBitCast(builder, res, bld_base->uint_bld.vec_type, "");
 949       } else if (stype == TGSI_TYPE_SIGNED) {
 950          res = LLVMBuildBitCast(builder, res, bld_base->int_bld.vec_type, "");
 951       }
 952    }
 953
 954    return res;
 955 }
 956
 957 /**
 958  * Register fetch with derivatives.
 959  */
 960 static void
 961 emit_fetch_deriv(
 962    struct lp_build_tgsi_soa_context *bld,
 963    LLVMValueRef src,
 964    LLVMValueRef *res,
 965    LLVMValueRef *ddx,
 966    LLVMValueRef *ddy)
 967 {
 968    if(res)
 969       *res = src;
 970
 971    /* TODO: use interpolation coeffs for inputs */
 972
 973    if(ddx)
 974       *ddx = lp_build_ddx(&bld->bld_base.base, src);
 975
 976    if(ddy)
 977       *ddy = lp_build_ddy(&bld->bld_base.base, src);
 978 }
 979
 980
 981 /**
 982  * Predicate.
 983  */
 984 static void
 985 emit_fetch_predicate(
 986    struct lp_build_tgsi_soa_context *bld,
 987    const struct tgsi_full_instruction *inst,
 988    LLVMValueRef *pred)
 989 {
 990    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
 991    unsigned index;
 992    unsigned char swizzles[4];
 993    LLVMValueRef unswizzled[4] = {NULL, NULL, NULL, NULL};
 994    LLVMValueRef value;
 995    unsigned chan;
 996
 997    if (!inst->Instruction.Predicate) {
 998       TGSI_FOR_EACH_CHANNEL( chan ) {
 999          pred[chan] = NULL;
1000       }
1001       return;
1002    }
1003
1004    swizzles[0] = inst->Predicate.SwizzleX;
1005    swizzles[1] = inst->Predicate.SwizzleY;
1006    swizzles[2] = inst->Predicate.SwizzleZ;
1007    swizzles[3] = inst->Predicate.SwizzleW;
1008
1009    index = inst->Predicate.Index;
1010    assert(index < LP_MAX_TGSI_PREDS);
1011
1012    TGSI_FOR_EACH_CHANNEL( chan ) {
1013       unsigned swizzle = swizzles[chan];
1014
1015       /*
1016        * Only fetch the predicate register channels that are actually listed
1017        * in the swizzles
1018        */
1019       if (!unswizzled[swizzle]) {
1020          value = LLVMBuildLoad(builder,
1021                                bld->preds[index][swizzle], "");
1022
1023          /*
1024           * Convert the value to an integer mask.
1025           *
1026           * TODO: Short-circuit this comparison -- a D3D setp_xx instructions
1027           * is needlessly causing two comparisons due to storing the intermediate
1028           * result as float vector instead of an integer mask vector.
1029           */
1030          value = lp_build_compare(bld->bld_base.base.gallivm,
1031                                   bld->bld_base.base.type,
1032                                   PIPE_FUNC_NOTEQUAL,
1033                                   value,
1034                                   bld->bld_base.base.zero);
1035          if (inst->Predicate.Negate) {
1036             value = LLVMBuildNot(builder, value, "");
1037          }
1038
1039          unswizzled[swizzle] = value;
1040       } else {
1041          value = unswizzled[swizzle];
1042       }
1043
1044       pred[chan] = value;
1045    }
1046 }
1047
1048 /**
1049  * Register store.
1050  */
1051 static void
1052 emit_store_chan(
1053    struct lp_build_tgsi_context *bld_base,
1054    const struct tgsi_full_instruction *inst,
1055    unsigned index,
1056    unsigned chan_index,
1057    LLVMValueRef pred,
1058    LLVMValueRef value)
1059 {
1060    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1061    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
1062    LLVMBuilderRef builder = gallivm->builder;
1063    const struct tgsi_full_dst_register *reg = &inst->Dst[index];
1064    struct lp_build_context *uint_bld = &bld_base->uint_bld;
1065    LLVMValueRef indirect_index = NULL;
1066    struct lp_build_context *bld_store;
1067    enum tgsi_opcode_type dtype = tgsi_opcode_infer_dst_type(inst->Instruction.Opcode);
1068
1069    switch (dtype) {
1070    default:
1071    case TGSI_TYPE_FLOAT:
1072    case TGSI_TYPE_UNTYPED:
1073       bld_store = &bld_base->base;
1074       break;
1075    case TGSI_TYPE_UNSIGNED:
1076       bld_store = &bld_base->uint_bld;
1077       break;
1078    case TGSI_TYPE_SIGNED:
1079       bld_store = &bld_base->int_bld;
1080       break;
1081    case TGSI_TYPE_DOUBLE:
1082    case TGSI_TYPE_VOID:
1083       assert(0);
1084       bld_store = NULL;
1085       break;
1086    }
1087
1088    /* If the destination is untyped then the source can be anything,
1089     * but LLVM won't like if the types don't match so lets cast
1090     * to the correct destination type as expected by LLVM. */
1091    if (dtype == TGSI_TYPE_UNTYPED &&
1092        !lp_check_vec_type(bld_store->type, LLVMTypeOf(value))) {
1093       value = LLVMBuildBitCast(builder, value, bld_store->vec_type,
1094                                "src_casted");
1095    }
1096
1097    switch( inst->Instruction.Saturate ) {
1098    case TGSI_SAT_NONE:
1099       break;
1100
1101    case TGSI_SAT_ZERO_ONE:
1102       value = lp_build_max(&bld->bld_base.base, value, bld->bld_base.base.zero);
1103       value = lp_build_min(&bld->bld_base.base, value, bld->bld_base.base.one);
1104       break;
1105
1106    case TGSI_SAT_MINUS_PLUS_ONE:
1107       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));
1108       value = lp_build_min(&bld->bld_base.base, value, bld->bld_base.base.one);
1109       break;
1110
1111    default:
1112       assert(0);
1113    }
1114
1115    if (reg->Register.Indirect) {
1116       indirect_index = get_indirect_index(bld,
1117                                           reg->Register.File,
1118                                           reg->Register.Index,
1119                                           &reg->Indirect);
1120    } else {
1121       assert(reg->Register.Index <=
1122                              bld->bld_base.info->file_max[reg->Register.File]);
1123    }
1124
1125    switch( reg->Register.File ) {
1126    case TGSI_FILE_OUTPUT:
1127       if (reg->Register.Indirect) {
1128          LLVMValueRef chan_vec =
1129             lp_build_const_int_vec(gallivm, uint_bld->type, chan_index);
1130          LLVMValueRef length_vec =
1131             lp_build_const_int_vec(gallivm, uint_bld->type, bld->bld_base.base.type.length);
1132          LLVMValueRef index_vec;  /* indexes into the temp registers */
1133          LLVMValueRef outputs_array;
1134          LLVMValueRef pixel_offsets;
1135          LLVMTypeRef float_ptr_type;
1136          int i;
1137
1138          /* build pixel offset vector: {0, 1, 2, 3, ...} */
1139          pixel_offsets = uint_bld->undef;
1140          for (i = 0; i < bld->bld_base.base.type.length; i++) {
1141             LLVMValueRef ii = lp_build_const_int32(gallivm, i);
1142             pixel_offsets = LLVMBuildInsertElement(builder, pixel_offsets,
1143                                                    ii, ii, "");
1144          }
1145
1146          /* index_vec = (indirect_index * 4 + chan_index) * length + offsets */
1147          index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
1148          index_vec = lp_build_add(uint_bld, index_vec, chan_vec);
1149          index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
1150          index_vec = lp_build_add(uint_bld, index_vec, pixel_offsets);
1151
1152          float_ptr_type =
1153             LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
1154          outputs_array = LLVMBuildBitCast(builder, bld->outputs_array,
1155                                           float_ptr_type, "");
1156
1157          /* Scatter store values into temp registers */
1158          emit_mask_scatter(bld, outputs_array, index_vec, value,
1159                            &bld->exec_mask, pred);
1160       }
1161       else {
1162          LLVMValueRef out_ptr = lp_get_output_ptr(bld, reg->Register.Index,
1163                                                   chan_index);
1164          lp_exec_mask_store(&bld->exec_mask, bld_store, pred, value, out_ptr);
1165       }
1166       break;
1167
1168    case TGSI_FILE_TEMPORARY:
1169       if (reg->Register.Indirect) {
1170          LLVMValueRef chan_vec =
1171             lp_build_const_int_vec(gallivm, uint_bld->type, chan_index);
1172          LLVMValueRef length_vec =
1173             lp_build_const_int_vec(gallivm, uint_bld->type,
1174                                    bld->bld_base.base.type.length);
1175          LLVMValueRef index_vec;  /* indexes into the temp registers */
1176          LLVMValueRef temps_array;
1177          LLVMValueRef pixel_offsets;
1178          LLVMTypeRef float_ptr_type;
1179          int i;
1180
1181          /* build pixel offset vector: {0, 1, 2, 3, ...} */
1182          pixel_offsets = uint_bld->undef;
1183          for (i = 0; i < bld->bld_base.base.type.length; i++) {
1184             LLVMValueRef ii = lp_build_const_int32(gallivm, i);
1185             pixel_offsets = LLVMBuildInsertElement(builder, pixel_offsets,
1186                                                    ii, ii, "");
1187          }
1188
1189          /* index_vec = (indirect_index * 4 + chan_index) * length + offsets */
1190          index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
1191          index_vec = lp_build_add(uint_bld, index_vec, chan_vec);
1192          index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
1193          index_vec = lp_build_add(uint_bld, index_vec, pixel_offsets);
1194
1195          float_ptr_type =
1196             LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
1197          temps_array = LLVMBuildBitCast(builder, bld->temps_array,
1198                                         float_ptr_type, "");
1199
1200          /* Scatter store values into temp registers */
1201          emit_mask_scatter(bld, temps_array, index_vec, value,
1202                            &bld->exec_mask, pred);
1203       }
1204       else {
1205          LLVMValueRef temp_ptr;
1206
1207          switch (dtype) {
1208          case TGSI_TYPE_UNSIGNED:
1209          case TGSI_TYPE_SIGNED: {
1210             LLVMTypeRef itype = bld_base->int_bld.vec_type;
1211             LLVMTypeRef ivtype = LLVMPointerType(itype, 0);
1212             LLVMValueRef tint_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index,
1213                                                         chan_index);
1214             LLVMValueRef temp_value_ptr;
1215
1216             temp_ptr = LLVMBuildBitCast(builder, tint_ptr, ivtype, "");
1217             temp_value_ptr = LLVMBuildBitCast(builder, value, itype, "");
1218             value = temp_value_ptr;
1219             break;
1220          }
1221          default:
1222          case TGSI_TYPE_FLOAT:
1223          case TGSI_TYPE_UNTYPED:
1224             temp_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index,
1225                                            chan_index);
1226             break;
1227          }
1228
1229          lp_exec_mask_store(&bld->exec_mask, bld_store, pred, value, temp_ptr);
1230       }
1231       break;
1232
1233    case TGSI_FILE_ADDRESS:
1234       assert(dtype == TGSI_TYPE_SIGNED);
1235       assert(LLVMTypeOf(value) == bld_base->base.int_vec_type);
1236       lp_exec_mask_store(&bld->exec_mask, bld_store, pred, value,
1237                          bld->addr[reg->Register.Index][chan_index]);
1238       break;
1239
1240    case TGSI_FILE_PREDICATE:
1241       lp_exec_mask_store(&bld->exec_mask, bld_store, pred, value,
1242                          bld->preds[reg->Register.Index][chan_index]);
1243       break;
1244
1245    default:
1246       assert( 0 );
1247    }
1248 }
1249
1250 static void
1251 emit_store(
1252    struct lp_build_tgsi_context * bld_base,
1253    const struct tgsi_full_instruction * inst,
1254    const struct tgsi_opcode_info * info,
1255    LLVMValueRef dst[4])
1256
1257 {
1258    unsigned chan_index;
1259    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1260
1261    if(info->num_dst) {
1262       LLVMValueRef pred[TGSI_NUM_CHANNELS];
1263
1264       emit_fetch_predicate( bld, inst, pred );
1265
1266       TGSI_FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1267          emit_store_chan(bld_base, inst, 0, chan_index, pred[chan_index], dst[chan_index]);
1268       }
1269    }
1270 }
1271
1272 /**
1273  * High-level instruction translators.
1274  */
1275
1276 static void
1277 emit_tex( struct lp_build_tgsi_soa_context *bld,
1278           const struct tgsi_full_instruction *inst,
1279           enum lp_build_tex_modifier modifier,
1280           LLVMValueRef *texel)
1281 {
1282    unsigned unit;
1283    LLVMValueRef lod_bias, explicit_lod;
1284    LLVMValueRef oow = NULL;
1285    LLVMValueRef coords[4];
1286    LLVMValueRef offsets[3] = { NULL };
1287    struct lp_derivatives derivs;
1288    struct lp_derivatives *deriv_ptr = NULL;
1289    unsigned num_coords, num_derivs, num_offsets;
1290    unsigned i;
1291
1292    if (!bld->sampler) {
1293       _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
1294       for (i = 0; i < 4; i++) {
1295          texel[i] = bld->bld_base.base.undef;
1296       }
1297       return;
1298    }
1299
1300    switch (inst->Texture.Texture) {
1301    case TGSI_TEXTURE_1D:
1302       num_coords = 1;
1303       num_offsets = 1;
1304       num_derivs = 1;
1305       break;
1306    case TGSI_TEXTURE_1D_ARRAY:
1307       num_coords = 2;
1308       num_offsets = 1;
1309       num_derivs = 1;
1310       break;
1311    case TGSI_TEXTURE_2D:
1312    case TGSI_TEXTURE_RECT:
1313       num_coords = 2;
1314       num_offsets = 2;
1315       num_derivs = 2;
1316       break;
1317    case TGSI_TEXTURE_SHADOW1D:
1318    case TGSI_TEXTURE_SHADOW1D_ARRAY:
1319       num_coords = 3;
1320       num_offsets = 1;
1321       num_derivs = 1;
1322       break;
1323    case TGSI_TEXTURE_SHADOW2D:
1324    case TGSI_TEXTURE_SHADOWRECT:
1325    case TGSI_TEXTURE_2D_ARRAY:
1326       num_coords = 3;
1327       num_offsets = 2;
1328       num_derivs = 2;
1329       break;
1330    case TGSI_TEXTURE_CUBE:
1331       num_coords = 3;
1332       num_offsets = 2;
1333       num_derivs = 3;
1334       break;
1335    case TGSI_TEXTURE_3D:
1336       num_coords = 3;
1337       num_offsets = 3;
1338       num_derivs = 3;
1339       break;
1340    case TGSI_TEXTURE_SHADOW2D_ARRAY:
1341       num_coords = 4;
1342       num_offsets = 2;
1343       num_derivs = 2;
1344       break;
1345    case TGSI_TEXTURE_SHADOWCUBE:
1346       num_coords = 4;
1347       num_offsets = 2;
1348       num_derivs = 3;
1349       break;
1350    default:
1351       assert(0);
1352       return;
1353    }
1354
1355    /* Note lod and especially projected are illegal in a LOT of cases */
1356    if (modifier == LP_BLD_TEX_MODIFIER_LOD_BIAS) {
1357       assert(num_coords < 4);
1358       lod_bias = lp_build_emit_fetch( &bld->bld_base, inst, 0, 3 );
1359       explicit_lod = NULL;
1360    }
1361    else if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_LOD) {
1362       assert(num_coords < 4);
1363       lod_bias = NULL;
1364       explicit_lod = lp_build_emit_fetch( &bld->bld_base, inst, 0, 3 );
1365    }
1366    else {
1367       lod_bias = NULL;
1368       explicit_lod = NULL;
1369    }
1370
1371    if (modifier == LP_BLD_TEX_MODIFIER_PROJECTED) {
1372       assert(num_coords < 4);
1373       oow = lp_build_emit_fetch( &bld->bld_base, inst, 0, 3 );
1374       oow = lp_build_rcp(&bld->bld_base.base, oow);
1375    }
1376
1377    for (i = 0; i < num_coords; i++) {
1378       coords[i] = lp_build_emit_fetch( &bld->bld_base, inst, 0, i );
1379       if (modifier == LP_BLD_TEX_MODIFIER_PROJECTED)
1380          coords[i] = lp_build_mul(&bld->bld_base.base, coords[i], oow);
1381    }
1382    for (i = num_coords; i < 4; i++) {
1383       coords[i] = bld->bld_base.base.undef;
1384    }
1385
1386    if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV) {
1387       unsigned dim;
1388       for (dim = 0; dim < num_derivs; ++dim) {
1389          derivs.ddx[dim] = lp_build_emit_fetch( &bld->bld_base, inst, 1, dim );
1390          derivs.ddy[dim] = lp_build_emit_fetch( &bld->bld_base, inst, 2, dim );
1391       }
1392       deriv_ptr = &derivs;
1393       unit = inst->Src[3].Register.Index;
1394    } else {
1395       unit = inst->Src[1].Register.Index;
1396    }
1397
1398    /* some advanced gather instructions (txgo) would require 4 offsets */
1399    if (inst->Texture.NumOffsets == 1) {
1400       unsigned dim;
1401       for (dim = 0; dim < num_offsets; dim++) {
1402          offsets[dim] = lp_build_emit_fetch_texoffset(&bld->bld_base, inst, 0, dim );
1403       }
1404    }
1405
1406    bld->sampler->emit_fetch_texel(bld->sampler,
1407                                   bld->bld_base.base.gallivm,
1408                                   bld->bld_base.base.type,
1409                                   FALSE,
1410                                   unit, unit,
1411                                   coords,
1412                                   offsets,
1413                                   deriv_ptr,
1414                                   lod_bias, explicit_lod,
1415                                   texel);
1416 }
1417
1418 static void
1419 emit_sample(struct lp_build_tgsi_soa_context *bld,
1420             const struct tgsi_full_instruction *inst,
1421             enum lp_build_tex_modifier modifier,
1422             boolean compare,
1423             LLVMValueRef *texel)
1424 {
1425    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
1426    unsigned texture_unit, sampler_unit;
1427    LLVMValueRef lod_bias, explicit_lod;
1428    LLVMValueRef coords[4];
1429    LLVMValueRef offsets[3] = { NULL };
1430    struct lp_derivatives derivs;
1431    struct lp_derivatives *deriv_ptr = NULL;
1432    unsigned num_coords, num_offsets, num_derivs;
1433    unsigned i;
1434
1435    if (!bld->sampler) {
1436       _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
1437       for (i = 0; i < 4; i++) {
1438          texel[i] = bld->bld_base.base.undef;
1439       }
1440       return;
1441    }
1442
1443    /*
1444     * unlike old-style tex opcodes the texture/sampler indices
1445     * always come from src1 and src2 respectively.
1446     */
1447    texture_unit = inst->Src[1].Register.Index;
1448    sampler_unit = inst->Src[2].Register.Index;
1449
1450    /*
1451     * Note inst->Texture.Texture will contain the number of offsets,
1452     * however the target information is NOT there and comes from the
1453     * declared sampler views instead.
1454     */
1455    switch (bld->sv[texture_unit].Resource) {
1456    case TGSI_TEXTURE_1D:
1457       num_coords = 1;
1458       num_offsets = 1;
1459       num_derivs = 1;
1460       break;
1461    case TGSI_TEXTURE_1D_ARRAY:
1462       num_coords = 2;
1463       num_offsets = 1;
1464       num_derivs = 1;
1465       break;
1466    case TGSI_TEXTURE_2D:
1467    case TGSI_TEXTURE_RECT:
1468       num_coords = 2;
1469       num_offsets = 2;
1470       num_derivs = 2;
1471       break;
1472    case TGSI_TEXTURE_2D_ARRAY:
1473       num_coords = 3;
1474       num_offsets = 2;
1475       num_derivs = 2;
1476       break;
1477    case TGSI_TEXTURE_CUBE:
1478       num_coords = 3;
1479       num_offsets = 2;
1480       num_derivs = 3;
1481       break;
1482    case TGSI_TEXTURE_3D:
1483       num_coords = 3;
1484       num_offsets = 3;
1485       num_derivs = 3;
1486       break;
1487    case TGSI_TEXTURE_CUBE_ARRAY:
1488       num_coords = 4;
1489       num_offsets = 2;
1490       num_derivs = 3;
1491       break;
1492    default:
1493       assert(0);
1494       return;
1495    }
1496
1497    /*
1498     * unlike old-style tex opcodes the texture/sampler indices
1499     * always come from src1 and src2 respectively.
1500     */
1501    texture_unit = inst->Src[1].Register.Index;
1502    sampler_unit = inst->Src[2].Register.Index;
1503
1504    if (modifier == LP_BLD_TEX_MODIFIER_LOD_BIAS) {
1505       lod_bias = lp_build_emit_fetch( &bld->bld_base, inst, 3, 0 );
1506       explicit_lod = NULL;
1507    }
1508    else if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_LOD) {
1509       lod_bias = NULL;
1510       explicit_lod = lp_build_emit_fetch( &bld->bld_base, inst, 3, 0 );
1511    }
1512    else if (modifier == LP_BLD_TEX_MODIFIER_LOD_ZERO) {
1513       lod_bias = NULL;
1514       /* XXX might be better to explicitly pass the level zero information */
1515       explicit_lod = lp_build_const_vec(gallivm, bld->bld_base.base.type, 0.0F);
1516    }
1517    else {
1518       lod_bias = NULL;
1519       explicit_lod = NULL;
1520    }
1521
1522    for (i = 0; i < num_coords; i++) {
1523       coords[i] = lp_build_emit_fetch( &bld->bld_base, inst, 0, i );
1524    }
1525    for (i = num_coords; i < 4; i++) {
1526       coords[i] = bld->bld_base.base.undef;
1527    }
1528    /*
1529     * XXX: whack shadow comparison value into place.
1530     * Should probably fix the interface for separate value
1531     * (it will not work for cube arrays if it is part of coords).
1532     */
1533    if (compare) {
1534       unsigned c_coord = num_coords > 2 ? 3 : 2;
1535       assert(num_coords < 4);
1536       coords[c_coord] = lp_build_emit_fetch( &bld->bld_base, inst, 3, 0 );
1537    }
1538
1539    if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV) {
1540       unsigned dim;
1541       for (dim = 0; dim < num_derivs; ++dim) {
1542          derivs.ddx[dim] = lp_build_emit_fetch( &bld->bld_base, inst, 3, dim );
1543          derivs.ddy[dim] = lp_build_emit_fetch( &bld->bld_base, inst, 4, dim );
1544       }
1545       deriv_ptr = &derivs;
1546    }
1547
1548    /* some advanced gather instructions (txgo) would require 4 offsets */
1549    if (inst->Texture.NumOffsets == 1) {
1550       unsigned dim;
1551       for (dim = 0; dim < num_offsets; dim++) {
1552          offsets[dim] = lp_build_emit_fetch_texoffset(&bld->bld_base, inst, 0, dim );
1553       }
1554    }
1555
1556    bld->sampler->emit_fetch_texel(bld->sampler,
1557                                   bld->bld_base.base.gallivm,
1558                                   bld->bld_base.base.type,
1559                                   FALSE,
1560                                   texture_unit, sampler_unit,
1561                                   coords,
1562                                   offsets,
1563                                   deriv_ptr,
1564                                   lod_bias, explicit_lod,
1565                                   texel);
1566 }
1567
1568 static void
1569 emit_fetch_texels( struct lp_build_tgsi_soa_context *bld,
1570                    const struct tgsi_full_instruction *inst,
1571                    LLVMValueRef *texel,
1572                    boolean is_samplei)
1573 {
1574    unsigned unit, target;
1575    LLVMValueRef coord_undef = LLVMGetUndef(bld->bld_base.base.int_vec_type);
1576    LLVMValueRef explicit_lod = NULL;
1577    LLVMValueRef coords[3];
1578    LLVMValueRef offsets[3] = { NULL };
1579    unsigned num_coords;
1580    unsigned dims;
1581    unsigned i;
1582
1583    if (!bld->sampler) {
1584       _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
1585       for (i = 0; i < 4; i++) {
1586          texel[i] = coord_undef;
1587       }
1588       return;
1589    }
1590
1591    unit = inst->Src[1].Register.Index;
1592
1593    if (is_samplei) {
1594       target = bld->sv[unit].Resource;
1595    }
1596    else {
1597       target = inst->Texture.Texture;
1598    }
1599
1600    switch (target) {
1601    case TGSI_TEXTURE_1D:
1602    case TGSI_TEXTURE_BUFFER:
1603       num_coords = 1;
1604       dims = 1;
1605       break;
1606    case TGSI_TEXTURE_1D_ARRAY:
1607       num_coords = 2;
1608       dims = 1;
1609       break;
1610    case TGSI_TEXTURE_2D:
1611    case TGSI_TEXTURE_RECT:
1612       num_coords = 2;
1613       dims = 2;
1614       break;
1615    case TGSI_TEXTURE_2D_ARRAY:
1616       num_coords = 3;
1617       dims = 2;
1618       break;
1619    case TGSI_TEXTURE_3D:
1620       num_coords = 3;
1621       dims = 3;
1622       break;
1623    default:
1624       assert(0);
1625       return;
1626    }
1627
1628    /* always have lod except for buffers ? */
1629    if (target != TGSI_TEXTURE_BUFFER) {
1630       explicit_lod = lp_build_emit_fetch( &bld->bld_base, inst, 0, 3 );
1631    }
1632
1633    for (i = 0; i < num_coords; i++) {
1634       coords[i] = lp_build_emit_fetch( &bld->bld_base, inst, 0, i );
1635    }
1636    for (i = num_coords; i < 3; i++) {
1637       coords[i] = coord_undef;
1638    }
1639
1640    if (inst->Texture.NumOffsets == 1) {
1641       unsigned dim;
1642       for (dim = 0; dim < dims; dim++) {
1643          offsets[dim] = lp_build_emit_fetch_texoffset(&bld->bld_base, inst, 0, dim );
1644       }
1645    }
1646
1647    bld->sampler->emit_fetch_texel(bld->sampler,
1648                                   bld->bld_base.base.gallivm,
1649                                   bld->bld_base.base.type,
1650                                   TRUE,
1651                                   unit, unit,
1652                                   coords,
1653                                   offsets,
1654                                   NULL,
1655                                   NULL, explicit_lod,
1656                                   texel);
1657 }
1658
1659 static void
1660 emit_size_query( struct lp_build_tgsi_soa_context *bld,
1661                  const struct tgsi_full_instruction *inst,
1662                  LLVMValueRef *sizes_out,
1663                  boolean is_sviewinfo)
1664 {
1665    LLVMValueRef explicit_lod;
1666    unsigned has_lod;
1667    unsigned i;
1668    unsigned unit = inst->Src[1].Register.Index;
1669    unsigned target;
1670
1671    if (is_sviewinfo) {
1672       target = bld->sv[unit].Resource;
1673    }
1674    else {
1675       target = inst->Texture.Texture;
1676    }
1677    switch (target) {
1678    case TGSI_TEXTURE_BUFFER:
1679    case TGSI_TEXTURE_RECT:
1680    case TGSI_TEXTURE_SHADOWRECT:
1681       has_lod = 0;
1682       break;
1683    default:
1684       has_lod = 1;
1685       break;
1686    }
1687
1688    if (!bld->sampler) {
1689       _debug_printf("warning: found texture query instruction but no sampler generator supplied\n");
1690       for (i = 0; i < 4; i++)
1691          sizes_out[i] = bld->bld_base.int_bld.undef;
1692       return;
1693    }
1694
1695    if (has_lod)
1696       explicit_lod = lp_build_emit_fetch( &bld->bld_base, inst, 0, 0 );
1697    else
1698       explicit_lod = NULL;
1699
1700    bld->sampler->emit_size_query(bld->sampler,
1701                                  bld->bld_base.base.gallivm,
1702                                  bld->bld_base.int_bld.type,
1703                                  unit,
1704                                  is_sviewinfo,
1705                                  explicit_lod,
1706                                  sizes_out);
1707 }
1708
1709 static boolean
1710 near_end_of_shader(struct lp_build_tgsi_soa_context *bld,
1711                    int pc)
1712 {
1713    int i;
1714
1715    for (i = 0; i < 5; i++) {
1716       unsigned opcode;
1717
1718       if (pc + i >= bld->bld_base.info->num_instructions)
1719          return TRUE;
1720
1721       opcode = bld->bld_base.instructions[pc + i].Instruction.Opcode;
1722
1723       if (opcode == TGSI_OPCODE_END)
1724          return TRUE;
1725
1726       if (opcode == TGSI_OPCODE_TEX ||
1727           opcode == TGSI_OPCODE_TXP ||
1728           opcode == TGSI_OPCODE_TXD ||
1729           opcode == TGSI_OPCODE_TXB ||
1730           opcode == TGSI_OPCODE_TXL ||
1731           opcode == TGSI_OPCODE_TXF ||
1732           opcode == TGSI_OPCODE_TXQ ||
1733           opcode == TGSI_OPCODE_CAL ||
1734           opcode == TGSI_OPCODE_CALLNZ ||
1735           opcode == TGSI_OPCODE_IF ||
1736           opcode == TGSI_OPCODE_UIF ||
1737           opcode == TGSI_OPCODE_BGNLOOP ||
1738           opcode == TGSI_OPCODE_SWITCH)
1739          return FALSE;
1740    }
1741
1742    return TRUE;
1743 }
1744
1745
1746
1747 /**
1748  * Kill fragment if any of the src register values are negative.
1749  */
1750 static void
1751 emit_kil(
1752    struct lp_build_tgsi_soa_context *bld,
1753    const struct tgsi_full_instruction *inst,
1754    int pc)
1755 {
1756    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
1757    const struct tgsi_full_src_register *reg = &inst->Src[0];
1758    LLVMValueRef terms[TGSI_NUM_CHANNELS];
1759    LLVMValueRef mask;
1760    unsigned chan_index;
1761
1762    memset(&terms, 0, sizeof terms);
1763
1764    TGSI_FOR_EACH_CHANNEL( chan_index ) {
1765       unsigned swizzle;
1766
1767       /* Unswizzle channel */
1768       swizzle = tgsi_util_get_full_src_register_swizzle( reg, chan_index );
1769
1770       /* Check if the component has not been already tested. */
1771       assert(swizzle < TGSI_NUM_CHANNELS);
1772       if( !terms[swizzle] )
1773          /* TODO: change the comparison operator instead of setting the sign */
1774          terms[swizzle] =  lp_build_emit_fetch(&bld->bld_base, inst, 0, chan_index );
1775    }
1776
1777    mask = NULL;
1778    TGSI_FOR_EACH_CHANNEL( chan_index ) {
1779       if(terms[chan_index]) {
1780          LLVMValueRef chan_mask;
1781
1782          /*
1783           * If term < 0 then mask = 0 else mask = ~0.
1784           */
1785          chan_mask = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_GEQUAL, terms[chan_index], bld->bld_base.base.zero);
1786
1787          if(mask)
1788             mask = LLVMBuildAnd(builder, mask, chan_mask, "");
1789          else
1790             mask = chan_mask;
1791       }
1792    }
1793
1794    if(mask) {
1795       lp_build_mask_update(bld->mask, mask);
1796
1797       if (!near_end_of_shader(bld, pc))
1798          lp_build_mask_check(bld->mask);
1799    }
1800 }
1801
1802
1803 /**
1804  * Predicated fragment kill.
1805  * XXX Actually, we do an unconditional kill (as in tgsi_exec.c).
1806  * The only predication is the execution mask which will apply if
1807  * we're inside a loop or conditional.
1808  */
1809 static void
1810 emit_kilp(struct lp_build_tgsi_soa_context *bld,
1811           int pc)
1812 {
1813    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
1814    LLVMValueRef mask;
1815
1816    /* For those channels which are "alive", disable fragment shader
1817     * execution.
1818     */
1819    if (bld->exec_mask.has_mask) {
1820       mask = LLVMBuildNot(builder, bld->exec_mask.exec_mask, "kilp");
1821    }
1822    else {
1823       LLVMValueRef zero = LLVMConstNull(bld->bld_base.base.int_vec_type);
1824       mask = zero;
1825    }
1826
1827    lp_build_mask_update(bld->mask, mask);
1828
1829    if (!near_end_of_shader(bld, pc))
1830       lp_build_mask_check(bld->mask);
1831 }
1832
1833
1834 /**
1835  * Emit code which will dump the value of all the temporary registers
1836  * to stdout.
1837  */
1838 static void
1839 emit_dump_temps(struct lp_build_tgsi_soa_context *bld)
1840 {
1841    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
1842    LLVMBuilderRef builder = gallivm->builder;
1843    LLVMValueRef temp_ptr;
1844    LLVMValueRef i0 = lp_build_const_int32(gallivm, 0);
1845    LLVMValueRef i1 = lp_build_const_int32(gallivm, 1);
1846    LLVMValueRef i2 = lp_build_const_int32(gallivm, 2);
1847    LLVMValueRef i3 = lp_build_const_int32(gallivm, 3);
1848    int index;
1849    int n = bld->bld_base.info->file_max[TGSI_FILE_TEMPORARY];
1850
1851    for (index = 0; index < n; index++) {
1852       LLVMValueRef idx = lp_build_const_int32(gallivm, index);
1853       LLVMValueRef v[4][4], res;
1854       int chan;
1855
1856       lp_build_printf(gallivm, "TEMP[%d]:\n", idx);
1857
1858       for (chan = 0; chan < 4; chan++) {
1859          temp_ptr = lp_get_temp_ptr_soa(bld, index, chan);
1860          res = LLVMBuildLoad(builder, temp_ptr, "");
1861          v[chan][0] = LLVMBuildExtractElement(builder, res, i0, "");
1862          v[chan][1] = LLVMBuildExtractElement(builder, res, i1, "");
1863          v[chan][2] = LLVMBuildExtractElement(builder, res, i2, "");
1864          v[chan][3] = LLVMBuildExtractElement(builder, res, i3, "");
1865       }
1866
1867       lp_build_printf(gallivm, "  X: %f %f %f %f\n",
1868                       v[0][0], v[0][1], v[0][2], v[0][3]);
1869       lp_build_printf(gallivm, "  Y: %f %f %f %f\n",
1870                       v[1][0], v[1][1], v[1][2], v[1][3]);
1871       lp_build_printf(gallivm, "  Z: %f %f %f %f\n",
1872                       v[2][0], v[2][1], v[2][2], v[2][3]);
1873       lp_build_printf(gallivm, "  W: %f %f %f %f\n",
1874                       v[3][0], v[3][1], v[3][2], v[3][3]);
1875    }
1876 }
1877
1878
1879
1880 void
1881 lp_emit_declaration_soa(
1882    struct lp_build_tgsi_context *bld_base,
1883    const struct tgsi_full_declaration *decl)
1884 {
1885    struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
1886    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
1887    LLVMTypeRef vec_type = bld->bld_base.base.vec_type;
1888    const unsigned first = decl->Range.First;
1889    const unsigned last = decl->Range.Last;
1890    unsigned idx, i;
1891
1892    for (idx = first; idx <= last; ++idx) {
1893       assert(last <= bld->bld_base.info->file_max[decl->Declaration.File]);
1894       switch (decl->Declaration.File) {
1895       case TGSI_FILE_TEMPORARY:
1896          assert(idx < LP_MAX_TGSI_TEMPS);
1897          if (!(bld->indirect_files & (1 << TGSI_FILE_TEMPORARY))) {
1898             for (i = 0; i < TGSI_NUM_CHANNELS; i++)
1899                bld->temps[idx][i] = lp_build_alloca(gallivm, vec_type, "temp");
1900          }
1901          break;
1902
1903       case TGSI_FILE_OUTPUT:
1904          if (!(bld->indirect_files & (1 << TGSI_FILE_OUTPUT))) {
1905             for (i = 0; i < TGSI_NUM_CHANNELS; i++)
1906                bld->outputs[idx][i] = lp_build_alloca(gallivm,
1907                                                       vec_type, "output");
1908          }
1909          break;
1910
1911       case TGSI_FILE_ADDRESS:
1912          /* ADDR registers are only allocated with an integer LLVM IR type,
1913           * as they are guaranteed to always have integers.
1914           * XXX: Not sure if this exception is worthwhile (or the whole idea of
1915           * an ADDR register for that matter).
1916           */
1917          assert(idx < LP_MAX_TGSI_ADDRS);
1918          for (i = 0; i < TGSI_NUM_CHANNELS; i++)
1919             bld->addr[idx][i] = lp_build_alloca(gallivm, bld_base->base.int_vec_type, "addr");
1920          break;
1921
1922       case TGSI_FILE_PREDICATE:
1923          assert(idx < LP_MAX_TGSI_PREDS);
1924          for (i = 0; i < TGSI_NUM_CHANNELS; i++)
1925             bld->preds[idx][i] = lp_build_alloca(gallivm, vec_type,
1926                                                  "predicate");
1927          break;
1928
1929       case TGSI_FILE_SAMPLER_VIEW:
1930          /*
1931           * The target stored here MUST match whatever there actually
1932           * is in the set sampler views (what about return type?).
1933           */
1934          assert(idx < PIPE_MAX_SHADER_SAMPLER_VIEWS);
1935          bld->sv[idx] = decl->SamplerView;
1936          break;
1937
1938       default:
1939          /* don't need to declare other vars */
1940          break;
1941       }
1942    }
1943 }
1944
1945
1946 void lp_emit_immediate_soa(
1947    struct lp_build_tgsi_context *bld_base,
1948    const struct tgsi_full_immediate *imm)
1949 {
1950    struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
1951    struct gallivm_state * gallivm = bld_base->base.gallivm;
1952
1953    /* simply copy the immediate values into the next immediates[] slot */
1954    unsigned i;
1955    const uint size = imm->Immediate.NrTokens - 1;
1956    assert(size <= 4);
1957    assert(bld->num_immediates < LP_MAX_TGSI_IMMEDIATES);
1958    switch (imm->Immediate.DataType) {
1959    case TGSI_IMM_FLOAT32:
1960       for( i = 0; i < size; ++i )
1961          bld->immediates[bld->num_immediates][i] =
1962             lp_build_const_vec(gallivm, bld_base->base.type, imm->u[i].Float);
1963
1964       break;
1965    case TGSI_IMM_UINT32:
1966       for( i = 0; i < size; ++i ) {
1967          LLVMValueRef tmp = lp_build_const_vec(gallivm, bld_base->uint_bld.type, imm->u[i].Uint);
1968          bld->immediates[bld->num_immediates][i] =
1969             LLVMConstBitCast(tmp, bld_base->base.vec_type);
1970       }
1971
1972       break;
1973    case TGSI_IMM_INT32:
1974       for( i = 0; i < size; ++i ) {
1975          LLVMValueRef tmp = lp_build_const_vec(gallivm, bld_base->int_bld.type, imm->u[i].Int);
1976          bld->immediates[bld->num_immediates][i] =
1977             LLVMConstBitCast(tmp, bld_base->base.vec_type);
1978       }
1979
1980       break;
1981    }
1982    for( i = size; i < 4; ++i )
1983       bld->immediates[bld->num_immediates][i] = bld_base->base.undef;
1984
1985    bld->num_immediates++;
1986 }
1987
1988 static void
1989 ddx_emit(
1990    const struct lp_build_tgsi_action * action,
1991    struct lp_build_tgsi_context * bld_base,
1992    struct lp_build_emit_data * emit_data)
1993 {
1994    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1995
1996    emit_fetch_deriv(bld, emit_data->args[0], NULL,
1997                     &emit_data->output[emit_data->chan], NULL);
1998 }
1999
2000 static void
2001 ddy_emit(
2002    const struct lp_build_tgsi_action * action,
2003    struct lp_build_tgsi_context * bld_base,
2004    struct lp_build_emit_data * emit_data)
2005 {
2006    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2007
2008    emit_fetch_deriv(bld, emit_data->args[0], NULL, NULL,
2009                     &emit_data->output[emit_data->chan]);
2010 }
2011
2012 static void
2013 kilp_emit(
2014    const struct lp_build_tgsi_action * action,
2015    struct lp_build_tgsi_context * bld_base,
2016    struct lp_build_emit_data * emit_data)
2017 {
2018    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2019
2020    emit_kilp(bld, bld_base->pc - 1);
2021 }
2022
2023 static void
2024 kil_emit(
2025    const struct lp_build_tgsi_action * action,
2026    struct lp_build_tgsi_context * bld_base,
2027    struct lp_build_emit_data * emit_data)
2028 {
2029    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2030
2031    emit_kil(bld, emit_data->inst, bld_base->pc - 1);
2032 }
2033
2034 static void
2035 tex_emit(
2036    const struct lp_build_tgsi_action * action,
2037    struct lp_build_tgsi_context * bld_base,
2038    struct lp_build_emit_data * emit_data)
2039 {
2040    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2041
2042    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_NONE, emit_data->output);
2043 }
2044
2045 static void
2046 txb_emit(
2047    const struct lp_build_tgsi_action * action,
2048    struct lp_build_tgsi_context * bld_base,
2049    struct lp_build_emit_data * emit_data)
2050 {
2051    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2052
2053    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_LOD_BIAS,
2054             emit_data->output);
2055 }
2056
2057 static void
2058 txd_emit(
2059    const struct lp_build_tgsi_action * action,
2060    struct lp_build_tgsi_context * bld_base,
2061    struct lp_build_emit_data * emit_data)
2062 {
2063    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2064
2065    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV,
2066             emit_data->output);
2067 }
2068
2069 static void
2070 txl_emit(
2071    const struct lp_build_tgsi_action * action,
2072    struct lp_build_tgsi_context * bld_base,
2073    struct lp_build_emit_data * emit_data)
2074 {
2075    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2076
2077    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD,
2078             emit_data->output);
2079 }
2080
2081 static void
2082 txp_emit(
2083    const struct lp_build_tgsi_action * action,
2084    struct lp_build_tgsi_context * bld_base,
2085    struct lp_build_emit_data * emit_data)
2086 {
2087    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2088
2089    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_PROJECTED,
2090             emit_data->output);
2091 }
2092
2093 static void
2094 txq_emit(
2095    const struct lp_build_tgsi_action * action,
2096    struct lp_build_tgsi_context * bld_base,
2097    struct lp_build_emit_data * emit_data)
2098 {
2099    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2100
2101    emit_size_query(bld, emit_data->inst, emit_data->output, FALSE);
2102 }
2103
2104 static void
2105 txf_emit(
2106    const struct lp_build_tgsi_action * action,
2107    struct lp_build_tgsi_context * bld_base,
2108    struct lp_build_emit_data * emit_data)
2109 {
2110    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2111
2112    emit_fetch_texels(bld, emit_data->inst, emit_data->output, FALSE);
2113 }
2114
2115 static void
2116 sample_i_emit(
2117    const struct lp_build_tgsi_action * action,
2118    struct lp_build_tgsi_context * bld_base,
2119    struct lp_build_emit_data * emit_data)
2120 {
2121    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2122
2123    emit_fetch_texels(bld, emit_data->inst, emit_data->output, TRUE);
2124 }
2125
2126 static void
2127 sample_emit(
2128    const struct lp_build_tgsi_action * action,
2129    struct lp_build_tgsi_context * bld_base,
2130    struct lp_build_emit_data * emit_data)
2131 {
2132    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2133
2134    emit_sample(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_NONE,
2135                FALSE, emit_data->output);
2136 }
2137
2138 static void
2139 sample_b_emit(
2140    const struct lp_build_tgsi_action * action,
2141    struct lp_build_tgsi_context * bld_base,
2142    struct lp_build_emit_data * emit_data)
2143 {
2144    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2145
2146    emit_sample(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_LOD_BIAS,
2147                FALSE, emit_data->output);
2148 }
2149
2150 static void
2151 sample_c_emit(
2152    const struct lp_build_tgsi_action * action,
2153    struct lp_build_tgsi_context * bld_base,
2154    struct lp_build_emit_data * emit_data)
2155 {
2156    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2157
2158    emit_sample(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_NONE,
2159                TRUE, emit_data->output);
2160 }
2161
2162 static void
2163 sample_c_lz_emit(
2164    const struct lp_build_tgsi_action * action,
2165    struct lp_build_tgsi_context * bld_base,
2166    struct lp_build_emit_data * emit_data)
2167 {
2168    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2169
2170    emit_sample(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_LOD_ZERO,
2171                TRUE, emit_data->output);
2172 }
2173
2174 static void
2175 sample_d_emit(
2176    const struct lp_build_tgsi_action * action,
2177    struct lp_build_tgsi_context * bld_base,
2178    struct lp_build_emit_data * emit_data)
2179 {
2180    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2181
2182    emit_sample(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV,
2183                FALSE, emit_data->output);
2184 }
2185
2186 static void
2187 sample_l_emit(
2188    const struct lp_build_tgsi_action * action,
2189    struct lp_build_tgsi_context * bld_base,
2190    struct lp_build_emit_data * emit_data)
2191 {
2192    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2193
2194    emit_sample(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD,
2195                FALSE, emit_data->output);
2196 }
2197
2198 static void
2199 sviewinfo_emit(
2200    const struct lp_build_tgsi_action * action,
2201    struct lp_build_tgsi_context * bld_base,
2202    struct lp_build_emit_data * emit_data)
2203 {
2204    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2205
2206    emit_size_query(bld, emit_data->inst, emit_data->output, TRUE);
2207 }
2208
2209 static LLVMValueRef
2210 mask_to_one_vec(struct lp_build_tgsi_context *bld_base)
2211 {
2212    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2213    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
2214    LLVMValueRef one_vec = bld_base->int_bld.one;
2215    struct lp_exec_mask *exec_mask = &bld->exec_mask;
2216
2217    if (exec_mask->has_mask) {
2218       one_vec = LLVMBuildAnd(builder, one_vec, exec_mask->exec_mask, "");
2219    }
2220    one_vec = LLVMBuildAnd(builder, one_vec,
2221                           lp_build_mask_value(bld->mask), "");
2222    return one_vec;
2223 }
2224
2225 static void
2226 increment_vec_ptr_by_mask(struct lp_build_tgsi_context * bld_base,
2227                           LLVMValueRef ptr,
2228                           LLVMValueRef mask)
2229 {
2230    LLVMBuilderRef builder = bld_base->base.gallivm->builder;
2231
2232    LLVMValueRef current_vec = LLVMBuildLoad(builder, ptr, "");
2233
2234    current_vec = LLVMBuildAdd(builder, current_vec, mask, "");
2235
2236    LLVMBuildStore(builder, current_vec, ptr);
2237 }
2238
2239 static void
2240 clear_uint_vec_ptr_from_mask(struct lp_build_tgsi_context * bld_base,
2241                              LLVMValueRef ptr,
2242                              LLVMValueRef mask)
2243 {
2244    LLVMBuilderRef builder = bld_base->base.gallivm->builder;
2245
2246    LLVMValueRef current_vec = LLVMBuildLoad(builder, ptr, "");
2247    LLVMValueRef full_mask = lp_build_cmp(&bld_base->uint_bld,
2248                                          PIPE_FUNC_NOTEQUAL,
2249                                          mask,
2250                                          bld_base->uint_bld.zero);
2251
2252    current_vec = lp_build_select(&bld_base->uint_bld,
2253                                  full_mask,
2254                                  bld_base->uint_bld.zero,
2255                                  current_vec);
2256
2257    LLVMBuildStore(builder, current_vec, ptr);
2258 }
2259
2260 static LLVMValueRef
2261 clamp_mask_to_max_output_vertices(struct lp_build_tgsi_soa_context * bld,
2262                                   LLVMValueRef current_mask_vec,
2263                                   LLVMValueRef total_emitted_vertices_vec)
2264 {
2265    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
2266    struct lp_build_context *uint_bld = &bld->bld_base.uint_bld;
2267    LLVMValueRef max_mask = lp_build_cmp(uint_bld, PIPE_FUNC_LESS,
2268                                         total_emitted_vertices_vec,
2269                                         bld->max_output_vertices_vec);
2270
2271    return LLVMBuildAnd(builder, current_mask_vec, max_mask, "");
2272 }
2273
2274 static void
2275 emit_vertex(
2276    const struct lp_build_tgsi_action * action,
2277    struct lp_build_tgsi_context * bld_base,
2278    struct lp_build_emit_data * emit_data)
2279 {
2280    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2281    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
2282
2283    if (bld->gs_iface->emit_vertex) {
2284       LLVMValueRef masked_ones = mask_to_one_vec(bld_base);
2285       LLVMValueRef total_emitted_vertices_vec =
2286          LLVMBuildLoad(builder, bld->total_emitted_vertices_vec_ptr, "");
2287       masked_ones = clamp_mask_to_max_output_vertices(bld, masked_ones,
2288                                                       total_emitted_vertices_vec);
2289       gather_outputs(bld);
2290       bld->gs_iface->emit_vertex(bld->gs_iface, &bld->bld_base,
2291                                  bld->outputs,
2292                                  total_emitted_vertices_vec);
2293       increment_vec_ptr_by_mask(bld_base, bld->emitted_vertices_vec_ptr,
2294                                 masked_ones);
2295       increment_vec_ptr_by_mask(bld_base, bld->total_emitted_vertices_vec_ptr,
2296                                 masked_ones);
2297 #if DUMP_GS_EMITS
2298       lp_build_print_value(bld->bld_base.base.gallivm,
2299                            " +++ emit vertex masked ones = ",
2300                            masked_ones);
2301       lp_build_print_value(bld->bld_base.base.gallivm,
2302                            " +++ emit vertex emitted = ",
2303                            total_emitted_vertices_vec);
2304 #endif
2305    }
2306 }
2307
2308
2309 static void
2310 end_primitive_masked(struct lp_build_tgsi_context * bld_base,
2311                      LLVMValueRef masked_ones)
2312 {
2313    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2314    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
2315
2316    if (bld->gs_iface->end_primitive) {
2317       LLVMValueRef emitted_vertices_vec =
2318          LLVMBuildLoad(builder, bld->emitted_vertices_vec_ptr, "");
2319       LLVMValueRef emitted_prims_vec =
2320          LLVMBuildLoad(builder, bld->emitted_prims_vec_ptr, "");
2321
2322       bld->gs_iface->end_primitive(bld->gs_iface, &bld->bld_base,
2323                                    emitted_vertices_vec,
2324                                    emitted_prims_vec);
2325
2326 #if DUMP_GS_EMITS
2327       lp_build_print_value(bld->bld_base.base.gallivm,
2328                            " +++ end prim masked ones = ",
2329                            masked_ones);
2330       lp_build_print_value(bld->bld_base.base.gallivm,
2331                            " +++ end prim emitted verts1 = ",
2332                            emitted_vertices_vec);
2333       lp_build_print_value(bld->bld_base.base.gallivm,
2334                            " +++ end prim emitted prims1 = ",
2335                            LLVMBuildLoad(builder,
2336                                          bld->emitted_prims_vec_ptr, ""));
2337 #endif
2338       increment_vec_ptr_by_mask(bld_base, bld->emitted_prims_vec_ptr,
2339                                 masked_ones);
2340       clear_uint_vec_ptr_from_mask(bld_base, bld->emitted_vertices_vec_ptr,
2341                                    masked_ones);
2342 #if DUMP_GS_EMITS
2343       lp_build_print_value(bld->bld_base.base.gallivm,
2344                            " +++ end prim emitted verts2 = ",
2345                            LLVMBuildLoad(builder,
2346                                          bld->emitted_vertices_vec_ptr, ""));
2347 #endif
2348    }
2349
2350 }
2351
2352 static void
2353 end_primitive(
2354    const struct lp_build_tgsi_action * action,
2355    struct lp_build_tgsi_context * bld_base,
2356    struct lp_build_emit_data * emit_data)
2357 {
2358    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2359
2360    if (bld->gs_iface->end_primitive) {
2361       LLVMBuilderRef builder = bld_base->base.gallivm->builder;
2362       LLVMValueRef masked_ones = mask_to_one_vec(bld_base);
2363       struct lp_build_context *uint_bld = &bld_base->uint_bld;
2364       LLVMValueRef emitted_verts = LLVMBuildLoad(
2365          builder, bld->emitted_vertices_vec_ptr, "");
2366       LLVMValueRef emitted_mask = lp_build_cmp(uint_bld, PIPE_FUNC_NOTEQUAL,
2367                                                emitted_verts,
2368                                                uint_bld->zero);
2369       /* We need to combine the current execution mask with the mask
2370          telling us which, if any, execution slots actually have
2371          unemitted primitives, this way we make sure that end_primitives
2372          executes only on the paths that have unflushed vertices */
2373       masked_ones = LLVMBuildAnd(builder, masked_ones, emitted_mask, "");
2374
2375       end_primitive_masked(bld_base, masked_ones);
2376    }
2377 }
2378
2379 static void
2380 cal_emit(
2381    const struct lp_build_tgsi_action * action,
2382    struct lp_build_tgsi_context * bld_base,
2383    struct lp_build_emit_data * emit_data)
2384 {
2385    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2386
2387    lp_exec_mask_call(&bld->exec_mask, emit_data->inst->Label.Label,
2388                      &bld_base->pc);
2389 }
2390
2391 static void
2392 ret_emit(
2393    const struct lp_build_tgsi_action * action,
2394    struct lp_build_tgsi_context * bld_base,
2395    struct lp_build_emit_data * emit_data)
2396 {
2397    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2398
2399    lp_exec_mask_ret(&bld->exec_mask, &bld_base->pc);
2400 }
2401
2402 static void
2403 brk_emit(
2404    const struct lp_build_tgsi_action * action,
2405    struct lp_build_tgsi_context * bld_base,
2406    struct lp_build_emit_data * emit_data)
2407 {
2408    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2409
2410    lp_exec_break(&bld->exec_mask);
2411 }
2412
2413 static void
2414 breakc_emit(
2415    const struct lp_build_tgsi_action * action,
2416    struct lp_build_tgsi_context * bld_base,
2417    struct lp_build_emit_data * emit_data)
2418 {
2419    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2420    LLVMBuilderRef builder = bld_base->base.gallivm->builder;
2421    struct lp_build_context *uint_bld = &bld_base->uint_bld;
2422    LLVMValueRef unsigned_cond =
2423       LLVMBuildBitCast(builder, emit_data->args[0], uint_bld->vec_type, "");
2424    LLVMValueRef cond = lp_build_cmp(uint_bld, PIPE_FUNC_NOTEQUAL,
2425                                     unsigned_cond,
2426                                     uint_bld->zero);
2427
2428    lp_exec_break_condition(&bld->exec_mask, cond);
2429 }
2430
2431 static void
2432 if_emit(
2433    const struct lp_build_tgsi_action * action,
2434    struct lp_build_tgsi_context * bld_base,
2435    struct lp_build_emit_data * emit_data)
2436 {
2437    LLVMValueRef tmp;
2438    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2439
2440    tmp = lp_build_cmp(&bld_base->base, PIPE_FUNC_NOTEQUAL,
2441                       emit_data->args[0], bld->bld_base.base.zero);
2442    lp_exec_mask_cond_push(&bld->exec_mask, tmp);
2443 }
2444
2445 static void
2446 uif_emit(
2447    const struct lp_build_tgsi_action * action,
2448    struct lp_build_tgsi_context * bld_base,
2449    struct lp_build_emit_data * emit_data)
2450 {
2451    LLVMValueRef tmp;
2452    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2453    struct lp_build_context *uint_bld = &bld_base->uint_bld;
2454
2455    tmp = lp_build_cmp(uint_bld, PIPE_FUNC_NOTEQUAL,
2456                       emit_data->args[0], uint_bld->zero);
2457    lp_exec_mask_cond_push(&bld->exec_mask, tmp);
2458 }
2459
2460 static void
2461 bgnloop_emit(
2462    const struct lp_build_tgsi_action * action,
2463    struct lp_build_tgsi_context * bld_base,
2464    struct lp_build_emit_data * emit_data)
2465 {
2466    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2467
2468    lp_exec_bgnloop(&bld->exec_mask);
2469 }
2470
2471 static void
2472 bgnsub_emit(
2473    const struct lp_build_tgsi_action * action,
2474    struct lp_build_tgsi_context * bld_base,
2475    struct lp_build_emit_data * emit_data)
2476 {
2477    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2478
2479    lp_exec_mask_bgnsub(&bld->exec_mask);
2480 }
2481
2482 static void
2483 else_emit(
2484    const struct lp_build_tgsi_action * action,
2485    struct lp_build_tgsi_context * bld_base,
2486    struct lp_build_emit_data * emit_data)
2487 {
2488    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2489
2490    lp_exec_mask_cond_invert(&bld->exec_mask);
2491 }
2492
2493 static void
2494 endif_emit(
2495    const struct lp_build_tgsi_action * action,
2496    struct lp_build_tgsi_context * bld_base,
2497    struct lp_build_emit_data * emit_data)
2498 {
2499    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2500
2501    lp_exec_mask_cond_pop(&bld->exec_mask);
2502 }
2503
2504 static void
2505 endloop_emit(
2506    const struct lp_build_tgsi_action * action,
2507    struct lp_build_tgsi_context * bld_base,
2508    struct lp_build_emit_data * emit_data)
2509 {
2510    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2511
2512    lp_exec_endloop(bld_base->base.gallivm, &bld->exec_mask);
2513 }
2514
2515 static void
2516 endsub_emit(
2517    const struct lp_build_tgsi_action * action,
2518    struct lp_build_tgsi_context * bld_base,
2519    struct lp_build_emit_data * emit_data)
2520 {
2521    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2522
2523    lp_exec_mask_endsub(&bld->exec_mask, &bld_base->pc);
2524 }
2525
2526 static void
2527 cont_emit(
2528    const struct lp_build_tgsi_action * action,
2529    struct lp_build_tgsi_context * bld_base,
2530    struct lp_build_emit_data * emit_data)
2531 {
2532    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2533
2534    lp_exec_continue(&bld->exec_mask);
2535 }
2536
2537 /* XXX: Refactor and move it to lp_bld_tgsi_action.c
2538  *
2539  * XXX: What do the comments about xmm registers mean?  Maybe they are left over
2540  * from old code, but there is no garauntee that LLVM will use those registers
2541  * for this code.
2542  *
2543  * XXX: There should be no calls to lp_build_emit_fetch in this function.  This
2544  * should be handled by the emit_data->fetch_args function. */
2545 static void
2546 nrm_emit(
2547    const struct lp_build_tgsi_action * action,
2548    struct lp_build_tgsi_context * bld_base,
2549    struct lp_build_emit_data * emit_data)
2550 {
2551    LLVMValueRef tmp0, tmp1;
2552    LLVMValueRef tmp4 = NULL;
2553    LLVMValueRef tmp5 = NULL;
2554    LLVMValueRef tmp6 = NULL;
2555    LLVMValueRef tmp7 = NULL;
2556    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2557
2558    uint dims = (emit_data->inst->Instruction.Opcode == TGSI_OPCODE_NRM) ? 3 : 4;
2559
2560   if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_X) ||
2561       TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Y) ||
2562       TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Z) ||
2563       (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_W) && dims == 4)) {
2564
2565       /* NOTE: Cannot use xmm regs 2/3 here (see emit_rsqrt() above). */
2566
2567       /* xmm4 = src.x */
2568       /* xmm0 = src.x * src.x */
2569       tmp0 = lp_build_emit_fetch(&bld->bld_base, emit_data->inst, 0, TGSI_CHAN_X);
2570       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_X)) {
2571          tmp4 = tmp0;
2572       }
2573       tmp0 = lp_build_mul( &bld->bld_base.base, tmp0, tmp0);
2574
2575       /* xmm5 = src.y */
2576       /* xmm0 = xmm0 + src.y * src.y */
2577       tmp1 = lp_build_emit_fetch(&bld->bld_base, emit_data->inst, 0, TGSI_CHAN_Y);
2578       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Y)) {
2579          tmp5 = tmp1;
2580       }
2581       tmp1 = lp_build_mul( &bld->bld_base.base, tmp1, tmp1);
2582       tmp0 = lp_build_add( &bld->bld_base.base, tmp0, tmp1);
2583
2584       /* xmm6 = src.z */
2585       /* xmm0 = xmm0 + src.z * src.z */
2586       tmp1 = lp_build_emit_fetch(&bld->bld_base, emit_data->inst, 0, TGSI_CHAN_Z);
2587       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Z)) {
2588          tmp6 = tmp1;
2589       }
2590       tmp1 = lp_build_mul( &bld->bld_base.base, tmp1, tmp1);
2591       tmp0 = lp_build_add( &bld->bld_base.base, tmp0, tmp1);
2592
2593       if (dims == 4) {
2594          /* xmm7 = src.w */
2595          /* xmm0 = xmm0 + src.w * src.w */
2596          tmp1 = lp_build_emit_fetch(&bld->bld_base, emit_data->inst, 0, TGSI_CHAN_W);
2597          if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_W)) {
2598             tmp7 = tmp1;
2599          }
2600          tmp1 = lp_build_mul( &bld->bld_base.base, tmp1, tmp1);
2601          tmp0 = lp_build_add( &bld->bld_base.base, tmp0, tmp1);
2602       }
2603       /* xmm1 = 1 / sqrt(xmm0) */
2604       tmp1 = lp_build_rsqrt( &bld->bld_base.base, tmp0);
2605        /* dst.x = xmm1 * src.x */
2606       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_X)) {
2607          emit_data->output[TGSI_CHAN_X] = lp_build_mul( &bld->bld_base.base, tmp4, tmp1);
2608       }
2609       /* dst.y = xmm1 * src.y */
2610       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Y)) {
2611          emit_data->output[TGSI_CHAN_Y] = lp_build_mul( &bld->bld_base.base, tmp5, tmp1);
2612       }
2613
2614       /* dst.z = xmm1 * src.z */
2615       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Z)) {
2616          emit_data->output[TGSI_CHAN_Z] = lp_build_mul( &bld->bld_base.base, tmp6, tmp1);
2617       }
2618       /* dst.w = xmm1 * src.w */
2619       if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_X) && dims == 4) {
2620          emit_data->output[TGSI_CHAN_W] = lp_build_mul( &bld->bld_base.base, tmp7, tmp1);
2621       }
2622    }
2623
2624    /* dst.w = 1.0 */
2625    if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_W) && dims == 3) {
2626        emit_data->output[TGSI_CHAN_W] = bld->bld_base.base.one;
2627    }
2628 }
2629
2630 static void emit_prologue(struct lp_build_tgsi_context * bld_base)
2631 {
2632    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2633    struct gallivm_state * gallivm = bld_base->base.gallivm;
2634
2635    if (bld->indirect_files & (1 << TGSI_FILE_TEMPORARY)) {
2636       LLVMValueRef array_size =
2637          lp_build_const_int32(gallivm,
2638                          bld_base->info->file_max[TGSI_FILE_TEMPORARY] * 4 + 4);
2639       bld->temps_array = lp_build_array_alloca(gallivm,
2640                                               bld_base->base.vec_type, array_size,
2641                                               "temp_array");
2642    }
2643
2644    if (bld->indirect_files & (1 << TGSI_FILE_OUTPUT)) {
2645       LLVMValueRef array_size =
2646          lp_build_const_int32(gallivm,
2647                             bld_base->info->file_max[TGSI_FILE_OUTPUT] * 4 + 4);
2648       bld->outputs_array = lp_build_array_alloca(gallivm,
2649                                                 bld_base->base.vec_type, array_size,
2650                                                 "output_array");
2651    }
2652
2653    /* If we have indirect addressing in inputs we need to copy them into
2654     * our alloca array to be able to iterate over them */
2655    if (bld->indirect_files & (1 << TGSI_FILE_INPUT) && !bld->gs_iface) {
2656       unsigned index, chan;
2657       LLVMTypeRef vec_type = bld_base->base.vec_type;
2658       LLVMValueRef array_size = lp_build_const_int32(gallivm,
2659             bld_base->info->file_max[TGSI_FILE_INPUT]*4 + 4);
2660       bld->inputs_array = lp_build_array_alloca(gallivm,
2661                                                vec_type, array_size,
2662                                                "input_array");
2663
2664       assert(bld_base->info->num_inputs
2665                         <= bld_base->info->file_max[TGSI_FILE_INPUT] + 1);
2666
2667       for (index = 0; index < bld_base->info->num_inputs; ++index) {
2668          for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
2669             LLVMValueRef lindex =
2670                lp_build_const_int32(gallivm, index * 4 + chan);
2671             LLVMValueRef input_ptr =
2672                LLVMBuildGEP(gallivm->builder, bld->inputs_array,
2673                             &lindex, 1, "");
2674             LLVMValueRef value = bld->inputs[index][chan];
2675             if (value)
2676                LLVMBuildStore(gallivm->builder, value, input_ptr);
2677          }
2678       }
2679    }
2680
2681    if (bld->gs_iface) {
2682       struct lp_build_context *uint_bld = &bld->bld_base.uint_bld;
2683       bld->emitted_prims_vec_ptr =
2684          lp_build_alloca(gallivm,
2685                          uint_bld->vec_type,
2686                          "emitted_prims_ptr");
2687       bld->emitted_vertices_vec_ptr =
2688          lp_build_alloca(gallivm,
2689                          uint_bld->vec_type,
2690                          "emitted_vertices_ptr");
2691       bld->total_emitted_vertices_vec_ptr =
2692          lp_build_alloca(gallivm,
2693                          uint_bld->vec_type,
2694                          "total_emitted_vertices_ptr");
2695
2696       LLVMBuildStore(gallivm->builder, uint_bld->zero,
2697                      bld->emitted_prims_vec_ptr);
2698       LLVMBuildStore(gallivm->builder, uint_bld->zero,
2699                      bld->emitted_vertices_vec_ptr);
2700       LLVMBuildStore(gallivm->builder, uint_bld->zero,
2701                      bld->total_emitted_vertices_vec_ptr);
2702    }
2703 }
2704
2705 static void emit_epilogue(struct lp_build_tgsi_context * bld_base)
2706 {
2707    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2708    LLVMBuilderRef builder = bld_base->base.gallivm->builder;
2709
2710    if (0) {
2711       /* for debugging */
2712       emit_dump_temps(bld);
2713    }
2714
2715    /* If we have indirect addressing in outputs we need to copy our alloca array
2716     * to the outputs slots specified by the caller */
2717    if (bld->gs_iface) {
2718       LLVMValueRef total_emitted_vertices_vec;
2719       LLVMValueRef emitted_prims_vec;
2720       /* implicit end_primitives, needed in case there are any unflushed
2721          vertices in the cache */
2722       end_primitive(NULL, bld_base, NULL);
2723
2724       total_emitted_vertices_vec =
2725          LLVMBuildLoad(builder, bld->total_emitted_vertices_vec_ptr, "");
2726       emitted_prims_vec =
2727          LLVMBuildLoad(builder, bld->emitted_prims_vec_ptr, "");
2728
2729       bld->gs_iface->gs_epilogue(bld->gs_iface,
2730                                  &bld->bld_base,
2731                                  total_emitted_vertices_vec,
2732                                  emitted_prims_vec);
2733    } else {
2734       gather_outputs(bld);
2735    }
2736 }
2737
2738 void
2739 lp_build_tgsi_soa(struct gallivm_state *gallivm,
2740                   const struct tgsi_token *tokens,
2741                   struct lp_type type,
2742                   struct lp_build_mask_context *mask,
2743                   LLVMValueRef consts_ptr,
2744                   const struct lp_bld_tgsi_system_values *system_values,
2745                   const LLVMValueRef (*inputs)[TGSI_NUM_CHANNELS],
2746                   LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
2747                   struct lp_build_sampler_soa *sampler,
2748                   const struct tgsi_shader_info *info,
2749                   const struct lp_build_tgsi_gs_iface *gs_iface)
2750 {
2751    struct lp_build_tgsi_soa_context bld;
2752
2753    struct lp_type res_type;
2754
2755    assert(type.length <= LP_MAX_VECTOR_LENGTH);
2756    memset(&res_type, 0, sizeof res_type);
2757    res_type.width = type.width;
2758    res_type.length = type.length;
2759    res_type.sign = 1;
2760
2761    /* Setup build context */
2762    memset(&bld, 0, sizeof bld);
2763    lp_build_context_init(&bld.bld_base.base, gallivm, type);
2764    lp_build_context_init(&bld.bld_base.uint_bld, gallivm, lp_uint_type(type));
2765    lp_build_context_init(&bld.bld_base.int_bld, gallivm, lp_int_type(type));
2766    lp_build_context_init(&bld.elem_bld, gallivm, lp_elem_type(type));
2767    bld.mask = mask;
2768    bld.inputs = inputs;
2769    bld.outputs = outputs;
2770    bld.consts_ptr = consts_ptr;
2771    bld.sampler = sampler;
2772    bld.bld_base.info = info;
2773    bld.indirect_files = info->indirect_files;
2774
2775    bld.bld_base.soa = TRUE;
2776    bld.bld_base.emit_fetch_funcs[TGSI_FILE_CONSTANT] = emit_fetch_constant;
2777    bld.bld_base.emit_fetch_funcs[TGSI_FILE_IMMEDIATE] = emit_fetch_immediate;
2778    bld.bld_base.emit_fetch_funcs[TGSI_FILE_INPUT] = emit_fetch_input;
2779    bld.bld_base.emit_fetch_funcs[TGSI_FILE_TEMPORARY] = emit_fetch_temporary;
2780    bld.bld_base.emit_fetch_funcs[TGSI_FILE_SYSTEM_VALUE] = emit_fetch_system_value;
2781    bld.bld_base.emit_store = emit_store;
2782
2783    bld.bld_base.emit_declaration = lp_emit_declaration_soa;
2784    bld.bld_base.emit_immediate = lp_emit_immediate_soa;
2785
2786    bld.bld_base.emit_prologue = emit_prologue;
2787    bld.bld_base.emit_epilogue = emit_epilogue;
2788
2789    /* Set opcode actions */
2790    lp_set_default_actions_cpu(&bld.bld_base);
2791
2792    bld.bld_base.op_actions[TGSI_OPCODE_BGNLOOP].emit = bgnloop_emit;
2793    bld.bld_base.op_actions[TGSI_OPCODE_BGNSUB].emit = bgnsub_emit;
2794    bld.bld_base.op_actions[TGSI_OPCODE_BRK].emit = brk_emit;
2795    bld.bld_base.op_actions[TGSI_OPCODE_BREAKC].emit = breakc_emit;
2796    bld.bld_base.op_actions[TGSI_OPCODE_CAL].emit = cal_emit;
2797    bld.bld_base.op_actions[TGSI_OPCODE_CONT].emit = cont_emit;
2798    bld.bld_base.op_actions[TGSI_OPCODE_DDX].emit = ddx_emit;
2799    bld.bld_base.op_actions[TGSI_OPCODE_DDY].emit = ddy_emit;
2800    bld.bld_base.op_actions[TGSI_OPCODE_ELSE].emit = else_emit;
2801    bld.bld_base.op_actions[TGSI_OPCODE_ENDIF].emit = endif_emit;
2802    bld.bld_base.op_actions[TGSI_OPCODE_ENDLOOP].emit = endloop_emit;
2803    bld.bld_base.op_actions[TGSI_OPCODE_ENDSUB].emit = endsub_emit;
2804    bld.bld_base.op_actions[TGSI_OPCODE_IF].emit = if_emit;
2805    bld.bld_base.op_actions[TGSI_OPCODE_UIF].emit = uif_emit;
2806    bld.bld_base.op_actions[TGSI_OPCODE_KIL].emit = kil_emit;
2807    bld.bld_base.op_actions[TGSI_OPCODE_KILP].emit = kilp_emit;
2808    bld.bld_base.op_actions[TGSI_OPCODE_NRM].emit = nrm_emit;
2809    bld.bld_base.op_actions[TGSI_OPCODE_NRM4].emit = nrm_emit;
2810    bld.bld_base.op_actions[TGSI_OPCODE_RET].emit = ret_emit;
2811    bld.bld_base.op_actions[TGSI_OPCODE_TEX].emit = tex_emit;
2812    bld.bld_base.op_actions[TGSI_OPCODE_TXB].emit = txb_emit;
2813    bld.bld_base.op_actions[TGSI_OPCODE_TXD].emit = txd_emit;
2814    bld.bld_base.op_actions[TGSI_OPCODE_TXL].emit = txl_emit;
2815    bld.bld_base.op_actions[TGSI_OPCODE_TXP].emit = txp_emit;
2816    bld.bld_base.op_actions[TGSI_OPCODE_TXQ].emit = txq_emit;
2817    bld.bld_base.op_actions[TGSI_OPCODE_TXF].emit = txf_emit;
2818    /* DX10 sampling ops */
2819    bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE].emit = sample_emit;
2820    bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE_B].emit = sample_b_emit;
2821    bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE_C].emit = sample_c_emit;
2822    bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE_C_LZ].emit = sample_c_lz_emit;
2823    bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE_D].emit = sample_d_emit;
2824    bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE_I].emit = sample_i_emit;
2825    bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE_L].emit = sample_l_emit;
2826    bld.bld_base.op_actions[TGSI_OPCODE_SVIEWINFO].emit = sviewinfo_emit;
2827
2828    if (gs_iface) {
2829       /* There's no specific value for this because it should always
2830        * be set, but apps using ext_geometry_shader4 quite often
2831        * were forgetting so we're using MAX_VERTEX_VARYING from
2832        * that spec even though we could debug_assert if it's not
2833        * set, but that's a lot uglier. */
2834       uint max_output_vertices = 32;
2835       uint i = 0;
2836       /* inputs are always indirect with gs */
2837       bld.indirect_files |= (1 << TGSI_FILE_INPUT);
2838       bld.gs_iface = gs_iface;
2839       bld.bld_base.emit_fetch_funcs[TGSI_FILE_INPUT] = emit_fetch_gs_input;
2840       bld.bld_base.op_actions[TGSI_OPCODE_EMIT].emit = emit_vertex;
2841       bld.bld_base.op_actions[TGSI_OPCODE_ENDPRIM].emit = end_primitive;
2842
2843       for (i = 0; i < info->num_properties; ++i) {
2844          if (info->properties[i].name ==
2845              TGSI_PROPERTY_GS_MAX_OUTPUT_VERTICES) {
2846             max_output_vertices = info->properties[i].data[0];
2847          }
2848       }
2849       bld.max_output_vertices_vec =
2850          lp_build_const_int_vec(gallivm, bld.bld_base.uint_bld.type,
2851                                 max_output_vertices);
2852    }
2853
2854    lp_exec_mask_init(&bld.exec_mask, &bld.bld_base.int_bld);
2855
2856    bld.system_values = *system_values;
2857
2858    lp_build_tgsi_llvm(&bld.bld_base, tokens);
2859
2860    if (0) {
2861       LLVMBasicBlockRef block = LLVMGetInsertBlock(gallivm->builder);
2862       LLVMValueRef function = LLVMGetBasicBlockParent(block);
2863       debug_printf("11111111111111111111111111111 \n");
2864       tgsi_dump(tokens, 0);
2865       lp_debug_dump_value(function);
2866       debug_printf("2222222222222222222222222222 \n");
2867    }
2868
2869    if (0) {
2870       LLVMModuleRef module = LLVMGetGlobalParent(
2871          LLVMGetBasicBlockParent(LLVMGetInsertBlock(gallivm->builder)));
2872       LLVMDumpModule(module);
2873
2874    }
2875 }