src/gallium/auxiliary/gallivm/lp_bld_tgsi_soa.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2009 VMware, Inc.
   4  * Copyright 2007-2008 VMware, Inc.
   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 VMWARE 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 "tgsi/tgsi_strings.h"
  51 #include "lp_bld_tgsi_action.h"
  52 #include "lp_bld_type.h"
  53 #include "lp_bld_const.h"
  54 #include "lp_bld_arit.h"
  55 #include "lp_bld_bitarit.h"
  56 #include "lp_bld_gather.h"
  57 #include "lp_bld_init.h"
  58 #include "lp_bld_logic.h"
  59 #include "lp_bld_swizzle.h"
  60 #include "lp_bld_flow.h"
  61 #include "lp_bld_quad.h"
  62 #include "lp_bld_tgsi.h"
  63 #include "lp_bld_limits.h"
  64 #include "lp_bld_debug.h"
  65 #include "lp_bld_printf.h"
  66 #include "lp_bld_sample.h"
  67 #include "lp_bld_struct.h"
  68
  69 /* SM 4.0 says that subroutines can nest 32 deep and
  70  * we need one more for our main function */
  71 #define LP_MAX_NUM_FUNCS 33
  72
  73 #define DUMP_GS_EMITS 0
  74
  75 /*
  76  * If non-zero, the generated LLVM IR will print intermediate results on every TGSI
  77  * instruction.
  78  *
  79  * TODO:
  80  * - take execution masks in consideration
  81  * - debug control-flow instructions
  82  */
  83 #define DEBUG_EXECUTION 0
  84
  85
  86 /*
  87  * Emit code to print a register value.
  88  */
  89 static void
  90 emit_dump_reg(struct gallivm_state *gallivm,
  91               unsigned file,
  92               unsigned index,
  93               unsigned chan,
  94               LLVMValueRef value)
  95 {
  96    char buf[32];
  97
  98    util_snprintf(buf, sizeof buf, "    %s[%u].%c = ",
  99                  tgsi_file_name(file),
 100                  index, "xyzw"[chan]);
 101
 102    lp_build_print_value(gallivm, buf, value);
 103 }
 104
 105 /*
 106  * Return the context for the current function.
 107  * (always 'main', if shader doesn't do any function calls)
 108  */
 109 static INLINE struct function_ctx *
 110 func_ctx(struct lp_exec_mask *mask)
 111 {
 112    assert(mask->function_stack_size > 0);
 113    assert(mask->function_stack_size <= LP_MAX_NUM_FUNCS);
 114    return &mask->function_stack[mask->function_stack_size - 1];
 115 }
 116
 117 /*
 118  * Returns true if we're in a loop.
 119  * It's global, meaning that it returns true even if there's
 120  * no loop inside the current function, but we were inside
 121  * a loop inside another function, from which this one was called.
 122  */
 123 static INLINE boolean
 124 mask_has_loop(struct lp_exec_mask *mask)
 125 {
 126    int i;
 127    for (i = mask->function_stack_size - 1; i >= 0; --i) {
 128       const struct function_ctx *ctx = &mask->function_stack[i];
 129       if (ctx->loop_stack_size > 0)
 130          return TRUE;
 131    }
 132    return FALSE;
 133 }
 134
 135 /*
 136  * Returns true if we're inside a switch statement.
 137  * It's global, meaning that it returns true even if there's
 138  * no switch in the current function, but we were inside
 139  * a switch inside another function, from which this one was called.
 140  */
 141 static INLINE boolean
 142 mask_has_switch(struct lp_exec_mask *mask)
 143 {
 144    int i;
 145    for (i = mask->function_stack_size - 1; i >= 0; --i) {
 146       const struct function_ctx *ctx = &mask->function_stack[i];
 147       if (ctx->switch_stack_size > 0)
 148          return TRUE;
 149    }
 150    return FALSE;
 151 }
 152
 153 /*
 154  * Returns true if we're inside a conditional.
 155  * It's global, meaning that it returns true even if there's
 156  * no conditional in the current function, but we were inside
 157  * a conditional inside another function, from which this one was called.
 158  */
 159 static INLINE boolean
 160 mask_has_cond(struct lp_exec_mask *mask)
 161 {
 162    int i;
 163    for (i = mask->function_stack_size - 1; i >= 0; --i) {
 164       const struct function_ctx *ctx = &mask->function_stack[i];
 165       if (ctx->cond_stack_size > 0)
 166          return TRUE;
 167    }
 168    return FALSE;
 169 }
 170
 171
 172 /*
 173  * Initialize a function context at the specified index.
 174  */
 175 static void
 176 lp_exec_mask_function_init(struct lp_exec_mask *mask, int function_idx)
 177 {
 178    LLVMTypeRef int_type = LLVMInt32TypeInContext(mask->bld->gallivm->context);
 179    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 180    struct function_ctx *ctx =  &mask->function_stack[function_idx];
 181
 182    ctx->cond_stack_size = 0;
 183    ctx->loop_stack_size = 0;
 184    ctx->switch_stack_size = 0;
 185
 186    if (function_idx == 0) {
 187       ctx->ret_mask = mask->ret_mask;
 188    }
 189
 190    ctx->loop_limiter = lp_build_alloca(mask->bld->gallivm,
 191                                        int_type, "looplimiter");
 192    LLVMBuildStore(
 193       builder,
 194       LLVMConstInt(int_type, LP_MAX_TGSI_LOOP_ITERATIONS, false),
 195       ctx->loop_limiter);
 196 }
 197
 198 static void lp_exec_mask_init(struct lp_exec_mask *mask, struct lp_build_context *bld)
 199 {
 200    mask->bld = bld;
 201    mask->has_mask = FALSE;
 202    mask->ret_in_main = FALSE;
 203    /* For the main function */
 204    mask->function_stack_size = 1;
 205
 206    mask->int_vec_type = lp_build_int_vec_type(bld->gallivm, mask->bld->type);
 207    mask->exec_mask = mask->ret_mask = mask->break_mask = mask->cont_mask =
 208          mask->cond_mask = mask->switch_mask =
 209          LLVMConstAllOnes(mask->int_vec_type);
 210
 211    mask->function_stack = CALLOC(LP_MAX_NUM_FUNCS,
 212                                  sizeof(mask->function_stack[0]));
 213    lp_exec_mask_function_init(mask, 0);
 214 }
 215
 216 static void
 217 lp_exec_mask_fini(struct lp_exec_mask *mask)
 218 {
 219    FREE(mask->function_stack);
 220 }
 221
 222 static void lp_exec_mask_update(struct lp_exec_mask *mask)
 223 {
 224    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 225    boolean has_loop_mask = mask_has_loop(mask);
 226    boolean has_cond_mask = mask_has_cond(mask);
 227    boolean has_switch_mask = mask_has_switch(mask);
 228    boolean has_ret_mask = mask->function_stack_size > 1 ||
 229          mask->ret_in_main;
 230
 231    if (has_loop_mask) {
 232       /*for loops we need to update the entire mask at runtime */
 233       LLVMValueRef tmp;
 234       assert(mask->break_mask);
 235       tmp = LLVMBuildAnd(builder,
 236                          mask->cont_mask,
 237                          mask->break_mask,
 238                          "maskcb");
 239       mask->exec_mask = LLVMBuildAnd(builder,
 240                                      mask->cond_mask,
 241                                      tmp,
 242                                      "maskfull");
 243    } else
 244       mask->exec_mask = mask->cond_mask;
 245
 246    if (has_switch_mask) {
 247       mask->exec_mask = LLVMBuildAnd(builder,
 248                                      mask->exec_mask,
 249                                      mask->switch_mask,
 250                                      "switchmask");
 251    }
 252
 253    if (has_ret_mask) {
 254       mask->exec_mask = LLVMBuildAnd(builder,
 255                                      mask->exec_mask,
 256                                      mask->ret_mask,
 257                                      "callmask");
 258    }
 259
 260    mask->has_mask = (has_cond_mask ||
 261                      has_loop_mask ||
 262                      has_switch_mask ||
 263                      has_ret_mask);
 264 }
 265
 266 static void lp_exec_mask_cond_push(struct lp_exec_mask *mask,
 267                                    LLVMValueRef val)
 268 {
 269    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 270    struct function_ctx *ctx = func_ctx(mask);
 271
 272    if (ctx->cond_stack_size >= LP_MAX_TGSI_NESTING) {
 273       ctx->cond_stack_size++;
 274       return;
 275    }
 276    if (ctx->cond_stack_size == 0 && mask->function_stack_size == 1) {
 277       assert(mask->cond_mask == LLVMConstAllOnes(mask->int_vec_type));
 278    }
 279    ctx->cond_stack[ctx->cond_stack_size++] = mask->cond_mask;
 280    assert(LLVMTypeOf(val) == mask->int_vec_type);
 281    mask->cond_mask = LLVMBuildAnd(builder,
 282                                   mask->cond_mask,
 283                                   val,
 284                                   "");
 285    lp_exec_mask_update(mask);
 286 }
 287
 288 static void lp_exec_mask_cond_invert(struct lp_exec_mask *mask)
 289 {
 290    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 291    struct function_ctx *ctx = func_ctx(mask);
 292    LLVMValueRef prev_mask;
 293    LLVMValueRef inv_mask;
 294
 295    assert(ctx->cond_stack_size);
 296    if (ctx->cond_stack_size >= LP_MAX_TGSI_NESTING)
 297       return;
 298    prev_mask = ctx->cond_stack[ctx->cond_stack_size - 1];
 299    if (ctx->cond_stack_size == 1 && mask->function_stack_size == 1) {
 300       assert(prev_mask == LLVMConstAllOnes(mask->int_vec_type));
 301    }
 302
 303    inv_mask = LLVMBuildNot(builder, mask->cond_mask, "");
 304
 305    mask->cond_mask = LLVMBuildAnd(builder,
 306                                   inv_mask,
 307                                   prev_mask, "");
 308    lp_exec_mask_update(mask);
 309 }
 310
 311 static void lp_exec_mask_cond_pop(struct lp_exec_mask *mask)
 312 {
 313    struct function_ctx *ctx = func_ctx(mask);
 314    assert(ctx->cond_stack_size);
 315    --ctx->cond_stack_size;
 316    if (ctx->cond_stack_size >= LP_MAX_TGSI_NESTING)
 317       return;
 318    mask->cond_mask = ctx->cond_stack[ctx->cond_stack_size];
 319    lp_exec_mask_update(mask);
 320 }
 321
 322 static void lp_exec_bgnloop(struct lp_exec_mask *mask)
 323 {
 324    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 325    struct function_ctx *ctx = func_ctx(mask);
 326
 327    if (ctx->loop_stack_size >= LP_MAX_TGSI_NESTING) {
 328       ++ctx->loop_stack_size;
 329       return;
 330    }
 331
 332    ctx->break_type_stack[ctx->loop_stack_size + ctx->switch_stack_size] =
 333       ctx->break_type;
 334    ctx->break_type = LP_EXEC_MASK_BREAK_TYPE_LOOP;
 335
 336    ctx->loop_stack[ctx->loop_stack_size].loop_block = ctx->loop_block;
 337    ctx->loop_stack[ctx->loop_stack_size].cont_mask = mask->cont_mask;
 338    ctx->loop_stack[ctx->loop_stack_size].break_mask = mask->break_mask;
 339    ctx->loop_stack[ctx->loop_stack_size].break_var = ctx->break_var;
 340    ++ctx->loop_stack_size;
 341
 342    ctx->break_var = lp_build_alloca(mask->bld->gallivm, mask->int_vec_type, "");
 343    LLVMBuildStore(builder, mask->break_mask, ctx->break_var);
 344
 345    ctx->loop_block = lp_build_insert_new_block(mask->bld->gallivm, "bgnloop");
 346
 347    LLVMBuildBr(builder, ctx->loop_block);
 348    LLVMPositionBuilderAtEnd(builder, ctx->loop_block);
 349
 350    mask->break_mask = LLVMBuildLoad(builder, ctx->break_var, "");
 351
 352    lp_exec_mask_update(mask);
 353 }
 354
 355 static void lp_exec_break(struct lp_exec_mask *mask,
 356                           struct lp_build_tgsi_context * bld_base)
 357 {
 358    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 359    struct function_ctx *ctx = func_ctx(mask);
 360
 361    if (ctx->break_type == LP_EXEC_MASK_BREAK_TYPE_LOOP) {
 362       LLVMValueRef exec_mask = LLVMBuildNot(builder,
 363                                             mask->exec_mask,
 364                                             "break");
 365
 366       mask->break_mask = LLVMBuildAnd(builder,
 367                                       mask->break_mask,
 368                                       exec_mask, "break_full");
 369    }
 370    else {
 371       unsigned opcode = bld_base->instructions[bld_base->pc + 1].Instruction.Opcode;
 372       boolean break_always = (opcode == TGSI_OPCODE_ENDSWITCH ||
 373                               opcode == TGSI_OPCODE_CASE);
 374
 375
 376       if (ctx->switch_in_default) {
 377          /*
 378           * stop default execution but only if this is an unconditional switch.
 379           * (The condition here is not perfect since dead code after break is
 380           * allowed but should be sufficient since false negatives are just
 381           * unoptimized - so we don't have to pre-evaluate that).
 382           */
 383          if(break_always && ctx->switch_pc) {
 384             bld_base->pc = ctx->switch_pc;
 385             return;
 386          }
 387       }
 388
 389       if (break_always) {
 390          mask->switch_mask = LLVMConstNull(mask->bld->int_vec_type);
 391       }
 392       else {
 393          LLVMValueRef exec_mask = LLVMBuildNot(builder,
 394                                                mask->exec_mask,
 395                                                "break");
 396          mask->switch_mask = LLVMBuildAnd(builder,
 397                                           mask->switch_mask,
 398                                           exec_mask, "break_switch");
 399       }
 400    }
 401
 402    lp_exec_mask_update(mask);
 403 }
 404
 405 static void lp_exec_break_condition(struct lp_exec_mask *mask,
 406                                     LLVMValueRef cond)
 407 {
 408    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 409    struct function_ctx *ctx = func_ctx(mask);
 410    LLVMValueRef cond_mask = LLVMBuildAnd(builder,
 411                                          mask->exec_mask,
 412                                          cond, "cond_mask");
 413    cond_mask = LLVMBuildNot(builder, cond_mask, "break_cond");
 414
 415    if (ctx->break_type == LP_EXEC_MASK_BREAK_TYPE_LOOP) {
 416       mask->break_mask = LLVMBuildAnd(builder,
 417                                       mask->break_mask,
 418                                       cond_mask, "breakc_full");
 419    }
 420    else {
 421       mask->switch_mask = LLVMBuildAnd(builder,
 422                                        mask->switch_mask,
 423                                        cond_mask, "breakc_switch");
 424    }
 425
 426    lp_exec_mask_update(mask);
 427 }
 428
 429 static void lp_exec_continue(struct lp_exec_mask *mask)
 430 {
 431    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 432    LLVMValueRef exec_mask = LLVMBuildNot(builder,
 433                                          mask->exec_mask,
 434                                          "");
 435
 436    mask->cont_mask = LLVMBuildAnd(builder,
 437                                   mask->cont_mask,
 438                                   exec_mask, "");
 439
 440    lp_exec_mask_update(mask);
 441 }
 442
 443
 444 static void lp_exec_endloop(struct gallivm_state *gallivm,
 445                             struct lp_exec_mask *mask)
 446 {
 447    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 448    struct function_ctx *ctx = func_ctx(mask);
 449    LLVMBasicBlockRef endloop;
 450    LLVMTypeRef int_type = LLVMInt32TypeInContext(mask->bld->gallivm->context);
 451    LLVMTypeRef reg_type = LLVMIntTypeInContext(gallivm->context,
 452                                                mask->bld->type.width *
 453                                                mask->bld->type.length);
 454    LLVMValueRef i1cond, i2cond, icond, limiter;
 455
 456    assert(mask->break_mask);
 457
 458
 459    assert(ctx->loop_stack_size);
 460    if (ctx->loop_stack_size > LP_MAX_TGSI_NESTING) {
 461       --ctx->loop_stack_size;
 462       return;
 463    }
 464
 465    /*
 466     * Restore the cont_mask, but don't pop
 467     */
 468    mask->cont_mask = ctx->loop_stack[ctx->loop_stack_size - 1].cont_mask;
 469    lp_exec_mask_update(mask);
 470
 471    /*
 472     * Unlike the continue mask, the break_mask must be preserved across loop
 473     * iterations
 474     */
 475    LLVMBuildStore(builder, mask->break_mask, ctx->break_var);
 476
 477    /* Decrement the loop limiter */
 478    limiter = LLVMBuildLoad(builder, ctx->loop_limiter, "");
 479
 480    limiter = LLVMBuildSub(
 481       builder,
 482       limiter,
 483       LLVMConstInt(int_type, 1, false),
 484       "");
 485
 486    LLVMBuildStore(builder, limiter, ctx->loop_limiter);
 487
 488    /* i1cond = (mask != 0) */
 489    i1cond = LLVMBuildICmp(
 490       builder,
 491       LLVMIntNE,
 492       LLVMBuildBitCast(builder, mask->exec_mask, reg_type, ""),
 493       LLVMConstNull(reg_type), "i1cond");
 494
 495    /* i2cond = (looplimiter > 0) */
 496    i2cond = LLVMBuildICmp(
 497       builder,
 498       LLVMIntSGT,
 499       limiter,
 500       LLVMConstNull(int_type), "i2cond");
 501
 502    /* if( i1cond && i2cond ) */
 503    icond = LLVMBuildAnd(builder, i1cond, i2cond, "");
 504
 505    endloop = lp_build_insert_new_block(mask->bld->gallivm, "endloop");
 506
 507    LLVMBuildCondBr(builder,
 508                    icond, ctx->loop_block, endloop);
 509
 510    LLVMPositionBuilderAtEnd(builder, endloop);
 511
 512    assert(ctx->loop_stack_size);
 513    --ctx->loop_stack_size;
 514    mask->cont_mask = ctx->loop_stack[ctx->loop_stack_size].cont_mask;
 515    mask->break_mask = ctx->loop_stack[ctx->loop_stack_size].break_mask;
 516    ctx->loop_block = ctx->loop_stack[ctx->loop_stack_size].loop_block;
 517    ctx->break_var = ctx->loop_stack[ctx->loop_stack_size].break_var;
 518    ctx->break_type = ctx->break_type_stack[ctx->loop_stack_size +
 519          ctx->switch_stack_size];
 520
 521    lp_exec_mask_update(mask);
 522 }
 523
 524 static void lp_exec_switch(struct lp_exec_mask *mask,
 525                            LLVMValueRef switchval)
 526 {
 527    struct function_ctx *ctx = func_ctx(mask);
 528
 529    if (ctx->switch_stack_size >= LP_MAX_TGSI_NESTING ||
 530        ctx->loop_stack_size > LP_MAX_TGSI_NESTING) {
 531       ctx->switch_stack_size++;
 532       return;
 533    }
 534
 535    ctx->break_type_stack[ctx->loop_stack_size + ctx->switch_stack_size] =
 536       ctx->break_type;
 537    ctx->break_type = LP_EXEC_MASK_BREAK_TYPE_SWITCH;
 538
 539    ctx->switch_stack[ctx->switch_stack_size].switch_mask = mask->switch_mask;
 540    ctx->switch_stack[ctx->switch_stack_size].switch_val = ctx->switch_val;
 541    ctx->switch_stack[ctx->switch_stack_size].switch_mask_default = ctx->switch_mask_default;
 542    ctx->switch_stack[ctx->switch_stack_size].switch_in_default = ctx->switch_in_default;
 543    ctx->switch_stack[ctx->switch_stack_size].switch_pc = ctx->switch_pc;
 544    ctx->switch_stack_size++;
 545
 546    mask->switch_mask = LLVMConstNull(mask->int_vec_type);
 547    ctx->switch_val = switchval;
 548    ctx->switch_mask_default = LLVMConstNull(mask->int_vec_type);
 549    ctx->switch_in_default = false;
 550    ctx->switch_pc = 0;
 551
 552    lp_exec_mask_update(mask);
 553 }
 554
 555 static void lp_exec_endswitch(struct lp_exec_mask *mask,
 556                               struct lp_build_tgsi_context * bld_base)
 557 {
 558    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 559    struct function_ctx *ctx = func_ctx(mask);
 560
 561    if (ctx->switch_stack_size > LP_MAX_TGSI_NESTING) {
 562       ctx->switch_stack_size--;
 563       return;
 564    }
 565
 566    /* check if there's deferred default if so do it now */
 567    if (ctx->switch_pc && !ctx->switch_in_default) {
 568       LLVMValueRef prevmask, defaultmask;
 569       unsigned tmp_pc;
 570       prevmask = ctx->switch_stack[ctx->switch_stack_size - 1].switch_mask;
 571       defaultmask = LLVMBuildNot(builder, ctx->switch_mask_default, "sw_default_mask");
 572       mask->switch_mask = LLVMBuildAnd(builder, prevmask, defaultmask, "sw_mask");
 573       ctx->switch_in_default = true;
 574
 575       lp_exec_mask_update(mask);
 576
 577       assert(bld_base->instructions[ctx->switch_pc - 1].Instruction.Opcode ==
 578              TGSI_OPCODE_DEFAULT);
 579
 580       tmp_pc = bld_base->pc;
 581       bld_base->pc = ctx->switch_pc;
 582       /*
 583        * re-purpose switch_pc to point to here again, since we stop execution of
 584        * the deferred default after next break.
 585        */
 586       ctx->switch_pc = tmp_pc - 1;
 587
 588       return;
 589    }
 590
 591    else if (ctx->switch_pc && ctx->switch_in_default) {
 592       assert(bld_base->pc == ctx->switch_pc + 1);
 593    }
 594
 595    ctx->switch_stack_size--;
 596    mask->switch_mask = ctx->switch_stack[ctx->switch_stack_size].switch_mask;
 597    ctx->switch_val = ctx->switch_stack[ctx->switch_stack_size].switch_val;
 598    ctx->switch_mask_default = ctx->switch_stack[ctx->switch_stack_size].switch_mask_default;
 599    ctx->switch_in_default = ctx->switch_stack[ctx->switch_stack_size].switch_in_default;
 600    ctx->switch_pc = ctx->switch_stack[ctx->switch_stack_size].switch_pc;
 601
 602    ctx->break_type = ctx->break_type_stack[ctx->loop_stack_size + ctx->switch_stack_size];
 603
 604    lp_exec_mask_update(mask);
 605 }
 606
 607 static void lp_exec_case(struct lp_exec_mask *mask,
 608                          LLVMValueRef caseval)
 609 {
 610    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 611    struct function_ctx *ctx = func_ctx(mask);
 612
 613    LLVMValueRef casemask, prevmask;
 614
 615    if (ctx->switch_stack_size > LP_MAX_TGSI_NESTING) {
 616       return;
 617    }
 618
 619    /* skipping case mask evaluation here is NOT optional (not in all cases anyway). */
 620    if (!ctx->switch_in_default) {
 621       prevmask = ctx->switch_stack[ctx->switch_stack_size - 1].switch_mask;
 622       casemask = lp_build_cmp(mask->bld, PIPE_FUNC_EQUAL, caseval, ctx->switch_val);
 623       ctx->switch_mask_default = LLVMBuildOr(builder, casemask,
 624                                              ctx->switch_mask_default, "sw_default_mask");
 625       casemask = LLVMBuildOr(builder, casemask, mask->switch_mask, "");
 626       mask->switch_mask = LLVMBuildAnd(builder, casemask, prevmask, "sw_mask");
 627
 628       lp_exec_mask_update(mask);
 629    }
 630 }
 631
 632 /*
 633  * Analyse default statement in a switch.
 634  * \return true if default is last statement, false otherwise
 635  * \param default_pc_start contains pc of instruction to jump to
 636  *                         if default wasn't last but there's no
 637  *                         fallthrough into default.
 638  */
 639 static boolean default_analyse_is_last(struct lp_exec_mask *mask,
 640                                        struct lp_build_tgsi_context * bld_base,
 641                                        int *default_pc_start)
 642 {
 643    unsigned pc = bld_base->pc;
 644    struct function_ctx *ctx = func_ctx(mask);
 645    unsigned curr_switch_stack = ctx->switch_stack_size;
 646
 647    if (ctx->switch_stack_size > LP_MAX_TGSI_NESTING) {
 648       return false;
 649    }
 650
 651    /* skip over case statements which are together with default */
 652    while (bld_base->instructions[pc].Instruction.Opcode == TGSI_OPCODE_CASE) {
 653       pc++;
 654    }
 655
 656    while (pc != -1 && pc < bld_base->num_instructions) {
 657       unsigned opcode = bld_base->instructions[pc].Instruction.Opcode;
 658       switch (opcode) {
 659       case TGSI_OPCODE_CASE:
 660          if (curr_switch_stack == ctx->switch_stack_size) {
 661             *default_pc_start = pc - 1;
 662             return false;
 663          }
 664          break;
 665       case TGSI_OPCODE_SWITCH:
 666          curr_switch_stack++;
 667          break;
 668       case TGSI_OPCODE_ENDSWITCH:
 669          if (curr_switch_stack == ctx->switch_stack_size) {
 670             *default_pc_start = pc - 1;
 671             return true;
 672          }
 673          curr_switch_stack--;
 674          break;
 675       }
 676       pc++;
 677    }
 678    /* should never arrive here */
 679    assert(0);
 680    return true;
 681 }
 682
 683 static void lp_exec_default(struct lp_exec_mask *mask,
 684                             struct lp_build_tgsi_context * bld_base)
 685 {
 686    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 687    struct function_ctx *ctx = func_ctx(mask);
 688
 689    int default_exec_pc;
 690    boolean default_is_last;
 691
 692    if (ctx->switch_stack_size > LP_MAX_TGSI_NESTING) {
 693       return;
 694    }
 695
 696    /*
 697     * This is a messy opcode, because it may not be always at the end and
 698     * there can be fallthrough in and out of it.
 699     */
 700
 701    default_is_last = default_analyse_is_last(mask, bld_base, &default_exec_pc);
 702    /*
 703     * If it is last statement in switch (note that case statements appearing
 704     * "at the same time" as default don't change that) everything is just fine,
 705     * update switch mask and go on. This means we can handle default with
 706     * fallthrough INTO it without overhead, if it is last.
 707     */
 708    if (default_is_last) {
 709       LLVMValueRef prevmask, defaultmask;
 710       prevmask = ctx->switch_stack[ctx->switch_stack_size - 1].switch_mask;
 711       defaultmask = LLVMBuildNot(builder, ctx->switch_mask_default, "sw_default_mask");
 712       defaultmask = LLVMBuildOr(builder, defaultmask, mask->switch_mask, "");
 713       mask->switch_mask = LLVMBuildAnd(builder, prevmask, defaultmask, "sw_mask");
 714       ctx->switch_in_default = true;
 715
 716       lp_exec_mask_update(mask);
 717    }
 718    else {
 719       /*
 720        * Technically, "case" immediately before default isn't really a
 721        * fallthrough, however we still have to count them as such as we
 722        * already have updated the masks.
 723        * If that happens in practice could add a switch optimizer pass
 724        * which just gets rid of all case statements appearing together with
 725        * default (or could do switch analysis at switch start time instead).
 726        */
 727       unsigned opcode = bld_base->instructions[bld_base->pc - 1].Instruction.Opcode;
 728       boolean ft_into = (opcode != TGSI_OPCODE_BRK &&
 729                          opcode != TGSI_OPCODE_SWITCH);
 730       /*
 731        * If it is not last statement and there was no fallthrough into it,
 732        * we record the PC and continue execution at next case (again, those
 733        * case encountered at the same time don't count). At endswitch
 734        * time, we update switchmask, and go back executing the code we skipped
 735        * until the next break (possibly re-executing some code with changed mask
 736        * if there was a fallthrough out of default).
 737        * Finally, if it is not last statement and there was a fallthrough into it,
 738        * do the same as with the former case, except instead of skipping the code
 739        * just execute it without updating the mask, then go back and re-execute.
 740        */
 741       ctx->switch_pc = bld_base->pc;
 742       if (!ft_into) {
 743          bld_base->pc = default_exec_pc;
 744       }
 745    }
 746 }
 747
 748
 749 /* stores val into an address pointed to by dst_ptr.
 750  * mask->exec_mask is used to figure out which bits of val
 751  * should be stored into the address
 752  * (0 means don't store this bit, 1 means do store).
 753  */
 754 static void lp_exec_mask_store(struct lp_exec_mask *mask,
 755                                struct lp_build_context *bld_store,
 756                                LLVMValueRef pred,
 757                                LLVMValueRef val,
 758                                LLVMValueRef dst_ptr)
 759 {
 760    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 761
 762    assert(lp_check_value(bld_store->type, val));
 763    assert(LLVMGetTypeKind(LLVMTypeOf(dst_ptr)) == LLVMPointerTypeKind);
 764    assert(LLVMGetElementType(LLVMTypeOf(dst_ptr)) == LLVMTypeOf(val));
 765
 766    /* Mix the predicate and execution mask */
 767    if (mask->has_mask) {
 768       if (pred) {
 769          pred = LLVMBuildAnd(builder, pred, mask->exec_mask, "");
 770       } else {
 771          pred = mask->exec_mask;
 772       }
 773    }
 774
 775    if (pred) {
 776       LLVMValueRef res, dst;
 777
 778       dst = LLVMBuildLoad(builder, dst_ptr, "");
 779       res = lp_build_select(bld_store, pred, val, dst);
 780       LLVMBuildStore(builder, res, dst_ptr);
 781    } else
 782       LLVMBuildStore(builder, val, dst_ptr);
 783 }
 784
 785 static void lp_exec_mask_call(struct lp_exec_mask *mask,
 786                               int func,
 787                               int *pc)
 788 {
 789    if (mask->function_stack_size >= LP_MAX_NUM_FUNCS) {
 790       return;
 791    }
 792
 793    lp_exec_mask_function_init(mask, mask->function_stack_size);
 794    mask->function_stack[mask->function_stack_size].pc = *pc;
 795    mask->function_stack[mask->function_stack_size].ret_mask = mask->ret_mask;
 796    mask->function_stack_size++;
 797    *pc = func;
 798 }
 799
 800 static void lp_exec_mask_ret(struct lp_exec_mask *mask, int *pc)
 801 {
 802    LLVMBuilderRef builder = mask->bld->gallivm->builder;
 803    struct function_ctx *ctx = func_ctx(mask);
 804    LLVMValueRef exec_mask;
 805
 806    if (ctx->cond_stack_size == 0 &&
 807        ctx->loop_stack_size == 0 &&
 808        ctx->switch_stack_size == 0 &&
 809        mask->function_stack_size == 1) {
 810       /* returning from main() */
 811       *pc = -1;
 812       return;
 813    }
 814
 815    if (mask->function_stack_size == 1) {
 816       /*
 817        * This requires special handling since we need to ensure
 818        * we don't drop the mask even if we have no call stack
 819        * (e.g. after a ret in a if clause after the endif)
 820        */
 821       mask->ret_in_main = TRUE;
 822    }
 823
 824    exec_mask = LLVMBuildNot(builder,
 825                             mask->exec_mask,
 826                             "ret");
 827
 828    mask->ret_mask = LLVMBuildAnd(builder,
 829                                  mask->ret_mask,
 830                                  exec_mask, "ret_full");
 831
 832    lp_exec_mask_update(mask);
 833 }
 834
 835 static void lp_exec_mask_bgnsub(struct lp_exec_mask *mask)
 836 {
 837 }
 838
 839 static void lp_exec_mask_endsub(struct lp_exec_mask *mask, int *pc)
 840 {
 841    struct function_ctx *ctx;
 842
 843    assert(mask->function_stack_size > 1);
 844    assert(mask->function_stack_size <= LP_MAX_NUM_FUNCS);
 845
 846    ctx = func_ctx(mask);
 847    mask->function_stack_size--;
 848
 849    *pc = ctx->pc;
 850    mask->ret_mask = ctx->ret_mask;
 851
 852    lp_exec_mask_update(mask);
 853 }
 854
 855
 856 static LLVMValueRef
 857 get_file_ptr(struct lp_build_tgsi_soa_context *bld,
 858              unsigned file,
 859              unsigned index,
 860              unsigned chan)
 861 {
 862    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
 863    LLVMValueRef (*array_of_vars)[TGSI_NUM_CHANNELS];
 864    LLVMValueRef var_of_array;
 865
 866    switch (file) {
 867    case TGSI_FILE_TEMPORARY:
 868       array_of_vars = bld->temps;
 869       var_of_array = bld->temps_array;
 870       break;
 871    case TGSI_FILE_OUTPUT:
 872       array_of_vars = bld->outputs;
 873       var_of_array = bld->outputs_array;
 874       break;
 875    default:
 876       assert(0);
 877       return NULL;
 878    }
 879
 880    assert(chan < 4);
 881
 882    if (bld->indirect_files & (1 << file)) {
 883       LLVMValueRef lindex = lp_build_const_int32(bld->bld_base.base.gallivm, index * 4 + chan);
 884       return LLVMBuildGEP(builder, var_of_array, &lindex, 1, "");
 885    }
 886    else {
 887       assert(index <= bld->bld_base.info->file_max[file]);
 888       return array_of_vars[index][chan];
 889    }
 890 }
 891
 892
 893 /**
 894  * Return pointer to a temporary register channel (src or dest).
 895  * Note that indirect addressing cannot be handled here.
 896  * \param index  which temporary register
 897  * \param chan  which channel of the temp register.
 898  */
 899 LLVMValueRef
 900 lp_get_temp_ptr_soa(struct lp_build_tgsi_soa_context *bld,
 901              unsigned index,
 902              unsigned chan)
 903 {
 904    return get_file_ptr(bld, TGSI_FILE_TEMPORARY, index, chan);
 905 }
 906
 907 /**
 908  * Return pointer to a output register channel (src or dest).
 909  * Note that indirect addressing cannot be handled here.
 910  * \param index  which output register
 911  * \param chan  which channel of the output register.
 912  */
 913 LLVMValueRef
 914 lp_get_output_ptr(struct lp_build_tgsi_soa_context *bld,
 915                unsigned index,
 916                unsigned chan)
 917 {
 918    return get_file_ptr(bld, TGSI_FILE_OUTPUT, index, chan);
 919 }
 920
 921 /*
 922  * If we have indirect addressing in outputs copy our alloca array
 923  * to the outputs slots specified by the caller to make sure
 924  * our outputs are delivered consistently via the same interface.
 925  */
 926 static void
 927 gather_outputs(struct lp_build_tgsi_soa_context * bld)
 928 {
 929    if ((bld->indirect_files & (1 << TGSI_FILE_OUTPUT))) {
 930       unsigned index, chan;
 931       assert(bld->bld_base.info->num_outputs <=
 932              bld->bld_base.info->file_max[TGSI_FILE_OUTPUT] + 1);
 933       for (index = 0; index < bld->bld_base.info->num_outputs; ++index) {
 934          for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
 935             bld->outputs[index][chan] = lp_get_output_ptr(bld, index, chan);
 936          }
 937       }
 938    }
 939 }
 940
 941 /**
 942  * Gather vector.
 943  * XXX the lp_build_gather() function should be capable of doing this
 944  * with a little work.
 945  */
 946 static LLVMValueRef
 947 build_gather(struct lp_build_context *bld,
 948              LLVMValueRef base_ptr,
 949              LLVMValueRef indexes,
 950              LLVMValueRef *overflow_mask)
 951 {
 952    LLVMBuilderRef builder = bld->gallivm->builder;
 953    LLVMValueRef res = bld->undef;
 954    unsigned i;
 955    LLVMValueRef temp_ptr = NULL;
 956
 957    if (overflow_mask) {
 958       temp_ptr = lp_build_alloca(
 959          bld->gallivm,
 960          lp_build_vec_type(bld->gallivm, bld->type), "");
 961    }
 962
 963    /*
 964     * Loop over elements of index_vec, load scalar value, insert it into 'res'.
 965     */
 966    for (i = 0; i < bld->type.length; i++) {
 967       LLVMValueRef ii = lp_build_const_int32(bld->gallivm, i);
 968       LLVMValueRef index = LLVMBuildExtractElement(builder,
 969                                                    indexes, ii, "");
 970       LLVMValueRef scalar_ptr, scalar;
 971       LLVMValueRef overflow;
 972       struct lp_build_if_state if_ctx;
 973
 974       /*
 975        * overflow_mask is a boolean vector telling us which channels
 976        * in the vector overflowed. We use the overflow behavior for
 977        * constant buffers which is defined as:
 978        * Out of bounds access to constant buffer returns 0 in all
 979        * componenets. Out of bounds behavior is always with respect
 980        * to the size of the buffer bound at that slot.
 981        */
 982       if (overflow_mask) {
 983          overflow = LLVMBuildExtractElement(builder, *overflow_mask,
 984                                             ii, "");
 985          lp_build_if(&if_ctx, bld->gallivm, overflow);
 986          {
 987             LLVMValueRef val = LLVMBuildLoad(builder, temp_ptr, "");
 988             val = LLVMBuildInsertElement(
 989                builder, val,
 990                LLVMConstNull(LLVMFloatTypeInContext(bld->gallivm->context)),
 991                ii, "");
 992             LLVMBuildStore(builder, val, temp_ptr);
 993          }
 994          lp_build_else(&if_ctx);
 995          {
 996             LLVMValueRef val = LLVMBuildLoad(builder, temp_ptr, "");
 997
 998             scalar_ptr = LLVMBuildGEP(builder, base_ptr,
 999                                       &index, 1, "gather_ptr");
1000             scalar = LLVMBuildLoad(builder, scalar_ptr, "");
1001
1002             val = LLVMBuildInsertElement(builder, val, scalar, ii, "");
1003
1004             LLVMBuildStore(builder, val, temp_ptr);
1005          }
1006          lp_build_endif(&if_ctx);
1007       } else {
1008          scalar_ptr = LLVMBuildGEP(builder, base_ptr,
1009                                    &index, 1, "gather_ptr");
1010          scalar = LLVMBuildLoad(builder, scalar_ptr, "");
1011
1012          res = LLVMBuildInsertElement(builder, res, scalar, ii, "");
1013       }
1014    }
1015
1016    if (overflow_mask) {
1017       res = LLVMBuildLoad(builder, temp_ptr, "gather_val");
1018    }
1019
1020    return res;
1021 }
1022
1023
1024 /**
1025  * Scatter/store vector.
1026  */
1027 static void
1028 emit_mask_scatter(struct lp_build_tgsi_soa_context *bld,
1029                   LLVMValueRef base_ptr,
1030                   LLVMValueRef indexes,
1031                   LLVMValueRef values,
1032                   struct lp_exec_mask *mask,
1033                   LLVMValueRef pred)
1034 {
1035    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
1036    LLVMBuilderRef builder = gallivm->builder;
1037    unsigned i;
1038
1039    /* Mix the predicate and execution mask */
1040    if (mask->has_mask) {
1041       if (pred) {
1042          pred = LLVMBuildAnd(builder, pred, mask->exec_mask, "");
1043       }
1044       else {
1045          pred = mask->exec_mask;
1046       }
1047    }
1048
1049    /*
1050     * Loop over elements of index_vec, store scalar value.
1051     */
1052    for (i = 0; i < bld->bld_base.base.type.length; i++) {
1053       LLVMValueRef ii = lp_build_const_int32(gallivm, i);
1054       LLVMValueRef index = LLVMBuildExtractElement(builder, indexes, ii, "");
1055       LLVMValueRef scalar_ptr = LLVMBuildGEP(builder, base_ptr, &index, 1, "scatter_ptr");
1056       LLVMValueRef val = LLVMBuildExtractElement(builder, values, ii, "scatter_val");
1057       LLVMValueRef scalar_pred = pred ?
1058          LLVMBuildExtractElement(builder, pred, ii, "scatter_pred") : NULL;
1059
1060       if (0)
1061          lp_build_printf(gallivm, "scatter %d: val %f at %d %p\n",
1062                          ii, val, index, scalar_ptr);
1063
1064       if (scalar_pred) {
1065          LLVMValueRef real_val, dst_val;
1066          dst_val = LLVMBuildLoad(builder, scalar_ptr, "");
1067          real_val = lp_build_select(&bld->elem_bld, scalar_pred, val, dst_val);
1068          LLVMBuildStore(builder, real_val, scalar_ptr);
1069       }
1070       else {
1071          LLVMBuildStore(builder, val, scalar_ptr);
1072       }
1073    }
1074 }
1075
1076
1077 /**
1078  * Read the current value of the ADDR register, convert the floats to
1079  * ints, add the base index and return the vector of offsets.
1080  * The offsets will be used to index into the constant buffer or
1081  * temporary register file.
1082  */
1083 static LLVMValueRef
1084 get_indirect_index(struct lp_build_tgsi_soa_context *bld,
1085                    unsigned reg_file, unsigned reg_index,
1086                    const struct tgsi_ind_register *indirect_reg)
1087 {
1088    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
1089    struct lp_build_context *uint_bld = &bld->bld_base.uint_bld;
1090    /* always use X component of address register */
1091    unsigned swizzle = indirect_reg->Swizzle;
1092    LLVMValueRef base;
1093    LLVMValueRef rel;
1094    LLVMValueRef max_index;
1095    LLVMValueRef index;
1096
1097    assert(bld->indirect_files & (1 << reg_file));
1098
1099    base = lp_build_const_int_vec(bld->bld_base.base.gallivm, uint_bld->type, reg_index);
1100
1101    assert(swizzle < 4);
1102    switch (indirect_reg->File) {
1103    case TGSI_FILE_ADDRESS:
1104       rel = LLVMBuildLoad(builder,
1105                           bld->addr[indirect_reg->Index][swizzle],
1106                           "load addr reg");
1107       /* ADDR LLVM values already have LLVM integer type. */
1108       break;
1109    case TGSI_FILE_TEMPORARY:
1110       rel = lp_get_temp_ptr_soa(bld, indirect_reg->Index, swizzle);
1111       rel = LLVMBuildLoad(builder, rel, "load temp reg");
1112       /* TEMP LLVM values always have LLVM float type, but for indirection, the
1113        * value actually stored is expected to be an integer */
1114       rel = LLVMBuildBitCast(builder, rel, uint_bld->vec_type, "");
1115       break;
1116    default:
1117       assert(0);
1118       rel = uint_bld->zero;
1119    }
1120
1121    index = lp_build_add(uint_bld, base, rel);
1122
1123    /*
1124     * emit_fetch_constant handles constant buffer overflow so this code
1125     * is pointless for them.
1126     * Furthermore the D3D10 spec in section 6.5 says:
1127     * If the constant buffer bound to a slot is larger than the size
1128     * declared in the shader for that slot, implementations are allowed
1129     * to return incorrect data (not necessarily 0) for indices that are
1130     * larger than the declared size but smaller than the buffer size.
1131     */
1132    if (reg_file != TGSI_FILE_CONSTANT) {
1133       max_index = lp_build_const_int_vec(bld->bld_base.base.gallivm,
1134                                          uint_bld->type,
1135                                          bld->bld_base.info->file_max[reg_file]);
1136
1137       assert(!uint_bld->type.sign);
1138       index = lp_build_min(uint_bld, index, max_index);
1139    }
1140
1141    return index;
1142 }
1143
1144 static struct lp_build_context *
1145 stype_to_fetch(struct lp_build_tgsi_context * bld_base,
1146                enum tgsi_opcode_type stype)
1147 {
1148    struct lp_build_context *bld_fetch;
1149
1150    switch (stype) {
1151    case TGSI_TYPE_FLOAT:
1152    case TGSI_TYPE_UNTYPED:
1153       bld_fetch = &bld_base->base;
1154       break;
1155    case TGSI_TYPE_UNSIGNED:
1156       bld_fetch = &bld_base->uint_bld;
1157       break;
1158    case TGSI_TYPE_SIGNED:
1159       bld_fetch = &bld_base->int_bld;
1160       break;
1161    case TGSI_TYPE_VOID:
1162    case TGSI_TYPE_DOUBLE:
1163    default:
1164       assert(0);
1165       bld_fetch = NULL;
1166       break;
1167    }
1168    return bld_fetch;
1169 }
1170
1171 static LLVMValueRef
1172 get_soa_array_offsets(struct lp_build_context *uint_bld,
1173                       LLVMValueRef indirect_index,
1174                       unsigned chan_index,
1175                       boolean need_perelement_offset)
1176 {
1177    struct gallivm_state *gallivm = uint_bld->gallivm;
1178    LLVMValueRef chan_vec =
1179       lp_build_const_int_vec(uint_bld->gallivm, uint_bld->type, chan_index);
1180    LLVMValueRef length_vec =
1181       lp_build_const_int_vec(gallivm, uint_bld->type, uint_bld->type.length);
1182    LLVMValueRef index_vec;
1183
1184    /* index_vec = (indirect_index * 4 + chan_index) * length + offsets */
1185    index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
1186    index_vec = lp_build_add(uint_bld, index_vec, chan_vec);
1187    index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
1188
1189    if (need_perelement_offset) {
1190       LLVMValueRef pixel_offsets;
1191       int i;
1192      /* build pixel offset vector: {0, 1, 2, 3, ...} */
1193       pixel_offsets = uint_bld->undef;
1194       for (i = 0; i < uint_bld->type.length; i++) {
1195          LLVMValueRef ii = lp_build_const_int32(gallivm, i);
1196          pixel_offsets = LLVMBuildInsertElement(gallivm->builder, pixel_offsets,
1197                                                 ii, ii, "");
1198       }
1199       index_vec = lp_build_add(uint_bld, index_vec, pixel_offsets);
1200    }
1201    return index_vec;
1202 }
1203
1204 static LLVMValueRef
1205 emit_fetch_constant(
1206    struct lp_build_tgsi_context * bld_base,
1207    const struct tgsi_full_src_register * reg,
1208    enum tgsi_opcode_type stype,
1209    unsigned swizzle)
1210 {
1211    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1212    struct gallivm_state *gallivm = bld_base->base.gallivm;
1213    LLVMBuilderRef builder = gallivm->builder;
1214    struct lp_build_context *uint_bld = &bld_base->uint_bld;
1215    unsigned dimension = 0;
1216    LLVMValueRef consts_ptr;
1217    LLVMValueRef num_consts;
1218    LLVMValueRef res;
1219
1220    /* XXX: Handle fetching xyzw components as a vector */
1221    assert(swizzle != ~0);
1222
1223    if (reg->Register.Dimension) {
1224       assert(!reg->Dimension.Indirect);
1225       dimension = reg->Dimension.Index;
1226       assert(dimension < LP_MAX_TGSI_CONST_BUFFERS);
1227    }
1228
1229    consts_ptr = bld->consts[dimension];
1230    num_consts = bld->consts_sizes[dimension];
1231
1232    if (reg->Register.Indirect) {
1233       LLVMValueRef indirect_index;
1234       LLVMValueRef swizzle_vec =
1235          lp_build_const_int_vec(gallivm, uint_bld->type, swizzle);
1236       LLVMValueRef index_vec;  /* index into the const buffer */
1237       LLVMValueRef overflow_mask;
1238
1239       indirect_index = get_indirect_index(bld,
1240                                           reg->Register.File,
1241                                           reg->Register.Index,
1242                                           &reg->Indirect);
1243
1244       /* All fetches are from the same constant buffer, so
1245        * we need to propagate the size to a vector to do a
1246        * vector comparison */
1247       num_consts = lp_build_broadcast_scalar(uint_bld, num_consts);
1248       /* Construct a boolean vector telling us which channels
1249        * overflow the bound constant buffer */
1250       overflow_mask = LLVMBuildICmp(builder, LLVMIntUGE,
1251                                     indirect_index,
1252                                     num_consts, "");
1253
1254       /* index_vec = indirect_index * 4 + swizzle */
1255       index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
1256       index_vec = lp_build_add(uint_bld, index_vec, swizzle_vec);
1257
1258       /* Gather values from the constant buffer */
1259       res = build_gather(&bld_base->base, consts_ptr, index_vec,
1260                          &overflow_mask);
1261    }
1262    else {
1263       LLVMValueRef index;  /* index into the const buffer */
1264       LLVMValueRef scalar, scalar_ptr;
1265
1266       index = lp_build_const_int32(gallivm, reg->Register.Index * 4 + swizzle);
1267
1268       scalar_ptr = LLVMBuildGEP(builder, consts_ptr,
1269                                 &index, 1, "");
1270       scalar = LLVMBuildLoad(builder, scalar_ptr, "");
1271       res = lp_build_broadcast_scalar(&bld_base->base, scalar);
1272    }
1273
1274    if (stype == TGSI_TYPE_SIGNED || stype == TGSI_TYPE_UNSIGNED) {
1275       struct lp_build_context *bld_fetch = stype_to_fetch(bld_base, stype);
1276       res = LLVMBuildBitCast(builder, res, bld_fetch->vec_type, "");
1277    }
1278
1279    return res;
1280 }
1281
1282 static LLVMValueRef
1283 emit_fetch_immediate(
1284    struct lp_build_tgsi_context * bld_base,
1285    const struct tgsi_full_src_register * reg,
1286    enum tgsi_opcode_type stype,
1287    unsigned swizzle)
1288 {
1289    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1290    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
1291    LLVMBuilderRef builder = gallivm->builder;
1292    LLVMValueRef res = NULL;
1293
1294    if (bld->use_immediates_array || reg->Register.Indirect) {
1295       LLVMValueRef imms_array;
1296       LLVMTypeRef fptr_type;
1297
1298       /* cast imms_array pointer to float* */
1299       fptr_type = LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
1300       imms_array = LLVMBuildBitCast(builder, bld->imms_array, fptr_type, "");
1301
1302       if (reg->Register.Indirect) {
1303          LLVMValueRef indirect_index;
1304          LLVMValueRef index_vec;  /* index into the immediate register array */
1305
1306          indirect_index = get_indirect_index(bld,
1307                                              reg->Register.File,
1308                                              reg->Register.Index,
1309                                              &reg->Indirect);
1310          /*
1311           * Unlike for other reg classes, adding pixel offsets is unnecessary -
1312           * immediates are stored as full vectors (FIXME??? - might be better
1313           * to store them the same as constants) but all elements are the same
1314           * in any case.
1315           */
1316          index_vec = get_soa_array_offsets(&bld_base->uint_bld,
1317                                            indirect_index,
1318                                            swizzle,
1319                                            FALSE);
1320
1321          /* Gather values from the immediate register array */
1322          res = build_gather(&bld_base->base, imms_array, index_vec, NULL);
1323       } else {
1324          LLVMValueRef lindex = lp_build_const_int32(gallivm,
1325                                         reg->Register.Index * 4 + swizzle);
1326          LLVMValueRef imms_ptr =  LLVMBuildGEP(builder,
1327                                                 bld->imms_array, &lindex, 1, "");
1328          res = LLVMBuildLoad(builder, imms_ptr, "");
1329       }
1330    }
1331    else {
1332       res = bld->immediates[reg->Register.Index][swizzle];
1333    }
1334
1335    if (stype == TGSI_TYPE_UNSIGNED) {
1336       res = LLVMBuildBitCast(builder, res, bld_base->uint_bld.vec_type, "");
1337    } else if (stype == TGSI_TYPE_SIGNED) {
1338       res = LLVMBuildBitCast(builder, res, bld_base->int_bld.vec_type, "");
1339    }
1340    return res;
1341 }
1342
1343 static LLVMValueRef
1344 emit_fetch_input(
1345    struct lp_build_tgsi_context * bld_base,
1346    const struct tgsi_full_src_register * reg,
1347    enum tgsi_opcode_type stype,
1348    unsigned swizzle)
1349 {
1350    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1351    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
1352    LLVMBuilderRef builder = gallivm->builder;
1353    LLVMValueRef res;
1354
1355    if (reg->Register.Indirect) {
1356       LLVMValueRef indirect_index;
1357       LLVMValueRef index_vec;  /* index into the input reg array */
1358       LLVMValueRef inputs_array;
1359       LLVMTypeRef fptr_type;
1360
1361       indirect_index = get_indirect_index(bld,
1362                                           reg->Register.File,
1363                                           reg->Register.Index,
1364                                           &reg->Indirect);
1365
1366       index_vec = get_soa_array_offsets(&bld_base->uint_bld,
1367                                         indirect_index,
1368                                         swizzle,
1369                                         TRUE);
1370
1371       /* cast inputs_array pointer to float* */
1372       fptr_type = LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
1373       inputs_array = LLVMBuildBitCast(builder, bld->inputs_array, fptr_type, "");
1374
1375       /* Gather values from the input register array */
1376       res = build_gather(&bld_base->base, inputs_array, index_vec, NULL);
1377    } else {
1378       if (bld->indirect_files & (1 << TGSI_FILE_INPUT)) {
1379          LLVMValueRef lindex = lp_build_const_int32(gallivm,
1380                                         reg->Register.Index * 4 + swizzle);
1381          LLVMValueRef input_ptr =  LLVMBuildGEP(builder,
1382                                                 bld->inputs_array, &lindex, 1, "");
1383          res = LLVMBuildLoad(builder, input_ptr, "");
1384       }
1385       else {
1386          res = bld->inputs[reg->Register.Index][swizzle];
1387       }
1388    }
1389
1390    assert(res);
1391
1392    if (stype == TGSI_TYPE_UNSIGNED) {
1393       res = LLVMBuildBitCast(builder, res, bld_base->uint_bld.vec_type, "");
1394    } else if (stype == TGSI_TYPE_SIGNED) {
1395       res = LLVMBuildBitCast(builder, res, bld_base->int_bld.vec_type, "");
1396    }
1397
1398    return res;
1399 }
1400
1401
1402 static LLVMValueRef
1403 emit_fetch_gs_input(
1404    struct lp_build_tgsi_context * bld_base,
1405    const struct tgsi_full_src_register * reg,
1406    enum tgsi_opcode_type stype,
1407    unsigned swizzle)
1408 {
1409    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1410    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
1411    const struct tgsi_shader_info *info = bld->bld_base.info;
1412    LLVMBuilderRef builder = gallivm->builder;
1413    LLVMValueRef attrib_index = NULL;
1414    LLVMValueRef vertex_index = NULL;
1415    LLVMValueRef swizzle_index = lp_build_const_int32(gallivm, swizzle);
1416    LLVMValueRef res;
1417
1418    if (info->input_semantic_name[reg->Register.Index] == TGSI_SEMANTIC_PRIMID) {
1419       /* This is really a system value not a regular input */
1420       assert(!reg->Register.Indirect);
1421       assert(!reg->Dimension.Indirect);
1422       res = bld->system_values.prim_id;
1423       if (stype != TGSI_TYPE_UNSIGNED && stype != TGSI_TYPE_SIGNED) {
1424          res = LLVMBuildBitCast(builder, res, bld_base->base.vec_type, "");
1425       }
1426       return res;
1427    }
1428
1429    if (reg->Register.Indirect) {
1430       attrib_index = get_indirect_index(bld,
1431                                         reg->Register.File,
1432                                         reg->Register.Index,
1433                                         &reg->Indirect);
1434    } else {
1435       attrib_index = lp_build_const_int32(gallivm, reg->Register.Index);
1436    }
1437
1438    if (reg->Dimension.Indirect) {
1439       vertex_index = get_indirect_index(bld,
1440                                         reg->Register.File,
1441                                         reg->Dimension.Index,
1442                                         &reg->DimIndirect);
1443    } else {
1444       vertex_index = lp_build_const_int32(gallivm, reg->Dimension.Index);
1445    }
1446
1447    res = bld->gs_iface->fetch_input(bld->gs_iface, bld_base,
1448                                     reg->Dimension.Indirect,
1449                                     vertex_index,
1450                                     reg->Register.Indirect,
1451                                     attrib_index,
1452                                     swizzle_index);
1453
1454    assert(res);
1455
1456    if (stype == TGSI_TYPE_UNSIGNED) {
1457       res = LLVMBuildBitCast(builder, res, bld_base->uint_bld.vec_type, "");
1458    } else if (stype == TGSI_TYPE_SIGNED) {
1459       res = LLVMBuildBitCast(builder, res, bld_base->int_bld.vec_type, "");
1460    }
1461
1462    return res;
1463 }
1464
1465 static LLVMValueRef
1466 emit_fetch_temporary(
1467    struct lp_build_tgsi_context * bld_base,
1468    const struct tgsi_full_src_register * reg,
1469    enum tgsi_opcode_type stype,
1470    unsigned swizzle)
1471 {
1472    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1473    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
1474    LLVMBuilderRef builder = gallivm->builder;
1475    LLVMValueRef res;
1476
1477    if (reg->Register.Indirect) {
1478       LLVMValueRef indirect_index;
1479       LLVMValueRef index_vec;  /* index into the temp reg array */
1480       LLVMValueRef temps_array;
1481       LLVMTypeRef fptr_type;
1482
1483       indirect_index = get_indirect_index(bld,
1484                                           reg->Register.File,
1485                                           reg->Register.Index,
1486                                           &reg->Indirect);
1487
1488       index_vec = get_soa_array_offsets(&bld_base->uint_bld,
1489                                         indirect_index,
1490                                         swizzle,
1491                                         TRUE);
1492
1493       /* cast temps_array pointer to float* */
1494       fptr_type = LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
1495       temps_array = LLVMBuildBitCast(builder, bld->temps_array, fptr_type, "");
1496
1497       /* Gather values from the temporary register array */
1498       res = build_gather(&bld_base->base, temps_array, index_vec, NULL);
1499    }
1500    else {
1501       LLVMValueRef temp_ptr;
1502       temp_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index, swizzle);
1503       res = LLVMBuildLoad(builder, temp_ptr, "");
1504    }
1505
1506    if (stype == TGSI_TYPE_SIGNED || stype == TGSI_TYPE_UNSIGNED) {
1507       struct lp_build_context *bld_fetch = stype_to_fetch(bld_base, stype);
1508       res = LLVMBuildBitCast(builder, res, bld_fetch->vec_type, "");
1509    }
1510
1511    return res;
1512 }
1513
1514 static LLVMValueRef
1515 emit_fetch_system_value(
1516    struct lp_build_tgsi_context * bld_base,
1517    const struct tgsi_full_src_register * reg,
1518    enum tgsi_opcode_type stype,
1519    unsigned swizzle)
1520 {
1521    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1522    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
1523    const struct tgsi_shader_info *info = bld->bld_base.info;
1524    LLVMBuilderRef builder = gallivm->builder;
1525    LLVMValueRef res;
1526    enum tgsi_opcode_type atype; // Actual type of the value
1527
1528    assert(!reg->Register.Indirect);
1529
1530    switch (info->system_value_semantic_name[reg->Register.Index]) {
1531    case TGSI_SEMANTIC_INSTANCEID:
1532       res = lp_build_broadcast_scalar(&bld_base->uint_bld, bld->system_values.instance_id);
1533       atype = TGSI_TYPE_UNSIGNED;
1534       break;
1535
1536    case TGSI_SEMANTIC_VERTEXID:
1537       res = bld->system_values.vertex_id;
1538       atype = TGSI_TYPE_UNSIGNED;
1539       break;
1540
1541    case TGSI_SEMANTIC_VERTEXID_NOBASE:
1542       res = bld->system_values.vertex_id_nobase;
1543       atype = TGSI_TYPE_UNSIGNED;
1544       break;
1545
1546    case TGSI_SEMANTIC_BASEVERTEX:
1547       res = bld->system_values.basevertex;
1548       atype = TGSI_TYPE_UNSIGNED;
1549       break;
1550
1551    case TGSI_SEMANTIC_PRIMID:
1552       res = bld->system_values.prim_id;
1553       atype = TGSI_TYPE_UNSIGNED;
1554       break;
1555
1556    default:
1557       assert(!"unexpected semantic in emit_fetch_system_value");
1558       res = bld_base->base.zero;
1559       atype = TGSI_TYPE_FLOAT;
1560       break;
1561    }
1562
1563    if (atype != stype) {
1564       if (stype == TGSI_TYPE_FLOAT) {
1565          res = LLVMBuildBitCast(builder, res, bld_base->base.vec_type, "");
1566       } else if (stype == TGSI_TYPE_UNSIGNED) {
1567          res = LLVMBuildBitCast(builder, res, bld_base->uint_bld.vec_type, "");
1568       } else if (stype == TGSI_TYPE_SIGNED) {
1569          res = LLVMBuildBitCast(builder, res, bld_base->int_bld.vec_type, "");
1570       }
1571    }
1572
1573    return res;
1574 }
1575
1576 /**
1577  * Register fetch with derivatives.
1578  */
1579 static void
1580 emit_fetch_deriv(
1581    struct lp_build_tgsi_soa_context *bld,
1582    LLVMValueRef src,
1583    LLVMValueRef *res,
1584    LLVMValueRef *ddx,
1585    LLVMValueRef *ddy)
1586 {
1587    if(res)
1588       *res = src;
1589
1590    /* TODO: use interpolation coeffs for inputs */
1591
1592    if(ddx)
1593       *ddx = lp_build_ddx(&bld->bld_base.base, src);
1594
1595    if(ddy)
1596       *ddy = lp_build_ddy(&bld->bld_base.base, src);
1597 }
1598
1599
1600 /**
1601  * Predicate.
1602  */
1603 static void
1604 emit_fetch_predicate(
1605    struct lp_build_tgsi_soa_context *bld,
1606    const struct tgsi_full_instruction *inst,
1607    LLVMValueRef *pred)
1608 {
1609    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
1610    unsigned index;
1611    unsigned char swizzles[4];
1612    LLVMValueRef unswizzled[4] = {NULL, NULL, NULL, NULL};
1613    LLVMValueRef value;
1614    unsigned chan;
1615
1616    if (!inst->Instruction.Predicate) {
1617       TGSI_FOR_EACH_CHANNEL( chan ) {
1618          pred[chan] = NULL;
1619       }
1620       return;
1621    }
1622
1623    swizzles[0] = inst->Predicate.SwizzleX;
1624    swizzles[1] = inst->Predicate.SwizzleY;
1625    swizzles[2] = inst->Predicate.SwizzleZ;
1626    swizzles[3] = inst->Predicate.SwizzleW;
1627
1628    index = inst->Predicate.Index;
1629    assert(index < LP_MAX_TGSI_PREDS);
1630
1631    TGSI_FOR_EACH_CHANNEL( chan ) {
1632       unsigned swizzle = swizzles[chan];
1633
1634       /*
1635        * Only fetch the predicate register channels that are actually listed
1636        * in the swizzles
1637        */
1638       if (!unswizzled[swizzle]) {
1639          value = LLVMBuildLoad(builder,
1640                                bld->preds[index][swizzle], "");
1641
1642          /*
1643           * Convert the value to an integer mask.
1644           *
1645           * TODO: Short-circuit this comparison -- a D3D setp_xx instructions
1646           * is needlessly causing two comparisons due to storing the intermediate
1647           * result as float vector instead of an integer mask vector.
1648           */
1649          value = lp_build_compare(bld->bld_base.base.gallivm,
1650                                   bld->bld_base.base.type,
1651                                   PIPE_FUNC_NOTEQUAL,
1652                                   value,
1653                                   bld->bld_base.base.zero);
1654          if (inst->Predicate.Negate) {
1655             value = LLVMBuildNot(builder, value, "");
1656          }
1657
1658          unswizzled[swizzle] = value;
1659       } else {
1660          value = unswizzled[swizzle];
1661       }
1662
1663       pred[chan] = value;
1664    }
1665 }
1666
1667
1668 /**
1669  * Register store.
1670  */
1671 static void
1672 emit_store_chan(
1673    struct lp_build_tgsi_context *bld_base,
1674    const struct tgsi_full_instruction *inst,
1675    unsigned index,
1676    unsigned chan_index,
1677    LLVMValueRef pred,
1678    LLVMValueRef value)
1679 {
1680    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1681    struct gallivm_state *gallivm = bld_base->base.gallivm;
1682    LLVMBuilderRef builder = gallivm->builder;
1683    const struct tgsi_full_dst_register *reg = &inst->Dst[index];
1684    struct lp_build_context *float_bld = &bld_base->base;
1685    struct lp_build_context *int_bld = &bld_base->int_bld;
1686    LLVMValueRef indirect_index = NULL;
1687    enum tgsi_opcode_type dtype = tgsi_opcode_infer_dst_type(inst->Instruction.Opcode);
1688
1689    /*
1690     * Apply saturation.
1691     *
1692     * It is always assumed to be float.
1693     */
1694    switch( inst->Instruction.Saturate ) {
1695    case TGSI_SAT_NONE:
1696       break;
1697
1698    case TGSI_SAT_ZERO_ONE:
1699       assert(dtype == TGSI_TYPE_FLOAT ||
1700              dtype == TGSI_TYPE_UNTYPED);
1701       value = LLVMBuildBitCast(builder, value, float_bld->vec_type, "");
1702       value = lp_build_clamp_zero_one_nanzero(float_bld, value);
1703       break;
1704
1705    case TGSI_SAT_MINUS_PLUS_ONE:
1706       assert(dtype == TGSI_TYPE_FLOAT ||
1707              dtype == TGSI_TYPE_UNTYPED);
1708       value = LLVMBuildBitCast(builder, value, float_bld->vec_type, "");
1709       /* This will give -1.0 for NaN which is probably not what we want. */
1710       value = lp_build_max_ext(float_bld, value,
1711                                lp_build_const_vec(gallivm, float_bld->type, -1.0),
1712                                GALLIVM_NAN_RETURN_OTHER_SECOND_NONNAN);
1713       value = lp_build_min(float_bld, value, float_bld->one);
1714       break;
1715
1716    default:
1717       assert(0);
1718    }
1719
1720    if (reg->Register.Indirect) {
1721       indirect_index = get_indirect_index(bld,
1722                                           reg->Register.File,
1723                                           reg->Register.Index,
1724                                           &reg->Indirect);
1725    } else {
1726       assert(reg->Register.Index <=
1727                              bld_base->info->file_max[reg->Register.File]);
1728    }
1729
1730    if (DEBUG_EXECUTION) {
1731       emit_dump_reg(gallivm, reg->Register.File, reg->Register.Index, chan_index, value);
1732    }
1733
1734    switch( reg->Register.File ) {
1735    case TGSI_FILE_OUTPUT:
1736       /* Outputs are always stored as floats */
1737       value = LLVMBuildBitCast(builder, value, float_bld->vec_type, "");
1738
1739       if (reg->Register.Indirect) {
1740          LLVMValueRef index_vec;  /* indexes into the output registers */
1741          LLVMValueRef outputs_array;
1742          LLVMTypeRef fptr_type;
1743
1744          index_vec = get_soa_array_offsets(&bld_base->uint_bld,
1745                                            indirect_index,
1746                                            chan_index,
1747                                            TRUE);
1748
1749          fptr_type = LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
1750          outputs_array = LLVMBuildBitCast(builder, bld->outputs_array, fptr_type, "");
1751
1752          /* Scatter store values into output registers */
1753          emit_mask_scatter(bld, outputs_array, index_vec, value,
1754                            &bld->exec_mask, pred);
1755       }
1756       else {
1757          LLVMValueRef out_ptr = lp_get_output_ptr(bld, reg->Register.Index,
1758                                                   chan_index);
1759          lp_exec_mask_store(&bld->exec_mask, float_bld, pred, value, out_ptr);
1760       }
1761       break;
1762
1763    case TGSI_FILE_TEMPORARY:
1764       /* Temporaries are always stored as floats */
1765       value = LLVMBuildBitCast(builder, value, float_bld->vec_type, "");
1766
1767       if (reg->Register.Indirect) {
1768          LLVMValueRef index_vec;  /* indexes into the temp registers */
1769          LLVMValueRef temps_array;
1770          LLVMTypeRef fptr_type;
1771
1772          index_vec = get_soa_array_offsets(&bld_base->uint_bld,
1773                                            indirect_index,
1774                                            chan_index,
1775                                            TRUE);
1776
1777          fptr_type = LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
1778          temps_array = LLVMBuildBitCast(builder, bld->temps_array, fptr_type, "");
1779
1780          /* Scatter store values into temp registers */
1781          emit_mask_scatter(bld, temps_array, index_vec, value,
1782                            &bld->exec_mask, pred);
1783       }
1784       else {
1785          LLVMValueRef temp_ptr;
1786          temp_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index, chan_index);
1787          lp_exec_mask_store(&bld->exec_mask, float_bld, pred, value, temp_ptr);
1788       }
1789       break;
1790
1791    case TGSI_FILE_ADDRESS:
1792       assert(dtype == TGSI_TYPE_SIGNED);
1793       assert(LLVMTypeOf(value) == int_bld->vec_type);
1794       value = LLVMBuildBitCast(builder, value, int_bld->vec_type, "");
1795       lp_exec_mask_store(&bld->exec_mask, int_bld, pred, value,
1796                          bld->addr[reg->Register.Index][chan_index]);
1797       break;
1798
1799    case TGSI_FILE_PREDICATE:
1800       assert(LLVMTypeOf(value) == float_bld->vec_type);
1801       value = LLVMBuildBitCast(builder, value, float_bld->vec_type, "");
1802       lp_exec_mask_store(&bld->exec_mask, float_bld, pred, value,
1803                          bld->preds[reg->Register.Index][chan_index]);
1804       break;
1805
1806    default:
1807       assert( 0 );
1808    }
1809
1810    (void)dtype;
1811 }
1812
1813 /*
1814  * Called at the beginning of the translation of each TGSI instruction, to
1815  * emit some debug code.
1816  */
1817 static void
1818 emit_debug(
1819    struct lp_build_tgsi_context * bld_base,
1820    const struct tgsi_full_instruction * inst,
1821    const struct tgsi_opcode_info * info)
1822
1823 {
1824    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1825
1826    if (DEBUG_EXECUTION) {
1827       /*
1828        * Dump the TGSI instruction.
1829        */
1830
1831       struct gallivm_state *gallivm = bld_base->base.gallivm;
1832       char buf[512];
1833       buf[0] = '$';
1834       buf[1] = ' ';
1835       tgsi_dump_instruction_str(inst, bld_base->pc, &buf[2], sizeof buf - 2);
1836       lp_build_printf(gallivm, buf);
1837
1838       /* Dump the execution mask.
1839        */
1840       if (bld->exec_mask.has_mask) {
1841          lp_build_print_value(gallivm, "    mask = ", bld->exec_mask.exec_mask);
1842       }
1843    }
1844 }
1845
1846 static void
1847 emit_store(
1848    struct lp_build_tgsi_context * bld_base,
1849    const struct tgsi_full_instruction * inst,
1850    const struct tgsi_opcode_info * info,
1851    LLVMValueRef dst[4])
1852
1853 {
1854    unsigned chan_index;
1855    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1856
1857    if(info->num_dst) {
1858       LLVMValueRef pred[TGSI_NUM_CHANNELS];
1859
1860       emit_fetch_predicate( bld, inst, pred );
1861
1862       TGSI_FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1863          emit_store_chan(bld_base, inst, 0, chan_index, pred[chan_index], dst[chan_index]);
1864       }
1865    }
1866 }
1867
1868 static unsigned
1869 tgsi_to_pipe_tex_target(unsigned tgsi_target)
1870 {
1871    switch (tgsi_target) {
1872    case TGSI_TEXTURE_BUFFER:
1873       return PIPE_BUFFER;
1874    case TGSI_TEXTURE_1D:
1875    case TGSI_TEXTURE_SHADOW1D:
1876       return PIPE_TEXTURE_1D;
1877    case TGSI_TEXTURE_2D:
1878    case TGSI_TEXTURE_SHADOW2D:
1879    case TGSI_TEXTURE_2D_MSAA:
1880       return PIPE_TEXTURE_2D;
1881    case TGSI_TEXTURE_3D:
1882       return PIPE_TEXTURE_3D;
1883    case TGSI_TEXTURE_CUBE:
1884    case TGSI_TEXTURE_SHADOWCUBE:
1885       return PIPE_TEXTURE_CUBE;
1886    case TGSI_TEXTURE_RECT:
1887    case TGSI_TEXTURE_SHADOWRECT:
1888       return PIPE_TEXTURE_RECT;
1889    case TGSI_TEXTURE_1D_ARRAY:
1890    case TGSI_TEXTURE_SHADOW1D_ARRAY:
1891       return PIPE_TEXTURE_1D_ARRAY;
1892    case TGSI_TEXTURE_2D_ARRAY:
1893    case TGSI_TEXTURE_SHADOW2D_ARRAY:
1894    case TGSI_TEXTURE_2D_ARRAY_MSAA:
1895       return PIPE_TEXTURE_2D_ARRAY;
1896    case TGSI_TEXTURE_CUBE_ARRAY:
1897    case TGSI_TEXTURE_SHADOWCUBE_ARRAY:
1898       return PIPE_TEXTURE_CUBE_ARRAY;
1899    default:
1900       assert(0);
1901       return PIPE_BUFFER;
1902    }
1903 }
1904
1905
1906 static enum lp_sampler_lod_property
1907 lp_build_lod_property(
1908    struct lp_build_tgsi_context *bld_base,
1909    const struct tgsi_full_instruction *inst,
1910    unsigned src_op)
1911 {
1912    const struct tgsi_full_src_register *reg = &inst->Src[src_op];
1913    enum lp_sampler_lod_property lod_property;
1914
1915    /*
1916     * Not much we can do here. We could try catching inputs declared
1917     * with constant interpolation but not sure it's worth it - since for
1918     * TEX opcodes as well as FETCH/LD the lod comes from same reg as
1919     * the coords, so it could only work for SAMPLE/TXQ/SVIEWINFO), just
1920     * like the constant/immediate recognition below.
1921     * What seems to be of more value would be to recognize temps holding
1922     * broadcasted scalars but no way we can do it.
1923     * Tried asking llvm but without any success (using LLVMIsConstant
1924     * even though this isn't exactly what we'd need), even as simple as
1925     * IMM[0] UINT32 (0,-1,0,0)
1926     * MOV TEMP[0] IMM[0].yyyy
1927     * SVIEWINFO TEMP[1], TEMP[0].xxxx, SVIEWINFO[0]
1928     * doesn't work.
1929     * This means there's ZERO chance this will ever catch a scalar lod
1930     * with traditional tex opcodes as well as texel fetches, since the lod
1931     * comes from the same reg as coords (except some test shaders using
1932     * constant coords maybe).
1933     * There's at least hope for sample opcodes as well as size queries.
1934     */
1935    if (reg->Register.File == TGSI_FILE_CONSTANT ||
1936        reg->Register.File == TGSI_FILE_IMMEDIATE) {
1937       lod_property = LP_SAMPLER_LOD_SCALAR;
1938    }
1939    else if (bld_base->info->processor == TGSI_PROCESSOR_FRAGMENT) {
1940       if (gallivm_debug & GALLIVM_DEBUG_NO_QUAD_LOD) {
1941          lod_property = LP_SAMPLER_LOD_PER_ELEMENT;
1942       }
1943       else {
1944          lod_property = LP_SAMPLER_LOD_PER_QUAD;
1945       }
1946    }
1947    else {
1948       /* never use scalar (per-quad) lod the results are just too wrong. */
1949       lod_property = LP_SAMPLER_LOD_PER_ELEMENT;
1950    }
1951    return lod_property;
1952 }
1953
1954
1955 /**
1956  * High-level instruction translators.
1957  */
1958
1959 static void
1960 emit_tex( struct lp_build_tgsi_soa_context *bld,
1961           const struct tgsi_full_instruction *inst,
1962           enum lp_build_tex_modifier modifier,
1963           LLVMValueRef *texel,
1964           unsigned sampler_reg)
1965 {
1966    unsigned unit = inst->Src[sampler_reg].Register.Index;
1967    LLVMValueRef lod_bias, explicit_lod;
1968    LLVMValueRef oow = NULL;
1969    LLVMValueRef coords[5];
1970    LLVMValueRef offsets[3] = { NULL };
1971    struct lp_derivatives derivs;
1972    struct lp_derivatives *deriv_ptr = NULL;
1973    enum lp_sampler_lod_property lod_property = LP_SAMPLER_LOD_SCALAR;
1974    unsigned num_derivs, num_offsets, i;
1975    unsigned shadow_coord = 0;
1976    unsigned layer_coord = 0;
1977
1978    if (!bld->sampler) {
1979       _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
1980       for (i = 0; i < 4; i++) {
1981          texel[i] = bld->bld_base.base.undef;
1982       }
1983       return;
1984    }
1985
1986    switch (inst->Texture.Texture) {
1987    case TGSI_TEXTURE_1D_ARRAY:
1988       layer_coord = 1;
1989       /* fallthrough */
1990    case TGSI_TEXTURE_1D:
1991       num_offsets = 1;
1992       num_derivs = 1;
1993       break;
1994    case TGSI_TEXTURE_2D_ARRAY:
1995       layer_coord = 2;
1996       /* fallthrough */
1997    case TGSI_TEXTURE_2D:
1998    case TGSI_TEXTURE_RECT:
1999       num_offsets = 2;
2000       num_derivs = 2;
2001       break;
2002    case TGSI_TEXTURE_SHADOW1D_ARRAY:
2003       layer_coord = 1;
2004       /* fallthrough */
2005    case TGSI_TEXTURE_SHADOW1D:
2006       shadow_coord = 2;
2007       num_offsets = 1;
2008       num_derivs = 1;
2009       break;
2010    case TGSI_TEXTURE_SHADOW2D_ARRAY:
2011       layer_coord = 2;
2012       shadow_coord = 3;
2013       num_offsets = 2;
2014       num_derivs = 2;
2015       break;
2016    case TGSI_TEXTURE_SHADOW2D:
2017    case TGSI_TEXTURE_SHADOWRECT:
2018       shadow_coord = 2;
2019       num_offsets = 2;
2020       num_derivs = 2;
2021       break;
2022    case TGSI_TEXTURE_CUBE:
2023       num_offsets = 2;
2024       num_derivs = 3;
2025       break;
2026    case TGSI_TEXTURE_3D:
2027       num_offsets = 3;
2028       num_derivs = 3;
2029       break;
2030    case TGSI_TEXTURE_SHADOWCUBE:
2031       shadow_coord = 3;
2032       num_offsets = 2;
2033       num_derivs = 3;
2034       break;
2035    case TGSI_TEXTURE_CUBE_ARRAY:
2036       num_offsets = 2;
2037       num_derivs = 3;
2038       layer_coord = 3;
2039       break;
2040    case TGSI_TEXTURE_SHADOWCUBE_ARRAY:
2041       num_offsets = 2;
2042       num_derivs = 3;
2043       layer_coord = 3;
2044       shadow_coord = 4; /* shadow coord special different reg */
2045       break;
2046    case TGSI_TEXTURE_2D_MSAA:
2047    case TGSI_TEXTURE_2D_ARRAY_MSAA:
2048    default:
2049       assert(0);
2050       return;
2051    }
2052
2053    /* Note lod and especially projected are illegal in a LOT of cases */
2054    if (modifier == LP_BLD_TEX_MODIFIER_LOD_BIAS ||
2055        modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_LOD) {
2056       LLVMValueRef lod;
2057       if (inst->Texture.Texture == TGSI_TEXTURE_SHADOWCUBE ||
2058           inst->Texture.Texture == TGSI_TEXTURE_CUBE_ARRAY) {
2059          /* note that shadow cube array with bias/explicit lod does not exist */
2060          lod = lp_build_emit_fetch(&bld->bld_base, inst, 1, 0);
2061       }
2062       else {
2063          lod = lp_build_emit_fetch(&bld->bld_base, inst, 0, 3);
2064       }
2065       if (modifier == LP_BLD_TEX_MODIFIER_LOD_BIAS) {
2066          lod_bias = lod;
2067          explicit_lod = NULL;
2068       }
2069       else if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_LOD) {
2070          lod_bias = NULL;
2071          explicit_lod = lod;
2072       }
2073       lod_property = lp_build_lod_property(&bld->bld_base, inst, 0);
2074    }
2075    else {
2076       lod_bias = NULL;
2077       explicit_lod = NULL;
2078    }
2079
2080    if (modifier == LP_BLD_TEX_MODIFIER_PROJECTED) {
2081       oow = lp_build_emit_fetch(&bld->bld_base, inst, 0, 3);
2082       oow = lp_build_rcp(&bld->bld_base.base, oow);
2083    }
2084
2085    for (i = 0; i < num_derivs; i++) {
2086       coords[i] = lp_build_emit_fetch(&bld->bld_base, inst, 0, i);
2087       if (modifier == LP_BLD_TEX_MODIFIER_PROJECTED)
2088          coords[i] = lp_build_mul(&bld->bld_base.base, coords[i], oow);
2089    }
2090    for (i = num_derivs; i < 5; i++) {
2091       coords[i] = bld->bld_base.base.undef;
2092    }
2093
2094    /* Layer coord always goes into 3rd slot, except for cube map arrays */
2095    if (layer_coord) {
2096       if (layer_coord == 3) {
2097          coords[3] = lp_build_emit_fetch(&bld->bld_base, inst, 0, layer_coord);
2098       }
2099       else {
2100          coords[2] = lp_build_emit_fetch(&bld->bld_base, inst, 0, layer_coord);
2101       }
2102       if (modifier == LP_BLD_TEX_MODIFIER_PROJECTED)
2103          coords[2] = lp_build_mul(&bld->bld_base.base, coords[2], oow);
2104    }
2105    /* Shadow coord occupies always 5th slot. */
2106    if (shadow_coord) {
2107       if (shadow_coord == 4) {
2108          coords[4] = lp_build_emit_fetch(&bld->bld_base, inst, 1, 0);
2109       }
2110       else {
2111          coords[4] = lp_build_emit_fetch(&bld->bld_base, inst, 0, shadow_coord);
2112       }
2113       if (modifier == LP_BLD_TEX_MODIFIER_PROJECTED)
2114          coords[4] = lp_build_mul(&bld->bld_base.base, coords[4], oow);
2115    }
2116
2117    if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV) {
2118       unsigned dim;
2119       for (dim = 0; dim < num_derivs; ++dim) {
2120          derivs.ddx[dim] = lp_build_emit_fetch(&bld->bld_base, inst, 1, dim);
2121          derivs.ddy[dim] = lp_build_emit_fetch(&bld->bld_base, inst, 2, dim);
2122       }
2123       deriv_ptr = &derivs;
2124       /*
2125        * could also check all src regs if constant but I doubt such
2126        * cases exist in practice.
2127        */
2128       if (bld->bld_base.info->processor == TGSI_PROCESSOR_FRAGMENT) {
2129          if (gallivm_debug & GALLIVM_DEBUG_NO_QUAD_LOD) {
2130             lod_property = LP_SAMPLER_LOD_PER_ELEMENT;
2131          }
2132          else {
2133             lod_property = LP_SAMPLER_LOD_PER_QUAD;
2134          }
2135       }
2136       else {
2137          lod_property = LP_SAMPLER_LOD_PER_ELEMENT;
2138       }
2139    }
2140
2141    /* some advanced gather instructions (txgo) would require 4 offsets */
2142    if (inst->Texture.NumOffsets == 1) {
2143       unsigned dim;
2144       for (dim = 0; dim < num_offsets; dim++) {
2145          offsets[dim] = lp_build_emit_fetch_texoffset(&bld->bld_base, inst, 0, dim);
2146       }
2147    }
2148
2149    bld->sampler->emit_fetch_texel(bld->sampler,
2150                                   bld->bld_base.base.gallivm,
2151                                   bld->bld_base.base.type,
2152                                   FALSE,
2153                                   unit, unit,
2154                                   bld->context_ptr,
2155                                   coords,
2156                                   offsets,
2157                                   deriv_ptr,
2158                                   lod_bias, explicit_lod, lod_property,
2159                                   texel);
2160 }
2161
2162 static void
2163 emit_sample(struct lp_build_tgsi_soa_context *bld,
2164             const struct tgsi_full_instruction *inst,
2165             enum lp_build_tex_modifier modifier,
2166             boolean compare,
2167             LLVMValueRef *texel)
2168 {
2169    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
2170    unsigned texture_unit, sampler_unit;
2171    LLVMValueRef lod_bias, explicit_lod;
2172    LLVMValueRef coords[5];
2173    LLVMValueRef offsets[3] = { NULL };
2174    struct lp_derivatives derivs;
2175    struct lp_derivatives *deriv_ptr = NULL;
2176    enum lp_sampler_lod_property lod_property = LP_SAMPLER_LOD_SCALAR;
2177
2178    unsigned num_offsets, num_derivs, i;
2179    unsigned layer_coord = 0;
2180
2181    if (!bld->sampler) {
2182       _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
2183       for (i = 0; i < 4; i++) {
2184          texel[i] = bld->bld_base.base.undef;
2185       }
2186       return;
2187    }
2188
2189    /*
2190     * unlike old-style tex opcodes the texture/sampler indices
2191     * always come from src1 and src2 respectively.
2192     */
2193    texture_unit = inst->Src[1].Register.Index;
2194    sampler_unit = inst->Src[2].Register.Index;
2195
2196    /*
2197     * Note inst->Texture.Texture will contain the number of offsets,
2198     * however the target information is NOT there and comes from the
2199     * declared sampler views instead.
2200     */
2201    switch (bld->sv[texture_unit].Resource) {
2202    case TGSI_TEXTURE_1D:
2203       num_offsets = 1;
2204       num_derivs = 1;
2205       break;
2206    case TGSI_TEXTURE_1D_ARRAY:
2207       layer_coord = 1;
2208       num_offsets = 1;
2209       num_derivs = 1;
2210       break;
2211    case TGSI_TEXTURE_2D:
2212    case TGSI_TEXTURE_RECT:
2213       num_offsets = 2;
2214       num_derivs = 2;
2215       break;
2216    case TGSI_TEXTURE_2D_ARRAY:
2217       layer_coord = 2;
2218       num_offsets = 2;
2219       num_derivs = 2;
2220       break;
2221    case TGSI_TEXTURE_CUBE:
2222       num_offsets = 2;
2223       num_derivs = 3;
2224       break;
2225    case TGSI_TEXTURE_3D:
2226       num_offsets = 3;
2227       num_derivs = 3;
2228       break;
2229    case TGSI_TEXTURE_CUBE_ARRAY:
2230       layer_coord = 3;
2231       num_offsets = 2;
2232       num_derivs = 3;
2233       break;
2234    default:
2235       assert(0);
2236       return;
2237    }
2238
2239    if (modifier == LP_BLD_TEX_MODIFIER_LOD_BIAS ||
2240        modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_LOD) {
2241       LLVMValueRef lod = lp_build_emit_fetch(&bld->bld_base, inst, 3, 0);
2242       if (modifier == LP_BLD_TEX_MODIFIER_LOD_BIAS) {
2243          lod_bias = lod;
2244          explicit_lod = NULL;
2245       }
2246       else if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_LOD) {
2247          lod_bias = NULL;
2248          explicit_lod = lod;
2249       }
2250       lod_property = lp_build_lod_property(&bld->bld_base, inst, 0);
2251    }
2252    else if (modifier == LP_BLD_TEX_MODIFIER_LOD_ZERO) {
2253       lod_bias = NULL;
2254       /* XXX might be better to explicitly pass the level zero information */
2255       explicit_lod = lp_build_const_vec(gallivm, bld->bld_base.base.type, 0.0F);
2256    }
2257    else {
2258       lod_bias = NULL;
2259       explicit_lod = NULL;
2260    }
2261
2262    for (i = 0; i < num_derivs; i++) {
2263       coords[i] = lp_build_emit_fetch(&bld->bld_base, inst, 0, i);
2264    }
2265    for (i = num_derivs; i < 5; i++) {
2266       coords[i] = bld->bld_base.base.undef;
2267    }
2268
2269    /* Layer coord always goes into 3rd slot, except for cube map arrays */
2270    if (layer_coord) {
2271       if (layer_coord == 3)
2272          coords[3] = lp_build_emit_fetch(&bld->bld_base, inst, 0, layer_coord);
2273       else
2274          coords[2] = lp_build_emit_fetch(&bld->bld_base, inst, 0, layer_coord);
2275    }
2276    /* Shadow coord occupies always 5th slot. */
2277    if (compare) {
2278       coords[4] = lp_build_emit_fetch(&bld->bld_base, inst, 3, 0);
2279    }
2280
2281    if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV) {
2282       unsigned dim;
2283       for (dim = 0; dim < num_derivs; ++dim) {
2284          derivs.ddx[dim] = lp_build_emit_fetch(&bld->bld_base, inst, 3, dim);
2285          derivs.ddy[dim] = lp_build_emit_fetch(&bld->bld_base, inst, 4, dim);
2286       }
2287       deriv_ptr = &derivs;
2288       /*
2289        * could also check all src regs if constant but I doubt such
2290        * cases exist in practice.
2291        */
2292       if (bld->bld_base.info->processor == TGSI_PROCESSOR_FRAGMENT) {
2293          if (gallivm_debug & GALLIVM_DEBUG_NO_QUAD_LOD) {
2294             lod_property = LP_SAMPLER_LOD_PER_ELEMENT;
2295          }
2296          else {
2297             lod_property = LP_SAMPLER_LOD_PER_QUAD;
2298          }
2299       }
2300       else {
2301          lod_property = LP_SAMPLER_LOD_PER_ELEMENT;
2302       }
2303    }
2304
2305    /* some advanced gather instructions (txgo) would require 4 offsets */
2306    if (inst->Texture.NumOffsets == 1) {
2307       unsigned dim;
2308       for (dim = 0; dim < num_offsets; dim++) {
2309          offsets[dim] = lp_build_emit_fetch_texoffset(&bld->bld_base, inst, 0, dim);
2310       }
2311    }
2312
2313    bld->sampler->emit_fetch_texel(bld->sampler,
2314                                   bld->bld_base.base.gallivm,
2315                                   bld->bld_base.base.type,
2316                                   FALSE,
2317                                   texture_unit, sampler_unit,
2318                                   bld->context_ptr,
2319                                   coords,
2320                                   offsets,
2321                                   deriv_ptr,
2322                                   lod_bias, explicit_lod, lod_property,
2323                                   texel);
2324
2325    if (inst->Src[1].Register.SwizzleX != PIPE_SWIZZLE_RED ||
2326        inst->Src[1].Register.SwizzleY != PIPE_SWIZZLE_GREEN ||
2327        inst->Src[1].Register.SwizzleZ != PIPE_SWIZZLE_BLUE ||
2328        inst->Src[1].Register.SwizzleW != PIPE_SWIZZLE_ALPHA) {
2329       unsigned char swizzles[4];
2330       swizzles[0] = inst->Src[1].Register.SwizzleX;
2331       swizzles[1] = inst->Src[1].Register.SwizzleY;
2332       swizzles[2] = inst->Src[1].Register.SwizzleZ;
2333       swizzles[3] = inst->Src[1].Register.SwizzleW;
2334
2335       lp_build_swizzle_soa_inplace(&bld->bld_base.base, texel, swizzles);
2336    }
2337 }
2338
2339 static void
2340 emit_fetch_texels( struct lp_build_tgsi_soa_context *bld,
2341                    const struct tgsi_full_instruction *inst,
2342                    LLVMValueRef *texel,
2343                    boolean is_samplei)
2344 {
2345    unsigned unit, target;
2346    LLVMValueRef coord_undef = LLVMGetUndef(bld->bld_base.base.int_vec_type);
2347    LLVMValueRef explicit_lod = NULL;
2348    LLVMValueRef coords[5];
2349    LLVMValueRef offsets[3] = { NULL };
2350    enum lp_sampler_lod_property lod_property = LP_SAMPLER_LOD_SCALAR;
2351    unsigned dims, i;
2352    unsigned layer_coord = 0;
2353
2354    if (!bld->sampler) {
2355       _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
2356       for (i = 0; i < 4; i++) {
2357          texel[i] = coord_undef;
2358       }
2359       return;
2360    }
2361
2362    unit = inst->Src[1].Register.Index;
2363
2364    if (is_samplei) {
2365       target = bld->sv[unit].Resource;
2366    }
2367    else {
2368       target = inst->Texture.Texture;
2369    }
2370
2371    switch (target) {
2372    case TGSI_TEXTURE_1D:
2373    case TGSI_TEXTURE_BUFFER:
2374       dims = 1;
2375       break;
2376    case TGSI_TEXTURE_1D_ARRAY:
2377       layer_coord = 1;
2378       dims = 1;
2379       break;
2380    case TGSI_TEXTURE_2D:
2381    case TGSI_TEXTURE_RECT:
2382    case TGSI_TEXTURE_2D_MSAA:
2383       dims = 2;
2384       break;
2385    case TGSI_TEXTURE_2D_ARRAY:
2386    case TGSI_TEXTURE_2D_ARRAY_MSAA:
2387       layer_coord = 2;
2388       dims = 2;
2389       break;
2390    case TGSI_TEXTURE_3D:
2391       dims = 3;
2392       break;
2393    default:
2394       assert(0);
2395       return;
2396    }
2397
2398    /* always have lod except for buffers and msaa targets ? */
2399    if (target != TGSI_TEXTURE_BUFFER &&
2400        target != TGSI_TEXTURE_2D_MSAA &&
2401        target != TGSI_TEXTURE_2D_ARRAY_MSAA) {
2402       explicit_lod = lp_build_emit_fetch(&bld->bld_base, inst, 0, 3);
2403       lod_property = lp_build_lod_property(&bld->bld_base, inst, 0);
2404    }
2405    /* XXX: for real msaa support, the w component would be the sample index. */
2406
2407    for (i = 0; i < dims; i++) {
2408       coords[i] = lp_build_emit_fetch(&bld->bld_base, inst, 0, i);
2409    }
2410    /* never use more than 3 coords here but emit_fetch_texel copies all 5 anyway */
2411    for (i = dims; i < 5; i++) {
2412       coords[i] = coord_undef;
2413    }
2414    if (layer_coord)
2415       coords[2] = lp_build_emit_fetch(&bld->bld_base, inst, 0, layer_coord);
2416
2417    if (inst->Texture.NumOffsets == 1) {
2418       unsigned dim;
2419       for (dim = 0; dim < dims; dim++) {
2420          offsets[dim] = lp_build_emit_fetch_texoffset(&bld->bld_base, inst, 0, dim);
2421       }
2422    }
2423
2424    bld->sampler->emit_fetch_texel(bld->sampler,
2425                                   bld->bld_base.base.gallivm,
2426                                   bld->bld_base.base.type,
2427                                   TRUE,
2428                                   unit, unit,
2429                                   bld->context_ptr,
2430                                   coords,
2431                                   offsets,
2432                                   NULL,
2433                                   NULL, explicit_lod, lod_property,
2434                                   texel);
2435
2436    if (is_samplei &&
2437        (inst->Src[1].Register.SwizzleX != PIPE_SWIZZLE_RED ||
2438         inst->Src[1].Register.SwizzleY != PIPE_SWIZZLE_GREEN ||
2439         inst->Src[1].Register.SwizzleZ != PIPE_SWIZZLE_BLUE ||
2440         inst->Src[1].Register.SwizzleW != PIPE_SWIZZLE_ALPHA)) {
2441       unsigned char swizzles[4];
2442       swizzles[0] = inst->Src[1].Register.SwizzleX;
2443       swizzles[1] = inst->Src[1].Register.SwizzleY;
2444       swizzles[2] = inst->Src[1].Register.SwizzleZ;
2445       swizzles[3] = inst->Src[1].Register.SwizzleW;
2446
2447       lp_build_swizzle_soa_inplace(&bld->bld_base.base, texel, swizzles);
2448    }
2449 }
2450
2451 static void
2452 emit_size_query( struct lp_build_tgsi_soa_context *bld,
2453                  const struct tgsi_full_instruction *inst,
2454                  LLVMValueRef *sizes_out,
2455                  boolean is_sviewinfo)
2456 {
2457    LLVMValueRef explicit_lod;
2458    enum lp_sampler_lod_property lod_property;
2459    unsigned has_lod;
2460    unsigned i;
2461    unsigned unit = inst->Src[1].Register.Index;
2462    unsigned target, pipe_target;
2463
2464    if (is_sviewinfo) {
2465       target = bld->sv[unit].Resource;
2466    }
2467    else {
2468       target = inst->Texture.Texture;
2469    }
2470    switch (target) {
2471    case TGSI_TEXTURE_BUFFER:
2472    case TGSI_TEXTURE_RECT:
2473    case TGSI_TEXTURE_SHADOWRECT:
2474       has_lod = 0;
2475       break;
2476    default:
2477       has_lod = 1;
2478       break;
2479    }
2480
2481    if (!bld->sampler) {
2482       _debug_printf("warning: found texture query instruction but no sampler generator supplied\n");
2483       for (i = 0; i < 4; i++)
2484          sizes_out[i] = bld->bld_base.int_bld.undef;
2485       return;
2486    }
2487
2488    if (has_lod) {
2489       explicit_lod = lp_build_emit_fetch(&bld->bld_base, inst, 0, 0);
2490       lod_property = lp_build_lod_property(&bld->bld_base, inst, 0);
2491    }
2492    else {
2493       explicit_lod = NULL;
2494       lod_property = LP_SAMPLER_LOD_SCALAR;
2495    }
2496
2497
2498    pipe_target = tgsi_to_pipe_tex_target(target);
2499
2500    bld->sampler->emit_size_query(bld->sampler,
2501                                  bld->bld_base.base.gallivm,
2502                                  bld->bld_base.int_bld.type,
2503                                  unit, pipe_target,
2504                                  bld->context_ptr,
2505                                  TRUE,
2506                                  lod_property,
2507                                  explicit_lod,
2508                                  sizes_out);
2509 }
2510
2511 static boolean
2512 near_end_of_shader(struct lp_build_tgsi_soa_context *bld,
2513                    int pc)
2514 {
2515    int i;
2516
2517    for (i = 0; i < 5; i++) {
2518       unsigned opcode;
2519
2520       if (pc + i >= bld->bld_base.info->num_instructions)
2521          return TRUE;
2522
2523       opcode = bld->bld_base.instructions[pc + i].Instruction.Opcode;
2524
2525       if (opcode == TGSI_OPCODE_END)
2526          return TRUE;
2527
2528       if (opcode == TGSI_OPCODE_TEX ||
2529          opcode == TGSI_OPCODE_TXP ||
2530          opcode == TGSI_OPCODE_TXD ||
2531          opcode == TGSI_OPCODE_TXB ||
2532          opcode == TGSI_OPCODE_TXL ||
2533          opcode == TGSI_OPCODE_TXF ||
2534          opcode == TGSI_OPCODE_TXQ ||
2535          opcode == TGSI_OPCODE_TEX2 ||
2536          opcode == TGSI_OPCODE_TXB2 ||
2537          opcode == TGSI_OPCODE_TXL2 ||
2538          opcode == TGSI_OPCODE_SAMPLE ||
2539          opcode == TGSI_OPCODE_SAMPLE_B ||
2540          opcode == TGSI_OPCODE_SAMPLE_C ||
2541          opcode == TGSI_OPCODE_SAMPLE_C_LZ ||
2542          opcode == TGSI_OPCODE_SAMPLE_D ||
2543          opcode == TGSI_OPCODE_SAMPLE_I ||
2544          opcode == TGSI_OPCODE_SAMPLE_L ||
2545          opcode == TGSI_OPCODE_SVIEWINFO ||
2546          opcode == TGSI_OPCODE_CAL ||
2547          opcode == TGSI_OPCODE_CALLNZ ||
2548          opcode == TGSI_OPCODE_IF ||
2549          opcode == TGSI_OPCODE_UIF ||
2550          opcode == TGSI_OPCODE_BGNLOOP ||
2551          opcode == TGSI_OPCODE_SWITCH)
2552          return FALSE;
2553    }
2554
2555    return TRUE;
2556 }
2557
2558
2559
2560 /**
2561  * Kill fragment if any of the src register values are negative.
2562  */
2563 static void
2564 emit_kill_if(
2565    struct lp_build_tgsi_soa_context *bld,
2566    const struct tgsi_full_instruction *inst,
2567    int pc)
2568 {
2569    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
2570    const struct tgsi_full_src_register *reg = &inst->Src[0];
2571    LLVMValueRef terms[TGSI_NUM_CHANNELS];
2572    LLVMValueRef mask;
2573    unsigned chan_index;
2574
2575    memset(&terms, 0, sizeof terms);
2576
2577    TGSI_FOR_EACH_CHANNEL( chan_index ) {
2578       unsigned swizzle;
2579
2580       /* Unswizzle channel */
2581       swizzle = tgsi_util_get_full_src_register_swizzle( reg, chan_index );
2582
2583       /* Check if the component has not been already tested. */
2584       assert(swizzle < TGSI_NUM_CHANNELS);
2585       if( !terms[swizzle] )
2586          /* TODO: change the comparison operator instead of setting the sign */
2587          terms[swizzle] =  lp_build_emit_fetch(&bld->bld_base, inst, 0, chan_index );
2588    }
2589
2590    mask = NULL;
2591    TGSI_FOR_EACH_CHANNEL( chan_index ) {
2592       if(terms[chan_index]) {
2593          LLVMValueRef chan_mask;
2594
2595          /*
2596           * If term < 0 then mask = 0 else mask = ~0.
2597           */
2598          chan_mask = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_GEQUAL, terms[chan_index], bld->bld_base.base.zero);
2599
2600          if(mask)
2601             mask = LLVMBuildAnd(builder, mask, chan_mask, "");
2602          else
2603             mask = chan_mask;
2604       }
2605    }
2606
2607    if (bld->exec_mask.has_mask) {
2608       LLVMValueRef invmask;
2609       invmask = LLVMBuildNot(builder, bld->exec_mask.exec_mask, "kilp");
2610       mask = LLVMBuildOr(builder, mask, invmask, "");
2611    }
2612
2613    lp_build_mask_update(bld->mask, mask);
2614    if (!near_end_of_shader(bld, pc))
2615       lp_build_mask_check(bld->mask);
2616 }
2617
2618
2619 /**
2620  * Unconditional fragment kill.
2621  * The only predication is the execution mask which will apply if
2622  * we're inside a loop or conditional.
2623  */
2624 static void
2625 emit_kill(struct lp_build_tgsi_soa_context *bld,
2626           int pc)
2627 {
2628    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
2629    LLVMValueRef mask;
2630
2631    /* For those channels which are "alive", disable fragment shader
2632     * execution.
2633     */
2634    if (bld->exec_mask.has_mask) {
2635       mask = LLVMBuildNot(builder, bld->exec_mask.exec_mask, "kilp");
2636    }
2637    else {
2638       LLVMValueRef zero = LLVMConstNull(bld->bld_base.base.int_vec_type);
2639       mask = zero;
2640    }
2641
2642    lp_build_mask_update(bld->mask, mask);
2643
2644    if (!near_end_of_shader(bld, pc))
2645       lp_build_mask_check(bld->mask);
2646 }
2647
2648
2649 /**
2650  * Emit code which will dump the value of all the temporary registers
2651  * to stdout.
2652  */
2653 static void
2654 emit_dump_file(struct lp_build_tgsi_soa_context *bld,
2655                unsigned file)
2656 {
2657    const struct tgsi_shader_info *info = bld->bld_base.info;
2658    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
2659    LLVMBuilderRef builder = gallivm->builder;
2660    LLVMValueRef reg_ptr;
2661    int index;
2662    int max_index = info->file_max[file];
2663
2664    /*
2665     * Some register files, particularly constants, can be very large,
2666     * and dumping everything could make this unusably slow.
2667     */
2668    max_index = MIN2(max_index, 32);
2669
2670    for (index = 0; index <= max_index; index++) {
2671       LLVMValueRef res;
2672       unsigned mask;
2673       int chan;
2674
2675       if (index < 8 * sizeof(unsigned) &&
2676           (info->file_mask[file] & (1 << index)) == 0)  {
2677          /* This was not declared.*/
2678          continue;
2679       }
2680
2681       if (file == TGSI_FILE_INPUT) {
2682          mask = info->input_usage_mask[index];
2683       } else {
2684          mask = TGSI_WRITEMASK_XYZW;
2685       }
2686
2687       for (chan = 0; chan < 4; chan++) {
2688          if ((mask & (1 << chan)) == 0) {
2689             /* This channel is not used.*/
2690             continue;
2691          }
2692
2693          if (file == TGSI_FILE_CONSTANT) {
2694             struct tgsi_full_src_register reg;
2695             memset(&reg, 0, sizeof reg);
2696             reg.Register.File = file;
2697             reg.Register.Index = index;
2698             reg.Register.SwizzleX = 0;
2699             reg.Register.SwizzleY = 1;
2700             reg.Register.SwizzleZ = 2;
2701             reg.Register.SwizzleW = 3;
2702
2703             res = bld->bld_base.emit_fetch_funcs[file](&bld->bld_base, &reg, TGSI_TYPE_FLOAT, chan);
2704             if (!res) {
2705                continue;
2706             }
2707          } else if (file == TGSI_FILE_INPUT) {
2708             res = bld->inputs[index][chan];
2709             if (!res) {
2710                continue;
2711             }
2712          } else if (file == TGSI_FILE_TEMPORARY) {
2713             reg_ptr = lp_get_temp_ptr_soa(bld, index, chan);
2714             assert(reg_ptr);
2715             res = LLVMBuildLoad(builder, reg_ptr, "");
2716          } else if (file == TGSI_FILE_OUTPUT) {
2717             reg_ptr = lp_get_output_ptr(bld, index, chan);
2718             assert(reg_ptr);
2719             res = LLVMBuildLoad(builder, reg_ptr, "");
2720          } else {
2721             assert(0);
2722             continue;
2723          }
2724
2725          emit_dump_reg(gallivm, file, index, chan, res);
2726       }
2727    }
2728 }
2729
2730
2731
2732 void
2733 lp_emit_declaration_soa(
2734    struct lp_build_tgsi_context *bld_base,
2735    const struct tgsi_full_declaration *decl)
2736 {
2737    struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
2738    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
2739    LLVMTypeRef vec_type = bld->bld_base.base.vec_type;
2740    const unsigned first = decl->Range.First;
2741    const unsigned last = decl->Range.Last;
2742    unsigned idx, i;
2743
2744    assert(last <= bld->bld_base.info->file_max[decl->Declaration.File]);
2745
2746    switch (decl->Declaration.File) {
2747    case TGSI_FILE_TEMPORARY:
2748       if (!(bld->indirect_files & (1 << TGSI_FILE_TEMPORARY))) {
2749          assert(last < LP_MAX_INLINED_TEMPS);
2750          for (idx = first; idx <= last; ++idx) {
2751             for (i = 0; i < TGSI_NUM_CHANNELS; i++)
2752                bld->temps[idx][i] = lp_build_alloca(gallivm, vec_type, "temp");
2753          }
2754       }
2755       break;
2756
2757    case TGSI_FILE_OUTPUT:
2758       if (!(bld->indirect_files & (1 << TGSI_FILE_OUTPUT))) {
2759          for (idx = first; idx <= last; ++idx) {
2760             for (i = 0; i < TGSI_NUM_CHANNELS; i++)
2761                bld->outputs[idx][i] = lp_build_alloca(gallivm,
2762                                                       vec_type, "output");
2763          }
2764       }
2765       break;
2766
2767    case TGSI_FILE_ADDRESS:
2768       /* ADDR registers are only allocated with an integer LLVM IR type,
2769        * as they are guaranteed to always have integers.
2770        * XXX: Not sure if this exception is worthwhile (or the whole idea of
2771        * an ADDR register for that matter).
2772        */
2773       assert(last < LP_MAX_TGSI_ADDRS);
2774       for (idx = first; idx <= last; ++idx) {
2775          assert(idx < LP_MAX_TGSI_ADDRS);
2776          for (i = 0; i < TGSI_NUM_CHANNELS; i++)
2777             bld->addr[idx][i] = lp_build_alloca(gallivm, bld_base->base.int_vec_type, "addr");
2778       }
2779       break;
2780
2781    case TGSI_FILE_PREDICATE:
2782       assert(last < LP_MAX_TGSI_PREDS);
2783       for (idx = first; idx <= last; ++idx) {
2784          for (i = 0; i < TGSI_NUM_CHANNELS; i++)
2785             bld->preds[idx][i] = lp_build_alloca(gallivm, vec_type,
2786                                                  "predicate");
2787       }
2788       break;
2789
2790    case TGSI_FILE_SAMPLER_VIEW:
2791       /*
2792        * The target stored here MUST match whatever there actually
2793        * is in the set sampler views (what about return type?).
2794        */
2795       assert(last < PIPE_MAX_SHADER_SAMPLER_VIEWS);
2796       for (idx = first; idx <= last; ++idx) {
2797          bld->sv[idx] = decl->SamplerView;
2798       }
2799       break;
2800
2801    case TGSI_FILE_CONSTANT:
2802    {
2803       /*
2804        * We could trivially fetch the per-buffer pointer when fetching the
2805        * constant, relying on llvm to figure out it's always the same pointer
2806        * anyway. However, doing so results in a huge (more than factor of 10)
2807        * slowdown in llvm compilation times for some (but not all) shaders
2808        * (more specifically, the IR optimization spends way more time in
2809        * DominatorTree::dominates). At least with llvm versions 3.1, 3.3.
2810        */
2811       unsigned idx2D = decl->Dim.Index2D;
2812       LLVMValueRef index2D = lp_build_const_int32(gallivm, idx2D);
2813       assert(idx2D < LP_MAX_TGSI_CONST_BUFFERS);
2814       bld->consts[idx2D] =
2815          lp_build_array_get(gallivm, bld->consts_ptr, index2D);
2816       bld->consts_sizes[idx2D] =
2817          lp_build_array_get(gallivm, bld->const_sizes_ptr, index2D);
2818    }
2819       break;
2820
2821    default:
2822       /* don't need to declare other vars */
2823       break;
2824    }
2825 }
2826
2827
2828 void lp_emit_immediate_soa(
2829    struct lp_build_tgsi_context *bld_base,
2830    const struct tgsi_full_immediate *imm)
2831 {
2832    struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
2833    struct gallivm_state * gallivm = bld_base->base.gallivm;
2834    LLVMValueRef imms[4];
2835    unsigned i;
2836    const uint size = imm->Immediate.NrTokens - 1;
2837    assert(size <= 4);
2838    switch (imm->Immediate.DataType) {
2839    case TGSI_IMM_FLOAT32:
2840       for( i = 0; i < size; ++i )
2841          imms[i] =
2842                lp_build_const_vec(gallivm, bld_base->base.type, imm->u[i].Float);
2843
2844       break;
2845    case TGSI_IMM_UINT32:
2846       for( i = 0; i < size; ++i ) {
2847          LLVMValueRef tmp = lp_build_const_vec(gallivm, bld_base->uint_bld.type, imm->u[i].Uint);
2848          imms[i] = LLVMConstBitCast(tmp, bld_base->base.vec_type);
2849       }
2850
2851       break;
2852    case TGSI_IMM_INT32:
2853       for( i = 0; i < size; ++i ) {
2854          LLVMValueRef tmp = lp_build_const_vec(gallivm, bld_base->int_bld.type, imm->u[i].Int);
2855          imms[i] = LLVMConstBitCast(tmp, bld_base->base.vec_type);
2856       }
2857
2858       break;
2859    }
2860    for( i = size; i < 4; ++i )
2861       imms[i] = bld_base->base.undef;
2862
2863    if (bld->use_immediates_array) {
2864       unsigned index = bld->num_immediates;
2865       struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
2866       LLVMBuilderRef builder = gallivm->builder;
2867
2868       assert(bld->indirect_files & (1 << TGSI_FILE_IMMEDIATE));
2869       for (i = 0; i < 4; ++i ) {
2870          LLVMValueRef lindex = lp_build_const_int32(
2871                   bld->bld_base.base.gallivm, index * 4 + i);
2872          LLVMValueRef imm_ptr = LLVMBuildGEP(builder,
2873                                              bld->imms_array, &lindex, 1, "");
2874          LLVMBuildStore(builder, imms[i], imm_ptr);
2875       }
2876    } else {
2877       /* simply copy the immediate values into the next immediates[] slot */
2878       unsigned i;
2879       const uint size = imm->Immediate.NrTokens - 1;
2880       assert(size <= 4);
2881       assert(bld->num_immediates < LP_MAX_INLINED_IMMEDIATES);
2882
2883       for(i = 0; i < 4; ++i )
2884          bld->immediates[bld->num_immediates][i] = imms[i];
2885
2886       if (bld->indirect_files & (1 << TGSI_FILE_IMMEDIATE)) {
2887          unsigned index = bld->num_immediates;
2888          struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
2889          LLVMBuilderRef builder = gallivm->builder;
2890          for (i = 0; i < 4; ++i ) {
2891             LLVMValueRef lindex = lp_build_const_int32(
2892                      bld->bld_base.base.gallivm, index * 4 + i);
2893             LLVMValueRef imm_ptr = LLVMBuildGEP(builder,
2894                                                 bld->imms_array, &lindex, 1, "");
2895             LLVMBuildStore(builder,
2896                            bld->immediates[index][i],
2897                            imm_ptr);
2898          }
2899       }
2900    }
2901
2902    bld->num_immediates++;
2903 }
2904
2905 static void
2906 ddx_emit(
2907    const struct lp_build_tgsi_action * action,
2908    struct lp_build_tgsi_context * bld_base,
2909    struct lp_build_emit_data * emit_data)
2910 {
2911    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2912
2913    emit_fetch_deriv(bld, emit_data->args[0], NULL,
2914                     &emit_data->output[emit_data->chan], NULL);
2915 }
2916
2917 static void
2918 ddy_emit(
2919    const struct lp_build_tgsi_action * action,
2920    struct lp_build_tgsi_context * bld_base,
2921    struct lp_build_emit_data * emit_data)
2922 {
2923    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2924
2925    emit_fetch_deriv(bld, emit_data->args[0], NULL, NULL,
2926                     &emit_data->output[emit_data->chan]);
2927 }
2928
2929 static void
2930 kill_emit(
2931    const struct lp_build_tgsi_action * action,
2932    struct lp_build_tgsi_context * bld_base,
2933    struct lp_build_emit_data * emit_data)
2934 {
2935    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2936
2937    emit_kill(bld, bld_base->pc - 1);
2938 }
2939
2940 static void
2941 kill_if_emit(
2942    const struct lp_build_tgsi_action * action,
2943    struct lp_build_tgsi_context * bld_base,
2944    struct lp_build_emit_data * emit_data)
2945 {
2946    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2947
2948    emit_kill_if(bld, emit_data->inst, bld_base->pc - 1);
2949 }
2950
2951 static void
2952 tex_emit(
2953    const struct lp_build_tgsi_action * action,
2954    struct lp_build_tgsi_context * bld_base,
2955    struct lp_build_emit_data * emit_data)
2956 {
2957    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2958
2959    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_NONE,
2960             emit_data->output, 1);
2961 }
2962
2963 static void
2964 tex2_emit(
2965    const struct lp_build_tgsi_action * action,
2966    struct lp_build_tgsi_context * bld_base,
2967    struct lp_build_emit_data * emit_data)
2968 {
2969    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2970
2971    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_NONE,
2972             emit_data->output, 2);
2973 }
2974
2975 static void
2976 txb_emit(
2977    const struct lp_build_tgsi_action * action,
2978    struct lp_build_tgsi_context * bld_base,
2979    struct lp_build_emit_data * emit_data)
2980 {
2981    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2982
2983    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_LOD_BIAS,
2984             emit_data->output, 1);
2985 }
2986
2987 static void
2988 txb2_emit(
2989    const struct lp_build_tgsi_action * action,
2990    struct lp_build_tgsi_context * bld_base,
2991    struct lp_build_emit_data * emit_data)
2992 {
2993    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2994
2995    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_LOD_BIAS,
2996             emit_data->output, 2);
2997 }
2998
2999 static void
3000 txd_emit(
3001    const struct lp_build_tgsi_action * action,
3002    struct lp_build_tgsi_context * bld_base,
3003    struct lp_build_emit_data * emit_data)
3004 {
3005    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3006
3007    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV,
3008             emit_data->output, 3);
3009 }
3010
3011 static void
3012 txl_emit(
3013    const struct lp_build_tgsi_action * action,
3014    struct lp_build_tgsi_context * bld_base,
3015    struct lp_build_emit_data * emit_data)
3016 {
3017    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3018
3019    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD,
3020             emit_data->output, 1);
3021 }
3022
3023 static void
3024 txl2_emit(
3025    const struct lp_build_tgsi_action * action,
3026    struct lp_build_tgsi_context * bld_base,
3027    struct lp_build_emit_data * emit_data)
3028 {
3029    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3030
3031    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD,
3032             emit_data->output, 2);
3033 }
3034
3035 static void
3036 txp_emit(
3037    const struct lp_build_tgsi_action * action,
3038    struct lp_build_tgsi_context * bld_base,
3039    struct lp_build_emit_data * emit_data)
3040 {
3041    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3042
3043    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_PROJECTED,
3044             emit_data->output, 1);
3045 }
3046
3047 static void
3048 txq_emit(
3049    const struct lp_build_tgsi_action * action,
3050    struct lp_build_tgsi_context * bld_base,
3051    struct lp_build_emit_data * emit_data)
3052 {
3053    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3054
3055    emit_size_query(bld, emit_data->inst, emit_data->output, FALSE);
3056 }
3057
3058 static void
3059 txf_emit(
3060    const struct lp_build_tgsi_action * action,
3061    struct lp_build_tgsi_context * bld_base,
3062    struct lp_build_emit_data * emit_data)
3063 {
3064    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3065
3066    emit_fetch_texels(bld, emit_data->inst, emit_data->output, FALSE);
3067 }
3068
3069 static void
3070 sample_i_emit(
3071    const struct lp_build_tgsi_action * action,
3072    struct lp_build_tgsi_context * bld_base,
3073    struct lp_build_emit_data * emit_data)
3074 {
3075    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3076
3077    emit_fetch_texels(bld, emit_data->inst, emit_data->output, TRUE);
3078 }
3079
3080 static void
3081 sample_emit(
3082    const struct lp_build_tgsi_action * action,
3083    struct lp_build_tgsi_context * bld_base,
3084    struct lp_build_emit_data * emit_data)
3085 {
3086    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3087
3088    emit_sample(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_NONE,
3089                FALSE, emit_data->output);
3090 }
3091
3092 static void
3093 sample_b_emit(
3094    const struct lp_build_tgsi_action * action,
3095    struct lp_build_tgsi_context * bld_base,
3096    struct lp_build_emit_data * emit_data)
3097 {
3098    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3099
3100    emit_sample(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_LOD_BIAS,
3101                FALSE, emit_data->output);
3102 }
3103
3104 static void
3105 sample_c_emit(
3106    const struct lp_build_tgsi_action * action,
3107    struct lp_build_tgsi_context * bld_base,
3108    struct lp_build_emit_data * emit_data)
3109 {
3110    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3111
3112    emit_sample(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_NONE,
3113                TRUE, emit_data->output);
3114 }
3115
3116 static void
3117 sample_c_lz_emit(
3118    const struct lp_build_tgsi_action * action,
3119    struct lp_build_tgsi_context * bld_base,
3120    struct lp_build_emit_data * emit_data)
3121 {
3122    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3123
3124    emit_sample(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_LOD_ZERO,
3125                TRUE, emit_data->output);
3126 }
3127
3128 static void
3129 sample_d_emit(
3130    const struct lp_build_tgsi_action * action,
3131    struct lp_build_tgsi_context * bld_base,
3132    struct lp_build_emit_data * emit_data)
3133 {
3134    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3135
3136    emit_sample(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV,
3137                FALSE, emit_data->output);
3138 }
3139
3140 static void
3141 sample_l_emit(
3142    const struct lp_build_tgsi_action * action,
3143    struct lp_build_tgsi_context * bld_base,
3144    struct lp_build_emit_data * emit_data)
3145 {
3146    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3147
3148    emit_sample(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD,
3149                FALSE, emit_data->output);
3150 }
3151
3152 static void
3153 sviewinfo_emit(
3154    const struct lp_build_tgsi_action * action,
3155    struct lp_build_tgsi_context * bld_base,
3156    struct lp_build_emit_data * emit_data)
3157 {
3158    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3159
3160    emit_size_query(bld, emit_data->inst, emit_data->output, TRUE);
3161 }
3162
3163 static LLVMValueRef
3164 mask_vec(struct lp_build_tgsi_context *bld_base)
3165 {
3166    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3167    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
3168    struct lp_exec_mask *exec_mask = &bld->exec_mask;
3169
3170    if (!exec_mask->has_mask) {
3171       return lp_build_mask_value(bld->mask);
3172    }
3173    return LLVMBuildAnd(builder, lp_build_mask_value(bld->mask),
3174                        exec_mask->exec_mask, "");
3175 }
3176
3177 static void
3178 increment_vec_ptr_by_mask(struct lp_build_tgsi_context * bld_base,
3179                           LLVMValueRef ptr,
3180                           LLVMValueRef mask)
3181 {
3182    LLVMBuilderRef builder = bld_base->base.gallivm->builder;
3183    LLVMValueRef current_vec = LLVMBuildLoad(builder, ptr, "");
3184
3185    current_vec = LLVMBuildSub(builder, current_vec, mask, "");
3186
3187    LLVMBuildStore(builder, current_vec, ptr);
3188 }
3189
3190 static void
3191 clear_uint_vec_ptr_from_mask(struct lp_build_tgsi_context * bld_base,
3192                              LLVMValueRef ptr,
3193                              LLVMValueRef mask)
3194 {
3195    LLVMBuilderRef builder = bld_base->base.gallivm->builder;
3196    LLVMValueRef current_vec = LLVMBuildLoad(builder, ptr, "");
3197
3198    current_vec = lp_build_select(&bld_base->uint_bld,
3199                                  mask,
3200                                  bld_base->uint_bld.zero,
3201                                  current_vec);
3202
3203    LLVMBuildStore(builder, current_vec, ptr);
3204 }
3205
3206 static LLVMValueRef
3207 clamp_mask_to_max_output_vertices(struct lp_build_tgsi_soa_context * bld,
3208                                   LLVMValueRef current_mask_vec,
3209                                   LLVMValueRef total_emitted_vertices_vec)
3210 {
3211    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
3212    struct lp_build_context *int_bld = &bld->bld_base.int_bld;
3213    LLVMValueRef max_mask = lp_build_cmp(int_bld, PIPE_FUNC_LESS,
3214                                         total_emitted_vertices_vec,
3215                                         bld->max_output_vertices_vec);
3216
3217    return LLVMBuildAnd(builder, current_mask_vec, max_mask, "");
3218 }
3219
3220 static void
3221 emit_vertex(
3222    const struct lp_build_tgsi_action * action,
3223    struct lp_build_tgsi_context * bld_base,
3224    struct lp_build_emit_data * emit_data)
3225 {
3226    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3227    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
3228
3229    if (bld->gs_iface->emit_vertex) {
3230       LLVMValueRef mask = mask_vec(bld_base);
3231       LLVMValueRef total_emitted_vertices_vec =
3232          LLVMBuildLoad(builder, bld->total_emitted_vertices_vec_ptr, "");
3233       mask = clamp_mask_to_max_output_vertices(bld, mask,
3234                                                total_emitted_vertices_vec);
3235       gather_outputs(bld);
3236       bld->gs_iface->emit_vertex(bld->gs_iface, &bld->bld_base,
3237                                  bld->outputs,
3238                                  total_emitted_vertices_vec);
3239       increment_vec_ptr_by_mask(bld_base, bld->emitted_vertices_vec_ptr,
3240                                 mask);
3241       increment_vec_ptr_by_mask(bld_base, bld->total_emitted_vertices_vec_ptr,
3242                                 mask);
3243 #if DUMP_GS_EMITS
3244       lp_build_print_value(bld->bld_base.base.gallivm,
3245                            " +++ emit vertex masked ones = ",
3246                            mask);
3247       lp_build_print_value(bld->bld_base.base.gallivm,
3248                            " +++ emit vertex emitted = ",
3249                            total_emitted_vertices_vec);
3250 #endif
3251    }
3252 }
3253
3254
3255 static void
3256 end_primitive_masked(struct lp_build_tgsi_context * bld_base,
3257                      LLVMValueRef mask)
3258 {
3259    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3260    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
3261
3262    if (bld->gs_iface->end_primitive) {
3263       struct lp_build_context *uint_bld = &bld_base->uint_bld;
3264       LLVMValueRef emitted_vertices_vec =
3265          LLVMBuildLoad(builder, bld->emitted_vertices_vec_ptr, "");
3266       LLVMValueRef emitted_prims_vec =
3267          LLVMBuildLoad(builder, bld->emitted_prims_vec_ptr, "");
3268
3269       LLVMValueRef emitted_mask = lp_build_cmp(uint_bld, PIPE_FUNC_NOTEQUAL,
3270                                                emitted_vertices_vec,
3271                                                uint_bld->zero);
3272       /* We need to combine the current execution mask with the mask
3273          telling us which, if any, execution slots actually have
3274          unemitted primitives, this way we make sure that end_primitives
3275          executes only on the paths that have unflushed vertices */
3276       mask = LLVMBuildAnd(builder, mask, emitted_mask, "");
3277
3278       bld->gs_iface->end_primitive(bld->gs_iface, &bld->bld_base,
3279                                    emitted_vertices_vec,
3280                                    emitted_prims_vec);
3281
3282 #if DUMP_GS_EMITS
3283       lp_build_print_value(bld->bld_base.base.gallivm,
3284                            " +++ end prim masked ones = ",
3285                            mask);
3286       lp_build_print_value(bld->bld_base.base.gallivm,
3287                            " +++ end prim emitted verts1 = ",
3288                            emitted_vertices_vec);
3289       lp_build_print_value(bld->bld_base.base.gallivm,
3290                            " +++ end prim emitted prims1 = ",
3291                            LLVMBuildLoad(builder,
3292                                          bld->emitted_prims_vec_ptr, ""));
3293 #endif
3294       increment_vec_ptr_by_mask(bld_base, bld->emitted_prims_vec_ptr,
3295                                 mask);
3296       clear_uint_vec_ptr_from_mask(bld_base, bld->emitted_vertices_vec_ptr,
3297                                    mask);
3298 #if DUMP_GS_EMITS
3299       lp_build_print_value(bld->bld_base.base.gallivm,
3300                            " +++ end prim emitted verts2 = ",
3301                            LLVMBuildLoad(builder,
3302                                          bld->emitted_vertices_vec_ptr, ""));
3303 #endif
3304    }
3305
3306 }
3307
3308 static void
3309 end_primitive(
3310    const struct lp_build_tgsi_action * action,
3311    struct lp_build_tgsi_context * bld_base,
3312    struct lp_build_emit_data * emit_data)
3313 {
3314    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3315
3316    if (bld->gs_iface->end_primitive) {
3317       LLVMValueRef mask = mask_vec(bld_base);
3318       end_primitive_masked(bld_base, mask);
3319    }
3320 }
3321
3322 static void
3323 cal_emit(
3324    const struct lp_build_tgsi_action * action,
3325    struct lp_build_tgsi_context * bld_base,
3326    struct lp_build_emit_data * emit_data)
3327 {
3328    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3329
3330    lp_exec_mask_call(&bld->exec_mask, emit_data->inst->Label.Label,
3331                      &bld_base->pc);
3332 }
3333
3334 static void
3335 ret_emit(
3336    const struct lp_build_tgsi_action * action,
3337    struct lp_build_tgsi_context * bld_base,
3338    struct lp_build_emit_data * emit_data)
3339 {
3340    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3341
3342    lp_exec_mask_ret(&bld->exec_mask, &bld_base->pc);
3343 }
3344
3345 static void
3346 brk_emit(
3347    const struct lp_build_tgsi_action * action,
3348    struct lp_build_tgsi_context * bld_base,
3349    struct lp_build_emit_data * emit_data)
3350 {
3351    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3352
3353    lp_exec_break(&bld->exec_mask, bld_base);
3354 }
3355
3356 static void
3357 breakc_emit(
3358    const struct lp_build_tgsi_action * action,
3359    struct lp_build_tgsi_context * bld_base,
3360    struct lp_build_emit_data * emit_data)
3361 {
3362    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3363    LLVMBuilderRef builder = bld_base->base.gallivm->builder;
3364    struct lp_build_context *uint_bld = &bld_base->uint_bld;
3365    LLVMValueRef unsigned_cond =
3366       LLVMBuildBitCast(builder, emit_data->args[0], uint_bld->vec_type, "");
3367    LLVMValueRef cond = lp_build_cmp(uint_bld, PIPE_FUNC_NOTEQUAL,
3368                                     unsigned_cond,
3369                                     uint_bld->zero);
3370
3371    lp_exec_break_condition(&bld->exec_mask, cond);
3372 }
3373
3374 static void
3375 if_emit(
3376    const struct lp_build_tgsi_action * action,
3377    struct lp_build_tgsi_context * bld_base,
3378    struct lp_build_emit_data * emit_data)
3379 {
3380    LLVMValueRef tmp;
3381    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3382
3383    tmp = lp_build_cmp(&bld_base->base, PIPE_FUNC_NOTEQUAL,
3384                       emit_data->args[0], bld->bld_base.base.zero);
3385    lp_exec_mask_cond_push(&bld->exec_mask, tmp);
3386 }
3387
3388 static void
3389 uif_emit(
3390    const struct lp_build_tgsi_action * action,
3391    struct lp_build_tgsi_context * bld_base,
3392    struct lp_build_emit_data * emit_data)
3393 {
3394    LLVMValueRef tmp;
3395    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3396    struct lp_build_context *uint_bld = &bld_base->uint_bld;
3397
3398    tmp = lp_build_cmp(uint_bld, PIPE_FUNC_NOTEQUAL,
3399                       emit_data->args[0], uint_bld->zero);
3400    lp_exec_mask_cond_push(&bld->exec_mask, tmp);
3401 }
3402
3403 static void
3404 case_emit(
3405    const struct lp_build_tgsi_action * action,
3406    struct lp_build_tgsi_context * bld_base,
3407    struct lp_build_emit_data * emit_data)
3408 {
3409    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3410
3411    lp_exec_case(&bld->exec_mask, emit_data->args[0]);
3412 }
3413
3414 static void
3415 default_emit(
3416    const struct lp_build_tgsi_action * action,
3417    struct lp_build_tgsi_context * bld_base,
3418    struct lp_build_emit_data * emit_data)
3419 {
3420    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3421
3422    lp_exec_default(&bld->exec_mask, bld_base);
3423 }
3424
3425 static void
3426 switch_emit(
3427    const struct lp_build_tgsi_action * action,
3428    struct lp_build_tgsi_context * bld_base,
3429    struct lp_build_emit_data * emit_data)
3430 {
3431    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3432
3433    lp_exec_switch(&bld->exec_mask, emit_data->args[0]);
3434 }
3435
3436 static void
3437 endswitch_emit(
3438    const struct lp_build_tgsi_action * action,
3439    struct lp_build_tgsi_context * bld_base,
3440    struct lp_build_emit_data * emit_data)
3441 {
3442    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3443
3444    lp_exec_endswitch(&bld->exec_mask, bld_base);
3445 }
3446
3447 static void
3448 bgnloop_emit(
3449    const struct lp_build_tgsi_action * action,
3450    struct lp_build_tgsi_context * bld_base,
3451    struct lp_build_emit_data * emit_data)
3452 {
3453    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3454
3455    lp_exec_bgnloop(&bld->exec_mask);
3456 }
3457
3458 static void
3459 bgnsub_emit(
3460    const struct lp_build_tgsi_action * action,
3461    struct lp_build_tgsi_context * bld_base,
3462    struct lp_build_emit_data * emit_data)
3463 {
3464    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3465
3466    lp_exec_mask_bgnsub(&bld->exec_mask);
3467 }
3468
3469 static void
3470 else_emit(
3471    const struct lp_build_tgsi_action * action,
3472    struct lp_build_tgsi_context * bld_base,
3473    struct lp_build_emit_data * emit_data)
3474 {
3475    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3476
3477    lp_exec_mask_cond_invert(&bld->exec_mask);
3478 }
3479
3480 static void
3481 endif_emit(
3482    const struct lp_build_tgsi_action * action,
3483    struct lp_build_tgsi_context * bld_base,
3484    struct lp_build_emit_data * emit_data)
3485 {
3486    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3487
3488    lp_exec_mask_cond_pop(&bld->exec_mask);
3489 }
3490
3491 static void
3492 endloop_emit(
3493    const struct lp_build_tgsi_action * action,
3494    struct lp_build_tgsi_context * bld_base,
3495    struct lp_build_emit_data * emit_data)
3496 {
3497    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3498
3499    lp_exec_endloop(bld_base->base.gallivm, &bld->exec_mask);
3500 }
3501
3502 static void
3503 endsub_emit(
3504    const struct lp_build_tgsi_action * action,
3505    struct lp_build_tgsi_context * bld_base,
3506    struct lp_build_emit_data * emit_data)
3507 {
3508    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3509
3510    lp_exec_mask_endsub(&bld->exec_mask, &bld_base->pc);
3511 }
3512
3513 static void
3514 cont_emit(
3515    const struct lp_build_tgsi_action * action,
3516    struct lp_build_tgsi_context * bld_base,
3517    struct lp_build_emit_data * emit_data)
3518 {
3519    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3520
3521    lp_exec_continue(&bld->exec_mask);
3522 }
3523
3524 static void emit_prologue(struct lp_build_tgsi_context * bld_base)
3525 {
3526    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3527    struct gallivm_state * gallivm = bld_base->base.gallivm;
3528
3529    if (bld->indirect_files & (1 << TGSI_FILE_TEMPORARY)) {
3530       LLVMValueRef array_size =
3531          lp_build_const_int32(gallivm,
3532                          bld_base->info->file_max[TGSI_FILE_TEMPORARY] * 4 + 4);
3533       bld->temps_array = lp_build_array_alloca(gallivm,
3534                                               bld_base->base.vec_type, array_size,
3535                                               "temp_array");
3536    }
3537
3538    if (bld->indirect_files & (1 << TGSI_FILE_OUTPUT)) {
3539       LLVMValueRef array_size =
3540          lp_build_const_int32(gallivm,
3541                             bld_base->info->file_max[TGSI_FILE_OUTPUT] * 4 + 4);
3542       bld->outputs_array = lp_build_array_alloca(gallivm,
3543                                                 bld_base->base.vec_type, array_size,
3544                                                 "output_array");
3545    }
3546
3547    if (bld->indirect_files & (1 << TGSI_FILE_IMMEDIATE)) {
3548       LLVMValueRef array_size =
3549          lp_build_const_int32(gallivm,
3550                          bld_base->info->file_max[TGSI_FILE_IMMEDIATE] * 4 + 4);
3551       bld->imms_array = lp_build_array_alloca(gallivm,
3552                                               bld_base->base.vec_type, array_size,
3553                                               "imms_array");
3554    }
3555
3556    /* If we have indirect addressing in inputs we need to copy them into
3557     * our alloca array to be able to iterate over them */
3558    if (bld->indirect_files & (1 << TGSI_FILE_INPUT) && !bld->gs_iface) {
3559       unsigned index, chan;
3560       LLVMTypeRef vec_type = bld_base->base.vec_type;
3561       LLVMValueRef array_size = lp_build_const_int32(gallivm,
3562             bld_base->info->file_max[TGSI_FILE_INPUT]*4 + 4);
3563       bld->inputs_array = lp_build_array_alloca(gallivm,
3564                                                vec_type, array_size,
3565                                                "input_array");
3566
3567       assert(bld_base->info->num_inputs
3568                         <= bld_base->info->file_max[TGSI_FILE_INPUT] + 1);
3569
3570       for (index = 0; index < bld_base->info->num_inputs; ++index) {
3571          for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
3572             LLVMValueRef lindex =
3573                lp_build_const_int32(gallivm, index * 4 + chan);
3574             LLVMValueRef input_ptr =
3575                LLVMBuildGEP(gallivm->builder, bld->inputs_array,
3576                             &lindex, 1, "");
3577             LLVMValueRef value = bld->inputs[index][chan];
3578             if (value)
3579                LLVMBuildStore(gallivm->builder, value, input_ptr);
3580          }
3581       }
3582    }
3583
3584    if (bld->gs_iface) {
3585       struct lp_build_context *uint_bld = &bld->bld_base.uint_bld;
3586       bld->emitted_prims_vec_ptr =
3587          lp_build_alloca(gallivm,
3588                          uint_bld->vec_type,
3589                          "emitted_prims_ptr");
3590       bld->emitted_vertices_vec_ptr =
3591          lp_build_alloca(gallivm,
3592                          uint_bld->vec_type,
3593                          "emitted_vertices_ptr");
3594       bld->total_emitted_vertices_vec_ptr =
3595          lp_build_alloca(gallivm,
3596                          uint_bld->vec_type,
3597                          "total_emitted_vertices_ptr");
3598
3599       LLVMBuildStore(gallivm->builder, uint_bld->zero,
3600                      bld->emitted_prims_vec_ptr);
3601       LLVMBuildStore(gallivm->builder, uint_bld->zero,
3602                      bld->emitted_vertices_vec_ptr);
3603       LLVMBuildStore(gallivm->builder, uint_bld->zero,
3604                      bld->total_emitted_vertices_vec_ptr);
3605    }
3606
3607    if (DEBUG_EXECUTION) {
3608       lp_build_printf(gallivm, "\n");
3609       emit_dump_file(bld, TGSI_FILE_CONSTANT);
3610       if (!bld->gs_iface)
3611          emit_dump_file(bld, TGSI_FILE_INPUT);
3612    }
3613 }
3614
3615 static void emit_epilogue(struct lp_build_tgsi_context * bld_base)
3616 {
3617    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3618    LLVMBuilderRef builder = bld_base->base.gallivm->builder;
3619
3620    if (DEBUG_EXECUTION) {
3621       /* for debugging */
3622       if (0) {
3623          emit_dump_file(bld, TGSI_FILE_TEMPORARY);
3624       }
3625       emit_dump_file(bld, TGSI_FILE_OUTPUT);
3626       lp_build_printf(bld_base->base.gallivm, "\n");
3627    }
3628
3629    /* If we have indirect addressing in outputs we need to copy our alloca array
3630     * to the outputs slots specified by the caller */
3631    if (bld->gs_iface) {
3632       LLVMValueRef total_emitted_vertices_vec;
3633       LLVMValueRef emitted_prims_vec;
3634       /* implicit end_primitives, needed in case there are any unflushed
3635          vertices in the cache. Note must not call end_primitive here
3636          since the exec_mask is not valid at this point. */
3637       end_primitive_masked(bld_base, lp_build_mask_value(bld->mask));
3638
3639       total_emitted_vertices_vec =
3640          LLVMBuildLoad(builder, bld->total_emitted_vertices_vec_ptr, "");
3641       emitted_prims_vec =
3642          LLVMBuildLoad(builder, bld->emitted_prims_vec_ptr, "");
3643
3644       bld->gs_iface->gs_epilogue(bld->gs_iface,
3645                                  &bld->bld_base,
3646                                  total_emitted_vertices_vec,
3647                                  emitted_prims_vec);
3648    } else {
3649       gather_outputs(bld);
3650    }
3651 }
3652
3653 void
3654 lp_build_tgsi_soa(struct gallivm_state *gallivm,
3655                   const struct tgsi_token *tokens,
3656                   struct lp_type type,
3657                   struct lp_build_mask_context *mask,
3658                   LLVMValueRef consts_ptr,
3659                   LLVMValueRef const_sizes_ptr,
3660                   const struct lp_bld_tgsi_system_values *system_values,
3661                   const LLVMValueRef (*inputs)[TGSI_NUM_CHANNELS],
3662                   LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
3663                   LLVMValueRef context_ptr,
3664                   struct lp_build_sampler_soa *sampler,
3665                   const struct tgsi_shader_info *info,
3666                   const struct lp_build_tgsi_gs_iface *gs_iface)
3667 {
3668    struct lp_build_tgsi_soa_context bld;
3669
3670    struct lp_type res_type;
3671
3672    assert(type.length <= LP_MAX_VECTOR_LENGTH);
3673    memset(&res_type, 0, sizeof res_type);
3674    res_type.width = type.width;
3675    res_type.length = type.length;
3676    res_type.sign = 1;
3677
3678    /* Setup build context */
3679    memset(&bld, 0, sizeof bld);
3680    lp_build_context_init(&bld.bld_base.base, gallivm, type);
3681    lp_build_context_init(&bld.bld_base.uint_bld, gallivm, lp_uint_type(type));
3682    lp_build_context_init(&bld.bld_base.int_bld, gallivm, lp_int_type(type));
3683    lp_build_context_init(&bld.elem_bld, gallivm, lp_elem_type(type));
3684    bld.mask = mask;
3685    bld.inputs = inputs;
3686    bld.outputs = outputs;
3687    bld.consts_ptr = consts_ptr;
3688    bld.const_sizes_ptr = const_sizes_ptr;
3689    bld.sampler = sampler;
3690    bld.bld_base.info = info;
3691    bld.indirect_files = info->indirect_files;
3692    bld.context_ptr = context_ptr;
3693
3694    /*
3695     * If the number of temporaries is rather large then we just
3696     * allocate them as an array right from the start and treat
3697     * like indirect temporaries.
3698     */
3699    if (info->file_max[TGSI_FILE_TEMPORARY] >= LP_MAX_INLINED_TEMPS) {
3700       bld.indirect_files |= (1 << TGSI_FILE_TEMPORARY);
3701    }
3702    /*
3703     * For performance reason immediates are always backed in a static
3704     * array, but if their number is too great, we have to use just
3705     * a dynamically allocated array.
3706     */
3707    bld.use_immediates_array =
3708          (info->file_max[TGSI_FILE_IMMEDIATE] >= LP_MAX_INLINED_IMMEDIATES);
3709    if (bld.use_immediates_array) {
3710       bld.indirect_files |= (1 << TGSI_FILE_IMMEDIATE);
3711    }
3712
3713
3714    bld.bld_base.soa = TRUE;
3715    bld.bld_base.emit_debug = emit_debug;
3716    bld.bld_base.emit_fetch_funcs[TGSI_FILE_CONSTANT] = emit_fetch_constant;
3717    bld.bld_base.emit_fetch_funcs[TGSI_FILE_IMMEDIATE] = emit_fetch_immediate;
3718    bld.bld_base.emit_fetch_funcs[TGSI_FILE_INPUT] = emit_fetch_input;
3719    bld.bld_base.emit_fetch_funcs[TGSI_FILE_TEMPORARY] = emit_fetch_temporary;
3720    bld.bld_base.emit_fetch_funcs[TGSI_FILE_SYSTEM_VALUE] = emit_fetch_system_value;
3721    bld.bld_base.emit_store = emit_store;
3722
3723    bld.bld_base.emit_declaration = lp_emit_declaration_soa;
3724    bld.bld_base.emit_immediate = lp_emit_immediate_soa;
3725
3726    bld.bld_base.emit_prologue = emit_prologue;
3727    bld.bld_base.emit_epilogue = emit_epilogue;
3728
3729    /* Set opcode actions */
3730    lp_set_default_actions_cpu(&bld.bld_base);
3731
3732    bld.bld_base.op_actions[TGSI_OPCODE_BGNLOOP].emit = bgnloop_emit;
3733    bld.bld_base.op_actions[TGSI_OPCODE_BGNSUB].emit = bgnsub_emit;
3734    bld.bld_base.op_actions[TGSI_OPCODE_BRK].emit = brk_emit;
3735    bld.bld_base.op_actions[TGSI_OPCODE_BREAKC].emit = breakc_emit;
3736    bld.bld_base.op_actions[TGSI_OPCODE_CAL].emit = cal_emit;
3737    bld.bld_base.op_actions[TGSI_OPCODE_CASE].emit = case_emit;
3738    bld.bld_base.op_actions[TGSI_OPCODE_CONT].emit = cont_emit;
3739    bld.bld_base.op_actions[TGSI_OPCODE_DDX].emit = ddx_emit;
3740    bld.bld_base.op_actions[TGSI_OPCODE_DDY].emit = ddy_emit;
3741    bld.bld_base.op_actions[TGSI_OPCODE_DEFAULT].emit = default_emit;
3742    bld.bld_base.op_actions[TGSI_OPCODE_ELSE].emit = else_emit;
3743    bld.bld_base.op_actions[TGSI_OPCODE_ENDIF].emit = endif_emit;
3744    bld.bld_base.op_actions[TGSI_OPCODE_ENDLOOP].emit = endloop_emit;
3745    bld.bld_base.op_actions[TGSI_OPCODE_ENDSUB].emit = endsub_emit;
3746    bld.bld_base.op_actions[TGSI_OPCODE_ENDSWITCH].emit = endswitch_emit;
3747    bld.bld_base.op_actions[TGSI_OPCODE_IF].emit = if_emit;
3748    bld.bld_base.op_actions[TGSI_OPCODE_UIF].emit = uif_emit;
3749    bld.bld_base.op_actions[TGSI_OPCODE_KILL_IF].emit = kill_if_emit;
3750    bld.bld_base.op_actions[TGSI_OPCODE_KILL].emit = kill_emit;
3751    bld.bld_base.op_actions[TGSI_OPCODE_RET].emit = ret_emit;
3752    bld.bld_base.op_actions[TGSI_OPCODE_SWITCH].emit = switch_emit;
3753    bld.bld_base.op_actions[TGSI_OPCODE_TEX].emit = tex_emit;
3754    bld.bld_base.op_actions[TGSI_OPCODE_TXB].emit = txb_emit;
3755    bld.bld_base.op_actions[TGSI_OPCODE_TXD].emit = txd_emit;
3756    bld.bld_base.op_actions[TGSI_OPCODE_TXL].emit = txl_emit;
3757    bld.bld_base.op_actions[TGSI_OPCODE_TXP].emit = txp_emit;
3758    bld.bld_base.op_actions[TGSI_OPCODE_TXQ].emit = txq_emit;
3759    bld.bld_base.op_actions[TGSI_OPCODE_TXF].emit = txf_emit;
3760    bld.bld_base.op_actions[TGSI_OPCODE_TEX2].emit = tex2_emit;
3761    bld.bld_base.op_actions[TGSI_OPCODE_TXB2].emit = txb2_emit;
3762    bld.bld_base.op_actions[TGSI_OPCODE_TXL2].emit = txl2_emit;
3763    /* DX10 sampling ops */
3764    bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE].emit = sample_emit;
3765    bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE_B].emit = sample_b_emit;
3766    bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE_C].emit = sample_c_emit;
3767    bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE_C_LZ].emit = sample_c_lz_emit;
3768    bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE_D].emit = sample_d_emit;
3769    bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE_I].emit = sample_i_emit;
3770    bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE_L].emit = sample_l_emit;
3771    bld.bld_base.op_actions[TGSI_OPCODE_SVIEWINFO].emit = sviewinfo_emit;
3772
3773    if (gs_iface) {
3774       /* There's no specific value for this because it should always
3775        * be set, but apps using ext_geometry_shader4 quite often
3776        * were forgetting so we're using MAX_VERTEX_VARYING from
3777        * that spec even though we could debug_assert if it's not
3778        * set, but that's a lot uglier. */
3779       uint max_output_vertices;
3780
3781       /* inputs are always indirect with gs */
3782       bld.indirect_files |= (1 << TGSI_FILE_INPUT);
3783       bld.gs_iface = gs_iface;
3784       bld.bld_base.emit_fetch_funcs[TGSI_FILE_INPUT] = emit_fetch_gs_input;
3785       bld.bld_base.op_actions[TGSI_OPCODE_EMIT].emit = emit_vertex;
3786       bld.bld_base.op_actions[TGSI_OPCODE_ENDPRIM].emit = end_primitive;
3787
3788       max_output_vertices =
3789             info->properties[TGSI_PROPERTY_GS_MAX_OUTPUT_VERTICES];
3790       if (!max_output_vertices)
3791          max_output_vertices = 32;
3792
3793       bld.max_output_vertices_vec =
3794          lp_build_const_int_vec(gallivm, bld.bld_base.int_bld.type,
3795                                 max_output_vertices);
3796    }
3797
3798    lp_exec_mask_init(&bld.exec_mask, &bld.bld_base.int_bld);
3799
3800    bld.system_values = *system_values;
3801
3802    lp_build_tgsi_llvm(&bld.bld_base, tokens);
3803
3804    if (0) {
3805       LLVMBasicBlockRef block = LLVMGetInsertBlock(gallivm->builder);
3806       LLVMValueRef function = LLVMGetBasicBlockParent(block);
3807       debug_printf("11111111111111111111111111111 \n");
3808       tgsi_dump(tokens, 0);
3809       lp_debug_dump_value(function);
3810       debug_printf("2222222222222222222222222222 \n");
3811    }
3812
3813    if (0) {
3814       LLVMModuleRef module = LLVMGetGlobalParent(
3815          LLVMGetBasicBlockParent(LLVMGetInsertBlock(gallivm->builder)));
3816       LLVMDumpModule(module);
3817
3818    }
3819    lp_exec_mask_fini(&bld.exec_mask);
3820 }