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 oow = NULL;
1968    LLVMValueRef lod = NULL;
1969    LLVMValueRef coords[5];
1970    LLVMValueRef offsets[3] = { NULL };
1971    struct lp_derivatives derivs;
1972    struct lp_sampler_params params;
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    unsigned sample_key = 0;
1978
1979    memset(&params, 0, sizeof(params));
1980
1981    if (!bld->sampler) {
1982       _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
1983       for (i = 0; i < 4; i++) {
1984          texel[i] = bld->bld_base.base.undef;
1985       }
1986       return;
1987    }
1988
1989    switch (inst->Texture.Texture) {
1990    case TGSI_TEXTURE_1D_ARRAY:
1991       layer_coord = 1;
1992       /* fallthrough */
1993    case TGSI_TEXTURE_1D:
1994       num_offsets = 1;
1995       num_derivs = 1;
1996       break;
1997    case TGSI_TEXTURE_2D_ARRAY:
1998       layer_coord = 2;
1999       /* fallthrough */
2000    case TGSI_TEXTURE_2D:
2001    case TGSI_TEXTURE_RECT:
2002       num_offsets = 2;
2003       num_derivs = 2;
2004       break;
2005    case TGSI_TEXTURE_SHADOW1D_ARRAY:
2006       layer_coord = 1;
2007       /* fallthrough */
2008    case TGSI_TEXTURE_SHADOW1D:
2009       shadow_coord = 2;
2010       num_offsets = 1;
2011       num_derivs = 1;
2012       break;
2013    case TGSI_TEXTURE_SHADOW2D_ARRAY:
2014       layer_coord = 2;
2015       shadow_coord = 3;
2016       num_offsets = 2;
2017       num_derivs = 2;
2018       break;
2019    case TGSI_TEXTURE_SHADOW2D:
2020    case TGSI_TEXTURE_SHADOWRECT:
2021       shadow_coord = 2;
2022       num_offsets = 2;
2023       num_derivs = 2;
2024       break;
2025    case TGSI_TEXTURE_CUBE:
2026       num_offsets = 2;
2027       num_derivs = 3;
2028       break;
2029    case TGSI_TEXTURE_3D:
2030       num_offsets = 3;
2031       num_derivs = 3;
2032       break;
2033    case TGSI_TEXTURE_SHADOWCUBE:
2034       shadow_coord = 3;
2035       num_offsets = 2;
2036       num_derivs = 3;
2037       break;
2038    case TGSI_TEXTURE_CUBE_ARRAY:
2039       num_offsets = 2;
2040       num_derivs = 3;
2041       layer_coord = 3;
2042       break;
2043    case TGSI_TEXTURE_SHADOWCUBE_ARRAY:
2044       num_offsets = 2;
2045       num_derivs = 3;
2046       layer_coord = 3;
2047       shadow_coord = 4; /* shadow coord special different reg */
2048       break;
2049    case TGSI_TEXTURE_2D_MSAA:
2050    case TGSI_TEXTURE_2D_ARRAY_MSAA:
2051    default:
2052       assert(0);
2053       return;
2054    }
2055
2056    /* Note lod and especially projected are illegal in a LOT of cases */
2057    if (modifier == LP_BLD_TEX_MODIFIER_LOD_BIAS ||
2058        modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_LOD) {
2059       if (inst->Texture.Texture == TGSI_TEXTURE_SHADOWCUBE ||
2060           inst->Texture.Texture == TGSI_TEXTURE_CUBE_ARRAY) {
2061          /* note that shadow cube array with bias/explicit lod does not exist */
2062          lod = lp_build_emit_fetch(&bld->bld_base, inst, 1, 0);
2063       }
2064       else {
2065          lod = lp_build_emit_fetch(&bld->bld_base, inst, 0, 3);
2066       }
2067       if (modifier == LP_BLD_TEX_MODIFIER_LOD_BIAS) {
2068          sample_key |= LP_SAMPLER_LOD_BIAS << LP_SAMPLER_LOD_CONTROL_SHIFT;
2069       }
2070       else if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_LOD) {
2071          sample_key |= LP_SAMPLER_LOD_EXPLICIT << LP_SAMPLER_LOD_CONTROL_SHIFT;
2072       }
2073       lod_property = lp_build_lod_property(&bld->bld_base, inst, 0);
2074    }
2075
2076    if (modifier == LP_BLD_TEX_MODIFIER_PROJECTED) {
2077       oow = lp_build_emit_fetch(&bld->bld_base, inst, 0, 3);
2078       oow = lp_build_rcp(&bld->bld_base.base, oow);
2079    }
2080
2081    for (i = 0; i < num_derivs; i++) {
2082       coords[i] = lp_build_emit_fetch(&bld->bld_base, inst, 0, i);
2083       if (modifier == LP_BLD_TEX_MODIFIER_PROJECTED)
2084          coords[i] = lp_build_mul(&bld->bld_base.base, coords[i], oow);
2085    }
2086    for (i = num_derivs; i < 5; i++) {
2087       coords[i] = bld->bld_base.base.undef;
2088    }
2089
2090    /* Layer coord always goes into 3rd slot, except for cube map arrays */
2091    if (layer_coord) {
2092       if (layer_coord == 3) {
2093          coords[3] = lp_build_emit_fetch(&bld->bld_base, inst, 0, layer_coord);
2094       }
2095       else {
2096          coords[2] = lp_build_emit_fetch(&bld->bld_base, inst, 0, layer_coord);
2097       }
2098       if (modifier == LP_BLD_TEX_MODIFIER_PROJECTED)
2099          coords[2] = lp_build_mul(&bld->bld_base.base, coords[2], oow);
2100    }
2101    /* Shadow coord occupies always 5th slot. */
2102    if (shadow_coord) {
2103       sample_key |= LP_SAMPLER_SHADOW;
2104       if (shadow_coord == 4) {
2105          coords[4] = lp_build_emit_fetch(&bld->bld_base, inst, 1, 0);
2106       }
2107       else {
2108          coords[4] = lp_build_emit_fetch(&bld->bld_base, inst, 0, shadow_coord);
2109       }
2110       if (modifier == LP_BLD_TEX_MODIFIER_PROJECTED)
2111          coords[4] = lp_build_mul(&bld->bld_base.base, coords[4], oow);
2112    }
2113
2114    if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV) {
2115       unsigned dim;
2116       sample_key |= LP_SAMPLER_LOD_DERIVATIVES << LP_SAMPLER_LOD_CONTROL_SHIFT;
2117       for (dim = 0; dim < num_derivs; ++dim) {
2118          derivs.ddx[dim] = lp_build_emit_fetch(&bld->bld_base, inst, 1, dim);
2119          derivs.ddy[dim] = lp_build_emit_fetch(&bld->bld_base, inst, 2, dim);
2120       }
2121       params.derivs = &derivs;
2122       /*
2123        * could also check all src regs if constant but I doubt such
2124        * cases exist in practice.
2125        */
2126       if (bld->bld_base.info->processor == TGSI_PROCESSOR_FRAGMENT) {
2127          if (gallivm_debug & GALLIVM_DEBUG_NO_QUAD_LOD) {
2128             lod_property = LP_SAMPLER_LOD_PER_ELEMENT;
2129          }
2130          else {
2131             lod_property = LP_SAMPLER_LOD_PER_QUAD;
2132          }
2133       }
2134       else {
2135          lod_property = LP_SAMPLER_LOD_PER_ELEMENT;
2136       }
2137    }
2138    sample_key |= lod_property << LP_SAMPLER_LOD_PROPERTY_SHIFT;
2139
2140    /* some advanced gather instructions (txgo) would require 4 offsets */
2141    if (inst->Texture.NumOffsets == 1) {
2142       unsigned dim;
2143       sample_key |= LP_SAMPLER_OFFSETS;
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    params.type = bld->bld_base.base.type;
2150    params.sample_key = sample_key;
2151    params.texture_index = unit;
2152    params.sampler_index = unit;
2153    params.context_ptr = bld->context_ptr;
2154    params.coords = coords;
2155    params.offsets = offsets;
2156    params.lod = lod;
2157    params.texel = texel;
2158
2159    bld->sampler->emit_tex_sample(bld->sampler,
2160                                  bld->bld_base.base.gallivm,
2161                                  &params);
2162 }
2163
2164 static void
2165 emit_sample(struct lp_build_tgsi_soa_context *bld,
2166             const struct tgsi_full_instruction *inst,
2167             enum lp_build_tex_modifier modifier,
2168             boolean compare,
2169             LLVMValueRef *texel)
2170 {
2171    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
2172    unsigned texture_unit, sampler_unit;
2173    LLVMValueRef lod = NULL;
2174    LLVMValueRef coords[5];
2175    LLVMValueRef offsets[3] = { NULL };
2176    struct lp_derivatives derivs;
2177    struct lp_sampler_params params;
2178    enum lp_sampler_lod_property lod_property = LP_SAMPLER_LOD_SCALAR;
2179
2180    unsigned num_offsets, num_derivs, i;
2181    unsigned layer_coord = 0;
2182    unsigned sample_key = 0;
2183
2184    memset(&params, 0, sizeof(params));
2185
2186    if (!bld->sampler) {
2187       _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
2188       for (i = 0; i < 4; i++) {
2189          texel[i] = bld->bld_base.base.undef;
2190       }
2191       return;
2192    }
2193
2194    /*
2195     * unlike old-style tex opcodes the texture/sampler indices
2196     * always come from src1 and src2 respectively.
2197     */
2198    texture_unit = inst->Src[1].Register.Index;
2199    sampler_unit = inst->Src[2].Register.Index;
2200
2201    /*
2202     * Note inst->Texture.Texture will contain the number of offsets,
2203     * however the target information is NOT there and comes from the
2204     * declared sampler views instead.
2205     */
2206    switch (bld->sv[texture_unit].Resource) {
2207    case TGSI_TEXTURE_1D:
2208       num_offsets = 1;
2209       num_derivs = 1;
2210       break;
2211    case TGSI_TEXTURE_1D_ARRAY:
2212       layer_coord = 1;
2213       num_offsets = 1;
2214       num_derivs = 1;
2215       break;
2216    case TGSI_TEXTURE_2D:
2217    case TGSI_TEXTURE_RECT:
2218       num_offsets = 2;
2219       num_derivs = 2;
2220       break;
2221    case TGSI_TEXTURE_2D_ARRAY:
2222       layer_coord = 2;
2223       num_offsets = 2;
2224       num_derivs = 2;
2225       break;
2226    case TGSI_TEXTURE_CUBE:
2227       num_offsets = 2;
2228       num_derivs = 3;
2229       break;
2230    case TGSI_TEXTURE_3D:
2231       num_offsets = 3;
2232       num_derivs = 3;
2233       break;
2234    case TGSI_TEXTURE_CUBE_ARRAY:
2235       layer_coord = 3;
2236       num_offsets = 2;
2237       num_derivs = 3;
2238       break;
2239    default:
2240       assert(0);
2241       return;
2242    }
2243
2244    if (modifier == LP_BLD_TEX_MODIFIER_LOD_BIAS ||
2245        modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_LOD) {
2246       lod = lp_build_emit_fetch(&bld->bld_base, inst, 3, 0);
2247       if (modifier == LP_BLD_TEX_MODIFIER_LOD_BIAS) {
2248          sample_key |= LP_SAMPLER_LOD_BIAS << LP_SAMPLER_LOD_CONTROL_SHIFT;
2249       }
2250       else if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_LOD) {
2251          sample_key |= LP_SAMPLER_LOD_EXPLICIT << LP_SAMPLER_LOD_CONTROL_SHIFT;
2252       }
2253       lod_property = lp_build_lod_property(&bld->bld_base, inst, 0);
2254    }
2255    else if (modifier == LP_BLD_TEX_MODIFIER_LOD_ZERO) {
2256       /* XXX might be better to explicitly pass the level zero information */
2257       sample_key |= LP_SAMPLER_LOD_EXPLICIT << LP_SAMPLER_LOD_CONTROL_SHIFT;
2258       lod = lp_build_const_vec(gallivm, bld->bld_base.base.type, 0.0F);
2259    }
2260
2261    for (i = 0; i < num_derivs; i++) {
2262       coords[i] = lp_build_emit_fetch(&bld->bld_base, inst, 0, i);
2263    }
2264    for (i = num_derivs; i < 5; i++) {
2265       coords[i] = bld->bld_base.base.undef;
2266    }
2267
2268    /* Layer coord always goes into 3rd slot, except for cube map arrays */
2269    if (layer_coord) {
2270       if (layer_coord == 3)
2271          coords[3] = lp_build_emit_fetch(&bld->bld_base, inst, 0, layer_coord);
2272       else
2273          coords[2] = lp_build_emit_fetch(&bld->bld_base, inst, 0, layer_coord);
2274    }
2275    /* Shadow coord occupies always 5th slot. */
2276    if (compare) {
2277       sample_key |= LP_SAMPLER_SHADOW;
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       sample_key |= LP_SAMPLER_LOD_DERIVATIVES << LP_SAMPLER_LOD_CONTROL_SHIFT;
2284       for (dim = 0; dim < num_derivs; ++dim) {
2285          derivs.ddx[dim] = lp_build_emit_fetch(&bld->bld_base, inst, 3, dim);
2286          derivs.ddy[dim] = lp_build_emit_fetch(&bld->bld_base, inst, 4, dim);
2287       }
2288       params.derivs = &derivs;
2289       /*
2290        * could also check all src regs if constant but I doubt such
2291        * cases exist in practice.
2292        */
2293       if (bld->bld_base.info->processor == TGSI_PROCESSOR_FRAGMENT) {
2294          if (gallivm_debug & GALLIVM_DEBUG_NO_QUAD_LOD) {
2295             lod_property = LP_SAMPLER_LOD_PER_ELEMENT;
2296          }
2297          else {
2298             lod_property = LP_SAMPLER_LOD_PER_QUAD;
2299          }
2300       }
2301       else {
2302          lod_property = LP_SAMPLER_LOD_PER_ELEMENT;
2303       }
2304    }
2305
2306    /* some advanced gather instructions (txgo) would require 4 offsets */
2307    if (inst->Texture.NumOffsets == 1) {
2308       unsigned dim;
2309       sample_key |= LP_SAMPLER_OFFSETS;
2310       for (dim = 0; dim < num_offsets; dim++) {
2311          offsets[dim] = lp_build_emit_fetch_texoffset(&bld->bld_base, inst, 0, dim);
2312       }
2313    }
2314    sample_key |= lod_property << LP_SAMPLER_LOD_PROPERTY_SHIFT;
2315
2316    params.type = bld->bld_base.base.type;
2317    params.sample_key = sample_key;
2318    params.texture_index = texture_unit;
2319    params.sampler_index = sampler_unit;
2320    params.context_ptr = bld->context_ptr;
2321    params.coords = coords;
2322    params.offsets = offsets;
2323    params.lod = lod;
2324    params.texel = texel;
2325
2326    bld->sampler->emit_tex_sample(bld->sampler,
2327                                  bld->bld_base.base.gallivm,
2328                                  &params);
2329
2330    if (inst->Src[1].Register.SwizzleX != PIPE_SWIZZLE_RED ||
2331        inst->Src[1].Register.SwizzleY != PIPE_SWIZZLE_GREEN ||
2332        inst->Src[1].Register.SwizzleZ != PIPE_SWIZZLE_BLUE ||
2333        inst->Src[1].Register.SwizzleW != PIPE_SWIZZLE_ALPHA) {
2334       unsigned char swizzles[4];
2335       swizzles[0] = inst->Src[1].Register.SwizzleX;
2336       swizzles[1] = inst->Src[1].Register.SwizzleY;
2337       swizzles[2] = inst->Src[1].Register.SwizzleZ;
2338       swizzles[3] = inst->Src[1].Register.SwizzleW;
2339
2340       lp_build_swizzle_soa_inplace(&bld->bld_base.base, texel, swizzles);
2341    }
2342 }
2343
2344 static void
2345 emit_fetch_texels( struct lp_build_tgsi_soa_context *bld,
2346                    const struct tgsi_full_instruction *inst,
2347                    LLVMValueRef *texel,
2348                    boolean is_samplei)
2349 {
2350    unsigned unit, target;
2351    LLVMValueRef coord_undef = LLVMGetUndef(bld->bld_base.base.int_vec_type);
2352    LLVMValueRef explicit_lod = NULL;
2353    LLVMValueRef coords[5];
2354    LLVMValueRef offsets[3] = { NULL };
2355    struct lp_sampler_params params;
2356    enum lp_sampler_lod_property lod_property = LP_SAMPLER_LOD_SCALAR;
2357    unsigned dims, i;
2358    unsigned layer_coord = 0;
2359    unsigned sample_key = LP_SAMPLER_FETCH;
2360
2361    memset(&params, 0, sizeof(params));
2362
2363    if (!bld->sampler) {
2364       _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
2365       for (i = 0; i < 4; i++) {
2366          texel[i] = coord_undef;
2367       }
2368       return;
2369    }
2370
2371    unit = inst->Src[1].Register.Index;
2372
2373    if (is_samplei) {
2374       target = bld->sv[unit].Resource;
2375    }
2376    else {
2377       target = inst->Texture.Texture;
2378    }
2379
2380    switch (target) {
2381    case TGSI_TEXTURE_1D:
2382    case TGSI_TEXTURE_BUFFER:
2383       dims = 1;
2384       break;
2385    case TGSI_TEXTURE_1D_ARRAY:
2386       layer_coord = 1;
2387       dims = 1;
2388       break;
2389    case TGSI_TEXTURE_2D:
2390    case TGSI_TEXTURE_RECT:
2391    case TGSI_TEXTURE_2D_MSAA:
2392       dims = 2;
2393       break;
2394    case TGSI_TEXTURE_2D_ARRAY:
2395    case TGSI_TEXTURE_2D_ARRAY_MSAA:
2396       layer_coord = 2;
2397       dims = 2;
2398       break;
2399    case TGSI_TEXTURE_3D:
2400       dims = 3;
2401       break;
2402    default:
2403       assert(0);
2404       return;
2405    }
2406
2407    /* always have lod except for buffers and msaa targets ? */
2408    if (target != TGSI_TEXTURE_BUFFER &&
2409        target != TGSI_TEXTURE_2D_MSAA &&
2410        target != TGSI_TEXTURE_2D_ARRAY_MSAA) {
2411       sample_key |= LP_SAMPLER_LOD_EXPLICIT << LP_SAMPLER_LOD_CONTROL_SHIFT;
2412       explicit_lod = lp_build_emit_fetch(&bld->bld_base, inst, 0, 3);
2413       lod_property = lp_build_lod_property(&bld->bld_base, inst, 0);
2414    }
2415    /* XXX: for real msaa support, the w component would be the sample index. */
2416
2417    for (i = 0; i < dims; i++) {
2418       coords[i] = lp_build_emit_fetch(&bld->bld_base, inst, 0, i);
2419    }
2420    /* never use more than 3 coords here but emit_fetch_texel copies all 5 anyway */
2421    for (i = dims; i < 5; i++) {
2422       coords[i] = coord_undef;
2423    }
2424    if (layer_coord)
2425       coords[2] = lp_build_emit_fetch(&bld->bld_base, inst, 0, layer_coord);
2426
2427    if (inst->Texture.NumOffsets == 1) {
2428       unsigned dim;
2429       sample_key |= LP_SAMPLER_OFFSETS;
2430       for (dim = 0; dim < dims; dim++) {
2431          offsets[dim] = lp_build_emit_fetch_texoffset(&bld->bld_base, inst, 0, dim);
2432       }
2433    }
2434    sample_key |= lod_property << LP_SAMPLER_LOD_PROPERTY_SHIFT;
2435
2436    params.type = bld->bld_base.base.type;
2437    params.sample_key = sample_key;
2438    params.texture_index = unit;
2439    params.sampler_index = unit;
2440    params.context_ptr = bld->context_ptr;
2441    params.coords = coords;
2442    params.offsets = offsets;
2443    params.derivs = NULL;
2444    params.lod = explicit_lod;
2445    params.texel = texel;
2446
2447    bld->sampler->emit_tex_sample(bld->sampler,
2448                                  bld->bld_base.base.gallivm,
2449                                  &params);
2450
2451    if (is_samplei &&
2452        (inst->Src[1].Register.SwizzleX != PIPE_SWIZZLE_RED ||
2453         inst->Src[1].Register.SwizzleY != PIPE_SWIZZLE_GREEN ||
2454         inst->Src[1].Register.SwizzleZ != PIPE_SWIZZLE_BLUE ||
2455         inst->Src[1].Register.SwizzleW != PIPE_SWIZZLE_ALPHA)) {
2456       unsigned char swizzles[4];
2457       swizzles[0] = inst->Src[1].Register.SwizzleX;
2458       swizzles[1] = inst->Src[1].Register.SwizzleY;
2459       swizzles[2] = inst->Src[1].Register.SwizzleZ;
2460       swizzles[3] = inst->Src[1].Register.SwizzleW;
2461
2462       lp_build_swizzle_soa_inplace(&bld->bld_base.base, texel, swizzles);
2463    }
2464 }
2465
2466 static void
2467 emit_size_query( struct lp_build_tgsi_soa_context *bld,
2468                  const struct tgsi_full_instruction *inst,
2469                  LLVMValueRef *sizes_out,
2470                  boolean is_sviewinfo)
2471 {
2472    LLVMValueRef explicit_lod;
2473    enum lp_sampler_lod_property lod_property;
2474    unsigned has_lod;
2475    unsigned i;
2476    unsigned unit = inst->Src[1].Register.Index;
2477    unsigned target, pipe_target;
2478
2479    if (is_sviewinfo) {
2480       target = bld->sv[unit].Resource;
2481    }
2482    else {
2483       target = inst->Texture.Texture;
2484    }
2485    switch (target) {
2486    case TGSI_TEXTURE_BUFFER:
2487    case TGSI_TEXTURE_RECT:
2488    case TGSI_TEXTURE_SHADOWRECT:
2489       has_lod = 0;
2490       break;
2491    default:
2492       has_lod = 1;
2493       break;
2494    }
2495
2496    if (!bld->sampler) {
2497       _debug_printf("warning: found texture query instruction but no sampler generator supplied\n");
2498       for (i = 0; i < 4; i++)
2499          sizes_out[i] = bld->bld_base.int_bld.undef;
2500       return;
2501    }
2502
2503    if (has_lod) {
2504       explicit_lod = lp_build_emit_fetch(&bld->bld_base, inst, 0, 0);
2505       lod_property = lp_build_lod_property(&bld->bld_base, inst, 0);
2506    }
2507    else {
2508       explicit_lod = NULL;
2509       lod_property = LP_SAMPLER_LOD_SCALAR;
2510    }
2511
2512
2513    pipe_target = tgsi_to_pipe_tex_target(target);
2514
2515    bld->sampler->emit_size_query(bld->sampler,
2516                                  bld->bld_base.base.gallivm,
2517                                  bld->bld_base.int_bld.type,
2518                                  unit, pipe_target,
2519                                  bld->context_ptr,
2520                                  TRUE,
2521                                  lod_property,
2522                                  explicit_lod,
2523                                  sizes_out);
2524 }
2525
2526 static boolean
2527 near_end_of_shader(struct lp_build_tgsi_soa_context *bld,
2528                    int pc)
2529 {
2530    int i;
2531
2532    for (i = 0; i < 5; i++) {
2533       unsigned opcode;
2534
2535       if (pc + i >= bld->bld_base.info->num_instructions)
2536          return TRUE;
2537
2538       opcode = bld->bld_base.instructions[pc + i].Instruction.Opcode;
2539
2540       if (opcode == TGSI_OPCODE_END)
2541          return TRUE;
2542
2543       if (opcode == TGSI_OPCODE_TEX ||
2544          opcode == TGSI_OPCODE_TXP ||
2545          opcode == TGSI_OPCODE_TXD ||
2546          opcode == TGSI_OPCODE_TXB ||
2547          opcode == TGSI_OPCODE_TXL ||
2548          opcode == TGSI_OPCODE_TXF ||
2549          opcode == TGSI_OPCODE_TXQ ||
2550          opcode == TGSI_OPCODE_TEX2 ||
2551          opcode == TGSI_OPCODE_TXB2 ||
2552          opcode == TGSI_OPCODE_TXL2 ||
2553          opcode == TGSI_OPCODE_SAMPLE ||
2554          opcode == TGSI_OPCODE_SAMPLE_B ||
2555          opcode == TGSI_OPCODE_SAMPLE_C ||
2556          opcode == TGSI_OPCODE_SAMPLE_C_LZ ||
2557          opcode == TGSI_OPCODE_SAMPLE_D ||
2558          opcode == TGSI_OPCODE_SAMPLE_I ||
2559          opcode == TGSI_OPCODE_SAMPLE_L ||
2560          opcode == TGSI_OPCODE_SVIEWINFO ||
2561          opcode == TGSI_OPCODE_CAL ||
2562          opcode == TGSI_OPCODE_CALLNZ ||
2563          opcode == TGSI_OPCODE_IF ||
2564          opcode == TGSI_OPCODE_UIF ||
2565          opcode == TGSI_OPCODE_BGNLOOP ||
2566          opcode == TGSI_OPCODE_SWITCH)
2567          return FALSE;
2568    }
2569
2570    return TRUE;
2571 }
2572
2573
2574
2575 /**
2576  * Kill fragment if any of the src register values are negative.
2577  */
2578 static void
2579 emit_kill_if(
2580    struct lp_build_tgsi_soa_context *bld,
2581    const struct tgsi_full_instruction *inst,
2582    int pc)
2583 {
2584    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
2585    const struct tgsi_full_src_register *reg = &inst->Src[0];
2586    LLVMValueRef terms[TGSI_NUM_CHANNELS];
2587    LLVMValueRef mask;
2588    unsigned chan_index;
2589
2590    memset(&terms, 0, sizeof terms);
2591
2592    TGSI_FOR_EACH_CHANNEL( chan_index ) {
2593       unsigned swizzle;
2594
2595       /* Unswizzle channel */
2596       swizzle = tgsi_util_get_full_src_register_swizzle( reg, chan_index );
2597
2598       /* Check if the component has not been already tested. */
2599       assert(swizzle < TGSI_NUM_CHANNELS);
2600       if( !terms[swizzle] )
2601          /* TODO: change the comparison operator instead of setting the sign */
2602          terms[swizzle] =  lp_build_emit_fetch(&bld->bld_base, inst, 0, chan_index );
2603    }
2604
2605    mask = NULL;
2606    TGSI_FOR_EACH_CHANNEL( chan_index ) {
2607       if(terms[chan_index]) {
2608          LLVMValueRef chan_mask;
2609
2610          /*
2611           * If term < 0 then mask = 0 else mask = ~0.
2612           */
2613          chan_mask = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_GEQUAL, terms[chan_index], bld->bld_base.base.zero);
2614
2615          if(mask)
2616             mask = LLVMBuildAnd(builder, mask, chan_mask, "");
2617          else
2618             mask = chan_mask;
2619       }
2620    }
2621
2622    if (bld->exec_mask.has_mask) {
2623       LLVMValueRef invmask;
2624       invmask = LLVMBuildNot(builder, bld->exec_mask.exec_mask, "kilp");
2625       mask = LLVMBuildOr(builder, mask, invmask, "");
2626    }
2627
2628    lp_build_mask_update(bld->mask, mask);
2629    if (!near_end_of_shader(bld, pc))
2630       lp_build_mask_check(bld->mask);
2631 }
2632
2633
2634 /**
2635  * Unconditional fragment kill.
2636  * The only predication is the execution mask which will apply if
2637  * we're inside a loop or conditional.
2638  */
2639 static void
2640 emit_kill(struct lp_build_tgsi_soa_context *bld,
2641           int pc)
2642 {
2643    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
2644    LLVMValueRef mask;
2645
2646    /* For those channels which are "alive", disable fragment shader
2647     * execution.
2648     */
2649    if (bld->exec_mask.has_mask) {
2650       mask = LLVMBuildNot(builder, bld->exec_mask.exec_mask, "kilp");
2651    }
2652    else {
2653       LLVMValueRef zero = LLVMConstNull(bld->bld_base.base.int_vec_type);
2654       mask = zero;
2655    }
2656
2657    lp_build_mask_update(bld->mask, mask);
2658
2659    if (!near_end_of_shader(bld, pc))
2660       lp_build_mask_check(bld->mask);
2661 }
2662
2663
2664 /**
2665  * Emit code which will dump the value of all the temporary registers
2666  * to stdout.
2667  */
2668 static void
2669 emit_dump_file(struct lp_build_tgsi_soa_context *bld,
2670                unsigned file)
2671 {
2672    const struct tgsi_shader_info *info = bld->bld_base.info;
2673    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
2674    LLVMBuilderRef builder = gallivm->builder;
2675    LLVMValueRef reg_ptr;
2676    int index;
2677    int max_index = info->file_max[file];
2678
2679    /*
2680     * Some register files, particularly constants, can be very large,
2681     * and dumping everything could make this unusably slow.
2682     */
2683    max_index = MIN2(max_index, 32);
2684
2685    for (index = 0; index <= max_index; index++) {
2686       LLVMValueRef res;
2687       unsigned mask;
2688       int chan;
2689
2690       if (index < 8 * sizeof(unsigned) &&
2691           (info->file_mask[file] & (1 << index)) == 0)  {
2692          /* This was not declared.*/
2693          continue;
2694       }
2695
2696       if (file == TGSI_FILE_INPUT) {
2697          mask = info->input_usage_mask[index];
2698       } else {
2699          mask = TGSI_WRITEMASK_XYZW;
2700       }
2701
2702       for (chan = 0; chan < 4; chan++) {
2703          if ((mask & (1 << chan)) == 0) {
2704             /* This channel is not used.*/
2705             continue;
2706          }
2707
2708          if (file == TGSI_FILE_CONSTANT) {
2709             struct tgsi_full_src_register reg;
2710             memset(&reg, 0, sizeof reg);
2711             reg.Register.File = file;
2712             reg.Register.Index = index;
2713             reg.Register.SwizzleX = 0;
2714             reg.Register.SwizzleY = 1;
2715             reg.Register.SwizzleZ = 2;
2716             reg.Register.SwizzleW = 3;
2717
2718             res = bld->bld_base.emit_fetch_funcs[file](&bld->bld_base, &reg, TGSI_TYPE_FLOAT, chan);
2719             if (!res) {
2720                continue;
2721             }
2722          } else if (file == TGSI_FILE_INPUT) {
2723             res = bld->inputs[index][chan];
2724             if (!res) {
2725                continue;
2726             }
2727          } else if (file == TGSI_FILE_TEMPORARY) {
2728             reg_ptr = lp_get_temp_ptr_soa(bld, index, chan);
2729             assert(reg_ptr);
2730             res = LLVMBuildLoad(builder, reg_ptr, "");
2731          } else if (file == TGSI_FILE_OUTPUT) {
2732             reg_ptr = lp_get_output_ptr(bld, index, chan);
2733             assert(reg_ptr);
2734             res = LLVMBuildLoad(builder, reg_ptr, "");
2735          } else {
2736             assert(0);
2737             continue;
2738          }
2739
2740          emit_dump_reg(gallivm, file, index, chan, res);
2741       }
2742    }
2743 }
2744
2745
2746
2747 void
2748 lp_emit_declaration_soa(
2749    struct lp_build_tgsi_context *bld_base,
2750    const struct tgsi_full_declaration *decl)
2751 {
2752    struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
2753    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
2754    LLVMTypeRef vec_type = bld->bld_base.base.vec_type;
2755    const unsigned first = decl->Range.First;
2756    const unsigned last = decl->Range.Last;
2757    unsigned idx, i;
2758
2759    assert(last <= bld->bld_base.info->file_max[decl->Declaration.File]);
2760
2761    switch (decl->Declaration.File) {
2762    case TGSI_FILE_TEMPORARY:
2763       if (!(bld->indirect_files & (1 << TGSI_FILE_TEMPORARY))) {
2764          assert(last < LP_MAX_INLINED_TEMPS);
2765          for (idx = first; idx <= last; ++idx) {
2766             for (i = 0; i < TGSI_NUM_CHANNELS; i++)
2767                bld->temps[idx][i] = lp_build_alloca(gallivm, vec_type, "temp");
2768          }
2769       }
2770       break;
2771
2772    case TGSI_FILE_OUTPUT:
2773       if (!(bld->indirect_files & (1 << TGSI_FILE_OUTPUT))) {
2774          for (idx = first; idx <= last; ++idx) {
2775             for (i = 0; i < TGSI_NUM_CHANNELS; i++)
2776                bld->outputs[idx][i] = lp_build_alloca(gallivm,
2777                                                       vec_type, "output");
2778          }
2779       }
2780       break;
2781
2782    case TGSI_FILE_ADDRESS:
2783       /* ADDR registers are only allocated with an integer LLVM IR type,
2784        * as they are guaranteed to always have integers.
2785        * XXX: Not sure if this exception is worthwhile (or the whole idea of
2786        * an ADDR register for that matter).
2787        */
2788       assert(last < LP_MAX_TGSI_ADDRS);
2789       for (idx = first; idx <= last; ++idx) {
2790          assert(idx < LP_MAX_TGSI_ADDRS);
2791          for (i = 0; i < TGSI_NUM_CHANNELS; i++)
2792             bld->addr[idx][i] = lp_build_alloca(gallivm, bld_base->base.int_vec_type, "addr");
2793       }
2794       break;
2795
2796    case TGSI_FILE_PREDICATE:
2797       assert(last < LP_MAX_TGSI_PREDS);
2798       for (idx = first; idx <= last; ++idx) {
2799          for (i = 0; i < TGSI_NUM_CHANNELS; i++)
2800             bld->preds[idx][i] = lp_build_alloca(gallivm, vec_type,
2801                                                  "predicate");
2802       }
2803       break;
2804
2805    case TGSI_FILE_SAMPLER_VIEW:
2806       /*
2807        * The target stored here MUST match whatever there actually
2808        * is in the set sampler views (what about return type?).
2809        */
2810       assert(last < PIPE_MAX_SHADER_SAMPLER_VIEWS);
2811       for (idx = first; idx <= last; ++idx) {
2812          bld->sv[idx] = decl->SamplerView;
2813       }
2814       break;
2815
2816    case TGSI_FILE_CONSTANT:
2817    {
2818       /*
2819        * We could trivially fetch the per-buffer pointer when fetching the
2820        * constant, relying on llvm to figure out it's always the same pointer
2821        * anyway. However, doing so results in a huge (more than factor of 10)
2822        * slowdown in llvm compilation times for some (but not all) shaders
2823        * (more specifically, the IR optimization spends way more time in
2824        * DominatorTree::dominates). At least with llvm versions 3.1, 3.3.
2825        */
2826       unsigned idx2D = decl->Dim.Index2D;
2827       LLVMValueRef index2D = lp_build_const_int32(gallivm, idx2D);
2828       assert(idx2D < LP_MAX_TGSI_CONST_BUFFERS);
2829       bld->consts[idx2D] =
2830          lp_build_array_get(gallivm, bld->consts_ptr, index2D);
2831       bld->consts_sizes[idx2D] =
2832          lp_build_array_get(gallivm, bld->const_sizes_ptr, index2D);
2833    }
2834       break;
2835
2836    default:
2837       /* don't need to declare other vars */
2838       break;
2839    }
2840 }
2841
2842
2843 void lp_emit_immediate_soa(
2844    struct lp_build_tgsi_context *bld_base,
2845    const struct tgsi_full_immediate *imm)
2846 {
2847    struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
2848    struct gallivm_state * gallivm = bld_base->base.gallivm;
2849    LLVMValueRef imms[4];
2850    unsigned i;
2851    const uint size = imm->Immediate.NrTokens - 1;
2852    assert(size <= 4);
2853    switch (imm->Immediate.DataType) {
2854    case TGSI_IMM_FLOAT32:
2855       for( i = 0; i < size; ++i )
2856          imms[i] =
2857                lp_build_const_vec(gallivm, bld_base->base.type, imm->u[i].Float);
2858
2859       break;
2860    case TGSI_IMM_UINT32:
2861       for( i = 0; i < size; ++i ) {
2862          LLVMValueRef tmp = lp_build_const_vec(gallivm, bld_base->uint_bld.type, imm->u[i].Uint);
2863          imms[i] = LLVMConstBitCast(tmp, bld_base->base.vec_type);
2864       }
2865
2866       break;
2867    case TGSI_IMM_INT32:
2868       for( i = 0; i < size; ++i ) {
2869          LLVMValueRef tmp = lp_build_const_vec(gallivm, bld_base->int_bld.type, imm->u[i].Int);
2870          imms[i] = LLVMConstBitCast(tmp, bld_base->base.vec_type);
2871       }
2872
2873       break;
2874    }
2875    for( i = size; i < 4; ++i )
2876       imms[i] = bld_base->base.undef;
2877
2878    if (bld->use_immediates_array) {
2879       unsigned index = bld->num_immediates;
2880       struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
2881       LLVMBuilderRef builder = gallivm->builder;
2882
2883       assert(bld->indirect_files & (1 << TGSI_FILE_IMMEDIATE));
2884       for (i = 0; i < 4; ++i ) {
2885          LLVMValueRef lindex = lp_build_const_int32(
2886                   bld->bld_base.base.gallivm, index * 4 + i);
2887          LLVMValueRef imm_ptr = LLVMBuildGEP(builder,
2888                                              bld->imms_array, &lindex, 1, "");
2889          LLVMBuildStore(builder, imms[i], imm_ptr);
2890       }
2891    } else {
2892       /* simply copy the immediate values into the next immediates[] slot */
2893       unsigned i;
2894       const uint size = imm->Immediate.NrTokens - 1;
2895       assert(size <= 4);
2896       assert(bld->num_immediates < LP_MAX_INLINED_IMMEDIATES);
2897
2898       for(i = 0; i < 4; ++i )
2899          bld->immediates[bld->num_immediates][i] = imms[i];
2900
2901       if (bld->indirect_files & (1 << TGSI_FILE_IMMEDIATE)) {
2902          unsigned index = bld->num_immediates;
2903          struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
2904          LLVMBuilderRef builder = gallivm->builder;
2905          for (i = 0; i < 4; ++i ) {
2906             LLVMValueRef lindex = lp_build_const_int32(
2907                      bld->bld_base.base.gallivm, index * 4 + i);
2908             LLVMValueRef imm_ptr = LLVMBuildGEP(builder,
2909                                                 bld->imms_array, &lindex, 1, "");
2910             LLVMBuildStore(builder,
2911                            bld->immediates[index][i],
2912                            imm_ptr);
2913          }
2914       }
2915    }
2916
2917    bld->num_immediates++;
2918 }
2919
2920 static void
2921 ddx_emit(
2922    const struct lp_build_tgsi_action * action,
2923    struct lp_build_tgsi_context * bld_base,
2924    struct lp_build_emit_data * emit_data)
2925 {
2926    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2927
2928    emit_fetch_deriv(bld, emit_data->args[0], NULL,
2929                     &emit_data->output[emit_data->chan], NULL);
2930 }
2931
2932 static void
2933 ddy_emit(
2934    const struct lp_build_tgsi_action * action,
2935    struct lp_build_tgsi_context * bld_base,
2936    struct lp_build_emit_data * emit_data)
2937 {
2938    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2939
2940    emit_fetch_deriv(bld, emit_data->args[0], NULL, NULL,
2941                     &emit_data->output[emit_data->chan]);
2942 }
2943
2944 static void
2945 kill_emit(
2946    const struct lp_build_tgsi_action * action,
2947    struct lp_build_tgsi_context * bld_base,
2948    struct lp_build_emit_data * emit_data)
2949 {
2950    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2951
2952    emit_kill(bld, bld_base->pc - 1);
2953 }
2954
2955 static void
2956 kill_if_emit(
2957    const struct lp_build_tgsi_action * action,
2958    struct lp_build_tgsi_context * bld_base,
2959    struct lp_build_emit_data * emit_data)
2960 {
2961    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2962
2963    emit_kill_if(bld, emit_data->inst, bld_base->pc - 1);
2964 }
2965
2966 static void
2967 tex_emit(
2968    const struct lp_build_tgsi_action * action,
2969    struct lp_build_tgsi_context * bld_base,
2970    struct lp_build_emit_data * emit_data)
2971 {
2972    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2973
2974    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_NONE,
2975             emit_data->output, 1);
2976 }
2977
2978 static void
2979 tex2_emit(
2980    const struct lp_build_tgsi_action * action,
2981    struct lp_build_tgsi_context * bld_base,
2982    struct lp_build_emit_data * emit_data)
2983 {
2984    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2985
2986    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_NONE,
2987             emit_data->output, 2);
2988 }
2989
2990 static void
2991 txb_emit(
2992    const struct lp_build_tgsi_action * action,
2993    struct lp_build_tgsi_context * bld_base,
2994    struct lp_build_emit_data * emit_data)
2995 {
2996    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2997
2998    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_LOD_BIAS,
2999             emit_data->output, 1);
3000 }
3001
3002 static void
3003 txb2_emit(
3004    const struct lp_build_tgsi_action * action,
3005    struct lp_build_tgsi_context * bld_base,
3006    struct lp_build_emit_data * emit_data)
3007 {
3008    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3009
3010    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_LOD_BIAS,
3011             emit_data->output, 2);
3012 }
3013
3014 static void
3015 txd_emit(
3016    const struct lp_build_tgsi_action * action,
3017    struct lp_build_tgsi_context * bld_base,
3018    struct lp_build_emit_data * emit_data)
3019 {
3020    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3021
3022    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV,
3023             emit_data->output, 3);
3024 }
3025
3026 static void
3027 txl_emit(
3028    const struct lp_build_tgsi_action * action,
3029    struct lp_build_tgsi_context * bld_base,
3030    struct lp_build_emit_data * emit_data)
3031 {
3032    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3033
3034    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD,
3035             emit_data->output, 1);
3036 }
3037
3038 static void
3039 txl2_emit(
3040    const struct lp_build_tgsi_action * action,
3041    struct lp_build_tgsi_context * bld_base,
3042    struct lp_build_emit_data * emit_data)
3043 {
3044    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3045
3046    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD,
3047             emit_data->output, 2);
3048 }
3049
3050 static void
3051 txp_emit(
3052    const struct lp_build_tgsi_action * action,
3053    struct lp_build_tgsi_context * bld_base,
3054    struct lp_build_emit_data * emit_data)
3055 {
3056    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3057
3058    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_PROJECTED,
3059             emit_data->output, 1);
3060 }
3061
3062 static void
3063 txq_emit(
3064    const struct lp_build_tgsi_action * action,
3065    struct lp_build_tgsi_context * bld_base,
3066    struct lp_build_emit_data * emit_data)
3067 {
3068    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3069
3070    emit_size_query(bld, emit_data->inst, emit_data->output, FALSE);
3071 }
3072
3073 static void
3074 txf_emit(
3075    const struct lp_build_tgsi_action * action,
3076    struct lp_build_tgsi_context * bld_base,
3077    struct lp_build_emit_data * emit_data)
3078 {
3079    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3080
3081    emit_fetch_texels(bld, emit_data->inst, emit_data->output, FALSE);
3082 }
3083
3084 static void
3085 sample_i_emit(
3086    const struct lp_build_tgsi_action * action,
3087    struct lp_build_tgsi_context * bld_base,
3088    struct lp_build_emit_data * emit_data)
3089 {
3090    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3091
3092    emit_fetch_texels(bld, emit_data->inst, emit_data->output, TRUE);
3093 }
3094
3095 static void
3096 sample_emit(
3097    const struct lp_build_tgsi_action * action,
3098    struct lp_build_tgsi_context * bld_base,
3099    struct lp_build_emit_data * emit_data)
3100 {
3101    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3102
3103    emit_sample(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_NONE,
3104                FALSE, emit_data->output);
3105 }
3106
3107 static void
3108 sample_b_emit(
3109    const struct lp_build_tgsi_action * action,
3110    struct lp_build_tgsi_context * bld_base,
3111    struct lp_build_emit_data * emit_data)
3112 {
3113    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3114
3115    emit_sample(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_LOD_BIAS,
3116                FALSE, emit_data->output);
3117 }
3118
3119 static void
3120 sample_c_emit(
3121    const struct lp_build_tgsi_action * action,
3122    struct lp_build_tgsi_context * bld_base,
3123    struct lp_build_emit_data * emit_data)
3124 {
3125    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3126
3127    emit_sample(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_NONE,
3128                TRUE, emit_data->output);
3129 }
3130
3131 static void
3132 sample_c_lz_emit(
3133    const struct lp_build_tgsi_action * action,
3134    struct lp_build_tgsi_context * bld_base,
3135    struct lp_build_emit_data * emit_data)
3136 {
3137    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3138
3139    emit_sample(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_LOD_ZERO,
3140                TRUE, emit_data->output);
3141 }
3142
3143 static void
3144 sample_d_emit(
3145    const struct lp_build_tgsi_action * action,
3146    struct lp_build_tgsi_context * bld_base,
3147    struct lp_build_emit_data * emit_data)
3148 {
3149    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3150
3151    emit_sample(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV,
3152                FALSE, emit_data->output);
3153 }
3154
3155 static void
3156 sample_l_emit(
3157    const struct lp_build_tgsi_action * action,
3158    struct lp_build_tgsi_context * bld_base,
3159    struct lp_build_emit_data * emit_data)
3160 {
3161    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3162
3163    emit_sample(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD,
3164                FALSE, emit_data->output);
3165 }
3166
3167 static void
3168 sviewinfo_emit(
3169    const struct lp_build_tgsi_action * action,
3170    struct lp_build_tgsi_context * bld_base,
3171    struct lp_build_emit_data * emit_data)
3172 {
3173    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3174
3175    emit_size_query(bld, emit_data->inst, emit_data->output, TRUE);
3176 }
3177
3178 static LLVMValueRef
3179 mask_vec(struct lp_build_tgsi_context *bld_base)
3180 {
3181    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3182    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
3183    struct lp_exec_mask *exec_mask = &bld->exec_mask;
3184
3185    if (!exec_mask->has_mask) {
3186       return lp_build_mask_value(bld->mask);
3187    }
3188    return LLVMBuildAnd(builder, lp_build_mask_value(bld->mask),
3189                        exec_mask->exec_mask, "");
3190 }
3191
3192 static void
3193 increment_vec_ptr_by_mask(struct lp_build_tgsi_context * bld_base,
3194                           LLVMValueRef ptr,
3195                           LLVMValueRef mask)
3196 {
3197    LLVMBuilderRef builder = bld_base->base.gallivm->builder;
3198    LLVMValueRef current_vec = LLVMBuildLoad(builder, ptr, "");
3199
3200    current_vec = LLVMBuildSub(builder, current_vec, mask, "");
3201
3202    LLVMBuildStore(builder, current_vec, ptr);
3203 }
3204
3205 static void
3206 clear_uint_vec_ptr_from_mask(struct lp_build_tgsi_context * bld_base,
3207                              LLVMValueRef ptr,
3208                              LLVMValueRef mask)
3209 {
3210    LLVMBuilderRef builder = bld_base->base.gallivm->builder;
3211    LLVMValueRef current_vec = LLVMBuildLoad(builder, ptr, "");
3212
3213    current_vec = lp_build_select(&bld_base->uint_bld,
3214                                  mask,
3215                                  bld_base->uint_bld.zero,
3216                                  current_vec);
3217
3218    LLVMBuildStore(builder, current_vec, ptr);
3219 }
3220
3221 static LLVMValueRef
3222 clamp_mask_to_max_output_vertices(struct lp_build_tgsi_soa_context * bld,
3223                                   LLVMValueRef current_mask_vec,
3224                                   LLVMValueRef total_emitted_vertices_vec)
3225 {
3226    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
3227    struct lp_build_context *int_bld = &bld->bld_base.int_bld;
3228    LLVMValueRef max_mask = lp_build_cmp(int_bld, PIPE_FUNC_LESS,
3229                                         total_emitted_vertices_vec,
3230                                         bld->max_output_vertices_vec);
3231
3232    return LLVMBuildAnd(builder, current_mask_vec, max_mask, "");
3233 }
3234
3235 static void
3236 emit_vertex(
3237    const struct lp_build_tgsi_action * action,
3238    struct lp_build_tgsi_context * bld_base,
3239    struct lp_build_emit_data * emit_data)
3240 {
3241    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3242    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
3243
3244    if (bld->gs_iface->emit_vertex) {
3245       LLVMValueRef mask = mask_vec(bld_base);
3246       LLVMValueRef total_emitted_vertices_vec =
3247          LLVMBuildLoad(builder, bld->total_emitted_vertices_vec_ptr, "");
3248       mask = clamp_mask_to_max_output_vertices(bld, mask,
3249                                                total_emitted_vertices_vec);
3250       gather_outputs(bld);
3251       bld->gs_iface->emit_vertex(bld->gs_iface, &bld->bld_base,
3252                                  bld->outputs,
3253                                  total_emitted_vertices_vec);
3254       increment_vec_ptr_by_mask(bld_base, bld->emitted_vertices_vec_ptr,
3255                                 mask);
3256       increment_vec_ptr_by_mask(bld_base, bld->total_emitted_vertices_vec_ptr,
3257                                 mask);
3258 #if DUMP_GS_EMITS
3259       lp_build_print_value(bld->bld_base.base.gallivm,
3260                            " +++ emit vertex masked ones = ",
3261                            mask);
3262       lp_build_print_value(bld->bld_base.base.gallivm,
3263                            " +++ emit vertex emitted = ",
3264                            total_emitted_vertices_vec);
3265 #endif
3266    }
3267 }
3268
3269
3270 static void
3271 end_primitive_masked(struct lp_build_tgsi_context * bld_base,
3272                      LLVMValueRef mask)
3273 {
3274    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3275    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
3276
3277    if (bld->gs_iface->end_primitive) {
3278       struct lp_build_context *uint_bld = &bld_base->uint_bld;
3279       LLVMValueRef emitted_vertices_vec =
3280          LLVMBuildLoad(builder, bld->emitted_vertices_vec_ptr, "");
3281       LLVMValueRef emitted_prims_vec =
3282          LLVMBuildLoad(builder, bld->emitted_prims_vec_ptr, "");
3283
3284       LLVMValueRef emitted_mask = lp_build_cmp(uint_bld, PIPE_FUNC_NOTEQUAL,
3285                                                emitted_vertices_vec,
3286                                                uint_bld->zero);
3287       /* We need to combine the current execution mask with the mask
3288          telling us which, if any, execution slots actually have
3289          unemitted primitives, this way we make sure that end_primitives
3290          executes only on the paths that have unflushed vertices */
3291       mask = LLVMBuildAnd(builder, mask, emitted_mask, "");
3292
3293       bld->gs_iface->end_primitive(bld->gs_iface, &bld->bld_base,
3294                                    emitted_vertices_vec,
3295                                    emitted_prims_vec);
3296
3297 #if DUMP_GS_EMITS
3298       lp_build_print_value(bld->bld_base.base.gallivm,
3299                            " +++ end prim masked ones = ",
3300                            mask);
3301       lp_build_print_value(bld->bld_base.base.gallivm,
3302                            " +++ end prim emitted verts1 = ",
3303                            emitted_vertices_vec);
3304       lp_build_print_value(bld->bld_base.base.gallivm,
3305                            " +++ end prim emitted prims1 = ",
3306                            LLVMBuildLoad(builder,
3307                                          bld->emitted_prims_vec_ptr, ""));
3308 #endif
3309       increment_vec_ptr_by_mask(bld_base, bld->emitted_prims_vec_ptr,
3310                                 mask);
3311       clear_uint_vec_ptr_from_mask(bld_base, bld->emitted_vertices_vec_ptr,
3312                                    mask);
3313 #if DUMP_GS_EMITS
3314       lp_build_print_value(bld->bld_base.base.gallivm,
3315                            " +++ end prim emitted verts2 = ",
3316                            LLVMBuildLoad(builder,
3317                                          bld->emitted_vertices_vec_ptr, ""));
3318 #endif
3319    }
3320
3321 }
3322
3323 static void
3324 end_primitive(
3325    const struct lp_build_tgsi_action * action,
3326    struct lp_build_tgsi_context * bld_base,
3327    struct lp_build_emit_data * emit_data)
3328 {
3329    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3330
3331    if (bld->gs_iface->end_primitive) {
3332       LLVMValueRef mask = mask_vec(bld_base);
3333       end_primitive_masked(bld_base, mask);
3334    }
3335 }
3336
3337 static void
3338 cal_emit(
3339    const struct lp_build_tgsi_action * action,
3340    struct lp_build_tgsi_context * bld_base,
3341    struct lp_build_emit_data * emit_data)
3342 {
3343    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3344
3345    lp_exec_mask_call(&bld->exec_mask, emit_data->inst->Label.Label,
3346                      &bld_base->pc);
3347 }
3348
3349 static void
3350 ret_emit(
3351    const struct lp_build_tgsi_action * action,
3352    struct lp_build_tgsi_context * bld_base,
3353    struct lp_build_emit_data * emit_data)
3354 {
3355    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3356
3357    lp_exec_mask_ret(&bld->exec_mask, &bld_base->pc);
3358 }
3359
3360 static void
3361 brk_emit(
3362    const struct lp_build_tgsi_action * action,
3363    struct lp_build_tgsi_context * bld_base,
3364    struct lp_build_emit_data * emit_data)
3365 {
3366    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3367
3368    lp_exec_break(&bld->exec_mask, bld_base);
3369 }
3370
3371 static void
3372 breakc_emit(
3373    const struct lp_build_tgsi_action * action,
3374    struct lp_build_tgsi_context * bld_base,
3375    struct lp_build_emit_data * emit_data)
3376 {
3377    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3378    LLVMBuilderRef builder = bld_base->base.gallivm->builder;
3379    struct lp_build_context *uint_bld = &bld_base->uint_bld;
3380    LLVMValueRef unsigned_cond =
3381       LLVMBuildBitCast(builder, emit_data->args[0], uint_bld->vec_type, "");
3382    LLVMValueRef cond = lp_build_cmp(uint_bld, PIPE_FUNC_NOTEQUAL,
3383                                     unsigned_cond,
3384                                     uint_bld->zero);
3385
3386    lp_exec_break_condition(&bld->exec_mask, cond);
3387 }
3388
3389 static void
3390 if_emit(
3391    const struct lp_build_tgsi_action * action,
3392    struct lp_build_tgsi_context * bld_base,
3393    struct lp_build_emit_data * emit_data)
3394 {
3395    LLVMValueRef tmp;
3396    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3397
3398    tmp = lp_build_cmp(&bld_base->base, PIPE_FUNC_NOTEQUAL,
3399                       emit_data->args[0], bld->bld_base.base.zero);
3400    lp_exec_mask_cond_push(&bld->exec_mask, tmp);
3401 }
3402
3403 static void
3404 uif_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    LLVMValueRef tmp;
3410    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3411    struct lp_build_context *uint_bld = &bld_base->uint_bld;
3412
3413    tmp = lp_build_cmp(uint_bld, PIPE_FUNC_NOTEQUAL,
3414                       emit_data->args[0], uint_bld->zero);
3415    lp_exec_mask_cond_push(&bld->exec_mask, tmp);
3416 }
3417
3418 static void
3419 case_emit(
3420    const struct lp_build_tgsi_action * action,
3421    struct lp_build_tgsi_context * bld_base,
3422    struct lp_build_emit_data * emit_data)
3423 {
3424    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3425
3426    lp_exec_case(&bld->exec_mask, emit_data->args[0]);
3427 }
3428
3429 static void
3430 default_emit(
3431    const struct lp_build_tgsi_action * action,
3432    struct lp_build_tgsi_context * bld_base,
3433    struct lp_build_emit_data * emit_data)
3434 {
3435    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3436
3437    lp_exec_default(&bld->exec_mask, bld_base);
3438 }
3439
3440 static void
3441 switch_emit(
3442    const struct lp_build_tgsi_action * action,
3443    struct lp_build_tgsi_context * bld_base,
3444    struct lp_build_emit_data * emit_data)
3445 {
3446    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3447
3448    lp_exec_switch(&bld->exec_mask, emit_data->args[0]);
3449 }
3450
3451 static void
3452 endswitch_emit(
3453    const struct lp_build_tgsi_action * action,
3454    struct lp_build_tgsi_context * bld_base,
3455    struct lp_build_emit_data * emit_data)
3456 {
3457    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3458
3459    lp_exec_endswitch(&bld->exec_mask, bld_base);
3460 }
3461
3462 static void
3463 bgnloop_emit(
3464    const struct lp_build_tgsi_action * action,
3465    struct lp_build_tgsi_context * bld_base,
3466    struct lp_build_emit_data * emit_data)
3467 {
3468    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3469
3470    lp_exec_bgnloop(&bld->exec_mask);
3471 }
3472
3473 static void
3474 bgnsub_emit(
3475    const struct lp_build_tgsi_action * action,
3476    struct lp_build_tgsi_context * bld_base,
3477    struct lp_build_emit_data * emit_data)
3478 {
3479    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3480
3481    lp_exec_mask_bgnsub(&bld->exec_mask);
3482 }
3483
3484 static void
3485 else_emit(
3486    const struct lp_build_tgsi_action * action,
3487    struct lp_build_tgsi_context * bld_base,
3488    struct lp_build_emit_data * emit_data)
3489 {
3490    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3491
3492    lp_exec_mask_cond_invert(&bld->exec_mask);
3493 }
3494
3495 static void
3496 endif_emit(
3497    const struct lp_build_tgsi_action * action,
3498    struct lp_build_tgsi_context * bld_base,
3499    struct lp_build_emit_data * emit_data)
3500 {
3501    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3502
3503    lp_exec_mask_cond_pop(&bld->exec_mask);
3504 }
3505
3506 static void
3507 endloop_emit(
3508    const struct lp_build_tgsi_action * action,
3509    struct lp_build_tgsi_context * bld_base,
3510    struct lp_build_emit_data * emit_data)
3511 {
3512    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3513
3514    lp_exec_endloop(bld_base->base.gallivm, &bld->exec_mask);
3515 }
3516
3517 static void
3518 endsub_emit(
3519    const struct lp_build_tgsi_action * action,
3520    struct lp_build_tgsi_context * bld_base,
3521    struct lp_build_emit_data * emit_data)
3522 {
3523    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3524
3525    lp_exec_mask_endsub(&bld->exec_mask, &bld_base->pc);
3526 }
3527
3528 static void
3529 cont_emit(
3530    const struct lp_build_tgsi_action * action,
3531    struct lp_build_tgsi_context * bld_base,
3532    struct lp_build_emit_data * emit_data)
3533 {
3534    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3535
3536    lp_exec_continue(&bld->exec_mask);
3537 }
3538
3539 static void emit_prologue(struct lp_build_tgsi_context * bld_base)
3540 {
3541    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3542    struct gallivm_state * gallivm = bld_base->base.gallivm;
3543
3544    if (bld->indirect_files & (1 << TGSI_FILE_TEMPORARY)) {
3545       LLVMValueRef array_size =
3546          lp_build_const_int32(gallivm,
3547                          bld_base->info->file_max[TGSI_FILE_TEMPORARY] * 4 + 4);
3548       bld->temps_array = lp_build_array_alloca(gallivm,
3549                                               bld_base->base.vec_type, array_size,
3550                                               "temp_array");
3551    }
3552
3553    if (bld->indirect_files & (1 << TGSI_FILE_OUTPUT)) {
3554       LLVMValueRef array_size =
3555          lp_build_const_int32(gallivm,
3556                             bld_base->info->file_max[TGSI_FILE_OUTPUT] * 4 + 4);
3557       bld->outputs_array = lp_build_array_alloca(gallivm,
3558                                                 bld_base->base.vec_type, array_size,
3559                                                 "output_array");
3560    }
3561
3562    if (bld->indirect_files & (1 << TGSI_FILE_IMMEDIATE)) {
3563       LLVMValueRef array_size =
3564          lp_build_const_int32(gallivm,
3565                          bld_base->info->file_max[TGSI_FILE_IMMEDIATE] * 4 + 4);
3566       bld->imms_array = lp_build_array_alloca(gallivm,
3567                                               bld_base->base.vec_type, array_size,
3568                                               "imms_array");
3569    }
3570
3571    /* If we have indirect addressing in inputs we need to copy them into
3572     * our alloca array to be able to iterate over them */
3573    if (bld->indirect_files & (1 << TGSI_FILE_INPUT) && !bld->gs_iface) {
3574       unsigned index, chan;
3575       LLVMTypeRef vec_type = bld_base->base.vec_type;
3576       LLVMValueRef array_size = lp_build_const_int32(gallivm,
3577             bld_base->info->file_max[TGSI_FILE_INPUT]*4 + 4);
3578       bld->inputs_array = lp_build_array_alloca(gallivm,
3579                                                vec_type, array_size,
3580                                                "input_array");
3581
3582       assert(bld_base->info->num_inputs
3583                         <= bld_base->info->file_max[TGSI_FILE_INPUT] + 1);
3584
3585       for (index = 0; index < bld_base->info->num_inputs; ++index) {
3586          for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
3587             LLVMValueRef lindex =
3588                lp_build_const_int32(gallivm, index * 4 + chan);
3589             LLVMValueRef input_ptr =
3590                LLVMBuildGEP(gallivm->builder, bld->inputs_array,
3591                             &lindex, 1, "");
3592             LLVMValueRef value = bld->inputs[index][chan];
3593             if (value)
3594                LLVMBuildStore(gallivm->builder, value, input_ptr);
3595          }
3596       }
3597    }
3598
3599    if (bld->gs_iface) {
3600       struct lp_build_context *uint_bld = &bld->bld_base.uint_bld;
3601       bld->emitted_prims_vec_ptr =
3602          lp_build_alloca(gallivm,
3603                          uint_bld->vec_type,
3604                          "emitted_prims_ptr");
3605       bld->emitted_vertices_vec_ptr =
3606          lp_build_alloca(gallivm,
3607                          uint_bld->vec_type,
3608                          "emitted_vertices_ptr");
3609       bld->total_emitted_vertices_vec_ptr =
3610          lp_build_alloca(gallivm,
3611                          uint_bld->vec_type,
3612                          "total_emitted_vertices_ptr");
3613
3614       LLVMBuildStore(gallivm->builder, uint_bld->zero,
3615                      bld->emitted_prims_vec_ptr);
3616       LLVMBuildStore(gallivm->builder, uint_bld->zero,
3617                      bld->emitted_vertices_vec_ptr);
3618       LLVMBuildStore(gallivm->builder, uint_bld->zero,
3619                      bld->total_emitted_vertices_vec_ptr);
3620    }
3621
3622    if (DEBUG_EXECUTION) {
3623       lp_build_printf(gallivm, "\n");
3624       emit_dump_file(bld, TGSI_FILE_CONSTANT);
3625       if (!bld->gs_iface)
3626          emit_dump_file(bld, TGSI_FILE_INPUT);
3627    }
3628 }
3629
3630 static void emit_epilogue(struct lp_build_tgsi_context * bld_base)
3631 {
3632    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3633    LLVMBuilderRef builder = bld_base->base.gallivm->builder;
3634
3635    if (DEBUG_EXECUTION) {
3636       /* for debugging */
3637       if (0) {
3638          emit_dump_file(bld, TGSI_FILE_TEMPORARY);
3639       }
3640       emit_dump_file(bld, TGSI_FILE_OUTPUT);
3641       lp_build_printf(bld_base->base.gallivm, "\n");
3642    }
3643
3644    /* If we have indirect addressing in outputs we need to copy our alloca array
3645     * to the outputs slots specified by the caller */
3646    if (bld->gs_iface) {
3647       LLVMValueRef total_emitted_vertices_vec;
3648       LLVMValueRef emitted_prims_vec;
3649       /* implicit end_primitives, needed in case there are any unflushed
3650          vertices in the cache. Note must not call end_primitive here
3651          since the exec_mask is not valid at this point. */
3652       end_primitive_masked(bld_base, lp_build_mask_value(bld->mask));
3653
3654       total_emitted_vertices_vec =
3655          LLVMBuildLoad(builder, bld->total_emitted_vertices_vec_ptr, "");
3656       emitted_prims_vec =
3657          LLVMBuildLoad(builder, bld->emitted_prims_vec_ptr, "");
3658
3659       bld->gs_iface->gs_epilogue(bld->gs_iface,
3660                                  &bld->bld_base,
3661                                  total_emitted_vertices_vec,
3662                                  emitted_prims_vec);
3663    } else {
3664       gather_outputs(bld);
3665    }
3666 }
3667
3668 void
3669 lp_build_tgsi_soa(struct gallivm_state *gallivm,
3670                   const struct tgsi_token *tokens,
3671                   struct lp_type type,
3672                   struct lp_build_mask_context *mask,
3673                   LLVMValueRef consts_ptr,
3674                   LLVMValueRef const_sizes_ptr,
3675                   const struct lp_bld_tgsi_system_values *system_values,
3676                   const LLVMValueRef (*inputs)[TGSI_NUM_CHANNELS],
3677                   LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
3678                   LLVMValueRef context_ptr,
3679                   struct lp_build_sampler_soa *sampler,
3680                   const struct tgsi_shader_info *info,
3681                   const struct lp_build_tgsi_gs_iface *gs_iface)
3682 {
3683    struct lp_build_tgsi_soa_context bld;
3684
3685    struct lp_type res_type;
3686
3687    assert(type.length <= LP_MAX_VECTOR_LENGTH);
3688    memset(&res_type, 0, sizeof res_type);
3689    res_type.width = type.width;
3690    res_type.length = type.length;
3691    res_type.sign = 1;
3692
3693    /* Setup build context */
3694    memset(&bld, 0, sizeof bld);
3695    lp_build_context_init(&bld.bld_base.base, gallivm, type);
3696    lp_build_context_init(&bld.bld_base.uint_bld, gallivm, lp_uint_type(type));
3697    lp_build_context_init(&bld.bld_base.int_bld, gallivm, lp_int_type(type));
3698    lp_build_context_init(&bld.elem_bld, gallivm, lp_elem_type(type));
3699    bld.mask = mask;
3700    bld.inputs = inputs;
3701    bld.outputs = outputs;
3702    bld.consts_ptr = consts_ptr;
3703    bld.const_sizes_ptr = const_sizes_ptr;
3704    bld.sampler = sampler;
3705    bld.bld_base.info = info;
3706    bld.indirect_files = info->indirect_files;
3707    bld.context_ptr = context_ptr;
3708
3709    /*
3710     * If the number of temporaries is rather large then we just
3711     * allocate them as an array right from the start and treat
3712     * like indirect temporaries.
3713     */
3714    if (info->file_max[TGSI_FILE_TEMPORARY] >= LP_MAX_INLINED_TEMPS) {
3715       bld.indirect_files |= (1 << TGSI_FILE_TEMPORARY);
3716    }
3717    /*
3718     * For performance reason immediates are always backed in a static
3719     * array, but if their number is too great, we have to use just
3720     * a dynamically allocated array.
3721     */
3722    bld.use_immediates_array =
3723          (info->file_max[TGSI_FILE_IMMEDIATE] >= LP_MAX_INLINED_IMMEDIATES);
3724    if (bld.use_immediates_array) {
3725       bld.indirect_files |= (1 << TGSI_FILE_IMMEDIATE);
3726    }
3727
3728
3729    bld.bld_base.soa = TRUE;
3730    bld.bld_base.emit_debug = emit_debug;
3731    bld.bld_base.emit_fetch_funcs[TGSI_FILE_CONSTANT] = emit_fetch_constant;
3732    bld.bld_base.emit_fetch_funcs[TGSI_FILE_IMMEDIATE] = emit_fetch_immediate;
3733    bld.bld_base.emit_fetch_funcs[TGSI_FILE_INPUT] = emit_fetch_input;
3734    bld.bld_base.emit_fetch_funcs[TGSI_FILE_TEMPORARY] = emit_fetch_temporary;
3735    bld.bld_base.emit_fetch_funcs[TGSI_FILE_SYSTEM_VALUE] = emit_fetch_system_value;
3736    bld.bld_base.emit_store = emit_store;
3737
3738    bld.bld_base.emit_declaration = lp_emit_declaration_soa;
3739    bld.bld_base.emit_immediate = lp_emit_immediate_soa;
3740
3741    bld.bld_base.emit_prologue = emit_prologue;
3742    bld.bld_base.emit_epilogue = emit_epilogue;
3743
3744    /* Set opcode actions */
3745    lp_set_default_actions_cpu(&bld.bld_base);
3746
3747    bld.bld_base.op_actions[TGSI_OPCODE_BGNLOOP].emit = bgnloop_emit;
3748    bld.bld_base.op_actions[TGSI_OPCODE_BGNSUB].emit = bgnsub_emit;
3749    bld.bld_base.op_actions[TGSI_OPCODE_BRK].emit = brk_emit;
3750    bld.bld_base.op_actions[TGSI_OPCODE_BREAKC].emit = breakc_emit;
3751    bld.bld_base.op_actions[TGSI_OPCODE_CAL].emit = cal_emit;
3752    bld.bld_base.op_actions[TGSI_OPCODE_CASE].emit = case_emit;
3753    bld.bld_base.op_actions[TGSI_OPCODE_CONT].emit = cont_emit;
3754    bld.bld_base.op_actions[TGSI_OPCODE_DDX].emit = ddx_emit;
3755    bld.bld_base.op_actions[TGSI_OPCODE_DDY].emit = ddy_emit;
3756    bld.bld_base.op_actions[TGSI_OPCODE_DEFAULT].emit = default_emit;
3757    bld.bld_base.op_actions[TGSI_OPCODE_ELSE].emit = else_emit;
3758    bld.bld_base.op_actions[TGSI_OPCODE_ENDIF].emit = endif_emit;
3759    bld.bld_base.op_actions[TGSI_OPCODE_ENDLOOP].emit = endloop_emit;
3760    bld.bld_base.op_actions[TGSI_OPCODE_ENDSUB].emit = endsub_emit;
3761    bld.bld_base.op_actions[TGSI_OPCODE_ENDSWITCH].emit = endswitch_emit;
3762    bld.bld_base.op_actions[TGSI_OPCODE_IF].emit = if_emit;
3763    bld.bld_base.op_actions[TGSI_OPCODE_UIF].emit = uif_emit;
3764    bld.bld_base.op_actions[TGSI_OPCODE_KILL_IF].emit = kill_if_emit;
3765    bld.bld_base.op_actions[TGSI_OPCODE_KILL].emit = kill_emit;
3766    bld.bld_base.op_actions[TGSI_OPCODE_RET].emit = ret_emit;
3767    bld.bld_base.op_actions[TGSI_OPCODE_SWITCH].emit = switch_emit;
3768    bld.bld_base.op_actions[TGSI_OPCODE_TEX].emit = tex_emit;
3769    bld.bld_base.op_actions[TGSI_OPCODE_TXB].emit = txb_emit;
3770    bld.bld_base.op_actions[TGSI_OPCODE_TXD].emit = txd_emit;
3771    bld.bld_base.op_actions[TGSI_OPCODE_TXL].emit = txl_emit;
3772    bld.bld_base.op_actions[TGSI_OPCODE_TXP].emit = txp_emit;
3773    bld.bld_base.op_actions[TGSI_OPCODE_TXQ].emit = txq_emit;
3774    bld.bld_base.op_actions[TGSI_OPCODE_TXF].emit = txf_emit;
3775    bld.bld_base.op_actions[TGSI_OPCODE_TEX2].emit = tex2_emit;
3776    bld.bld_base.op_actions[TGSI_OPCODE_TXB2].emit = txb2_emit;
3777    bld.bld_base.op_actions[TGSI_OPCODE_TXL2].emit = txl2_emit;
3778    /* DX10 sampling ops */
3779    bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE].emit = sample_emit;
3780    bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE_B].emit = sample_b_emit;
3781    bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE_C].emit = sample_c_emit;
3782    bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE_C_LZ].emit = sample_c_lz_emit;
3783    bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE_D].emit = sample_d_emit;
3784    bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE_I].emit = sample_i_emit;
3785    bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE_L].emit = sample_l_emit;
3786    bld.bld_base.op_actions[TGSI_OPCODE_SVIEWINFO].emit = sviewinfo_emit;
3787
3788    if (gs_iface) {
3789       /* There's no specific value for this because it should always
3790        * be set, but apps using ext_geometry_shader4 quite often
3791        * were forgetting so we're using MAX_VERTEX_VARYING from
3792        * that spec even though we could debug_assert if it's not
3793        * set, but that's a lot uglier. */
3794       uint max_output_vertices;
3795
3796       /* inputs are always indirect with gs */
3797       bld.indirect_files |= (1 << TGSI_FILE_INPUT);
3798       bld.gs_iface = gs_iface;
3799       bld.bld_base.emit_fetch_funcs[TGSI_FILE_INPUT] = emit_fetch_gs_input;
3800       bld.bld_base.op_actions[TGSI_OPCODE_EMIT].emit = emit_vertex;
3801       bld.bld_base.op_actions[TGSI_OPCODE_ENDPRIM].emit = end_primitive;
3802
3803       max_output_vertices =
3804             info->properties[TGSI_PROPERTY_GS_MAX_OUTPUT_VERTICES];
3805       if (!max_output_vertices)
3806          max_output_vertices = 32;
3807
3808       bld.max_output_vertices_vec =
3809          lp_build_const_int_vec(gallivm, bld.bld_base.int_bld.type,
3810                                 max_output_vertices);
3811    }
3812
3813    lp_exec_mask_init(&bld.exec_mask, &bld.bld_base.int_bld);
3814
3815    bld.system_values = *system_values;
3816
3817    lp_build_tgsi_llvm(&bld.bld_base, tokens);
3818
3819    if (0) {
3820       LLVMBasicBlockRef block = LLVMGetInsertBlock(gallivm->builder);
3821       LLVMValueRef function = LLVMGetBasicBlockParent(block);
3822       debug_printf("11111111111111111111111111111 \n");
3823       tgsi_dump(tokens, 0);
3824       lp_debug_dump_value(function);
3825       debug_printf("2222222222222222222222222222 \n");
3826    }
3827
3828    if (0) {
3829       LLVMModuleRef module = LLVMGetGlobalParent(
3830          LLVMGetBasicBlockParent(LLVMGetInsertBlock(gallivm->builder)));
3831       LLVMDumpModule(module);
3832
3833    }
3834    lp_exec_mask_fini(&bld.exec_mask);
3835 }