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_tgsi_context *bld_base,
 948              LLVMValueRef base_ptr,
 949              LLVMValueRef indexes,
 950              LLVMValueRef overflow_mask)
 951 {
 952    struct gallivm_state *gallivm = bld_base->base.gallivm;
 953    LLVMBuilderRef builder = gallivm->builder;
 954    struct lp_build_context *uint_bld = &bld_base->uint_bld;
 955    struct lp_build_context *bld = &bld_base->base;
 956    LLVMValueRef res = bld->undef;
 957    unsigned i;
 958
 959    /*
 960     * overflow_mask is a vector telling us which channels
 961     * in the vector overflowed. We use the overflow behavior for
 962     * constant buffers which is defined as:
 963     * Out of bounds access to constant buffer returns 0 in all
 964     * components. Out of bounds behavior is always with respect
 965     * to the size of the buffer bound at that slot.
 966     */
 967
 968    if (overflow_mask) {
 969       /*
 970        * We avoid per-element control flow here (also due to llvm going crazy,
 971        * though I suspect it's better anyway since overflow is likely rare).
 972        * Note that since we still fetch from buffers even if num_elements was
 973        * zero (in this case we'll fetch from index zero) the jit func callers
 974        * MUST provide valid fake constant buffers of size 4x32 (the values do
 975        * not matter), otherwise we'd still need (not per element though)
 976        * control flow.
 977        */
 978       indexes = lp_build_select(uint_bld, overflow_mask, uint_bld->zero, indexes);
 979    }
 980
 981    /*
 982     * Loop over elements of index_vec, load scalar value, insert it into 'res'.
 983     */
 984    for (i = 0; i < bld->type.length; i++) {
 985       LLVMValueRef ii = lp_build_const_int32(bld->gallivm, i);
 986       LLVMValueRef index = LLVMBuildExtractElement(builder,
 987                                                    indexes, ii, "");
 988       LLVMValueRef scalar_ptr, scalar;
 989
 990       scalar_ptr = LLVMBuildGEP(builder, base_ptr,
 991                                 &index, 1, "gather_ptr");
 992       scalar = LLVMBuildLoad(builder, scalar_ptr, "");
 993
 994       res = LLVMBuildInsertElement(builder, res, scalar, ii, "");
 995    }
 996
 997    if (overflow_mask) {
 998       res = lp_build_select(bld, overflow_mask, bld->zero, res);
 999    }
1000
1001    return res;
1002 }
1003
1004
1005 /**
1006  * Scatter/store vector.
1007  */
1008 static void
1009 emit_mask_scatter(struct lp_build_tgsi_soa_context *bld,
1010                   LLVMValueRef base_ptr,
1011                   LLVMValueRef indexes,
1012                   LLVMValueRef values,
1013                   struct lp_exec_mask *mask,
1014                   LLVMValueRef pred)
1015 {
1016    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
1017    LLVMBuilderRef builder = gallivm->builder;
1018    unsigned i;
1019
1020    /* Mix the predicate and execution mask */
1021    if (mask->has_mask) {
1022       if (pred) {
1023          pred = LLVMBuildAnd(builder, pred, mask->exec_mask, "");
1024       }
1025       else {
1026          pred = mask->exec_mask;
1027       }
1028    }
1029
1030    /*
1031     * Loop over elements of index_vec, store scalar value.
1032     */
1033    for (i = 0; i < bld->bld_base.base.type.length; i++) {
1034       LLVMValueRef ii = lp_build_const_int32(gallivm, i);
1035       LLVMValueRef index = LLVMBuildExtractElement(builder, indexes, ii, "");
1036       LLVMValueRef scalar_ptr = LLVMBuildGEP(builder, base_ptr, &index, 1, "scatter_ptr");
1037       LLVMValueRef val = LLVMBuildExtractElement(builder, values, ii, "scatter_val");
1038       LLVMValueRef scalar_pred = pred ?
1039          LLVMBuildExtractElement(builder, pred, ii, "scatter_pred") : NULL;
1040
1041       if (0)
1042          lp_build_printf(gallivm, "scatter %d: val %f at %d %p\n",
1043                          ii, val, index, scalar_ptr);
1044
1045       if (scalar_pred) {
1046          LLVMValueRef real_val, dst_val;
1047          dst_val = LLVMBuildLoad(builder, scalar_ptr, "");
1048          real_val = lp_build_select(&bld->elem_bld, scalar_pred, val, dst_val);
1049          LLVMBuildStore(builder, real_val, scalar_ptr);
1050       }
1051       else {
1052          LLVMBuildStore(builder, val, scalar_ptr);
1053       }
1054    }
1055 }
1056
1057
1058 /**
1059  * Read the current value of the ADDR register, convert the floats to
1060  * ints, add the base index and return the vector of offsets.
1061  * The offsets will be used to index into the constant buffer or
1062  * temporary register file.
1063  */
1064 static LLVMValueRef
1065 get_indirect_index(struct lp_build_tgsi_soa_context *bld,
1066                    unsigned reg_file, unsigned reg_index,
1067                    const struct tgsi_ind_register *indirect_reg)
1068 {
1069    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
1070    struct lp_build_context *uint_bld = &bld->bld_base.uint_bld;
1071    /* always use X component of address register */
1072    unsigned swizzle = indirect_reg->Swizzle;
1073    LLVMValueRef base;
1074    LLVMValueRef rel;
1075    LLVMValueRef max_index;
1076    LLVMValueRef index;
1077
1078    assert(bld->indirect_files & (1 << reg_file));
1079
1080    base = lp_build_const_int_vec(bld->bld_base.base.gallivm, uint_bld->type, reg_index);
1081
1082    assert(swizzle < 4);
1083    switch (indirect_reg->File) {
1084    case TGSI_FILE_ADDRESS:
1085       rel = LLVMBuildLoad(builder,
1086                           bld->addr[indirect_reg->Index][swizzle],
1087                           "load addr reg");
1088       /* ADDR LLVM values already have LLVM integer type. */
1089       break;
1090    case TGSI_FILE_TEMPORARY:
1091       rel = lp_get_temp_ptr_soa(bld, indirect_reg->Index, swizzle);
1092       rel = LLVMBuildLoad(builder, rel, "load temp reg");
1093       /* TEMP LLVM values always have LLVM float type, but for indirection, the
1094        * value actually stored is expected to be an integer */
1095       rel = LLVMBuildBitCast(builder, rel, uint_bld->vec_type, "");
1096       break;
1097    default:
1098       assert(0);
1099       rel = uint_bld->zero;
1100    }
1101
1102    index = lp_build_add(uint_bld, base, rel);
1103
1104    /*
1105     * emit_fetch_constant handles constant buffer overflow so this code
1106     * is pointless for them.
1107     * Furthermore the D3D10 spec in section 6.5 says:
1108     * If the constant buffer bound to a slot is larger than the size
1109     * declared in the shader for that slot, implementations are allowed
1110     * to return incorrect data (not necessarily 0) for indices that are
1111     * larger than the declared size but smaller than the buffer size.
1112     */
1113    if (reg_file != TGSI_FILE_CONSTANT) {
1114       max_index = lp_build_const_int_vec(bld->bld_base.base.gallivm,
1115                                          uint_bld->type,
1116                                          bld->bld_base.info->file_max[reg_file]);
1117
1118       assert(!uint_bld->type.sign);
1119       index = lp_build_min(uint_bld, index, max_index);
1120    }
1121
1122    return index;
1123 }
1124
1125 static struct lp_build_context *
1126 stype_to_fetch(struct lp_build_tgsi_context * bld_base,
1127                enum tgsi_opcode_type stype)
1128 {
1129    struct lp_build_context *bld_fetch;
1130
1131    switch (stype) {
1132    case TGSI_TYPE_FLOAT:
1133    case TGSI_TYPE_UNTYPED:
1134       bld_fetch = &bld_base->base;
1135       break;
1136    case TGSI_TYPE_UNSIGNED:
1137       bld_fetch = &bld_base->uint_bld;
1138       break;
1139    case TGSI_TYPE_SIGNED:
1140       bld_fetch = &bld_base->int_bld;
1141       break;
1142    case TGSI_TYPE_VOID:
1143    case TGSI_TYPE_DOUBLE:
1144    default:
1145       assert(0);
1146       bld_fetch = NULL;
1147       break;
1148    }
1149    return bld_fetch;
1150 }
1151
1152 static LLVMValueRef
1153 get_soa_array_offsets(struct lp_build_context *uint_bld,
1154                       LLVMValueRef indirect_index,
1155                       unsigned chan_index,
1156                       boolean need_perelement_offset)
1157 {
1158    struct gallivm_state *gallivm = uint_bld->gallivm;
1159    LLVMValueRef chan_vec =
1160       lp_build_const_int_vec(uint_bld->gallivm, uint_bld->type, chan_index);
1161    LLVMValueRef length_vec =
1162       lp_build_const_int_vec(gallivm, uint_bld->type, uint_bld->type.length);
1163    LLVMValueRef index_vec;
1164
1165    /* index_vec = (indirect_index * 4 + chan_index) * length + offsets */
1166    index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
1167    index_vec = lp_build_add(uint_bld, index_vec, chan_vec);
1168    index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
1169
1170    if (need_perelement_offset) {
1171       LLVMValueRef pixel_offsets;
1172       int i;
1173      /* build pixel offset vector: {0, 1, 2, 3, ...} */
1174       pixel_offsets = uint_bld->undef;
1175       for (i = 0; i < uint_bld->type.length; i++) {
1176          LLVMValueRef ii = lp_build_const_int32(gallivm, i);
1177          pixel_offsets = LLVMBuildInsertElement(gallivm->builder, pixel_offsets,
1178                                                 ii, ii, "");
1179       }
1180       index_vec = lp_build_add(uint_bld, index_vec, pixel_offsets);
1181    }
1182    return index_vec;
1183 }
1184
1185 static LLVMValueRef
1186 emit_fetch_constant(
1187    struct lp_build_tgsi_context * bld_base,
1188    const struct tgsi_full_src_register * reg,
1189    enum tgsi_opcode_type stype,
1190    unsigned swizzle)
1191 {
1192    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1193    struct gallivm_state *gallivm = bld_base->base.gallivm;
1194    LLVMBuilderRef builder = gallivm->builder;
1195    struct lp_build_context *uint_bld = &bld_base->uint_bld;
1196    unsigned dimension = 0;
1197    LLVMValueRef consts_ptr;
1198    LLVMValueRef num_consts;
1199    LLVMValueRef res;
1200
1201    /* XXX: Handle fetching xyzw components as a vector */
1202    assert(swizzle != ~0);
1203
1204    if (reg->Register.Dimension) {
1205       assert(!reg->Dimension.Indirect);
1206       dimension = reg->Dimension.Index;
1207       assert(dimension < LP_MAX_TGSI_CONST_BUFFERS);
1208    }
1209
1210    consts_ptr = bld->consts[dimension];
1211    num_consts = bld->consts_sizes[dimension];
1212
1213    if (reg->Register.Indirect) {
1214       LLVMValueRef indirect_index;
1215       LLVMValueRef swizzle_vec =
1216          lp_build_const_int_vec(gallivm, uint_bld->type, swizzle);
1217       LLVMValueRef index_vec;  /* index into the const buffer */
1218       LLVMValueRef overflow_mask;
1219
1220       indirect_index = get_indirect_index(bld,
1221                                           reg->Register.File,
1222                                           reg->Register.Index,
1223                                           &reg->Indirect);
1224
1225       /* All fetches are from the same constant buffer, so
1226        * we need to propagate the size to a vector to do a
1227        * vector comparison */
1228       num_consts = lp_build_broadcast_scalar(uint_bld, num_consts);
1229       /* Construct a boolean vector telling us which channels
1230        * overflow the bound constant buffer */
1231       overflow_mask = lp_build_compare(gallivm, uint_bld->type, PIPE_FUNC_GEQUAL,
1232                                        indirect_index, num_consts);
1233
1234       /* index_vec = indirect_index * 4 + swizzle */
1235       index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
1236       index_vec = lp_build_add(uint_bld, index_vec, swizzle_vec);
1237
1238       /* Gather values from the constant buffer */
1239       res = build_gather(bld_base, consts_ptr, index_vec, overflow_mask);
1240    }
1241    else {
1242       LLVMValueRef index;  /* index into the const buffer */
1243       LLVMValueRef scalar, scalar_ptr;
1244
1245       index = lp_build_const_int32(gallivm, reg->Register.Index * 4 + swizzle);
1246
1247       scalar_ptr = LLVMBuildGEP(builder, consts_ptr,
1248                                 &index, 1, "");
1249       scalar = LLVMBuildLoad(builder, scalar_ptr, "");
1250       res = lp_build_broadcast_scalar(&bld_base->base, scalar);
1251    }
1252
1253    if (stype == TGSI_TYPE_SIGNED || stype == TGSI_TYPE_UNSIGNED) {
1254       struct lp_build_context *bld_fetch = stype_to_fetch(bld_base, stype);
1255       res = LLVMBuildBitCast(builder, res, bld_fetch->vec_type, "");
1256    }
1257
1258    return res;
1259 }
1260
1261 static LLVMValueRef
1262 emit_fetch_immediate(
1263    struct lp_build_tgsi_context * bld_base,
1264    const struct tgsi_full_src_register * reg,
1265    enum tgsi_opcode_type stype,
1266    unsigned swizzle)
1267 {
1268    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1269    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
1270    LLVMBuilderRef builder = gallivm->builder;
1271    LLVMValueRef res = NULL;
1272
1273    if (bld->use_immediates_array || reg->Register.Indirect) {
1274       LLVMValueRef imms_array;
1275       LLVMTypeRef fptr_type;
1276
1277       /* cast imms_array pointer to float* */
1278       fptr_type = LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
1279       imms_array = LLVMBuildBitCast(builder, bld->imms_array, fptr_type, "");
1280
1281       if (reg->Register.Indirect) {
1282          LLVMValueRef indirect_index;
1283          LLVMValueRef index_vec;  /* index into the immediate register array */
1284
1285          indirect_index = get_indirect_index(bld,
1286                                              reg->Register.File,
1287                                              reg->Register.Index,
1288                                              &reg->Indirect);
1289          /*
1290           * Unlike for other reg classes, adding pixel offsets is unnecessary -
1291           * immediates are stored as full vectors (FIXME??? - might be better
1292           * to store them the same as constants) but all elements are the same
1293           * in any case.
1294           */
1295          index_vec = get_soa_array_offsets(&bld_base->uint_bld,
1296                                            indirect_index,
1297                                            swizzle,
1298                                            FALSE);
1299
1300          /* Gather values from the immediate register array */
1301          res = build_gather(bld_base, imms_array, index_vec, NULL);
1302       } else {
1303          LLVMValueRef lindex = lp_build_const_int32(gallivm,
1304                                         reg->Register.Index * 4 + swizzle);
1305          LLVMValueRef imms_ptr =  LLVMBuildGEP(builder,
1306                                                 bld->imms_array, &lindex, 1, "");
1307          res = LLVMBuildLoad(builder, imms_ptr, "");
1308       }
1309    }
1310    else {
1311       res = bld->immediates[reg->Register.Index][swizzle];
1312    }
1313
1314    if (stype == TGSI_TYPE_UNSIGNED) {
1315       res = LLVMBuildBitCast(builder, res, bld_base->uint_bld.vec_type, "");
1316    } else if (stype == TGSI_TYPE_SIGNED) {
1317       res = LLVMBuildBitCast(builder, res, bld_base->int_bld.vec_type, "");
1318    }
1319    return res;
1320 }
1321
1322 static LLVMValueRef
1323 emit_fetch_input(
1324    struct lp_build_tgsi_context * bld_base,
1325    const struct tgsi_full_src_register * reg,
1326    enum tgsi_opcode_type stype,
1327    unsigned swizzle)
1328 {
1329    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1330    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
1331    LLVMBuilderRef builder = gallivm->builder;
1332    LLVMValueRef res;
1333
1334    if (reg->Register.Indirect) {
1335       LLVMValueRef indirect_index;
1336       LLVMValueRef index_vec;  /* index into the input reg array */
1337       LLVMValueRef inputs_array;
1338       LLVMTypeRef fptr_type;
1339
1340       indirect_index = get_indirect_index(bld,
1341                                           reg->Register.File,
1342                                           reg->Register.Index,
1343                                           &reg->Indirect);
1344
1345       index_vec = get_soa_array_offsets(&bld_base->uint_bld,
1346                                         indirect_index,
1347                                         swizzle,
1348                                         TRUE);
1349
1350       /* cast inputs_array pointer to float* */
1351       fptr_type = LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
1352       inputs_array = LLVMBuildBitCast(builder, bld->inputs_array, fptr_type, "");
1353
1354       /* Gather values from the input register array */
1355       res = build_gather(bld_base, inputs_array, index_vec, NULL);
1356    } else {
1357       if (bld->indirect_files & (1 << TGSI_FILE_INPUT)) {
1358          LLVMValueRef lindex = lp_build_const_int32(gallivm,
1359                                         reg->Register.Index * 4 + swizzle);
1360          LLVMValueRef input_ptr =  LLVMBuildGEP(builder,
1361                                                 bld->inputs_array, &lindex, 1, "");
1362          res = LLVMBuildLoad(builder, input_ptr, "");
1363       }
1364       else {
1365          res = bld->inputs[reg->Register.Index][swizzle];
1366       }
1367    }
1368
1369    assert(res);
1370
1371    if (stype == TGSI_TYPE_UNSIGNED) {
1372       res = LLVMBuildBitCast(builder, res, bld_base->uint_bld.vec_type, "");
1373    } else if (stype == TGSI_TYPE_SIGNED) {
1374       res = LLVMBuildBitCast(builder, res, bld_base->int_bld.vec_type, "");
1375    }
1376
1377    return res;
1378 }
1379
1380
1381 static LLVMValueRef
1382 emit_fetch_gs_input(
1383    struct lp_build_tgsi_context * bld_base,
1384    const struct tgsi_full_src_register * reg,
1385    enum tgsi_opcode_type stype,
1386    unsigned swizzle)
1387 {
1388    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1389    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
1390    const struct tgsi_shader_info *info = bld->bld_base.info;
1391    LLVMBuilderRef builder = gallivm->builder;
1392    LLVMValueRef attrib_index = NULL;
1393    LLVMValueRef vertex_index = NULL;
1394    LLVMValueRef swizzle_index = lp_build_const_int32(gallivm, swizzle);
1395    LLVMValueRef res;
1396
1397    if (info->input_semantic_name[reg->Register.Index] == TGSI_SEMANTIC_PRIMID) {
1398       /* This is really a system value not a regular input */
1399       assert(!reg->Register.Indirect);
1400       assert(!reg->Dimension.Indirect);
1401       res = bld->system_values.prim_id;
1402       if (stype != TGSI_TYPE_UNSIGNED && stype != TGSI_TYPE_SIGNED) {
1403          res = LLVMBuildBitCast(builder, res, bld_base->base.vec_type, "");
1404       }
1405       return res;
1406    }
1407
1408    if (reg->Register.Indirect) {
1409       attrib_index = get_indirect_index(bld,
1410                                         reg->Register.File,
1411                                         reg->Register.Index,
1412                                         &reg->Indirect);
1413    } else {
1414       attrib_index = lp_build_const_int32(gallivm, reg->Register.Index);
1415    }
1416
1417    if (reg->Dimension.Indirect) {
1418       vertex_index = get_indirect_index(bld,
1419                                         reg->Register.File,
1420                                         reg->Dimension.Index,
1421                                         &reg->DimIndirect);
1422    } else {
1423       vertex_index = lp_build_const_int32(gallivm, reg->Dimension.Index);
1424    }
1425
1426    res = bld->gs_iface->fetch_input(bld->gs_iface, bld_base,
1427                                     reg->Dimension.Indirect,
1428                                     vertex_index,
1429                                     reg->Register.Indirect,
1430                                     attrib_index,
1431                                     swizzle_index);
1432
1433    assert(res);
1434
1435    if (stype == TGSI_TYPE_UNSIGNED) {
1436       res = LLVMBuildBitCast(builder, res, bld_base->uint_bld.vec_type, "");
1437    } else if (stype == TGSI_TYPE_SIGNED) {
1438       res = LLVMBuildBitCast(builder, res, bld_base->int_bld.vec_type, "");
1439    }
1440
1441    return res;
1442 }
1443
1444 static LLVMValueRef
1445 emit_fetch_temporary(
1446    struct lp_build_tgsi_context * bld_base,
1447    const struct tgsi_full_src_register * reg,
1448    enum tgsi_opcode_type stype,
1449    unsigned swizzle)
1450 {
1451    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1452    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
1453    LLVMBuilderRef builder = gallivm->builder;
1454    LLVMValueRef res;
1455
1456    if (reg->Register.Indirect) {
1457       LLVMValueRef indirect_index;
1458       LLVMValueRef index_vec;  /* index into the temp reg array */
1459       LLVMValueRef temps_array;
1460       LLVMTypeRef fptr_type;
1461
1462       indirect_index = get_indirect_index(bld,
1463                                           reg->Register.File,
1464                                           reg->Register.Index,
1465                                           &reg->Indirect);
1466
1467       index_vec = get_soa_array_offsets(&bld_base->uint_bld,
1468                                         indirect_index,
1469                                         swizzle,
1470                                         TRUE);
1471
1472       /* cast temps_array pointer to float* */
1473       fptr_type = LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
1474       temps_array = LLVMBuildBitCast(builder, bld->temps_array, fptr_type, "");
1475
1476       /* Gather values from the temporary register array */
1477       res = build_gather(bld_base, temps_array, index_vec, NULL);
1478    }
1479    else {
1480       LLVMValueRef temp_ptr;
1481       temp_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index, swizzle);
1482       res = LLVMBuildLoad(builder, temp_ptr, "");
1483    }
1484
1485    if (stype == TGSI_TYPE_SIGNED || stype == TGSI_TYPE_UNSIGNED) {
1486       struct lp_build_context *bld_fetch = stype_to_fetch(bld_base, stype);
1487       res = LLVMBuildBitCast(builder, res, bld_fetch->vec_type, "");
1488    }
1489
1490    return res;
1491 }
1492
1493 static LLVMValueRef
1494 emit_fetch_system_value(
1495    struct lp_build_tgsi_context * bld_base,
1496    const struct tgsi_full_src_register * reg,
1497    enum tgsi_opcode_type stype,
1498    unsigned swizzle)
1499 {
1500    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1501    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
1502    const struct tgsi_shader_info *info = bld->bld_base.info;
1503    LLVMBuilderRef builder = gallivm->builder;
1504    LLVMValueRef res;
1505    enum tgsi_opcode_type atype; // Actual type of the value
1506
1507    assert(!reg->Register.Indirect);
1508
1509    switch (info->system_value_semantic_name[reg->Register.Index]) {
1510    case TGSI_SEMANTIC_INSTANCEID:
1511       res = lp_build_broadcast_scalar(&bld_base->uint_bld, bld->system_values.instance_id);
1512       atype = TGSI_TYPE_UNSIGNED;
1513       break;
1514
1515    case TGSI_SEMANTIC_VERTEXID:
1516       res = bld->system_values.vertex_id;
1517       atype = TGSI_TYPE_UNSIGNED;
1518       break;
1519
1520    case TGSI_SEMANTIC_VERTEXID_NOBASE:
1521       res = bld->system_values.vertex_id_nobase;
1522       atype = TGSI_TYPE_UNSIGNED;
1523       break;
1524
1525    case TGSI_SEMANTIC_BASEVERTEX:
1526       res = bld->system_values.basevertex;
1527       atype = TGSI_TYPE_UNSIGNED;
1528       break;
1529
1530    case TGSI_SEMANTIC_PRIMID:
1531       res = bld->system_values.prim_id;
1532       atype = TGSI_TYPE_UNSIGNED;
1533       break;
1534
1535    case TGSI_SEMANTIC_INVOCATIONID:
1536       res = lp_build_broadcast_scalar(&bld_base->uint_bld, bld->system_values.invocation_id);
1537       atype = TGSI_TYPE_UNSIGNED;
1538       break;
1539
1540    default:
1541       assert(!"unexpected semantic in emit_fetch_system_value");
1542       res = bld_base->base.zero;
1543       atype = TGSI_TYPE_FLOAT;
1544       break;
1545    }
1546
1547    if (atype != stype) {
1548       if (stype == TGSI_TYPE_FLOAT) {
1549          res = LLVMBuildBitCast(builder, res, bld_base->base.vec_type, "");
1550       } else if (stype == TGSI_TYPE_UNSIGNED) {
1551          res = LLVMBuildBitCast(builder, res, bld_base->uint_bld.vec_type, "");
1552       } else if (stype == TGSI_TYPE_SIGNED) {
1553          res = LLVMBuildBitCast(builder, res, bld_base->int_bld.vec_type, "");
1554       }
1555    }
1556
1557    return res;
1558 }
1559
1560 /**
1561  * Register fetch with derivatives.
1562  */
1563 static void
1564 emit_fetch_deriv(
1565    struct lp_build_tgsi_soa_context *bld,
1566    LLVMValueRef src,
1567    LLVMValueRef *res,
1568    LLVMValueRef *ddx,
1569    LLVMValueRef *ddy)
1570 {
1571    if(res)
1572       *res = src;
1573
1574    /* TODO: use interpolation coeffs for inputs */
1575
1576    if(ddx)
1577       *ddx = lp_build_ddx(&bld->bld_base.base, src);
1578
1579    if(ddy)
1580       *ddy = lp_build_ddy(&bld->bld_base.base, src);
1581 }
1582
1583
1584 /**
1585  * Predicate.
1586  */
1587 static void
1588 emit_fetch_predicate(
1589    struct lp_build_tgsi_soa_context *bld,
1590    const struct tgsi_full_instruction *inst,
1591    LLVMValueRef *pred)
1592 {
1593    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
1594    unsigned index;
1595    unsigned char swizzles[4];
1596    LLVMValueRef unswizzled[4] = {NULL, NULL, NULL, NULL};
1597    LLVMValueRef value;
1598    unsigned chan;
1599
1600    if (!inst->Instruction.Predicate) {
1601       TGSI_FOR_EACH_CHANNEL( chan ) {
1602          pred[chan] = NULL;
1603       }
1604       return;
1605    }
1606
1607    swizzles[0] = inst->Predicate.SwizzleX;
1608    swizzles[1] = inst->Predicate.SwizzleY;
1609    swizzles[2] = inst->Predicate.SwizzleZ;
1610    swizzles[3] = inst->Predicate.SwizzleW;
1611
1612    index = inst->Predicate.Index;
1613    assert(index < LP_MAX_TGSI_PREDS);
1614
1615    TGSI_FOR_EACH_CHANNEL( chan ) {
1616       unsigned swizzle = swizzles[chan];
1617
1618       /*
1619        * Only fetch the predicate register channels that are actually listed
1620        * in the swizzles
1621        */
1622       if (!unswizzled[swizzle]) {
1623          value = LLVMBuildLoad(builder,
1624                                bld->preds[index][swizzle], "");
1625
1626          /*
1627           * Convert the value to an integer mask.
1628           *
1629           * TODO: Short-circuit this comparison -- a D3D setp_xx instructions
1630           * is needlessly causing two comparisons due to storing the intermediate
1631           * result as float vector instead of an integer mask vector.
1632           */
1633          value = lp_build_compare(bld->bld_base.base.gallivm,
1634                                   bld->bld_base.base.type,
1635                                   PIPE_FUNC_NOTEQUAL,
1636                                   value,
1637                                   bld->bld_base.base.zero);
1638          if (inst->Predicate.Negate) {
1639             value = LLVMBuildNot(builder, value, "");
1640          }
1641
1642          unswizzled[swizzle] = value;
1643       } else {
1644          value = unswizzled[swizzle];
1645       }
1646
1647       pred[chan] = value;
1648    }
1649 }
1650
1651
1652 /**
1653  * Register store.
1654  */
1655 static void
1656 emit_store_chan(
1657    struct lp_build_tgsi_context *bld_base,
1658    const struct tgsi_full_instruction *inst,
1659    unsigned index,
1660    unsigned chan_index,
1661    LLVMValueRef pred,
1662    LLVMValueRef value)
1663 {
1664    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1665    struct gallivm_state *gallivm = bld_base->base.gallivm;
1666    LLVMBuilderRef builder = gallivm->builder;
1667    const struct tgsi_full_dst_register *reg = &inst->Dst[index];
1668    struct lp_build_context *float_bld = &bld_base->base;
1669    struct lp_build_context *int_bld = &bld_base->int_bld;
1670    LLVMValueRef indirect_index = NULL;
1671    enum tgsi_opcode_type dtype = tgsi_opcode_infer_dst_type(inst->Instruction.Opcode);
1672
1673    /*
1674     * Apply saturation.
1675     *
1676     * It is always assumed to be float.
1677     */
1678    if (inst->Instruction.Saturate) {
1679       assert(dtype == TGSI_TYPE_FLOAT ||
1680              dtype == TGSI_TYPE_UNTYPED);
1681       value = LLVMBuildBitCast(builder, value, float_bld->vec_type, "");
1682       value = lp_build_clamp_zero_one_nanzero(float_bld, value);
1683    }
1684
1685    if (reg->Register.Indirect) {
1686       indirect_index = get_indirect_index(bld,
1687                                           reg->Register.File,
1688                                           reg->Register.Index,
1689                                           &reg->Indirect);
1690    } else {
1691       assert(reg->Register.Index <=
1692                              bld_base->info->file_max[reg->Register.File]);
1693    }
1694
1695    if (DEBUG_EXECUTION) {
1696       emit_dump_reg(gallivm, reg->Register.File, reg->Register.Index, chan_index, value);
1697    }
1698
1699    switch( reg->Register.File ) {
1700    case TGSI_FILE_OUTPUT:
1701       /* Outputs are always stored as floats */
1702       value = LLVMBuildBitCast(builder, value, float_bld->vec_type, "");
1703
1704       if (reg->Register.Indirect) {
1705          LLVMValueRef index_vec;  /* indexes into the output registers */
1706          LLVMValueRef outputs_array;
1707          LLVMTypeRef fptr_type;
1708
1709          index_vec = get_soa_array_offsets(&bld_base->uint_bld,
1710                                            indirect_index,
1711                                            chan_index,
1712                                            TRUE);
1713
1714          fptr_type = LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
1715          outputs_array = LLVMBuildBitCast(builder, bld->outputs_array, fptr_type, "");
1716
1717          /* Scatter store values into output registers */
1718          emit_mask_scatter(bld, outputs_array, index_vec, value,
1719                            &bld->exec_mask, pred);
1720       }
1721       else {
1722          LLVMValueRef out_ptr = lp_get_output_ptr(bld, reg->Register.Index,
1723                                                   chan_index);
1724          lp_exec_mask_store(&bld->exec_mask, float_bld, pred, value, out_ptr);
1725       }
1726       break;
1727
1728    case TGSI_FILE_TEMPORARY:
1729       /* Temporaries are always stored as floats */
1730       value = LLVMBuildBitCast(builder, value, float_bld->vec_type, "");
1731
1732       if (reg->Register.Indirect) {
1733          LLVMValueRef index_vec;  /* indexes into the temp registers */
1734          LLVMValueRef temps_array;
1735          LLVMTypeRef fptr_type;
1736
1737          index_vec = get_soa_array_offsets(&bld_base->uint_bld,
1738                                            indirect_index,
1739                                            chan_index,
1740                                            TRUE);
1741
1742          fptr_type = LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
1743          temps_array = LLVMBuildBitCast(builder, bld->temps_array, fptr_type, "");
1744
1745          /* Scatter store values into temp registers */
1746          emit_mask_scatter(bld, temps_array, index_vec, value,
1747                            &bld->exec_mask, pred);
1748       }
1749       else {
1750          LLVMValueRef temp_ptr;
1751          temp_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index, chan_index);
1752          lp_exec_mask_store(&bld->exec_mask, float_bld, pred, value, temp_ptr);
1753       }
1754       break;
1755
1756    case TGSI_FILE_ADDRESS:
1757       assert(dtype == TGSI_TYPE_SIGNED);
1758       assert(LLVMTypeOf(value) == int_bld->vec_type);
1759       value = LLVMBuildBitCast(builder, value, int_bld->vec_type, "");
1760       lp_exec_mask_store(&bld->exec_mask, int_bld, pred, value,
1761                          bld->addr[reg->Register.Index][chan_index]);
1762       break;
1763
1764    case TGSI_FILE_PREDICATE:
1765       assert(LLVMTypeOf(value) == float_bld->vec_type);
1766       value = LLVMBuildBitCast(builder, value, float_bld->vec_type, "");
1767       lp_exec_mask_store(&bld->exec_mask, float_bld, pred, value,
1768                          bld->preds[reg->Register.Index][chan_index]);
1769       break;
1770
1771    default:
1772       assert( 0 );
1773    }
1774
1775    (void)dtype;
1776 }
1777
1778 /*
1779  * Called at the beginning of the translation of each TGSI instruction, to
1780  * emit some debug code.
1781  */
1782 static void
1783 emit_debug(
1784    struct lp_build_tgsi_context * bld_base,
1785    const struct tgsi_full_instruction * inst,
1786    const struct tgsi_opcode_info * info)
1787
1788 {
1789    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1790
1791    if (DEBUG_EXECUTION) {
1792       /*
1793        * Dump the TGSI instruction.
1794        */
1795
1796       struct gallivm_state *gallivm = bld_base->base.gallivm;
1797       char buf[512];
1798       buf[0] = '$';
1799       buf[1] = ' ';
1800       tgsi_dump_instruction_str(inst, bld_base->pc, &buf[2], sizeof buf - 2);
1801       lp_build_printf(gallivm, buf);
1802
1803       /* Dump the execution mask.
1804        */
1805       if (bld->exec_mask.has_mask) {
1806          lp_build_print_value(gallivm, "    mask = ", bld->exec_mask.exec_mask);
1807       }
1808    }
1809 }
1810
1811 static void
1812 emit_store(
1813    struct lp_build_tgsi_context * bld_base,
1814    const struct tgsi_full_instruction * inst,
1815    const struct tgsi_opcode_info * info,
1816    LLVMValueRef dst[4])
1817
1818 {
1819    unsigned chan_index;
1820    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1821
1822    if(info->num_dst) {
1823       LLVMValueRef pred[TGSI_NUM_CHANNELS];
1824
1825       emit_fetch_predicate( bld, inst, pred );
1826
1827       TGSI_FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1828          emit_store_chan(bld_base, inst, 0, chan_index, pred[chan_index], dst[chan_index]);
1829       }
1830    }
1831 }
1832
1833 static unsigned
1834 tgsi_to_pipe_tex_target(unsigned tgsi_target)
1835 {
1836    switch (tgsi_target) {
1837    case TGSI_TEXTURE_BUFFER:
1838       return PIPE_BUFFER;
1839    case TGSI_TEXTURE_1D:
1840    case TGSI_TEXTURE_SHADOW1D:
1841       return PIPE_TEXTURE_1D;
1842    case TGSI_TEXTURE_2D:
1843    case TGSI_TEXTURE_SHADOW2D:
1844    case TGSI_TEXTURE_2D_MSAA:
1845       return PIPE_TEXTURE_2D;
1846    case TGSI_TEXTURE_3D:
1847       return PIPE_TEXTURE_3D;
1848    case TGSI_TEXTURE_CUBE:
1849    case TGSI_TEXTURE_SHADOWCUBE:
1850       return PIPE_TEXTURE_CUBE;
1851    case TGSI_TEXTURE_RECT:
1852    case TGSI_TEXTURE_SHADOWRECT:
1853       return PIPE_TEXTURE_RECT;
1854    case TGSI_TEXTURE_1D_ARRAY:
1855    case TGSI_TEXTURE_SHADOW1D_ARRAY:
1856       return PIPE_TEXTURE_1D_ARRAY;
1857    case TGSI_TEXTURE_2D_ARRAY:
1858    case TGSI_TEXTURE_SHADOW2D_ARRAY:
1859    case TGSI_TEXTURE_2D_ARRAY_MSAA:
1860       return PIPE_TEXTURE_2D_ARRAY;
1861    case TGSI_TEXTURE_CUBE_ARRAY:
1862    case TGSI_TEXTURE_SHADOWCUBE_ARRAY:
1863       return PIPE_TEXTURE_CUBE_ARRAY;
1864    default:
1865       assert(0);
1866       return PIPE_BUFFER;
1867    }
1868 }
1869
1870
1871 static enum lp_sampler_lod_property
1872 lp_build_lod_property(
1873    struct lp_build_tgsi_context *bld_base,
1874    const struct tgsi_full_instruction *inst,
1875    unsigned src_op)
1876 {
1877    const struct tgsi_full_src_register *reg = &inst->Src[src_op];
1878    enum lp_sampler_lod_property lod_property;
1879
1880    /*
1881     * Not much we can do here. We could try catching inputs declared
1882     * with constant interpolation but not sure it's worth it - since for
1883     * TEX opcodes as well as FETCH/LD the lod comes from same reg as
1884     * the coords, so it could only work for SAMPLE/TXQ/SVIEWINFO), just
1885     * like the constant/immediate recognition below.
1886     * What seems to be of more value would be to recognize temps holding
1887     * broadcasted scalars but no way we can do it.
1888     * Tried asking llvm but without any success (using LLVMIsConstant
1889     * even though this isn't exactly what we'd need), even as simple as
1890     * IMM[0] UINT32 (0,-1,0,0)
1891     * MOV TEMP[0] IMM[0].yyyy
1892     * SVIEWINFO TEMP[1], TEMP[0].xxxx, SVIEWINFO[0]
1893     * doesn't work.
1894     * This means there's ZERO chance this will ever catch a scalar lod
1895     * with traditional tex opcodes as well as texel fetches, since the lod
1896     * comes from the same reg as coords (except some test shaders using
1897     * constant coords maybe).
1898     * There's at least hope for sample opcodes as well as size queries.
1899     */
1900    if (reg->Register.File == TGSI_FILE_CONSTANT ||
1901        reg->Register.File == TGSI_FILE_IMMEDIATE) {
1902       lod_property = LP_SAMPLER_LOD_SCALAR;
1903    }
1904    else if (bld_base->info->processor == TGSI_PROCESSOR_FRAGMENT) {
1905       if (gallivm_debug & GALLIVM_DEBUG_NO_QUAD_LOD) {
1906          lod_property = LP_SAMPLER_LOD_PER_ELEMENT;
1907       }
1908       else {
1909          lod_property = LP_SAMPLER_LOD_PER_QUAD;
1910       }
1911    }
1912    else {
1913       /* never use scalar (per-quad) lod the results are just too wrong. */
1914       lod_property = LP_SAMPLER_LOD_PER_ELEMENT;
1915    }
1916    return lod_property;
1917 }
1918
1919
1920 /**
1921  * High-level instruction translators.
1922  */
1923
1924 static void
1925 emit_tex( struct lp_build_tgsi_soa_context *bld,
1926           const struct tgsi_full_instruction *inst,
1927           enum lp_build_tex_modifier modifier,
1928           LLVMValueRef *texel,
1929           unsigned sampler_reg,
1930           enum lp_sampler_op_type sampler_op)
1931 {
1932    unsigned unit = inst->Src[sampler_reg].Register.Index;
1933    LLVMValueRef oow = NULL;
1934    LLVMValueRef lod = NULL;
1935    LLVMValueRef coords[5];
1936    LLVMValueRef offsets[3] = { NULL };
1937    struct lp_derivatives derivs;
1938    struct lp_sampler_params params;
1939    enum lp_sampler_lod_property lod_property = LP_SAMPLER_LOD_SCALAR;
1940    unsigned num_derivs, num_offsets, i;
1941    unsigned shadow_coord = 0;
1942    unsigned layer_coord = 0;
1943    unsigned sample_key = sampler_op << LP_SAMPLER_OP_TYPE_SHIFT;
1944
1945    memset(&params, 0, sizeof(params));
1946
1947    if (!bld->sampler) {
1948       _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
1949       for (i = 0; i < 4; i++) {
1950          texel[i] = bld->bld_base.base.undef;
1951       }
1952       return;
1953    }
1954
1955    switch (inst->Texture.Texture) {
1956    case TGSI_TEXTURE_1D_ARRAY:
1957       layer_coord = 1;
1958       /* fallthrough */
1959    case TGSI_TEXTURE_1D:
1960       num_offsets = 1;
1961       num_derivs = 1;
1962       break;
1963    case TGSI_TEXTURE_2D_ARRAY:
1964       layer_coord = 2;
1965       /* fallthrough */
1966    case TGSI_TEXTURE_2D:
1967    case TGSI_TEXTURE_RECT:
1968       num_offsets = 2;
1969       num_derivs = 2;
1970       break;
1971    case TGSI_TEXTURE_SHADOW1D_ARRAY:
1972       layer_coord = 1;
1973       /* fallthrough */
1974    case TGSI_TEXTURE_SHADOW1D:
1975       shadow_coord = 2;
1976       num_offsets = 1;
1977       num_derivs = 1;
1978       break;
1979    case TGSI_TEXTURE_SHADOW2D_ARRAY:
1980       layer_coord = 2;
1981       shadow_coord = 3;
1982       num_offsets = 2;
1983       num_derivs = 2;
1984       break;
1985    case TGSI_TEXTURE_SHADOW2D:
1986    case TGSI_TEXTURE_SHADOWRECT:
1987       shadow_coord = 2;
1988       num_offsets = 2;
1989       num_derivs = 2;
1990       break;
1991    case TGSI_TEXTURE_CUBE:
1992       num_offsets = 2;
1993       num_derivs = 3;
1994       break;
1995    case TGSI_TEXTURE_3D:
1996       num_offsets = 3;
1997       num_derivs = 3;
1998       break;
1999    case TGSI_TEXTURE_SHADOWCUBE:
2000       shadow_coord = 3;
2001       num_offsets = 2;
2002       num_derivs = 3;
2003       break;
2004    case TGSI_TEXTURE_CUBE_ARRAY:
2005       num_offsets = 2;
2006       num_derivs = 3;
2007       layer_coord = 3;
2008       break;
2009    case TGSI_TEXTURE_SHADOWCUBE_ARRAY:
2010       num_offsets = 2;
2011       num_derivs = 3;
2012       layer_coord = 3;
2013       shadow_coord = 4; /* shadow coord special different reg */
2014       break;
2015    case TGSI_TEXTURE_2D_MSAA:
2016    case TGSI_TEXTURE_2D_ARRAY_MSAA:
2017    default:
2018       assert(0);
2019       return;
2020    }
2021
2022    /* Note lod and especially projected are illegal in a LOT of cases */
2023    if (modifier == LP_BLD_TEX_MODIFIER_LOD_BIAS ||
2024        modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_LOD) {
2025       if (inst->Texture.Texture == TGSI_TEXTURE_SHADOWCUBE ||
2026           inst->Texture.Texture == TGSI_TEXTURE_CUBE_ARRAY) {
2027          /* note that shadow cube array with bias/explicit lod does not exist */
2028          lod = lp_build_emit_fetch(&bld->bld_base, inst, 1, 0);
2029       }
2030       else {
2031          lod = lp_build_emit_fetch(&bld->bld_base, inst, 0, 3);
2032       }
2033       if (modifier == LP_BLD_TEX_MODIFIER_LOD_BIAS) {
2034          sample_key |= LP_SAMPLER_LOD_BIAS << LP_SAMPLER_LOD_CONTROL_SHIFT;
2035       }
2036       else if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_LOD) {
2037          sample_key |= LP_SAMPLER_LOD_EXPLICIT << LP_SAMPLER_LOD_CONTROL_SHIFT;
2038       }
2039       lod_property = lp_build_lod_property(&bld->bld_base, inst, 0);
2040    }
2041
2042    if (modifier == LP_BLD_TEX_MODIFIER_PROJECTED) {
2043       oow = lp_build_emit_fetch(&bld->bld_base, inst, 0, 3);
2044       oow = lp_build_rcp(&bld->bld_base.base, oow);
2045    }
2046
2047    for (i = 0; i < num_derivs; i++) {
2048       coords[i] = lp_build_emit_fetch(&bld->bld_base, inst, 0, i);
2049       if (modifier == LP_BLD_TEX_MODIFIER_PROJECTED)
2050          coords[i] = lp_build_mul(&bld->bld_base.base, coords[i], oow);
2051    }
2052    for (i = num_derivs; i < 5; i++) {
2053       coords[i] = bld->bld_base.base.undef;
2054    }
2055
2056    /* Layer coord always goes into 3rd slot, except for cube map arrays */
2057    if (layer_coord) {
2058       if (layer_coord == 3) {
2059          coords[3] = lp_build_emit_fetch(&bld->bld_base, inst, 0, layer_coord);
2060       }
2061       else {
2062          coords[2] = lp_build_emit_fetch(&bld->bld_base, inst, 0, layer_coord);
2063       }
2064       if (modifier == LP_BLD_TEX_MODIFIER_PROJECTED)
2065          coords[2] = lp_build_mul(&bld->bld_base.base, coords[2], oow);
2066    }
2067    /* Shadow coord occupies always 5th slot. */
2068    if (shadow_coord) {
2069       sample_key |= LP_SAMPLER_SHADOW;
2070       if (shadow_coord == 4) {
2071          coords[4] = lp_build_emit_fetch(&bld->bld_base, inst, 1, 0);
2072       }
2073       else {
2074          coords[4] = lp_build_emit_fetch(&bld->bld_base, inst, 0, shadow_coord);
2075       }
2076       if (modifier == LP_BLD_TEX_MODIFIER_PROJECTED)
2077          coords[4] = lp_build_mul(&bld->bld_base.base, coords[4], oow);
2078    }
2079
2080    if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV) {
2081       unsigned dim;
2082       sample_key |= LP_SAMPLER_LOD_DERIVATIVES << LP_SAMPLER_LOD_CONTROL_SHIFT;
2083       for (dim = 0; dim < num_derivs; ++dim) {
2084          derivs.ddx[dim] = lp_build_emit_fetch(&bld->bld_base, inst, 1, dim);
2085          derivs.ddy[dim] = lp_build_emit_fetch(&bld->bld_base, inst, 2, dim);
2086       }
2087       params.derivs = &derivs;
2088       /*
2089        * could also check all src regs if constant but I doubt such
2090        * cases exist in practice.
2091        */
2092       if (bld->bld_base.info->processor == TGSI_PROCESSOR_FRAGMENT) {
2093          if (gallivm_debug & GALLIVM_DEBUG_NO_QUAD_LOD) {
2094             lod_property = LP_SAMPLER_LOD_PER_ELEMENT;
2095          }
2096          else {
2097             lod_property = LP_SAMPLER_LOD_PER_QUAD;
2098          }
2099       }
2100       else {
2101          lod_property = LP_SAMPLER_LOD_PER_ELEMENT;
2102       }
2103    }
2104    sample_key |= lod_property << LP_SAMPLER_LOD_PROPERTY_SHIFT;
2105
2106    /* we don't handle the 4 offset version of tg4 */
2107    if (inst->Texture.NumOffsets == 1) {
2108       unsigned dim;
2109       sample_key |= LP_SAMPLER_OFFSETS;
2110       for (dim = 0; dim < num_offsets; dim++) {
2111          offsets[dim] = lp_build_emit_fetch_texoffset(&bld->bld_base, inst, 0, dim);
2112       }
2113    }
2114
2115    params.type = bld->bld_base.base.type;
2116    params.sample_key = sample_key;
2117    params.texture_index = unit;
2118    params.sampler_index = unit;
2119    params.context_ptr = bld->context_ptr;
2120    params.coords = coords;
2121    params.offsets = offsets;
2122    params.lod = lod;
2123    params.texel = texel;
2124
2125    bld->sampler->emit_tex_sample(bld->sampler,
2126                                  bld->bld_base.base.gallivm,
2127                                  &params);
2128 }
2129
2130 static void
2131 emit_sample(struct lp_build_tgsi_soa_context *bld,
2132             const struct tgsi_full_instruction *inst,
2133             enum lp_build_tex_modifier modifier,
2134             boolean compare,
2135             LLVMValueRef *texel)
2136 {
2137    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
2138    unsigned texture_unit, sampler_unit;
2139    LLVMValueRef lod = NULL;
2140    LLVMValueRef coords[5];
2141    LLVMValueRef offsets[3] = { NULL };
2142    struct lp_derivatives derivs;
2143    struct lp_sampler_params params;
2144    enum lp_sampler_lod_property lod_property = LP_SAMPLER_LOD_SCALAR;
2145
2146    unsigned num_offsets, num_derivs, i;
2147    unsigned layer_coord = 0;
2148    unsigned sample_key = LP_SAMPLER_OP_TEXTURE << LP_SAMPLER_OP_TYPE_SHIFT;
2149
2150    memset(&params, 0, sizeof(params));
2151
2152    if (!bld->sampler) {
2153       _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
2154       for (i = 0; i < 4; i++) {
2155          texel[i] = bld->bld_base.base.undef;
2156       }
2157       return;
2158    }
2159
2160    /*
2161     * unlike old-style tex opcodes the texture/sampler indices
2162     * always come from src1 and src2 respectively.
2163     */
2164    texture_unit = inst->Src[1].Register.Index;
2165    sampler_unit = inst->Src[2].Register.Index;
2166
2167    /*
2168     * Note inst->Texture.Texture will contain the number of offsets,
2169     * however the target information is NOT there and comes from the
2170     * declared sampler views instead.
2171     */
2172    switch (bld->sv[texture_unit].Resource) {
2173    case TGSI_TEXTURE_1D:
2174       num_offsets = 1;
2175       num_derivs = 1;
2176       break;
2177    case TGSI_TEXTURE_1D_ARRAY:
2178       layer_coord = 1;
2179       num_offsets = 1;
2180       num_derivs = 1;
2181       break;
2182    case TGSI_TEXTURE_2D:
2183    case TGSI_TEXTURE_RECT:
2184       num_offsets = 2;
2185       num_derivs = 2;
2186       break;
2187    case TGSI_TEXTURE_2D_ARRAY:
2188       layer_coord = 2;
2189       num_offsets = 2;
2190       num_derivs = 2;
2191       break;
2192    case TGSI_TEXTURE_CUBE:
2193       num_offsets = 2;
2194       num_derivs = 3;
2195       break;
2196    case TGSI_TEXTURE_3D:
2197       num_offsets = 3;
2198       num_derivs = 3;
2199       break;
2200    case TGSI_TEXTURE_CUBE_ARRAY:
2201       layer_coord = 3;
2202       num_offsets = 2;
2203       num_derivs = 3;
2204       break;
2205    default:
2206       assert(0);
2207       return;
2208    }
2209
2210    if (modifier == LP_BLD_TEX_MODIFIER_LOD_BIAS ||
2211        modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_LOD) {
2212       lod = lp_build_emit_fetch(&bld->bld_base, inst, 3, 0);
2213       if (modifier == LP_BLD_TEX_MODIFIER_LOD_BIAS) {
2214          sample_key |= LP_SAMPLER_LOD_BIAS << LP_SAMPLER_LOD_CONTROL_SHIFT;
2215       }
2216       else if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_LOD) {
2217          sample_key |= LP_SAMPLER_LOD_EXPLICIT << LP_SAMPLER_LOD_CONTROL_SHIFT;
2218       }
2219       lod_property = lp_build_lod_property(&bld->bld_base, inst, 0);
2220    }
2221    else if (modifier == LP_BLD_TEX_MODIFIER_LOD_ZERO) {
2222       /* XXX might be better to explicitly pass the level zero information */
2223       sample_key |= LP_SAMPLER_LOD_EXPLICIT << LP_SAMPLER_LOD_CONTROL_SHIFT;
2224       lod = lp_build_const_vec(gallivm, bld->bld_base.base.type, 0.0F);
2225    }
2226
2227    for (i = 0; i < num_derivs; i++) {
2228       coords[i] = lp_build_emit_fetch(&bld->bld_base, inst, 0, i);
2229    }
2230    for (i = num_derivs; i < 5; i++) {
2231       coords[i] = bld->bld_base.base.undef;
2232    }
2233
2234    /* Layer coord always goes into 3rd slot, except for cube map arrays */
2235    if (layer_coord) {
2236       if (layer_coord == 3)
2237          coords[3] = lp_build_emit_fetch(&bld->bld_base, inst, 0, layer_coord);
2238       else
2239          coords[2] = lp_build_emit_fetch(&bld->bld_base, inst, 0, layer_coord);
2240    }
2241    /* Shadow coord occupies always 5th slot. */
2242    if (compare) {
2243       sample_key |= LP_SAMPLER_SHADOW;
2244       coords[4] = lp_build_emit_fetch(&bld->bld_base, inst, 3, 0);
2245    }
2246
2247    if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV) {
2248       unsigned dim;
2249       sample_key |= LP_SAMPLER_LOD_DERIVATIVES << LP_SAMPLER_LOD_CONTROL_SHIFT;
2250       for (dim = 0; dim < num_derivs; ++dim) {
2251          derivs.ddx[dim] = lp_build_emit_fetch(&bld->bld_base, inst, 3, dim);
2252          derivs.ddy[dim] = lp_build_emit_fetch(&bld->bld_base, inst, 4, dim);
2253       }
2254       params.derivs = &derivs;
2255       /*
2256        * could also check all src regs if constant but I doubt such
2257        * cases exist in practice.
2258        */
2259       if (bld->bld_base.info->processor == TGSI_PROCESSOR_FRAGMENT) {
2260          if (gallivm_debug & GALLIVM_DEBUG_NO_QUAD_LOD) {
2261             lod_property = LP_SAMPLER_LOD_PER_ELEMENT;
2262          }
2263          else {
2264             lod_property = LP_SAMPLER_LOD_PER_QUAD;
2265          }
2266       }
2267       else {
2268          lod_property = LP_SAMPLER_LOD_PER_ELEMENT;
2269       }
2270    }
2271
2272    /* some advanced gather instructions (txgo) would require 4 offsets */
2273    if (inst->Texture.NumOffsets == 1) {
2274       unsigned dim;
2275       sample_key |= LP_SAMPLER_OFFSETS;
2276       for (dim = 0; dim < num_offsets; dim++) {
2277          offsets[dim] = lp_build_emit_fetch_texoffset(&bld->bld_base, inst, 0, dim);
2278       }
2279    }
2280    sample_key |= lod_property << LP_SAMPLER_LOD_PROPERTY_SHIFT;
2281
2282    params.type = bld->bld_base.base.type;
2283    params.sample_key = sample_key;
2284    params.texture_index = texture_unit;
2285    params.sampler_index = sampler_unit;
2286    params.context_ptr = bld->context_ptr;
2287    params.coords = coords;
2288    params.offsets = offsets;
2289    params.lod = lod;
2290    params.texel = texel;
2291
2292    bld->sampler->emit_tex_sample(bld->sampler,
2293                                  bld->bld_base.base.gallivm,
2294                                  &params);
2295
2296    if (inst->Src[1].Register.SwizzleX != PIPE_SWIZZLE_RED ||
2297        inst->Src[1].Register.SwizzleY != PIPE_SWIZZLE_GREEN ||
2298        inst->Src[1].Register.SwizzleZ != PIPE_SWIZZLE_BLUE ||
2299        inst->Src[1].Register.SwizzleW != PIPE_SWIZZLE_ALPHA) {
2300       unsigned char swizzles[4];
2301       swizzles[0] = inst->Src[1].Register.SwizzleX;
2302       swizzles[1] = inst->Src[1].Register.SwizzleY;
2303       swizzles[2] = inst->Src[1].Register.SwizzleZ;
2304       swizzles[3] = inst->Src[1].Register.SwizzleW;
2305
2306       lp_build_swizzle_soa_inplace(&bld->bld_base.base, texel, swizzles);
2307    }
2308 }
2309
2310 static void
2311 emit_fetch_texels( struct lp_build_tgsi_soa_context *bld,
2312                    const struct tgsi_full_instruction *inst,
2313                    LLVMValueRef *texel,
2314                    boolean is_samplei)
2315 {
2316    unsigned unit, target;
2317    LLVMValueRef coord_undef = LLVMGetUndef(bld->bld_base.base.int_vec_type);
2318    LLVMValueRef explicit_lod = NULL;
2319    LLVMValueRef coords[5];
2320    LLVMValueRef offsets[3] = { NULL };
2321    struct lp_sampler_params params;
2322    enum lp_sampler_lod_property lod_property = LP_SAMPLER_LOD_SCALAR;
2323    unsigned dims, i;
2324    unsigned layer_coord = 0;
2325    unsigned sample_key = LP_SAMPLER_OP_FETCH << LP_SAMPLER_OP_TYPE_SHIFT;
2326
2327    memset(&params, 0, sizeof(params));
2328
2329    if (!bld->sampler) {
2330       _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
2331       for (i = 0; i < 4; i++) {
2332          texel[i] = coord_undef;
2333       }
2334       return;
2335    }
2336
2337    unit = inst->Src[1].Register.Index;
2338
2339    if (is_samplei) {
2340       target = bld->sv[unit].Resource;
2341    }
2342    else {
2343       target = inst->Texture.Texture;
2344    }
2345
2346    switch (target) {
2347    case TGSI_TEXTURE_1D:
2348    case TGSI_TEXTURE_BUFFER:
2349       dims = 1;
2350       break;
2351    case TGSI_TEXTURE_1D_ARRAY:
2352       layer_coord = 1;
2353       dims = 1;
2354       break;
2355    case TGSI_TEXTURE_2D:
2356    case TGSI_TEXTURE_RECT:
2357    case TGSI_TEXTURE_2D_MSAA:
2358       dims = 2;
2359       break;
2360    case TGSI_TEXTURE_2D_ARRAY:
2361    case TGSI_TEXTURE_2D_ARRAY_MSAA:
2362       layer_coord = 2;
2363       dims = 2;
2364       break;
2365    case TGSI_TEXTURE_3D:
2366       dims = 3;
2367       break;
2368    default:
2369       assert(0);
2370       return;
2371    }
2372
2373    /* always have lod except for buffers and msaa targets ? */
2374    if (target != TGSI_TEXTURE_BUFFER &&
2375        target != TGSI_TEXTURE_2D_MSAA &&
2376        target != TGSI_TEXTURE_2D_ARRAY_MSAA) {
2377       sample_key |= LP_SAMPLER_LOD_EXPLICIT << LP_SAMPLER_LOD_CONTROL_SHIFT;
2378       explicit_lod = lp_build_emit_fetch(&bld->bld_base, inst, 0, 3);
2379       lod_property = lp_build_lod_property(&bld->bld_base, inst, 0);
2380    }
2381    /* XXX: for real msaa support, the w component would be the sample index. */
2382
2383    for (i = 0; i < dims; i++) {
2384       coords[i] = lp_build_emit_fetch(&bld->bld_base, inst, 0, i);
2385    }
2386    /* never use more than 3 coords here but emit_fetch_texel copies all 5 anyway */
2387    for (i = dims; i < 5; i++) {
2388       coords[i] = coord_undef;
2389    }
2390    if (layer_coord)
2391       coords[2] = lp_build_emit_fetch(&bld->bld_base, inst, 0, layer_coord);
2392
2393    if (inst->Texture.NumOffsets == 1) {
2394       unsigned dim;
2395       sample_key |= LP_SAMPLER_OFFSETS;
2396       for (dim = 0; dim < dims; dim++) {
2397          offsets[dim] = lp_build_emit_fetch_texoffset(&bld->bld_base, inst, 0, dim);
2398       }
2399    }
2400    sample_key |= lod_property << LP_SAMPLER_LOD_PROPERTY_SHIFT;
2401
2402    params.type = bld->bld_base.base.type;
2403    params.sample_key = sample_key;
2404    params.texture_index = unit;
2405    params.sampler_index = unit;
2406    params.context_ptr = bld->context_ptr;
2407    params.coords = coords;
2408    params.offsets = offsets;
2409    params.derivs = NULL;
2410    params.lod = explicit_lod;
2411    params.texel = texel;
2412
2413    bld->sampler->emit_tex_sample(bld->sampler,
2414                                  bld->bld_base.base.gallivm,
2415                                  &params);
2416
2417    if (is_samplei &&
2418        (inst->Src[1].Register.SwizzleX != PIPE_SWIZZLE_RED ||
2419         inst->Src[1].Register.SwizzleY != PIPE_SWIZZLE_GREEN ||
2420         inst->Src[1].Register.SwizzleZ != PIPE_SWIZZLE_BLUE ||
2421         inst->Src[1].Register.SwizzleW != PIPE_SWIZZLE_ALPHA)) {
2422       unsigned char swizzles[4];
2423       swizzles[0] = inst->Src[1].Register.SwizzleX;
2424       swizzles[1] = inst->Src[1].Register.SwizzleY;
2425       swizzles[2] = inst->Src[1].Register.SwizzleZ;
2426       swizzles[3] = inst->Src[1].Register.SwizzleW;
2427
2428       lp_build_swizzle_soa_inplace(&bld->bld_base.base, texel, swizzles);
2429    }
2430 }
2431
2432 static void
2433 emit_size_query( struct lp_build_tgsi_soa_context *bld,
2434                  const struct tgsi_full_instruction *inst,
2435                  LLVMValueRef *sizes_out,
2436                  boolean is_sviewinfo)
2437 {
2438    LLVMValueRef explicit_lod;
2439    enum lp_sampler_lod_property lod_property;
2440    unsigned has_lod;
2441    unsigned i;
2442    unsigned unit = inst->Src[1].Register.Index;
2443    unsigned target, pipe_target;
2444
2445    if (is_sviewinfo) {
2446       target = bld->sv[unit].Resource;
2447    }
2448    else {
2449       target = inst->Texture.Texture;
2450    }
2451    switch (target) {
2452    case TGSI_TEXTURE_BUFFER:
2453    case TGSI_TEXTURE_RECT:
2454    case TGSI_TEXTURE_SHADOWRECT:
2455       has_lod = 0;
2456       break;
2457    default:
2458       has_lod = 1;
2459       break;
2460    }
2461
2462    if (!bld->sampler) {
2463       _debug_printf("warning: found texture query instruction but no sampler generator supplied\n");
2464       for (i = 0; i < 4; i++)
2465          sizes_out[i] = bld->bld_base.int_bld.undef;
2466       return;
2467    }
2468
2469    if (has_lod) {
2470       explicit_lod = lp_build_emit_fetch(&bld->bld_base, inst, 0, 0);
2471       lod_property = lp_build_lod_property(&bld->bld_base, inst, 0);
2472    }
2473    else {
2474       explicit_lod = NULL;
2475       lod_property = LP_SAMPLER_LOD_SCALAR;
2476    }
2477
2478
2479    pipe_target = tgsi_to_pipe_tex_target(target);
2480
2481    bld->sampler->emit_size_query(bld->sampler,
2482                                  bld->bld_base.base.gallivm,
2483                                  bld->bld_base.int_bld.type,
2484                                  unit, pipe_target,
2485                                  bld->context_ptr,
2486                                  TRUE,
2487                                  lod_property,
2488                                  explicit_lod,
2489                                  sizes_out);
2490 }
2491
2492 static boolean
2493 near_end_of_shader(struct lp_build_tgsi_soa_context *bld,
2494                    int pc)
2495 {
2496    int i;
2497
2498    for (i = 0; i < 5; i++) {
2499       unsigned opcode;
2500
2501       if (pc + i >= bld->bld_base.info->num_instructions)
2502          return TRUE;
2503
2504       opcode = bld->bld_base.instructions[pc + i].Instruction.Opcode;
2505
2506       if (opcode == TGSI_OPCODE_END)
2507          return TRUE;
2508
2509       if (opcode == TGSI_OPCODE_TEX ||
2510          opcode == TGSI_OPCODE_TXP ||
2511          opcode == TGSI_OPCODE_TXD ||
2512          opcode == TGSI_OPCODE_TXB ||
2513          opcode == TGSI_OPCODE_TXL ||
2514          opcode == TGSI_OPCODE_TXF ||
2515          opcode == TGSI_OPCODE_TXQ ||
2516          opcode == TGSI_OPCODE_TEX2 ||
2517          opcode == TGSI_OPCODE_TXB2 ||
2518          opcode == TGSI_OPCODE_TXL2 ||
2519          opcode == TGSI_OPCODE_SAMPLE ||
2520          opcode == TGSI_OPCODE_SAMPLE_B ||
2521          opcode == TGSI_OPCODE_SAMPLE_C ||
2522          opcode == TGSI_OPCODE_SAMPLE_C_LZ ||
2523          opcode == TGSI_OPCODE_SAMPLE_D ||
2524          opcode == TGSI_OPCODE_SAMPLE_I ||
2525          opcode == TGSI_OPCODE_SAMPLE_L ||
2526          opcode == TGSI_OPCODE_SVIEWINFO ||
2527          opcode == TGSI_OPCODE_CAL ||
2528          opcode == TGSI_OPCODE_CALLNZ ||
2529          opcode == TGSI_OPCODE_IF ||
2530          opcode == TGSI_OPCODE_UIF ||
2531          opcode == TGSI_OPCODE_BGNLOOP ||
2532          opcode == TGSI_OPCODE_SWITCH)
2533          return FALSE;
2534    }
2535
2536    return TRUE;
2537 }
2538
2539
2540
2541 /**
2542  * Kill fragment if any of the src register values are negative.
2543  */
2544 static void
2545 emit_kill_if(
2546    struct lp_build_tgsi_soa_context *bld,
2547    const struct tgsi_full_instruction *inst,
2548    int pc)
2549 {
2550    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
2551    const struct tgsi_full_src_register *reg = &inst->Src[0];
2552    LLVMValueRef terms[TGSI_NUM_CHANNELS];
2553    LLVMValueRef mask;
2554    unsigned chan_index;
2555
2556    memset(&terms, 0, sizeof terms);
2557
2558    TGSI_FOR_EACH_CHANNEL( chan_index ) {
2559       unsigned swizzle;
2560
2561       /* Unswizzle channel */
2562       swizzle = tgsi_util_get_full_src_register_swizzle( reg, chan_index );
2563
2564       /* Check if the component has not been already tested. */
2565       assert(swizzle < TGSI_NUM_CHANNELS);
2566       if( !terms[swizzle] )
2567          /* TODO: change the comparison operator instead of setting the sign */
2568          terms[swizzle] =  lp_build_emit_fetch(&bld->bld_base, inst, 0, chan_index );
2569    }
2570
2571    mask = NULL;
2572    TGSI_FOR_EACH_CHANNEL( chan_index ) {
2573       if(terms[chan_index]) {
2574          LLVMValueRef chan_mask;
2575
2576          /*
2577           * If term < 0 then mask = 0 else mask = ~0.
2578           */
2579          chan_mask = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_GEQUAL, terms[chan_index], bld->bld_base.base.zero);
2580
2581          if(mask)
2582             mask = LLVMBuildAnd(builder, mask, chan_mask, "");
2583          else
2584             mask = chan_mask;
2585       }
2586    }
2587
2588    if (bld->exec_mask.has_mask) {
2589       LLVMValueRef invmask;
2590       invmask = LLVMBuildNot(builder, bld->exec_mask.exec_mask, "kilp");
2591       mask = LLVMBuildOr(builder, mask, invmask, "");
2592    }
2593
2594    lp_build_mask_update(bld->mask, mask);
2595    if (!near_end_of_shader(bld, pc))
2596       lp_build_mask_check(bld->mask);
2597 }
2598
2599
2600 /**
2601  * Unconditional fragment kill.
2602  * The only predication is the execution mask which will apply if
2603  * we're inside a loop or conditional.
2604  */
2605 static void
2606 emit_kill(struct lp_build_tgsi_soa_context *bld,
2607           int pc)
2608 {
2609    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
2610    LLVMValueRef mask;
2611
2612    /* For those channels which are "alive", disable fragment shader
2613     * execution.
2614     */
2615    if (bld->exec_mask.has_mask) {
2616       mask = LLVMBuildNot(builder, bld->exec_mask.exec_mask, "kilp");
2617    }
2618    else {
2619       LLVMValueRef zero = LLVMConstNull(bld->bld_base.base.int_vec_type);
2620       mask = zero;
2621    }
2622
2623    lp_build_mask_update(bld->mask, mask);
2624
2625    if (!near_end_of_shader(bld, pc))
2626       lp_build_mask_check(bld->mask);
2627 }
2628
2629
2630 /**
2631  * Emit code which will dump the value of all the temporary registers
2632  * to stdout.
2633  */
2634 static void
2635 emit_dump_file(struct lp_build_tgsi_soa_context *bld,
2636                unsigned file)
2637 {
2638    const struct tgsi_shader_info *info = bld->bld_base.info;
2639    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
2640    LLVMBuilderRef builder = gallivm->builder;
2641    LLVMValueRef reg_ptr;
2642    int index;
2643    int max_index = info->file_max[file];
2644
2645    /*
2646     * Some register files, particularly constants, can be very large,
2647     * and dumping everything could make this unusably slow.
2648     */
2649    max_index = MIN2(max_index, 32);
2650
2651    for (index = 0; index <= max_index; index++) {
2652       LLVMValueRef res;
2653       unsigned mask;
2654       int chan;
2655
2656       if (index < 8 * sizeof(unsigned) &&
2657           (info->file_mask[file] & (1 << index)) == 0)  {
2658          /* This was not declared.*/
2659          continue;
2660       }
2661
2662       if (file == TGSI_FILE_INPUT) {
2663          mask = info->input_usage_mask[index];
2664       } else {
2665          mask = TGSI_WRITEMASK_XYZW;
2666       }
2667
2668       for (chan = 0; chan < 4; chan++) {
2669          if ((mask & (1 << chan)) == 0) {
2670             /* This channel is not used.*/
2671             continue;
2672          }
2673
2674          if (file == TGSI_FILE_CONSTANT) {
2675             struct tgsi_full_src_register reg;
2676             memset(&reg, 0, sizeof reg);
2677             reg.Register.File = file;
2678             reg.Register.Index = index;
2679             reg.Register.SwizzleX = 0;
2680             reg.Register.SwizzleY = 1;
2681             reg.Register.SwizzleZ = 2;
2682             reg.Register.SwizzleW = 3;
2683
2684             res = bld->bld_base.emit_fetch_funcs[file](&bld->bld_base, &reg, TGSI_TYPE_FLOAT, chan);
2685             if (!res) {
2686                continue;
2687             }
2688          } else if (file == TGSI_FILE_INPUT) {
2689             res = bld->inputs[index][chan];
2690             if (!res) {
2691                continue;
2692             }
2693          } else if (file == TGSI_FILE_TEMPORARY) {
2694             reg_ptr = lp_get_temp_ptr_soa(bld, index, chan);
2695             assert(reg_ptr);
2696             res = LLVMBuildLoad(builder, reg_ptr, "");
2697          } else if (file == TGSI_FILE_OUTPUT) {
2698             reg_ptr = lp_get_output_ptr(bld, index, chan);
2699             assert(reg_ptr);
2700             res = LLVMBuildLoad(builder, reg_ptr, "");
2701          } else {
2702             assert(0);
2703             continue;
2704          }
2705
2706          emit_dump_reg(gallivm, file, index, chan, res);
2707       }
2708    }
2709 }
2710
2711
2712
2713 void
2714 lp_emit_declaration_soa(
2715    struct lp_build_tgsi_context *bld_base,
2716    const struct tgsi_full_declaration *decl)
2717 {
2718    struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
2719    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
2720    LLVMTypeRef vec_type = bld->bld_base.base.vec_type;
2721    const unsigned first = decl->Range.First;
2722    const unsigned last = decl->Range.Last;
2723    unsigned idx, i;
2724
2725    assert(last <= bld->bld_base.info->file_max[decl->Declaration.File]);
2726
2727    switch (decl->Declaration.File) {
2728    case TGSI_FILE_TEMPORARY:
2729       if (!(bld->indirect_files & (1 << TGSI_FILE_TEMPORARY))) {
2730          assert(last < LP_MAX_INLINED_TEMPS);
2731          for (idx = first; idx <= last; ++idx) {
2732             for (i = 0; i < TGSI_NUM_CHANNELS; i++)
2733                bld->temps[idx][i] = lp_build_alloca(gallivm, vec_type, "temp");
2734          }
2735       }
2736       break;
2737
2738    case TGSI_FILE_OUTPUT:
2739       if (!(bld->indirect_files & (1 << TGSI_FILE_OUTPUT))) {
2740          for (idx = first; idx <= last; ++idx) {
2741             for (i = 0; i < TGSI_NUM_CHANNELS; i++)
2742                bld->outputs[idx][i] = lp_build_alloca(gallivm,
2743                                                       vec_type, "output");
2744          }
2745       }
2746       break;
2747
2748    case TGSI_FILE_ADDRESS:
2749       /* ADDR registers are only allocated with an integer LLVM IR type,
2750        * as they are guaranteed to always have integers.
2751        * XXX: Not sure if this exception is worthwhile (or the whole idea of
2752        * an ADDR register for that matter).
2753        */
2754       assert(last < LP_MAX_TGSI_ADDRS);
2755       for (idx = first; idx <= last; ++idx) {
2756          assert(idx < LP_MAX_TGSI_ADDRS);
2757          for (i = 0; i < TGSI_NUM_CHANNELS; i++)
2758             bld->addr[idx][i] = lp_build_alloca(gallivm, bld_base->base.int_vec_type, "addr");
2759       }
2760       break;
2761
2762    case TGSI_FILE_PREDICATE:
2763       assert(last < LP_MAX_TGSI_PREDS);
2764       for (idx = first; idx <= last; ++idx) {
2765          for (i = 0; i < TGSI_NUM_CHANNELS; i++)
2766             bld->preds[idx][i] = lp_build_alloca(gallivm, vec_type,
2767                                                  "predicate");
2768       }
2769       break;
2770
2771    case TGSI_FILE_SAMPLER_VIEW:
2772       /*
2773        * The target stored here MUST match whatever there actually
2774        * is in the set sampler views (what about return type?).
2775        */
2776       assert(last < PIPE_MAX_SHADER_SAMPLER_VIEWS);
2777       for (idx = first; idx <= last; ++idx) {
2778          bld->sv[idx] = decl->SamplerView;
2779       }
2780       break;
2781
2782    case TGSI_FILE_CONSTANT:
2783    {
2784       /*
2785        * We could trivially fetch the per-buffer pointer when fetching the
2786        * constant, relying on llvm to figure out it's always the same pointer
2787        * anyway. However, doing so results in a huge (more than factor of 10)
2788        * slowdown in llvm compilation times for some (but not all) shaders
2789        * (more specifically, the IR optimization spends way more time in
2790        * DominatorTree::dominates). At least with llvm versions 3.1, 3.3.
2791        */
2792       unsigned idx2D = decl->Dim.Index2D;
2793       LLVMValueRef index2D = lp_build_const_int32(gallivm, idx2D);
2794       assert(idx2D < LP_MAX_TGSI_CONST_BUFFERS);
2795       bld->consts[idx2D] =
2796          lp_build_array_get(gallivm, bld->consts_ptr, index2D);
2797       bld->consts_sizes[idx2D] =
2798          lp_build_array_get(gallivm, bld->const_sizes_ptr, index2D);
2799    }
2800       break;
2801
2802    default:
2803       /* don't need to declare other vars */
2804       break;
2805    }
2806 }
2807
2808
2809 void lp_emit_immediate_soa(
2810    struct lp_build_tgsi_context *bld_base,
2811    const struct tgsi_full_immediate *imm)
2812 {
2813    struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
2814    struct gallivm_state * gallivm = bld_base->base.gallivm;
2815    LLVMValueRef imms[4];
2816    unsigned i;
2817    const uint size = imm->Immediate.NrTokens - 1;
2818    assert(size <= 4);
2819    switch (imm->Immediate.DataType) {
2820    case TGSI_IMM_FLOAT32:
2821       for( i = 0; i < size; ++i )
2822          imms[i] =
2823                lp_build_const_vec(gallivm, bld_base->base.type, imm->u[i].Float);
2824
2825       break;
2826    case TGSI_IMM_UINT32:
2827       for( i = 0; i < size; ++i ) {
2828          LLVMValueRef tmp = lp_build_const_vec(gallivm, bld_base->uint_bld.type, imm->u[i].Uint);
2829          imms[i] = LLVMConstBitCast(tmp, bld_base->base.vec_type);
2830       }
2831
2832       break;
2833    case TGSI_IMM_INT32:
2834       for( i = 0; i < size; ++i ) {
2835          LLVMValueRef tmp = lp_build_const_vec(gallivm, bld_base->int_bld.type, imm->u[i].Int);
2836          imms[i] = LLVMConstBitCast(tmp, bld_base->base.vec_type);
2837       }
2838
2839       break;
2840    }
2841    for( i = size; i < 4; ++i )
2842       imms[i] = bld_base->base.undef;
2843
2844    if (bld->use_immediates_array) {
2845       unsigned index = bld->num_immediates;
2846       struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
2847       LLVMBuilderRef builder = gallivm->builder;
2848
2849       assert(bld->indirect_files & (1 << TGSI_FILE_IMMEDIATE));
2850       for (i = 0; i < 4; ++i ) {
2851          LLVMValueRef lindex = lp_build_const_int32(
2852                   bld->bld_base.base.gallivm, index * 4 + i);
2853          LLVMValueRef imm_ptr = LLVMBuildGEP(builder,
2854                                              bld->imms_array, &lindex, 1, "");
2855          LLVMBuildStore(builder, imms[i], imm_ptr);
2856       }
2857    } else {
2858       /* simply copy the immediate values into the next immediates[] slot */
2859       unsigned i;
2860       const uint size = imm->Immediate.NrTokens - 1;
2861       assert(size <= 4);
2862       assert(bld->num_immediates < LP_MAX_INLINED_IMMEDIATES);
2863
2864       for(i = 0; i < 4; ++i )
2865          bld->immediates[bld->num_immediates][i] = imms[i];
2866
2867       if (bld->indirect_files & (1 << TGSI_FILE_IMMEDIATE)) {
2868          unsigned index = bld->num_immediates;
2869          struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
2870          LLVMBuilderRef builder = gallivm->builder;
2871          for (i = 0; i < 4; ++i ) {
2872             LLVMValueRef lindex = lp_build_const_int32(
2873                      bld->bld_base.base.gallivm, index * 4 + i);
2874             LLVMValueRef imm_ptr = LLVMBuildGEP(builder,
2875                                                 bld->imms_array, &lindex, 1, "");
2876             LLVMBuildStore(builder,
2877                            bld->immediates[index][i],
2878                            imm_ptr);
2879          }
2880       }
2881    }
2882
2883    bld->num_immediates++;
2884 }
2885
2886 static void
2887 ddx_emit(
2888    const struct lp_build_tgsi_action * action,
2889    struct lp_build_tgsi_context * bld_base,
2890    struct lp_build_emit_data * emit_data)
2891 {
2892    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2893
2894    emit_fetch_deriv(bld, emit_data->args[0], NULL,
2895                     &emit_data->output[emit_data->chan], NULL);
2896 }
2897
2898 static void
2899 ddy_emit(
2900    const struct lp_build_tgsi_action * action,
2901    struct lp_build_tgsi_context * bld_base,
2902    struct lp_build_emit_data * emit_data)
2903 {
2904    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2905
2906    emit_fetch_deriv(bld, emit_data->args[0], NULL, NULL,
2907                     &emit_data->output[emit_data->chan]);
2908 }
2909
2910 static void
2911 kill_emit(
2912    const struct lp_build_tgsi_action * action,
2913    struct lp_build_tgsi_context * bld_base,
2914    struct lp_build_emit_data * emit_data)
2915 {
2916    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2917
2918    emit_kill(bld, bld_base->pc - 1);
2919 }
2920
2921 static void
2922 kill_if_emit(
2923    const struct lp_build_tgsi_action * action,
2924    struct lp_build_tgsi_context * bld_base,
2925    struct lp_build_emit_data * emit_data)
2926 {
2927    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2928
2929    emit_kill_if(bld, emit_data->inst, bld_base->pc - 1);
2930 }
2931
2932 static void
2933 tex_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_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_NONE,
2941             emit_data->output, 1, LP_SAMPLER_OP_TEXTURE);
2942 }
2943
2944 static void
2945 tex2_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_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_NONE,
2953             emit_data->output, 2, LP_SAMPLER_OP_TEXTURE);
2954 }
2955
2956 static void
2957 txb_emit(
2958    const struct lp_build_tgsi_action * action,
2959    struct lp_build_tgsi_context * bld_base,
2960    struct lp_build_emit_data * emit_data)
2961 {
2962    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2963
2964    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_LOD_BIAS,
2965             emit_data->output, 1, LP_SAMPLER_OP_TEXTURE);
2966 }
2967
2968 static void
2969 txb2_emit(
2970    const struct lp_build_tgsi_action * action,
2971    struct lp_build_tgsi_context * bld_base,
2972    struct lp_build_emit_data * emit_data)
2973 {
2974    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2975
2976    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_LOD_BIAS,
2977             emit_data->output, 2, LP_SAMPLER_OP_TEXTURE);
2978 }
2979
2980 static void
2981 txd_emit(
2982    const struct lp_build_tgsi_action * action,
2983    struct lp_build_tgsi_context * bld_base,
2984    struct lp_build_emit_data * emit_data)
2985 {
2986    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2987
2988    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV,
2989             emit_data->output, 3, LP_SAMPLER_OP_TEXTURE);
2990 }
2991
2992 static void
2993 txl_emit(
2994    const struct lp_build_tgsi_action * action,
2995    struct lp_build_tgsi_context * bld_base,
2996    struct lp_build_emit_data * emit_data)
2997 {
2998    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2999
3000    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD,
3001             emit_data->output, 1, LP_SAMPLER_OP_TEXTURE);
3002 }
3003
3004 static void
3005 txl2_emit(
3006    const struct lp_build_tgsi_action * action,
3007    struct lp_build_tgsi_context * bld_base,
3008    struct lp_build_emit_data * emit_data)
3009 {
3010    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3011
3012    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD,
3013             emit_data->output, 2, LP_SAMPLER_OP_TEXTURE);
3014 }
3015
3016 static void
3017 txp_emit(
3018    const struct lp_build_tgsi_action * action,
3019    struct lp_build_tgsi_context * bld_base,
3020    struct lp_build_emit_data * emit_data)
3021 {
3022    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3023
3024    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_PROJECTED,
3025             emit_data->output, 1, LP_SAMPLER_OP_TEXTURE);
3026 }
3027
3028 static void
3029 tg4_emit(
3030    const struct lp_build_tgsi_action * action,
3031    struct lp_build_tgsi_context * bld_base,
3032    struct lp_build_emit_data * emit_data)
3033 {
3034    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3035
3036    emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_NONE,
3037             emit_data->output, 2, LP_SAMPLER_OP_GATHER);
3038 }
3039
3040 static void
3041 txq_emit(
3042    const struct lp_build_tgsi_action * action,
3043    struct lp_build_tgsi_context * bld_base,
3044    struct lp_build_emit_data * emit_data)
3045 {
3046    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3047
3048    emit_size_query(bld, emit_data->inst, emit_data->output, FALSE);
3049 }
3050
3051 static void
3052 txf_emit(
3053    const struct lp_build_tgsi_action * action,
3054    struct lp_build_tgsi_context * bld_base,
3055    struct lp_build_emit_data * emit_data)
3056 {
3057    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3058
3059    emit_fetch_texels(bld, emit_data->inst, emit_data->output, FALSE);
3060 }
3061
3062 static void
3063 sample_i_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_fetch_texels(bld, emit_data->inst, emit_data->output, TRUE);
3071 }
3072
3073 static void
3074 sample_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_sample(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_NONE,
3082                FALSE, emit_data->output);
3083 }
3084
3085 static void
3086 sample_b_emit(
3087    const struct lp_build_tgsi_action * action,
3088    struct lp_build_tgsi_context * bld_base,
3089    struct lp_build_emit_data * emit_data)
3090 {
3091    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3092
3093    emit_sample(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_LOD_BIAS,
3094                FALSE, emit_data->output);
3095 }
3096
3097 static void
3098 sample_c_emit(
3099    const struct lp_build_tgsi_action * action,
3100    struct lp_build_tgsi_context * bld_base,
3101    struct lp_build_emit_data * emit_data)
3102 {
3103    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3104
3105    emit_sample(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_NONE,
3106                TRUE, emit_data->output);
3107 }
3108
3109 static void
3110 sample_c_lz_emit(
3111    const struct lp_build_tgsi_action * action,
3112    struct lp_build_tgsi_context * bld_base,
3113    struct lp_build_emit_data * emit_data)
3114 {
3115    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3116
3117    emit_sample(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_LOD_ZERO,
3118                TRUE, emit_data->output);
3119 }
3120
3121 static void
3122 sample_d_emit(
3123    const struct lp_build_tgsi_action * action,
3124    struct lp_build_tgsi_context * bld_base,
3125    struct lp_build_emit_data * emit_data)
3126 {
3127    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3128
3129    emit_sample(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV,
3130                FALSE, emit_data->output);
3131 }
3132
3133 static void
3134 sample_l_emit(
3135    const struct lp_build_tgsi_action * action,
3136    struct lp_build_tgsi_context * bld_base,
3137    struct lp_build_emit_data * emit_data)
3138 {
3139    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3140
3141    emit_sample(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD,
3142                FALSE, emit_data->output);
3143 }
3144
3145 static void
3146 sviewinfo_emit(
3147    const struct lp_build_tgsi_action * action,
3148    struct lp_build_tgsi_context * bld_base,
3149    struct lp_build_emit_data * emit_data)
3150 {
3151    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3152
3153    emit_size_query(bld, emit_data->inst, emit_data->output, TRUE);
3154 }
3155
3156 static LLVMValueRef
3157 mask_vec(struct lp_build_tgsi_context *bld_base)
3158 {
3159    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3160    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
3161    struct lp_exec_mask *exec_mask = &bld->exec_mask;
3162
3163    if (!exec_mask->has_mask) {
3164       return lp_build_mask_value(bld->mask);
3165    }
3166    return LLVMBuildAnd(builder, lp_build_mask_value(bld->mask),
3167                        exec_mask->exec_mask, "");
3168 }
3169
3170 static void
3171 increment_vec_ptr_by_mask(struct lp_build_tgsi_context * bld_base,
3172                           LLVMValueRef ptr,
3173                           LLVMValueRef mask)
3174 {
3175    LLVMBuilderRef builder = bld_base->base.gallivm->builder;
3176    LLVMValueRef current_vec = LLVMBuildLoad(builder, ptr, "");
3177
3178    current_vec = LLVMBuildSub(builder, current_vec, mask, "");
3179
3180    LLVMBuildStore(builder, current_vec, ptr);
3181 }
3182
3183 static void
3184 clear_uint_vec_ptr_from_mask(struct lp_build_tgsi_context * bld_base,
3185                              LLVMValueRef ptr,
3186                              LLVMValueRef mask)
3187 {
3188    LLVMBuilderRef builder = bld_base->base.gallivm->builder;
3189    LLVMValueRef current_vec = LLVMBuildLoad(builder, ptr, "");
3190
3191    current_vec = lp_build_select(&bld_base->uint_bld,
3192                                  mask,
3193                                  bld_base->uint_bld.zero,
3194                                  current_vec);
3195
3196    LLVMBuildStore(builder, current_vec, ptr);
3197 }
3198
3199 static LLVMValueRef
3200 clamp_mask_to_max_output_vertices(struct lp_build_tgsi_soa_context * bld,
3201                                   LLVMValueRef current_mask_vec,
3202                                   LLVMValueRef total_emitted_vertices_vec)
3203 {
3204    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
3205    struct lp_build_context *int_bld = &bld->bld_base.int_bld;
3206    LLVMValueRef max_mask = lp_build_cmp(int_bld, PIPE_FUNC_LESS,
3207                                         total_emitted_vertices_vec,
3208                                         bld->max_output_vertices_vec);
3209
3210    return LLVMBuildAnd(builder, current_mask_vec, max_mask, "");
3211 }
3212
3213 static void
3214 emit_vertex(
3215    const struct lp_build_tgsi_action * action,
3216    struct lp_build_tgsi_context * bld_base,
3217    struct lp_build_emit_data * emit_data)
3218 {
3219    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3220    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
3221
3222    if (bld->gs_iface->emit_vertex) {
3223       LLVMValueRef mask = mask_vec(bld_base);
3224       LLVMValueRef total_emitted_vertices_vec =
3225          LLVMBuildLoad(builder, bld->total_emitted_vertices_vec_ptr, "");
3226       mask = clamp_mask_to_max_output_vertices(bld, mask,
3227                                                total_emitted_vertices_vec);
3228       gather_outputs(bld);
3229       bld->gs_iface->emit_vertex(bld->gs_iface, &bld->bld_base,
3230                                  bld->outputs,
3231                                  total_emitted_vertices_vec);
3232       increment_vec_ptr_by_mask(bld_base, bld->emitted_vertices_vec_ptr,
3233                                 mask);
3234       increment_vec_ptr_by_mask(bld_base, bld->total_emitted_vertices_vec_ptr,
3235                                 mask);
3236 #if DUMP_GS_EMITS
3237       lp_build_print_value(bld->bld_base.base.gallivm,
3238                            " +++ emit vertex masked ones = ",
3239                            mask);
3240       lp_build_print_value(bld->bld_base.base.gallivm,
3241                            " +++ emit vertex emitted = ",
3242                            total_emitted_vertices_vec);
3243 #endif
3244    }
3245 }
3246
3247
3248 static void
3249 end_primitive_masked(struct lp_build_tgsi_context * bld_base,
3250                      LLVMValueRef mask)
3251 {
3252    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3253    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
3254
3255    if (bld->gs_iface->end_primitive) {
3256       struct lp_build_context *uint_bld = &bld_base->uint_bld;
3257       LLVMValueRef emitted_vertices_vec =
3258          LLVMBuildLoad(builder, bld->emitted_vertices_vec_ptr, "");
3259       LLVMValueRef emitted_prims_vec =
3260          LLVMBuildLoad(builder, bld->emitted_prims_vec_ptr, "");
3261
3262       LLVMValueRef emitted_mask = lp_build_cmp(uint_bld, PIPE_FUNC_NOTEQUAL,
3263                                                emitted_vertices_vec,
3264                                                uint_bld->zero);
3265       /* We need to combine the current execution mask with the mask
3266          telling us which, if any, execution slots actually have
3267          unemitted primitives, this way we make sure that end_primitives
3268          executes only on the paths that have unflushed vertices */
3269       mask = LLVMBuildAnd(builder, mask, emitted_mask, "");
3270
3271       bld->gs_iface->end_primitive(bld->gs_iface, &bld->bld_base,
3272                                    emitted_vertices_vec,
3273                                    emitted_prims_vec);
3274
3275 #if DUMP_GS_EMITS
3276       lp_build_print_value(bld->bld_base.base.gallivm,
3277                            " +++ end prim masked ones = ",
3278                            mask);
3279       lp_build_print_value(bld->bld_base.base.gallivm,
3280                            " +++ end prim emitted verts1 = ",
3281                            emitted_vertices_vec);
3282       lp_build_print_value(bld->bld_base.base.gallivm,
3283                            " +++ end prim emitted prims1 = ",
3284                            LLVMBuildLoad(builder,
3285                                          bld->emitted_prims_vec_ptr, ""));
3286 #endif
3287       increment_vec_ptr_by_mask(bld_base, bld->emitted_prims_vec_ptr,
3288                                 mask);
3289       clear_uint_vec_ptr_from_mask(bld_base, bld->emitted_vertices_vec_ptr,
3290                                    mask);
3291 #if DUMP_GS_EMITS
3292       lp_build_print_value(bld->bld_base.base.gallivm,
3293                            " +++ end prim emitted verts2 = ",
3294                            LLVMBuildLoad(builder,
3295                                          bld->emitted_vertices_vec_ptr, ""));
3296 #endif
3297    }
3298
3299 }
3300
3301 static void
3302 end_primitive(
3303    const struct lp_build_tgsi_action * action,
3304    struct lp_build_tgsi_context * bld_base,
3305    struct lp_build_emit_data * emit_data)
3306 {
3307    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3308
3309    if (bld->gs_iface->end_primitive) {
3310       LLVMValueRef mask = mask_vec(bld_base);
3311       end_primitive_masked(bld_base, mask);
3312    }
3313 }
3314
3315 static void
3316 cal_emit(
3317    const struct lp_build_tgsi_action * action,
3318    struct lp_build_tgsi_context * bld_base,
3319    struct lp_build_emit_data * emit_data)
3320 {
3321    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3322
3323    lp_exec_mask_call(&bld->exec_mask, emit_data->inst->Label.Label,
3324                      &bld_base->pc);
3325 }
3326
3327 static void
3328 ret_emit(
3329    const struct lp_build_tgsi_action * action,
3330    struct lp_build_tgsi_context * bld_base,
3331    struct lp_build_emit_data * emit_data)
3332 {
3333    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3334
3335    lp_exec_mask_ret(&bld->exec_mask, &bld_base->pc);
3336 }
3337
3338 static void
3339 brk_emit(
3340    const struct lp_build_tgsi_action * action,
3341    struct lp_build_tgsi_context * bld_base,
3342    struct lp_build_emit_data * emit_data)
3343 {
3344    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3345
3346    lp_exec_break(&bld->exec_mask, bld_base);
3347 }
3348
3349 static void
3350 breakc_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    LLVMBuilderRef builder = bld_base->base.gallivm->builder;
3357    struct lp_build_context *uint_bld = &bld_base->uint_bld;
3358    LLVMValueRef unsigned_cond =
3359       LLVMBuildBitCast(builder, emit_data->args[0], uint_bld->vec_type, "");
3360    LLVMValueRef cond = lp_build_cmp(uint_bld, PIPE_FUNC_NOTEQUAL,
3361                                     unsigned_cond,
3362                                     uint_bld->zero);
3363
3364    lp_exec_break_condition(&bld->exec_mask, cond);
3365 }
3366
3367 static void
3368 if_emit(
3369    const struct lp_build_tgsi_action * action,
3370    struct lp_build_tgsi_context * bld_base,
3371    struct lp_build_emit_data * emit_data)
3372 {
3373    LLVMValueRef tmp;
3374    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3375
3376    tmp = lp_build_cmp(&bld_base->base, PIPE_FUNC_NOTEQUAL,
3377                       emit_data->args[0], bld->bld_base.base.zero);
3378    lp_exec_mask_cond_push(&bld->exec_mask, tmp);
3379 }
3380
3381 static void
3382 uif_emit(
3383    const struct lp_build_tgsi_action * action,
3384    struct lp_build_tgsi_context * bld_base,
3385    struct lp_build_emit_data * emit_data)
3386 {
3387    LLVMValueRef tmp;
3388    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3389    struct lp_build_context *uint_bld = &bld_base->uint_bld;
3390
3391    tmp = lp_build_cmp(uint_bld, PIPE_FUNC_NOTEQUAL,
3392                       emit_data->args[0], uint_bld->zero);
3393    lp_exec_mask_cond_push(&bld->exec_mask, tmp);
3394 }
3395
3396 static void
3397 case_emit(
3398    const struct lp_build_tgsi_action * action,
3399    struct lp_build_tgsi_context * bld_base,
3400    struct lp_build_emit_data * emit_data)
3401 {
3402    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3403
3404    lp_exec_case(&bld->exec_mask, emit_data->args[0]);
3405 }
3406
3407 static void
3408 default_emit(
3409    const struct lp_build_tgsi_action * action,
3410    struct lp_build_tgsi_context * bld_base,
3411    struct lp_build_emit_data * emit_data)
3412 {
3413    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3414
3415    lp_exec_default(&bld->exec_mask, bld_base);
3416 }
3417
3418 static void
3419 switch_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_switch(&bld->exec_mask, emit_data->args[0]);
3427 }
3428
3429 static void
3430 endswitch_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_endswitch(&bld->exec_mask, bld_base);
3438 }
3439
3440 static void
3441 bgnloop_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_bgnloop(&bld->exec_mask);
3449 }
3450
3451 static void
3452 bgnsub_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_mask_bgnsub(&bld->exec_mask);
3460 }
3461
3462 static void
3463 else_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_mask_cond_invert(&bld->exec_mask);
3471 }
3472
3473 static void
3474 endif_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_cond_pop(&bld->exec_mask);
3482 }
3483
3484 static void
3485 endloop_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_endloop(bld_base->base.gallivm, &bld->exec_mask);
3493 }
3494
3495 static void
3496 endsub_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_endsub(&bld->exec_mask, &bld_base->pc);
3504 }
3505
3506 static void
3507 cont_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_continue(&bld->exec_mask);
3515 }
3516
3517 static void emit_prologue(struct lp_build_tgsi_context * bld_base)
3518 {
3519    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3520    struct gallivm_state * gallivm = bld_base->base.gallivm;
3521
3522    if (bld->indirect_files & (1 << TGSI_FILE_TEMPORARY)) {
3523       LLVMValueRef array_size =
3524          lp_build_const_int32(gallivm,
3525                          bld_base->info->file_max[TGSI_FILE_TEMPORARY] * 4 + 4);
3526       bld->temps_array = lp_build_array_alloca(gallivm,
3527                                               bld_base->base.vec_type, array_size,
3528                                               "temp_array");
3529    }
3530
3531    if (bld->indirect_files & (1 << TGSI_FILE_OUTPUT)) {
3532       LLVMValueRef array_size =
3533          lp_build_const_int32(gallivm,
3534                             bld_base->info->file_max[TGSI_FILE_OUTPUT] * 4 + 4);
3535       bld->outputs_array = lp_build_array_alloca(gallivm,
3536                                                 bld_base->base.vec_type, array_size,
3537                                                 "output_array");
3538    }
3539
3540    if (bld->indirect_files & (1 << TGSI_FILE_IMMEDIATE)) {
3541       LLVMValueRef array_size =
3542          lp_build_const_int32(gallivm,
3543                          bld_base->info->file_max[TGSI_FILE_IMMEDIATE] * 4 + 4);
3544       bld->imms_array = lp_build_array_alloca(gallivm,
3545                                               bld_base->base.vec_type, array_size,
3546                                               "imms_array");
3547    }
3548
3549    /* If we have indirect addressing in inputs we need to copy them into
3550     * our alloca array to be able to iterate over them */
3551    if (bld->indirect_files & (1 << TGSI_FILE_INPUT) && !bld->gs_iface) {
3552       unsigned index, chan;
3553       LLVMTypeRef vec_type = bld_base->base.vec_type;
3554       LLVMValueRef array_size = lp_build_const_int32(gallivm,
3555             bld_base->info->file_max[TGSI_FILE_INPUT]*4 + 4);
3556       bld->inputs_array = lp_build_array_alloca(gallivm,
3557                                                vec_type, array_size,
3558                                                "input_array");
3559
3560       assert(bld_base->info->num_inputs
3561                         <= bld_base->info->file_max[TGSI_FILE_INPUT] + 1);
3562
3563       for (index = 0; index < bld_base->info->num_inputs; ++index) {
3564          for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
3565             LLVMValueRef lindex =
3566                lp_build_const_int32(gallivm, index * 4 + chan);
3567             LLVMValueRef input_ptr =
3568                LLVMBuildGEP(gallivm->builder, bld->inputs_array,
3569                             &lindex, 1, "");
3570             LLVMValueRef value = bld->inputs[index][chan];
3571             if (value)
3572                LLVMBuildStore(gallivm->builder, value, input_ptr);
3573          }
3574       }
3575    }
3576
3577    if (bld->gs_iface) {
3578       struct lp_build_context *uint_bld = &bld->bld_base.uint_bld;
3579       bld->emitted_prims_vec_ptr =
3580          lp_build_alloca(gallivm,
3581                          uint_bld->vec_type,
3582                          "emitted_prims_ptr");
3583       bld->emitted_vertices_vec_ptr =
3584          lp_build_alloca(gallivm,
3585                          uint_bld->vec_type,
3586                          "emitted_vertices_ptr");
3587       bld->total_emitted_vertices_vec_ptr =
3588          lp_build_alloca(gallivm,
3589                          uint_bld->vec_type,
3590                          "total_emitted_vertices_ptr");
3591
3592       LLVMBuildStore(gallivm->builder, uint_bld->zero,
3593                      bld->emitted_prims_vec_ptr);
3594       LLVMBuildStore(gallivm->builder, uint_bld->zero,
3595                      bld->emitted_vertices_vec_ptr);
3596       LLVMBuildStore(gallivm->builder, uint_bld->zero,
3597                      bld->total_emitted_vertices_vec_ptr);
3598    }
3599
3600    if (DEBUG_EXECUTION) {
3601       lp_build_printf(gallivm, "\n");
3602       emit_dump_file(bld, TGSI_FILE_CONSTANT);
3603       if (!bld->gs_iface)
3604          emit_dump_file(bld, TGSI_FILE_INPUT);
3605    }
3606 }
3607
3608 static void emit_epilogue(struct lp_build_tgsi_context * bld_base)
3609 {
3610    struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3611    LLVMBuilderRef builder = bld_base->base.gallivm->builder;
3612
3613    if (DEBUG_EXECUTION) {
3614       /* for debugging */
3615       if (0) {
3616          emit_dump_file(bld, TGSI_FILE_TEMPORARY);
3617       }
3618       emit_dump_file(bld, TGSI_FILE_OUTPUT);
3619       lp_build_printf(bld_base->base.gallivm, "\n");
3620    }
3621
3622    /* If we have indirect addressing in outputs we need to copy our alloca array
3623     * to the outputs slots specified by the caller */
3624    if (bld->gs_iface) {
3625       LLVMValueRef total_emitted_vertices_vec;
3626       LLVMValueRef emitted_prims_vec;
3627       /* implicit end_primitives, needed in case there are any unflushed
3628          vertices in the cache. Note must not call end_primitive here
3629          since the exec_mask is not valid at this point. */
3630       end_primitive_masked(bld_base, lp_build_mask_value(bld->mask));
3631
3632       total_emitted_vertices_vec =
3633          LLVMBuildLoad(builder, bld->total_emitted_vertices_vec_ptr, "");
3634       emitted_prims_vec =
3635          LLVMBuildLoad(builder, bld->emitted_prims_vec_ptr, "");
3636
3637       bld->gs_iface->gs_epilogue(bld->gs_iface,
3638                                  &bld->bld_base,
3639                                  total_emitted_vertices_vec,
3640                                  emitted_prims_vec);
3641    } else {
3642       gather_outputs(bld);
3643    }
3644 }
3645
3646 void
3647 lp_build_tgsi_soa(struct gallivm_state *gallivm,
3648                   const struct tgsi_token *tokens,
3649                   struct lp_type type,
3650                   struct lp_build_mask_context *mask,
3651                   LLVMValueRef consts_ptr,
3652                   LLVMValueRef const_sizes_ptr,
3653                   const struct lp_bld_tgsi_system_values *system_values,
3654                   const LLVMValueRef (*inputs)[TGSI_NUM_CHANNELS],
3655                   LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
3656                   LLVMValueRef context_ptr,
3657                   struct lp_build_sampler_soa *sampler,
3658                   const struct tgsi_shader_info *info,
3659                   const struct lp_build_tgsi_gs_iface *gs_iface)
3660 {
3661    struct lp_build_tgsi_soa_context bld;
3662
3663    struct lp_type res_type;
3664
3665    assert(type.length <= LP_MAX_VECTOR_LENGTH);
3666    memset(&res_type, 0, sizeof res_type);
3667    res_type.width = type.width;
3668    res_type.length = type.length;
3669    res_type.sign = 1;
3670
3671    /* Setup build context */
3672    memset(&bld, 0, sizeof bld);
3673    lp_build_context_init(&bld.bld_base.base, gallivm, type);
3674    lp_build_context_init(&bld.bld_base.uint_bld, gallivm, lp_uint_type(type));
3675    lp_build_context_init(&bld.bld_base.int_bld, gallivm, lp_int_type(type));
3676    lp_build_context_init(&bld.elem_bld, gallivm, lp_elem_type(type));
3677    bld.mask = mask;
3678    bld.inputs = inputs;
3679    bld.outputs = outputs;
3680    bld.consts_ptr = consts_ptr;
3681    bld.const_sizes_ptr = const_sizes_ptr;
3682    bld.sampler = sampler;
3683    bld.bld_base.info = info;
3684    bld.indirect_files = info->indirect_files;
3685    bld.context_ptr = context_ptr;
3686
3687    /*
3688     * If the number of temporaries is rather large then we just
3689     * allocate them as an array right from the start and treat
3690     * like indirect temporaries.
3691     */
3692    if (info->file_max[TGSI_FILE_TEMPORARY] >= LP_MAX_INLINED_TEMPS) {
3693       bld.indirect_files |= (1 << TGSI_FILE_TEMPORARY);
3694    }
3695    /*
3696     * For performance reason immediates are always backed in a static
3697     * array, but if their number is too great, we have to use just
3698     * a dynamically allocated array.
3699     */
3700    bld.use_immediates_array =
3701          (info->file_max[TGSI_FILE_IMMEDIATE] >= LP_MAX_INLINED_IMMEDIATES);
3702    if (bld.use_immediates_array) {
3703       bld.indirect_files |= (1 << TGSI_FILE_IMMEDIATE);
3704    }
3705
3706
3707    bld.bld_base.soa = TRUE;
3708    bld.bld_base.emit_debug = emit_debug;
3709    bld.bld_base.emit_fetch_funcs[TGSI_FILE_CONSTANT] = emit_fetch_constant;
3710    bld.bld_base.emit_fetch_funcs[TGSI_FILE_IMMEDIATE] = emit_fetch_immediate;
3711    bld.bld_base.emit_fetch_funcs[TGSI_FILE_INPUT] = emit_fetch_input;
3712    bld.bld_base.emit_fetch_funcs[TGSI_FILE_TEMPORARY] = emit_fetch_temporary;
3713    bld.bld_base.emit_fetch_funcs[TGSI_FILE_SYSTEM_VALUE] = emit_fetch_system_value;
3714    bld.bld_base.emit_store = emit_store;
3715
3716    bld.bld_base.emit_declaration = lp_emit_declaration_soa;
3717    bld.bld_base.emit_immediate = lp_emit_immediate_soa;
3718
3719    bld.bld_base.emit_prologue = emit_prologue;
3720    bld.bld_base.emit_epilogue = emit_epilogue;
3721
3722    /* Set opcode actions */
3723    lp_set_default_actions_cpu(&bld.bld_base);
3724
3725    bld.bld_base.op_actions[TGSI_OPCODE_BGNLOOP].emit = bgnloop_emit;
3726    bld.bld_base.op_actions[TGSI_OPCODE_BGNSUB].emit = bgnsub_emit;
3727    bld.bld_base.op_actions[TGSI_OPCODE_BRK].emit = brk_emit;
3728    bld.bld_base.op_actions[TGSI_OPCODE_BREAKC].emit = breakc_emit;
3729    bld.bld_base.op_actions[TGSI_OPCODE_CAL].emit = cal_emit;
3730    bld.bld_base.op_actions[TGSI_OPCODE_CASE].emit = case_emit;
3731    bld.bld_base.op_actions[TGSI_OPCODE_CONT].emit = cont_emit;
3732    bld.bld_base.op_actions[TGSI_OPCODE_DDX].emit = ddx_emit;
3733    bld.bld_base.op_actions[TGSI_OPCODE_DDY].emit = ddy_emit;
3734    bld.bld_base.op_actions[TGSI_OPCODE_DEFAULT].emit = default_emit;
3735    bld.bld_base.op_actions[TGSI_OPCODE_ELSE].emit = else_emit;
3736    bld.bld_base.op_actions[TGSI_OPCODE_ENDIF].emit = endif_emit;
3737    bld.bld_base.op_actions[TGSI_OPCODE_ENDLOOP].emit = endloop_emit;
3738    bld.bld_base.op_actions[TGSI_OPCODE_ENDSUB].emit = endsub_emit;
3739    bld.bld_base.op_actions[TGSI_OPCODE_ENDSWITCH].emit = endswitch_emit;
3740    bld.bld_base.op_actions[TGSI_OPCODE_IF].emit = if_emit;
3741    bld.bld_base.op_actions[TGSI_OPCODE_UIF].emit = uif_emit;
3742    bld.bld_base.op_actions[TGSI_OPCODE_KILL_IF].emit = kill_if_emit;
3743    bld.bld_base.op_actions[TGSI_OPCODE_KILL].emit = kill_emit;
3744    bld.bld_base.op_actions[TGSI_OPCODE_RET].emit = ret_emit;
3745    bld.bld_base.op_actions[TGSI_OPCODE_SWITCH].emit = switch_emit;
3746    bld.bld_base.op_actions[TGSI_OPCODE_TEX].emit = tex_emit;
3747    bld.bld_base.op_actions[TGSI_OPCODE_TXB].emit = txb_emit;
3748    bld.bld_base.op_actions[TGSI_OPCODE_TXD].emit = txd_emit;
3749    bld.bld_base.op_actions[TGSI_OPCODE_TXL].emit = txl_emit;
3750    bld.bld_base.op_actions[TGSI_OPCODE_TXP].emit = txp_emit;
3751    bld.bld_base.op_actions[TGSI_OPCODE_TXQ].emit = txq_emit;
3752    bld.bld_base.op_actions[TGSI_OPCODE_TXF].emit = txf_emit;
3753    bld.bld_base.op_actions[TGSI_OPCODE_TEX2].emit = tex2_emit;
3754    bld.bld_base.op_actions[TGSI_OPCODE_TXB2].emit = txb2_emit;
3755    bld.bld_base.op_actions[TGSI_OPCODE_TXL2].emit = txl2_emit;
3756    bld.bld_base.op_actions[TGSI_OPCODE_TG4].emit = tg4_emit;
3757    /* DX10 sampling ops */
3758    bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE].emit = sample_emit;
3759    bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE_B].emit = sample_b_emit;
3760    bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE_C].emit = sample_c_emit;
3761    bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE_C_LZ].emit = sample_c_lz_emit;
3762    bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE_D].emit = sample_d_emit;
3763    bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE_I].emit = sample_i_emit;
3764    bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE_L].emit = sample_l_emit;
3765    bld.bld_base.op_actions[TGSI_OPCODE_SVIEWINFO].emit = sviewinfo_emit;
3766
3767    if (gs_iface) {
3768       /* There's no specific value for this because it should always
3769        * be set, but apps using ext_geometry_shader4 quite often
3770        * were forgetting so we're using MAX_VERTEX_VARYING from
3771        * that spec even though we could debug_assert if it's not
3772        * set, but that's a lot uglier. */
3773       uint max_output_vertices;
3774
3775       /* inputs are always indirect with gs */
3776       bld.indirect_files |= (1 << TGSI_FILE_INPUT);
3777       bld.gs_iface = gs_iface;
3778       bld.bld_base.emit_fetch_funcs[TGSI_FILE_INPUT] = emit_fetch_gs_input;
3779       bld.bld_base.op_actions[TGSI_OPCODE_EMIT].emit = emit_vertex;
3780       bld.bld_base.op_actions[TGSI_OPCODE_ENDPRIM].emit = end_primitive;
3781
3782       max_output_vertices =
3783             info->properties[TGSI_PROPERTY_GS_MAX_OUTPUT_VERTICES];
3784       if (!max_output_vertices)
3785          max_output_vertices = 32;
3786
3787       bld.max_output_vertices_vec =
3788          lp_build_const_int_vec(gallivm, bld.bld_base.int_bld.type,
3789                                 max_output_vertices);
3790    }
3791
3792    lp_exec_mask_init(&bld.exec_mask, &bld.bld_base.int_bld);
3793
3794    bld.system_values = *system_values;
3795
3796    lp_build_tgsi_llvm(&bld.bld_base, tokens);
3797
3798    if (0) {
3799       LLVMBasicBlockRef block = LLVMGetInsertBlock(gallivm->builder);
3800       LLVMValueRef function = LLVMGetBasicBlockParent(block);
3801       debug_printf("11111111111111111111111111111 \n");
3802       tgsi_dump(tokens, 0);
3803       lp_debug_dump_value(function);
3804       debug_printf("2222222222222222222222222222 \n");
3805    }
3806
3807    if (0) {
3808       LLVMModuleRef module = LLVMGetGlobalParent(
3809          LLVMGetBasicBlockParent(LLVMGetInsertBlock(gallivm->builder)));
3810       LLVMDumpModule(module);
3811
3812    }
3813    lp_exec_mask_fini(&bld.exec_mask);
3814 }