src/gallium/drivers/llvmpipe/lp_bld_tgsi_soa.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2009 VMware, Inc.
   4  * Copyright 2007-2008 Tungsten Graphics, Inc., Cedar Park, Texas.
   5  * All Rights Reserved.
   6  *
   7  * Permission is hereby granted, free of charge, to any person obtaining a
   8  * copy of this software and associated documentation files (the
   9  * "Software"), to deal in the Software without restriction, including
  10  * without limitation the rights to use, copy, modify, merge, publish,
  11  * distribute, sub license, and/or sell copies of the Software, and to
  12  * permit persons to whom the Software is furnished to do so, subject to
  13  * the following conditions:
  14  *
  15  * The above copyright notice and this permission notice (including the
  16  * next paragraph) shall be included in all copies or substantial portions
  17  * of the Software.
  18  *
  19  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  20  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  21  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  22  * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
  23  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  24  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  25  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  26  *
  27  **************************************************************************/
  28
  29 /**
  30  * @file
  31  * TGSI to LLVM IR translation -- SoA.
  32  *
  33  * @author Jose Fonseca <jfonseca@vmware.com>
  34  *
  35  * Based on tgsi_sse2.c code written by Michal Krol, Keith Whitwell,
  36  * Brian Paul, and others.
  37  */
  38
  39 #include "pipe/p_config.h"
  40 #include "pipe/p_shader_tokens.h"
  41 #include "util/u_debug.h"
  42 #include "util/u_math.h"
  43 #include "util/u_memory.h"
  44 #include "tgsi/tgsi_info.h"
  45 #include "tgsi/tgsi_parse.h"
  46 #include "tgsi/tgsi_util.h"
  47 #include "tgsi/tgsi_exec.h"
  48 #include "lp_bld_type.h"
  49 #include "lp_bld_const.h"
  50 #include "lp_bld_intr.h"
  51 #include "lp_bld_arit.h"
  52 #include "lp_bld_logic.h"
  53 #include "lp_bld_swizzle.h"
  54 #include "lp_bld_flow.h"
  55 #include "lp_bld_tgsi.h"
  56 #include "lp_bld_debug.h"
  57
  58
  59 #define LP_MAX_TEMPS 256
  60 #define LP_MAX_IMMEDIATES 256
  61
  62
  63 #define FOR_EACH_CHANNEL( CHAN )\
  64    for (CHAN = 0; CHAN < NUM_CHANNELS; CHAN++)
  65
  66 #define IS_DST0_CHANNEL_ENABLED( INST, CHAN )\
  67    ((INST)->FullDstRegisters[0].DstRegister.WriteMask & (1 << (CHAN)))
  68
  69 #define IF_IS_DST0_CHANNEL_ENABLED( INST, CHAN )\
  70    if (IS_DST0_CHANNEL_ENABLED( INST, CHAN ))
  71
  72 #define FOR_EACH_DST0_ENABLED_CHANNEL( INST, CHAN )\
  73    FOR_EACH_CHANNEL( CHAN )\
  74       IF_IS_DST0_CHANNEL_ENABLED( INST, CHAN )
  75
  76 #define CHAN_X 0
  77 #define CHAN_Y 1
  78 #define CHAN_Z 2
  79 #define CHAN_W 3
  80
  81
  82 struct lp_build_tgsi_soa_context
  83 {
  84    struct lp_build_context base;
  85
  86    LLVMValueRef consts_ptr;
  87    const LLVMValueRef *pos;
  88    const LLVMValueRef (*inputs)[NUM_CHANNELS];
  89    LLVMValueRef (*outputs)[NUM_CHANNELS];
  90
  91    struct lp_build_sampler_soa *sampler;
  92
  93    LLVMValueRef immediates[LP_MAX_IMMEDIATES][NUM_CHANNELS];
  94    LLVMValueRef temps[LP_MAX_TEMPS][NUM_CHANNELS];
  95
  96    struct lp_build_mask_context *mask;
  97 };
  98
  99
 100 /**
 101  * Register fetch.
 102  */
 103 static LLVMValueRef
 104 emit_fetch(
 105    struct lp_build_tgsi_soa_context *bld,
 106    const struct tgsi_full_instruction *inst,
 107    unsigned index,
 108    const unsigned chan_index )
 109 {
 110    const struct tgsi_full_src_register *reg = &inst->FullSrcRegisters[index];
 111    unsigned swizzle = tgsi_util_get_full_src_register_extswizzle( reg, chan_index );
 112    LLVMValueRef res;
 113
 114    switch (swizzle) {
 115    case TGSI_EXTSWIZZLE_X:
 116    case TGSI_EXTSWIZZLE_Y:
 117    case TGSI_EXTSWIZZLE_Z:
 118    case TGSI_EXTSWIZZLE_W:
 119
 120       switch (reg->SrcRegister.File) {
 121       case TGSI_FILE_CONSTANT: {
 122          LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), reg->SrcRegister.Index*4 + swizzle, 0);
 123          LLVMValueRef scalar_ptr = LLVMBuildGEP(bld->base.builder, bld->consts_ptr, &index, 1, "");
 124          LLVMValueRef scalar = LLVMBuildLoad(bld->base.builder, scalar_ptr, "");
 125          res = lp_build_broadcast_scalar(&bld->base, scalar);
 126          break;
 127       }
 128
 129       case TGSI_FILE_IMMEDIATE:
 130          res = bld->immediates[reg->SrcRegister.Index][swizzle];
 131          assert(res);
 132          break;
 133
 134       case TGSI_FILE_INPUT:
 135          res = bld->inputs[reg->SrcRegister.Index][swizzle];
 136          assert(res);
 137          break;
 138
 139       case TGSI_FILE_TEMPORARY:
 140          res = bld->temps[reg->SrcRegister.Index][swizzle];
 141          if(!res)
 142             return bld->base.undef;
 143          break;
 144
 145       default:
 146          assert( 0 );
 147          return bld->base.undef;
 148       }
 149       break;
 150
 151    case TGSI_EXTSWIZZLE_ZERO:
 152       res = bld->base.zero;
 153       break;
 154
 155    case TGSI_EXTSWIZZLE_ONE:
 156       res = bld->base.one;
 157       break;
 158
 159    default:
 160       assert( 0 );
 161       return bld->base.undef;
 162    }
 163
 164    switch( tgsi_util_get_full_src_register_sign_mode( reg, chan_index ) ) {
 165    case TGSI_UTIL_SIGN_CLEAR:
 166       res = lp_build_abs( &bld->base, res );
 167       break;
 168
 169    case TGSI_UTIL_SIGN_SET:
 170       res = lp_build_abs( &bld->base, res );
 171       res = LLVMBuildNeg( bld->base.builder, res, "" );
 172       break;
 173
 174    case TGSI_UTIL_SIGN_TOGGLE:
 175       res = LLVMBuildNeg( bld->base.builder, res, "" );
 176       break;
 177
 178    case TGSI_UTIL_SIGN_KEEP:
 179       break;
 180    }
 181
 182    return res;
 183 }
 184
 185
 186 /**
 187  * Register store.
 188  */
 189 static void
 190 emit_store(
 191    struct lp_build_tgsi_soa_context *bld,
 192    const struct tgsi_full_instruction *inst,
 193    unsigned index,
 194    unsigned chan_index,
 195    LLVMValueRef value)
 196 {
 197    const struct tgsi_full_dst_register *reg = &inst->FullDstRegisters[index];
 198
 199    switch( inst->Instruction.Saturate ) {
 200    case TGSI_SAT_NONE:
 201       break;
 202
 203    case TGSI_SAT_ZERO_ONE:
 204       value = lp_build_max(&bld->base, value, bld->base.zero);
 205       value = lp_build_min(&bld->base, value, bld->base.one);
 206       break;
 207
 208    case TGSI_SAT_MINUS_PLUS_ONE:
 209       value = lp_build_max(&bld->base, value, lp_build_const_scalar(bld->base.type, -1.0));
 210       value = lp_build_min(&bld->base, value, bld->base.one);
 211       break;
 212
 213    default:
 214       assert(0);
 215    }
 216
 217    switch( reg->DstRegister.File ) {
 218    case TGSI_FILE_OUTPUT:
 219       bld->outputs[reg->DstRegister.Index][chan_index] = value;
 220       break;
 221
 222    case TGSI_FILE_TEMPORARY:
 223       bld->temps[reg->DstRegister.Index][chan_index] = value;
 224       break;
 225
 226    case TGSI_FILE_ADDRESS:
 227       /* FIXME */
 228       assert(0);
 229       break;
 230
 231    default:
 232       assert( 0 );
 233    }
 234 }
 235
 236
 237 /**
 238  * High-level instruction translators.
 239  */
 240
 241 static void
 242 emit_tex( struct lp_build_tgsi_soa_context *bld,
 243           const struct tgsi_full_instruction *inst,
 244           boolean apply_lodbias,
 245           boolean projected)
 246 {
 247    const uint unit = inst->FullSrcRegisters[1].SrcRegister.Index;
 248    LLVMValueRef lodbias;
 249    LLVMValueRef oow;
 250    LLVMValueRef coords[3];
 251    LLVMValueRef texel[4];
 252    unsigned num_coords;
 253    unsigned i;
 254
 255    switch (inst->InstructionExtTexture.Texture) {
 256    case TGSI_TEXTURE_1D:
 257       num_coords = 1;
 258       break;
 259    case TGSI_TEXTURE_2D:
 260    case TGSI_TEXTURE_RECT:
 261       num_coords = 2;
 262       break;
 263    case TGSI_TEXTURE_SHADOW1D:
 264    case TGSI_TEXTURE_SHADOW2D:
 265    case TGSI_TEXTURE_SHADOWRECT:
 266    case TGSI_TEXTURE_3D:
 267    case TGSI_TEXTURE_CUBE:
 268       num_coords = 3;
 269       break;
 270    default:
 271       assert(0);
 272       return;
 273    }
 274
 275    if(apply_lodbias)
 276       lodbias = emit_fetch( bld, inst, 0, 3 );
 277    else
 278       lodbias = bld->base.zero;
 279
 280    if (projected) {
 281       oow = emit_fetch( bld, inst, 0, 3 );
 282       oow = lp_build_rcp(&bld->base, oow);
 283    }
 284
 285    for (i = 0; i < num_coords; i++) {
 286       coords[i] = emit_fetch( bld, inst, 0, i );
 287       if (projected)
 288          coords[i] = lp_build_mul(&bld->base, coords[i], oow);
 289    }
 290
 291    bld->sampler->emit_fetch_texel(bld->sampler,
 292                                   bld->base.builder,
 293                                   bld->base.type,
 294                                   unit, num_coords, coords, lodbias,
 295                                   texel);
 296
 297    FOR_EACH_DST0_ENABLED_CHANNEL( inst, i ) {
 298       emit_store( bld, inst, 0, i, texel[i] );
 299    }
 300 }
 301
 302
 303 static void
 304 emit_kil(
 305    struct lp_build_tgsi_soa_context *bld,
 306    const struct tgsi_full_instruction *inst )
 307 {
 308    const struct tgsi_full_src_register *reg = &inst->FullSrcRegisters[0];
 309    LLVMValueRef terms[NUM_CHANNELS];
 310    LLVMValueRef mask;
 311    unsigned chan_index;
 312
 313    memset(&terms, 0, sizeof terms);
 314
 315    FOR_EACH_CHANNEL( chan_index ) {
 316       unsigned swizzle;
 317
 318       /* Unswizzle channel */
 319       swizzle = tgsi_util_get_full_src_register_extswizzle( reg, chan_index );
 320
 321       /* Note that we test if the value is less than zero, so 1.0 and 0.0 need
 322        * not to be tested. */
 323       if(swizzle == TGSI_EXTSWIZZLE_ZERO || swizzle == TGSI_EXTSWIZZLE_ONE)
 324          continue;
 325
 326       /* Check if the component has not been already tested. */
 327       assert(swizzle < NUM_CHANNELS);
 328       if( !terms[swizzle] )
 329          /* TODO: change the comparison operator instead of setting the sign */
 330          terms[swizzle] =  emit_fetch(bld, inst, 0, chan_index );
 331    }
 332
 333    mask = NULL;
 334    FOR_EACH_CHANNEL( chan_index ) {
 335       if(terms[chan_index]) {
 336          LLVMValueRef chan_mask;
 337
 338          chan_mask = lp_build_cmp(&bld->base, PIPE_FUNC_GEQUAL, terms[chan_index], bld->base.zero);
 339
 340          if(mask)
 341             mask = LLVMBuildAnd(bld->base.builder, mask, chan_mask, "");
 342          else
 343             mask = chan_mask;
 344       }
 345    }
 346
 347    if(mask)
 348       lp_build_mask_update(bld->mask, mask);
 349 }
 350
 351
 352 static void
 353 emit_kilp(
 354    struct lp_build_tgsi_soa_context *bld )
 355 {
 356    /* XXX todo / fix me */
 357 }
 358
 359
 360 /**
 361  * Check if inst src/dest regs use indirect addressing into temporary
 362  * register file.
 363  */
 364 static boolean
 365 indirect_temp_reference(const struct tgsi_full_instruction *inst)
 366 {
 367    uint i;
 368    for (i = 0; i < inst->Instruction.NumSrcRegs; i++) {
 369       const struct tgsi_full_src_register *reg = &inst->FullSrcRegisters[i];
 370       if (reg->SrcRegister.File == TGSI_FILE_TEMPORARY &&
 371           reg->SrcRegister.Indirect)
 372          return TRUE;
 373    }
 374    for (i = 0; i < inst->Instruction.NumDstRegs; i++) {
 375       const struct tgsi_full_dst_register *reg = &inst->FullDstRegisters[i];
 376       if (reg->DstRegister.File == TGSI_FILE_TEMPORARY &&
 377           reg->DstRegister.Indirect)
 378          return TRUE;
 379    }
 380    return FALSE;
 381 }
 382
 383
 384 static int
 385 emit_instruction(
 386    struct lp_build_tgsi_soa_context *bld,
 387    struct tgsi_full_instruction *inst )
 388 {
 389    unsigned chan_index;
 390    LLVMValueRef src0, src1, src2;
 391    LLVMValueRef tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
 392    LLVMValueRef dst0;
 393
 394    /* we can't handle indirect addressing into temp register file yet */
 395    if (indirect_temp_reference(inst))
 396       return FALSE;
 397
 398    switch (inst->Instruction.Opcode) {
 399 #if 0
 400    case TGSI_OPCODE_ARL:
 401       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 402          tmp0 = emit_fetch( bld, inst, 0, chan_index );
 403          emit_flr(bld, 0, 0);
 404          emit_f2it( bld, 0 );
 405          emit_store( bld, inst, 0, chan_index, tmp0);
 406       }
 407       break;
 408 #endif
 409
 410    case TGSI_OPCODE_MOV:
 411    case TGSI_OPCODE_SWZ:
 412       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 413          tmp0 = emit_fetch( bld, inst, 0, chan_index );
 414          emit_store( bld, inst, 0, chan_index, tmp0);
 415       }
 416       break;
 417
 418    case TGSI_OPCODE_LIT:
 419       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ) {
 420          emit_store( bld, inst, 0, CHAN_X, bld->base.one);
 421       }
 422       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ) {
 423          src0 = emit_fetch( bld, inst, 0, CHAN_X );
 424          dst0 = lp_build_max( &bld->base, src0, bld->base.zero);
 425          emit_store( bld, inst, 0, CHAN_Y, dst0);
 426       }
 427       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) ) {
 428          /* XMM[1] = SrcReg[0].yyyy */
 429          tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );
 430          /* XMM[1] = max(XMM[1], 0) */
 431          tmp1 = lp_build_max( &bld->base, tmp1, bld->base.zero);
 432          /* XMM[2] = SrcReg[0].wwww */
 433          tmp2 = emit_fetch( bld, inst, 0, CHAN_W );
 434          tmp1 = lp_build_pow( &bld->base, tmp1, tmp2);
 435          tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
 436          tmp2 = lp_build_cmp(&bld->base, PIPE_FUNC_GREATER, tmp0, bld->base.zero);
 437          dst0 = lp_build_select(&bld->base, tmp2, tmp1, bld->base.zero);
 438          emit_store( bld, inst, 0, CHAN_Z, dst0);
 439       }
 440       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) ) {
 441          emit_store( bld, inst, 0, CHAN_W, bld->base.one);
 442       }
 443       break;
 444
 445    case TGSI_OPCODE_RCP:
 446    /* TGSI_OPCODE_RECIP */
 447       src0 = emit_fetch( bld, inst, 0, CHAN_X );
 448       dst0 = lp_build_rcp(&bld->base, src0);
 449       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 450          emit_store( bld, inst, 0, chan_index, dst0 );
 451       }
 452       break;
 453
 454    case TGSI_OPCODE_RSQ:
 455    /* TGSI_OPCODE_RECIPSQRT */
 456       src0 = emit_fetch( bld, inst, 0, CHAN_X );
 457       src0 = lp_build_abs(&bld->base, src0);
 458       dst0 = lp_build_rsqrt(&bld->base, src0);
 459       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 460          emit_store( bld, inst, 0, chan_index, dst0 );
 461       }
 462       break;
 463
 464    case TGSI_OPCODE_EXP:
 465       if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ||
 466           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ||
 467           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z )) {
 468          LLVMValueRef *p_exp2_int_part = NULL;
 469          LLVMValueRef *p_frac_part = NULL;
 470          LLVMValueRef *p_exp2 = NULL;
 471
 472          src0 = emit_fetch( bld, inst, 0, CHAN_X );
 473
 474          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ))
 475             p_exp2_int_part = &tmp0;
 476          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ))
 477             p_frac_part = &tmp1;
 478          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ))
 479             p_exp2 = &tmp2;
 480
 481          lp_build_exp2_approx(&bld->base, src0, p_exp2_int_part, p_frac_part, p_exp2);
 482
 483          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ))
 484             emit_store( bld, inst, 0, CHAN_X, tmp0);
 485          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ))
 486             emit_store( bld, inst, 0, CHAN_Y, tmp1);
 487          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ))
 488             emit_store( bld, inst, 0, CHAN_Z, tmp2);
 489       }
 490       /* dst.w = 1.0 */
 491       if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_W )) {
 492          tmp0 = bld->base.one;
 493          emit_store( bld, inst, 0, CHAN_W, tmp0);
 494       }
 495       break;
 496
 497    case TGSI_OPCODE_LOG:
 498       if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ||
 499           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ||
 500           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z )) {
 501          LLVMValueRef *p_floor_log2;
 502          LLVMValueRef *p_exp;
 503          LLVMValueRef *p_log2;
 504
 505          src0 = emit_fetch( bld, inst, 0, CHAN_X );
 506          src0 = lp_build_abs( &bld->base, src0 );
 507
 508          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ))
 509             p_floor_log2 = &tmp0;
 510          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ))
 511             p_exp = &tmp1;
 512          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ))
 513             p_log2 = &tmp2;
 514
 515          lp_build_log2_approx(&bld->base, src0, p_exp, p_floor_log2, p_log2);
 516
 517          /* dst.x = floor(lg2(abs(src.x))) */
 518          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ))
 519             emit_store( bld, inst, 0, CHAN_X, tmp0);
 520          /* dst.y = abs(src)/ex2(floor(lg2(abs(src.x)))) */
 521          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y )) {
 522             tmp1 = lp_build_div( &bld->base, src0, tmp1);
 523             emit_store( bld, inst, 0, CHAN_Y, tmp1);
 524          }
 525          /* dst.z = lg2(abs(src.x)) */
 526          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ))
 527             emit_store( bld, inst, 0, CHAN_Z, tmp2);
 528       }
 529       /* dst.w = 1.0 */
 530       if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_W )) {
 531          tmp0 = bld->base.one;
 532          emit_store( bld, inst, 0, CHAN_W, tmp0);
 533       }
 534       break;
 535
 536    case TGSI_OPCODE_MUL:
 537       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 538          src0 = emit_fetch( bld, inst, 0, chan_index );
 539          src1 = emit_fetch( bld, inst, 1, chan_index );
 540          dst0 = lp_build_mul(&bld->base, src0, src1);
 541          emit_store( bld, inst, 0, chan_index, dst0);
 542       }
 543       break;
 544
 545    case TGSI_OPCODE_ADD:
 546       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 547          src0 = emit_fetch( bld, inst, 0, chan_index );
 548          src1 = emit_fetch( bld, inst, 1, chan_index );
 549          dst0 = lp_build_add(&bld->base, src0, src1);
 550          emit_store( bld, inst, 0, chan_index, dst0);
 551       }
 552       break;
 553
 554    case TGSI_OPCODE_DP3:
 555    /* TGSI_OPCODE_DOT3 */
 556       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
 557       tmp1 = emit_fetch( bld, inst, 1, CHAN_X );
 558       tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);
 559       tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );
 560       tmp2 = emit_fetch( bld, inst, 1, CHAN_Y );
 561       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
 562       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
 563       tmp1 = emit_fetch( bld, inst, 0, CHAN_Z );
 564       tmp2 = emit_fetch( bld, inst, 1, CHAN_Z );
 565       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
 566       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
 567       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 568          emit_store( bld, inst, 0, chan_index, tmp0);
 569       }
 570       break;
 571
 572    case TGSI_OPCODE_DP4:
 573    /* TGSI_OPCODE_DOT4 */
 574       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
 575       tmp1 = emit_fetch( bld, inst, 1, CHAN_X );
 576       tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);
 577       tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );
 578       tmp2 = emit_fetch( bld, inst, 1, CHAN_Y );
 579       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
 580       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
 581       tmp1 = emit_fetch( bld, inst, 0, CHAN_Z );
 582       tmp2 = emit_fetch( bld, inst, 1, CHAN_Z );
 583       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
 584       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
 585       tmp1 = emit_fetch( bld, inst, 0, CHAN_W );
 586       tmp2 = emit_fetch( bld, inst, 1, CHAN_W );
 587       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
 588       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
 589       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 590          emit_store( bld, inst, 0, chan_index, tmp0);
 591       }
 592       break;
 593
 594    case TGSI_OPCODE_DST:
 595       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) {
 596          tmp0 = bld->base.one;
 597          emit_store( bld, inst, 0, CHAN_X, tmp0);
 598       }
 599       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) {
 600          tmp0 = emit_fetch( bld, inst, 0, CHAN_Y );
 601          tmp1 = emit_fetch( bld, inst, 1, CHAN_Y );
 602          tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);
 603          emit_store( bld, inst, 0, CHAN_Y, tmp0);
 604       }
 605       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) {
 606          tmp0 = emit_fetch( bld, inst, 0, CHAN_Z );
 607          emit_store( bld, inst, 0, CHAN_Z, tmp0);
 608       }
 609       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) {
 610          tmp0 = emit_fetch( bld, inst, 1, CHAN_W );
 611          emit_store( bld, inst, 0, CHAN_W, tmp0);
 612       }
 613       break;
 614
 615    case TGSI_OPCODE_MIN:
 616       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 617          src0 = emit_fetch( bld, inst, 0, chan_index );
 618          src1 = emit_fetch( bld, inst, 1, chan_index );
 619          dst0 = lp_build_min( &bld->base, src0, src1 );
 620          emit_store( bld, inst, 0, chan_index, dst0);
 621       }
 622       break;
 623
 624    case TGSI_OPCODE_MAX:
 625       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 626          src0 = emit_fetch( bld, inst, 0, chan_index );
 627          src1 = emit_fetch( bld, inst, 1, chan_index );
 628          dst0 = lp_build_max( &bld->base, src0, src1 );
 629          emit_store( bld, inst, 0, chan_index, dst0);
 630       }
 631       break;
 632
 633    case TGSI_OPCODE_SLT:
 634    /* TGSI_OPCODE_SETLT */
 635       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 636          src0 = emit_fetch( bld, inst, 0, chan_index );
 637          src1 = emit_fetch( bld, inst, 1, chan_index );
 638          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_LESS, src0, src1 );
 639          dst0 = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
 640          emit_store( bld, inst, 0, chan_index, dst0);
 641       }
 642       break;
 643
 644    case TGSI_OPCODE_SGE:
 645    /* TGSI_OPCODE_SETGE */
 646       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 647          src0 = emit_fetch( bld, inst, 0, chan_index );
 648          src1 = emit_fetch( bld, inst, 1, chan_index );
 649          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_GEQUAL, src0, src1 );
 650          dst0 = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
 651          emit_store( bld, inst, 0, chan_index, dst0);
 652       }
 653       break;
 654
 655    case TGSI_OPCODE_MAD:
 656    /* TGSI_OPCODE_MADD */
 657       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 658          tmp0 = emit_fetch( bld, inst, 0, chan_index );
 659          tmp1 = emit_fetch( bld, inst, 1, chan_index );
 660          tmp2 = emit_fetch( bld, inst, 2, chan_index );
 661          tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);
 662          tmp0 = lp_build_add( &bld->base, tmp0, tmp2);
 663          emit_store( bld, inst, 0, chan_index, tmp0);
 664       }
 665       break;
 666
 667    case TGSI_OPCODE_SUB:
 668       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 669          tmp0 = emit_fetch( bld, inst, 0, chan_index );
 670          tmp1 = emit_fetch( bld, inst, 1, chan_index );
 671          tmp0 = lp_build_sub( &bld->base, tmp0, tmp1);
 672          emit_store( bld, inst, 0, chan_index, tmp0);
 673       }
 674       break;
 675
 676    case TGSI_OPCODE_LRP:
 677       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 678          src0 = emit_fetch( bld, inst, 0, chan_index );
 679          src1 = emit_fetch( bld, inst, 1, chan_index );
 680          src2 = emit_fetch( bld, inst, 2, chan_index );
 681          tmp0 = lp_build_sub( &bld->base, src1, src2 );
 682          tmp0 = lp_build_mul( &bld->base, src0, tmp0 );
 683          dst0 = lp_build_add( &bld->base, tmp0, src2 );
 684          emit_store( bld, inst, 0, chan_index, dst0 );
 685       }
 686       break;
 687
 688    case TGSI_OPCODE_CND:
 689       return 0;
 690       break;
 691
 692    case TGSI_OPCODE_DP2A:
 693       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );  /* xmm0 = src[0].x */
 694       tmp1 = emit_fetch( bld, inst, 1, CHAN_X );  /* xmm1 = src[1].x */
 695       tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);              /* xmm0 = xmm0 * xmm1 */
 696       tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );  /* xmm1 = src[0].y */
 697       tmp2 = emit_fetch( bld, inst, 1, CHAN_Y );  /* xmm2 = src[1].y */
 698       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);              /* xmm1 = xmm1 * xmm2 */
 699       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);              /* xmm0 = xmm0 + xmm1 */
 700       tmp1 = emit_fetch( bld, inst, 2, CHAN_X );  /* xmm1 = src[2].x */
 701       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);              /* xmm0 = xmm0 + xmm1 */
 702       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 703          emit_store( bld, inst, 0, chan_index, tmp0);  /* dest[ch] = xmm0 */
 704       }
 705       break;
 706
 707 #if 0
 708    case TGSI_OPCODE_FRC:
 709       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 710          tmp0 = emit_fetch( bld, inst, 0, chan_index );
 711          emit_frc( bld, 0, 0 );
 712          emit_store( bld, inst, 0, chan_index, tmp0);
 713       }
 714       break;
 715
 716    case TGSI_OPCODE_CLAMP:
 717       return 0;
 718       break;
 719
 720    case TGSI_OPCODE_FLR:
 721       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 722          tmp0 = emit_fetch( bld, inst, 0, chan_index );
 723          emit_flr( bld, 0, 0 );
 724          emit_store( bld, inst, 0, chan_index, tmp0);
 725       }
 726       break;
 727
 728    case TGSI_OPCODE_ROUND:
 729       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 730          tmp0 = emit_fetch( bld, inst, 0, chan_index );
 731          emit_rnd( bld, 0, 0 );
 732          emit_store( bld, inst, 0, chan_index, tmp0);
 733       }
 734       break;
 735 #endif
 736
 737    case TGSI_OPCODE_EX2: {
 738       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
 739       tmp0 = lp_build_exp2( &bld->base, tmp0);
 740       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 741          emit_store( bld, inst, 0, chan_index, tmp0);
 742       }
 743       break;
 744    }
 745
 746    case TGSI_OPCODE_LG2:
 747       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
 748       tmp0 = lp_build_log2( &bld->base, tmp0);
 749       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 750          emit_store( bld, inst, 0, chan_index, tmp0);
 751       }
 752       break;
 753
 754    case TGSI_OPCODE_POW:
 755       src0 = emit_fetch( bld, inst, 0, CHAN_X );
 756       src1 = emit_fetch( bld, inst, 1, CHAN_X );
 757       dst0 = lp_build_pow( &bld->base, src0, src1 );
 758       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 759          emit_store( bld, inst, 0, chan_index, dst0 );
 760       }
 761       break;
 762
 763    case TGSI_OPCODE_XPD:
 764       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ||
 765           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ) {
 766          tmp1 = emit_fetch( bld, inst, 1, CHAN_Z );
 767          tmp3 = emit_fetch( bld, inst, 0, CHAN_Z );
 768       }
 769       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ||
 770           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) ) {
 771          tmp0 = emit_fetch( bld, inst, 0, CHAN_Y );
 772          tmp4 = emit_fetch( bld, inst, 1, CHAN_Y );
 773       }
 774       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) {
 775          tmp2 = tmp0;
 776          tmp2 = lp_build_mul( &bld->base, tmp2, tmp1);
 777          tmp5 = tmp3;
 778          tmp5 = lp_build_mul( &bld->base, tmp5, tmp4);
 779          tmp2 = lp_build_sub( &bld->base, tmp2, tmp5);
 780          emit_store( bld, inst, 0, CHAN_X, tmp2);
 781       }
 782       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ||
 783           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) ) {
 784          tmp2 = emit_fetch( bld, inst, 1, CHAN_X );
 785          tmp5 = emit_fetch( bld, inst, 0, CHAN_X );
 786       }
 787       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) {
 788          tmp3 = lp_build_mul( &bld->base, tmp3, tmp2);
 789          tmp1 = lp_build_mul( &bld->base, tmp1, tmp5);
 790          tmp3 = lp_build_sub( &bld->base, tmp3, tmp1);
 791          emit_store( bld, inst, 0, CHAN_Y, tmp3);
 792       }
 793       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) {
 794          tmp5 = lp_build_mul( &bld->base, tmp5, tmp4);
 795          tmp0 = lp_build_mul( &bld->base, tmp0, tmp2);
 796          tmp5 = lp_build_sub( &bld->base, tmp5, tmp0);
 797          emit_store( bld, inst, 0, CHAN_Z, tmp5);
 798       }
 799       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) {
 800          tmp0 = bld->base.one;
 801          emit_store( bld, inst, 0, CHAN_W, tmp0);
 802       }
 803       break;
 804
 805    case TGSI_OPCODE_ABS:
 806       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 807          tmp0 = emit_fetch( bld, inst, 0, chan_index );
 808          tmp0 = lp_build_abs( &bld->base, tmp0 ) ;
 809          emit_store( bld, inst, 0, chan_index, tmp0);
 810       }
 811       break;
 812
 813    case TGSI_OPCODE_RCC:
 814       return 0;
 815       break;
 816
 817    case TGSI_OPCODE_DPH:
 818       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
 819       tmp1 = emit_fetch( bld, inst, 1, CHAN_X );
 820       tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);
 821       tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );
 822       tmp2 = emit_fetch( bld, inst, 1, CHAN_Y );
 823       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
 824       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
 825       tmp1 = emit_fetch( bld, inst, 0, CHAN_Z );
 826       tmp2 = emit_fetch( bld, inst, 1, CHAN_Z );
 827       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
 828       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
 829       tmp1 = emit_fetch( bld, inst, 1, CHAN_W );
 830       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
 831       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 832          emit_store( bld, inst, 0, chan_index, tmp0);
 833       }
 834       break;
 835
 836    case TGSI_OPCODE_COS:
 837       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
 838       tmp0 = lp_build_cos( &bld->base, tmp0 );
 839       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 840          emit_store( bld, inst, 0, chan_index, tmp0);
 841       }
 842       break;
 843
 844    case TGSI_OPCODE_DDX:
 845       return 0;
 846       break;
 847
 848    case TGSI_OPCODE_DDY:
 849       return 0;
 850       break;
 851
 852 #if 0
 853    case TGSI_OPCODE_KILP:
 854       /* predicated kill */
 855       emit_kilp( bld );
 856       return 0; /* XXX fix me */
 857       break;
 858 #endif
 859
 860    case TGSI_OPCODE_KIL:
 861       /* conditional kill */
 862       emit_kil( bld, inst );
 863       break;
 864
 865    case TGSI_OPCODE_PK2H:
 866       return 0;
 867       break;
 868
 869    case TGSI_OPCODE_PK2US:
 870       return 0;
 871       break;
 872
 873    case TGSI_OPCODE_PK4B:
 874       return 0;
 875       break;
 876
 877    case TGSI_OPCODE_PK4UB:
 878       return 0;
 879       break;
 880
 881    case TGSI_OPCODE_RFL:
 882       return 0;
 883       break;
 884
 885    case TGSI_OPCODE_SEQ:
 886       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 887          src0 = emit_fetch( bld, inst, 0, chan_index );
 888          src1 = emit_fetch( bld, inst, 1, chan_index );
 889          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_EQUAL, src0, src1 );
 890          dst0 = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
 891          emit_store( bld, inst, 0, chan_index, dst0);
 892       }
 893       break;
 894
 895    case TGSI_OPCODE_SFL:
 896       return 0;
 897       break;
 898
 899    case TGSI_OPCODE_SGT:
 900       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 901          src0 = emit_fetch( bld, inst, 0, chan_index );
 902          src1 = emit_fetch( bld, inst, 1, chan_index );
 903          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_GREATER, src0, src1 );
 904          dst0 = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
 905          emit_store( bld, inst, 0, chan_index, dst0);
 906       }
 907       break;
 908
 909    case TGSI_OPCODE_SIN:
 910       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
 911       tmp0 = lp_build_sin( &bld->base, tmp0 );
 912       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 913          emit_store( bld, inst, 0, chan_index, tmp0);
 914       }
 915       break;
 916
 917    case TGSI_OPCODE_SLE:
 918       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 919          src0 = emit_fetch( bld, inst, 0, chan_index );
 920          src1 = emit_fetch( bld, inst, 1, chan_index );
 921          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_LEQUAL, src0, src1 );
 922          dst0 = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
 923          emit_store( bld, inst, 0, chan_index, dst0);
 924       }
 925       break;
 926
 927    case TGSI_OPCODE_SNE:
 928       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 929          src0 = emit_fetch( bld, inst, 0, chan_index );
 930          src1 = emit_fetch( bld, inst, 1, chan_index );
 931          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_NOTEQUAL, src0, src1 );
 932          dst0 = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
 933          emit_store( bld, inst, 0, chan_index, dst0);
 934       }
 935       break;
 936
 937    case TGSI_OPCODE_STR:
 938       return 0;
 939       break;
 940
 941    case TGSI_OPCODE_TEX:
 942       emit_tex( bld, inst, FALSE, FALSE );
 943       break;
 944
 945    case TGSI_OPCODE_TXD:
 946       return 0;
 947       break;
 948
 949    case TGSI_OPCODE_UP2H:
 950       return 0;
 951       break;
 952
 953    case TGSI_OPCODE_UP2US:
 954       return 0;
 955       break;
 956
 957    case TGSI_OPCODE_UP4B:
 958       return 0;
 959       break;
 960
 961    case TGSI_OPCODE_UP4UB:
 962       return 0;
 963       break;
 964
 965    case TGSI_OPCODE_X2D:
 966       return 0;
 967       break;
 968
 969    case TGSI_OPCODE_ARA:
 970       return 0;
 971       break;
 972
 973 #if 0
 974    case TGSI_OPCODE_ARR:
 975       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
 976          tmp0 = emit_fetch( bld, inst, 0, chan_index );
 977          emit_rnd( bld, 0, 0 );
 978          emit_f2it( bld, 0 );
 979          emit_store( bld, inst, 0, chan_index, tmp0);
 980       }
 981       break;
 982 #endif
 983
 984    case TGSI_OPCODE_BRA:
 985       return 0;
 986       break;
 987
 988    case TGSI_OPCODE_CAL:
 989       return 0;
 990       break;
 991
 992 #if 0
 993    case TGSI_OPCODE_RET:
 994       emit_ret( bld );
 995       break;
 996 #endif
 997
 998    case TGSI_OPCODE_END:
 999       break;
1000
1001 #if 0
1002    case TGSI_OPCODE_SSG:
1003    /* TGSI_OPCODE_SGN */
1004       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1005          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1006          emit_sgn( bld, 0, 0 );
1007          emit_store( bld, inst, 0, chan_index, tmp0);
1008       }
1009       break;
1010 #endif
1011
1012    case TGSI_OPCODE_CMP:
1013       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1014          src0 = emit_fetch( bld, inst, 0, chan_index );
1015          src1 = emit_fetch( bld, inst, 1, chan_index );
1016          src2 = emit_fetch( bld, inst, 2, chan_index );
1017          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_LESS, src0, bld->base.zero );
1018          dst0 = lp_build_select( &bld->base, tmp0, src1, src2);
1019          emit_store( bld, inst, 0, chan_index, dst0);
1020       }
1021       break;
1022
1023    case TGSI_OPCODE_SCS:
1024       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) {
1025          tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1026          tmp0 = lp_build_cos( &bld->base, tmp0 );
1027          emit_store( bld, inst, 0, CHAN_X, tmp0);
1028       }
1029       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) {
1030          tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1031          tmp0 = lp_build_sin( &bld->base, tmp0 );
1032          emit_store( bld, inst, 0, CHAN_Y, tmp0);
1033       }
1034       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) {
1035          tmp0 = bld->base.zero;
1036          emit_store( bld, inst, 0, CHAN_Z, tmp0);
1037       }
1038       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) {
1039          tmp0 = bld->base.one;
1040          emit_store( bld, inst, 0, CHAN_W, tmp0);
1041       }
1042       break;
1043
1044    case TGSI_OPCODE_TXB:
1045       emit_tex( bld, inst, TRUE, FALSE );
1046       break;
1047
1048    case TGSI_OPCODE_NRM:
1049       /* fall-through */
1050    case TGSI_OPCODE_NRM4:
1051       /* 3 or 4-component normalization */
1052       {
1053          uint dims = (inst->Instruction.Opcode == TGSI_OPCODE_NRM) ? 3 : 4;
1054
1055          if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X) ||
1056              IS_DST0_CHANNEL_ENABLED(inst, CHAN_Y) ||
1057              IS_DST0_CHANNEL_ENABLED(inst, CHAN_Z) ||
1058              (IS_DST0_CHANNEL_ENABLED(inst, CHAN_W) && dims == 4)) {
1059
1060             /* NOTE: Cannot use xmm regs 2/3 here (see emit_rsqrt() above). */
1061
1062             /* xmm4 = src.x */
1063             /* xmm0 = src.x * src.x */
1064             tmp0 = emit_fetch(bld, inst, 0, CHAN_X);
1065             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X)) {
1066                tmp4 = tmp0;
1067             }
1068             tmp0 = lp_build_mul( &bld->base, tmp0, tmp0);
1069
1070             /* xmm5 = src.y */
1071             /* xmm0 = xmm0 + src.y * src.y */
1072             tmp1 = emit_fetch(bld, inst, 0, CHAN_Y);
1073             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Y)) {
1074                tmp5 = tmp1;
1075             }
1076             tmp1 = lp_build_mul( &bld->base, tmp1, tmp1);
1077             tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1078
1079             /* xmm6 = src.z */
1080             /* xmm0 = xmm0 + src.z * src.z */
1081             tmp1 = emit_fetch(bld, inst, 0, CHAN_Z);
1082             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Z)) {
1083                tmp6 = tmp1;
1084             }
1085             tmp1 = lp_build_mul( &bld->base, tmp1, tmp1);
1086             tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1087
1088             if (dims == 4) {
1089                /* xmm7 = src.w */
1090                /* xmm0 = xmm0 + src.w * src.w */
1091                tmp1 = emit_fetch(bld, inst, 0, CHAN_W);
1092                if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_W)) {
1093                   tmp7 = tmp1;
1094                }
1095                tmp1 = lp_build_mul( &bld->base, tmp1, tmp1);
1096                tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1097             }
1098
1099             /* xmm1 = 1 / sqrt(xmm0) */
1100             tmp1 = lp_build_rsqrt( &bld->base, tmp0);
1101
1102             /* dst.x = xmm1 * src.x */
1103             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X)) {
1104                tmp4 = lp_build_mul( &bld->base, tmp4, tmp1);
1105                emit_store(bld, inst, 0, CHAN_X, tmp4);
1106             }
1107
1108             /* dst.y = xmm1 * src.y */
1109             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Y)) {
1110                tmp5 = lp_build_mul( &bld->base, tmp5, tmp1);
1111                emit_store(bld, inst, 0, CHAN_Y, tmp5);
1112             }
1113
1114             /* dst.z = xmm1 * src.z */
1115             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Z)) {
1116                tmp6 = lp_build_mul( &bld->base, tmp6, tmp1);
1117                emit_store(bld, inst, 0, CHAN_Z, tmp6);
1118             }
1119
1120             /* dst.w = xmm1 * src.w */
1121             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X) && dims == 4) {
1122                tmp7 = lp_build_mul( &bld->base, tmp7, tmp1);
1123                emit_store(bld, inst, 0, CHAN_W, tmp7);
1124             }
1125          }
1126
1127          /* dst0.w = 1.0 */
1128          if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_W) && dims == 3) {
1129             tmp0 = bld->base.one;
1130             emit_store(bld, inst, 0, CHAN_W, tmp0);
1131          }
1132       }
1133       break;
1134
1135    case TGSI_OPCODE_DIV:
1136       return 0;
1137       break;
1138
1139    case TGSI_OPCODE_DP2:
1140       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );  /* xmm0 = src[0].x */
1141       tmp1 = emit_fetch( bld, inst, 1, CHAN_X );  /* xmm1 = src[1].x */
1142       tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);              /* xmm0 = xmm0 * xmm1 */
1143       tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );  /* xmm1 = src[0].y */
1144       tmp2 = emit_fetch( bld, inst, 1, CHAN_Y );  /* xmm2 = src[1].y */
1145       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);              /* xmm1 = xmm1 * xmm2 */
1146       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);              /* xmm0 = xmm0 + xmm1 */
1147       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1148          emit_store( bld, inst, 0, chan_index, tmp0);  /* dest[ch] = xmm0 */
1149       }
1150       break;
1151
1152    case TGSI_OPCODE_TXL:
1153       emit_tex( bld, inst, TRUE, FALSE );
1154       break;
1155
1156    case TGSI_OPCODE_TXP:
1157       emit_tex( bld, inst, FALSE, TRUE );
1158       break;
1159
1160    case TGSI_OPCODE_BRK:
1161       return 0;
1162       break;
1163
1164    case TGSI_OPCODE_IF:
1165       return 0;
1166       break;
1167
1168    case TGSI_OPCODE_BGNFOR:
1169       return 0;
1170       break;
1171
1172    case TGSI_OPCODE_REP:
1173       return 0;
1174       break;
1175
1176    case TGSI_OPCODE_ELSE:
1177       return 0;
1178       break;
1179
1180    case TGSI_OPCODE_ENDIF:
1181       return 0;
1182       break;
1183
1184    case TGSI_OPCODE_ENDFOR:
1185       return 0;
1186       break;
1187
1188    case TGSI_OPCODE_ENDREP:
1189       return 0;
1190       break;
1191
1192    case TGSI_OPCODE_PUSHA:
1193       return 0;
1194       break;
1195
1196    case TGSI_OPCODE_POPA:
1197       return 0;
1198       break;
1199
1200    case TGSI_OPCODE_CEIL:
1201       return 0;
1202       break;
1203
1204    case TGSI_OPCODE_I2F:
1205       return 0;
1206       break;
1207
1208    case TGSI_OPCODE_NOT:
1209       return 0;
1210       break;
1211
1212 #if 0
1213    case TGSI_OPCODE_TRUNC:
1214       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1215          tmp0 = emit_fetch( bld, inst, 0, chan_index );
1216          emit_f2it( bld, 0 );
1217          emit_i2f( bld, 0 );
1218          emit_store( bld, inst, 0, chan_index, tmp0);
1219       }
1220       break;
1221 #endif
1222
1223    case TGSI_OPCODE_SHL:
1224       return 0;
1225       break;
1226
1227    case TGSI_OPCODE_SHR:
1228       return 0;
1229       break;
1230
1231    case TGSI_OPCODE_AND:
1232       return 0;
1233       break;
1234
1235    case TGSI_OPCODE_OR:
1236       return 0;
1237       break;
1238
1239    case TGSI_OPCODE_MOD:
1240       return 0;
1241       break;
1242
1243    case TGSI_OPCODE_XOR:
1244       return 0;
1245       break;
1246
1247    case TGSI_OPCODE_SAD:
1248       return 0;
1249       break;
1250
1251    case TGSI_OPCODE_TXF:
1252       return 0;
1253       break;
1254
1255    case TGSI_OPCODE_TXQ:
1256       return 0;
1257       break;
1258
1259    case TGSI_OPCODE_CONT:
1260       return 0;
1261       break;
1262
1263    case TGSI_OPCODE_EMIT:
1264       return 0;
1265       break;
1266
1267    case TGSI_OPCODE_ENDPRIM:
1268       return 0;
1269       break;
1270
1271    default:
1272       return 0;
1273    }
1274
1275    return 1;
1276 }
1277
1278
1279 void
1280 lp_build_tgsi_soa(LLVMBuilderRef builder,
1281                   const struct tgsi_token *tokens,
1282                   union lp_type type,
1283                   struct lp_build_mask_context *mask,
1284                   LLVMValueRef consts_ptr,
1285                   const LLVMValueRef *pos,
1286                   const LLVMValueRef (*inputs)[NUM_CHANNELS],
1287                   LLVMValueRef (*outputs)[NUM_CHANNELS],
1288                   struct lp_build_sampler_soa *sampler)
1289 {
1290    struct lp_build_tgsi_soa_context bld;
1291    struct tgsi_parse_context parse;
1292    uint num_immediates = 0;
1293    unsigned i;
1294
1295    /* Setup build context */
1296    memset(&bld, 0, sizeof bld);
1297    lp_build_context_init(&bld.base, builder, type);
1298    bld.mask = mask;
1299    bld.pos = pos;
1300    bld.inputs = inputs;
1301    bld.outputs = outputs;
1302    bld.consts_ptr = consts_ptr;
1303    bld.sampler = sampler;
1304
1305    tgsi_parse_init( &parse, tokens );
1306
1307    while( !tgsi_parse_end_of_tokens( &parse ) ) {
1308       tgsi_parse_token( &parse );
1309
1310       switch( parse.FullToken.Token.Type ) {
1311       case TGSI_TOKEN_TYPE_DECLARATION:
1312          /* Input already interpolated */
1313          break;
1314
1315       case TGSI_TOKEN_TYPE_INSTRUCTION:
1316          if (!emit_instruction( &bld, &parse.FullToken.FullInstruction )) {
1317             unsigned opcode = parse.FullToken.FullInstruction.Instruction.Opcode;
1318             const struct tgsi_opcode_info *info = tgsi_get_opcode_info(opcode);
1319             _debug_printf("warning: failed to translate tgsi opcode %s to LLVM\n",
1320                           info ? info->mnemonic : "<invalid>");
1321          }
1322          break;
1323
1324       case TGSI_TOKEN_TYPE_IMMEDIATE:
1325          /* simply copy the immediate values into the next immediates[] slot */
1326          {
1327             const uint size = parse.FullToken.FullImmediate.Immediate.NrTokens - 1;
1328             assert(size <= 4);
1329             assert(num_immediates < LP_MAX_IMMEDIATES);
1330             for( i = 0; i < size; ++i )
1331                bld.immediates[num_immediates][i] =
1332                   lp_build_const_scalar(type, parse.FullToken.FullImmediate.u[i].Float);
1333             for( i = size; i < 4; ++i )
1334                bld.immediates[num_immediates][i] = bld.base.undef;
1335             num_immediates++;
1336          }
1337          break;
1338
1339       default:
1340          assert( 0 );
1341       }
1342    }
1343
1344    tgsi_parse_free( &parse );
1345 }
1346