src/gallium/auxiliary/gallivm/lp_bld_logic.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2009 VMware, Inc.
   4  * All Rights Reserved.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the
   8  * "Software"), to deal in the Software without restriction, including
   9  * without limitation the rights to use, copy, modify, merge, publish,
  10  * distribute, sub license, and/or sell copies of the Software, and to
  11  * permit persons to whom the Software is furnished to do so, subject to
  12  * the following conditions:
  13  *
  14  * The above copyright notice and this permission notice (including the
  15  * next paragraph) shall be included in all copies or substantial portions
  16  * of the Software.
  17  *
  18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  19  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  20  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  21  * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
  22  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  23  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  24  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  25  *
  26  **************************************************************************/
  27
  28 /**
  29  * @file
  30  * Helper functions for logical operations.
  31  *
  32  * @author Jose Fonseca <jfonseca@vmware.com>
  33  */
  34
  35
  36 #include "util/u_cpu_detect.h"
  37 #include "util/u_debug.h"
  38
  39 #include "lp_bld_type.h"
  40 #include "lp_bld_const.h"
  41 #include "lp_bld_intr.h"
  42 #include "lp_bld_logic.h"
  43
  44
  45 /*
  46  * XXX
  47  *
  48  * Selection with vector conditional like
  49  *
  50  *    select <4 x i1> %C, %A, %B
  51  *
  52  * is valid IR (e.g. llvm/test/Assembler/vector-select.ll), but it is not
  53  * supported on any backend.
  54  *
  55  * Expanding the boolean vector to full SIMD register width, as in
  56  *
  57  *    sext <4 x i1> %C to <4 x i32>
  58  *
  59  * is valid and supported (e.g., llvm/test/CodeGen/X86/vec_compare.ll), but
  60  * it causes assertion failures in LLVM 2.6. It appears to work correctly on
  61  * LLVM 2.7.
  62  */
  63
  64
  65 /**
  66  * Build code to compare two values 'a' and 'b' of 'type' using the given func.
  67  * \param func  one of PIPE_FUNC_x
  68  * The result values will be 0 for false or ~0 for true.
  69  */
  70 LLVMValueRef
  71 lp_build_compare(LLVMBuilderRef builder,
  72                  const struct lp_type type,
  73                  unsigned func,
  74                  LLVMValueRef a,
  75                  LLVMValueRef b)
  76 {
  77    LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
  78    LLVMValueRef zeros = LLVMConstNull(int_vec_type);
  79    LLVMValueRef ones = LLVMConstAllOnes(int_vec_type);
  80    LLVMValueRef cond;
  81    LLVMValueRef res;
  82
  83    assert(func >= PIPE_FUNC_NEVER);
  84    assert(func <= PIPE_FUNC_ALWAYS);
  85
  86    if(func == PIPE_FUNC_NEVER)
  87       return zeros;
  88    if(func == PIPE_FUNC_ALWAYS)
  89       return ones;
  90
  91    /* TODO: optimize the constant case */
  92
  93    /* XXX: It is not clear if we should use the ordered or unordered operators */
  94
  95 #if HAVE_LLVM < 0x0207
  96 #if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
  97    if(type.width * type.length == 128) {
  98       if(type.floating && util_cpu_caps.has_sse) {
  99          /* float[4] comparison */
 100          LLVMTypeRef vec_type = lp_build_vec_type(type);
 101          LLVMValueRef args[3];
 102          unsigned cc;
 103          boolean swap;
 104
 105          swap = FALSE;
 106          switch(func) {
 107          case PIPE_FUNC_EQUAL:
 108             cc = 0;
 109             break;
 110          case PIPE_FUNC_NOTEQUAL:
 111             cc = 4;
 112             break;
 113          case PIPE_FUNC_LESS:
 114             cc = 1;
 115             break;
 116          case PIPE_FUNC_LEQUAL:
 117             cc = 2;
 118             break;
 119          case PIPE_FUNC_GREATER:
 120             cc = 1;
 121             swap = TRUE;
 122             break;
 123          case PIPE_FUNC_GEQUAL:
 124             cc = 2;
 125             swap = TRUE;
 126             break;
 127          default:
 128             assert(0);
 129             return lp_build_undef(type);
 130          }
 131
 132          if(swap) {
 133             args[0] = b;
 134             args[1] = a;
 135          }
 136          else {
 137             args[0] = a;
 138             args[1] = b;
 139          }
 140
 141          args[2] = LLVMConstInt(LLVMInt8Type(), cc, 0);
 142          res = lp_build_intrinsic(builder,
 143                                   "llvm.x86.sse.cmp.ps",
 144                                   vec_type,
 145                                   args, 3);
 146          res = LLVMBuildBitCast(builder, res, int_vec_type, "");
 147          return res;
 148       }
 149       else if(util_cpu_caps.has_sse2) {
 150          /* int[4] comparison */
 151          static const struct {
 152             unsigned swap:1;
 153             unsigned eq:1;
 154             unsigned gt:1;
 155             unsigned not:1;
 156          } table[] = {
 157             {0, 0, 0, 1}, /* PIPE_FUNC_NEVER */
 158             {1, 0, 1, 0}, /* PIPE_FUNC_LESS */
 159             {0, 1, 0, 0}, /* PIPE_FUNC_EQUAL */
 160             {0, 0, 1, 1}, /* PIPE_FUNC_LEQUAL */
 161             {0, 0, 1, 0}, /* PIPE_FUNC_GREATER */
 162             {0, 1, 0, 1}, /* PIPE_FUNC_NOTEQUAL */
 163             {1, 0, 1, 1}, /* PIPE_FUNC_GEQUAL */
 164             {0, 0, 0, 0}  /* PIPE_FUNC_ALWAYS */
 165          };
 166          const char *pcmpeq;
 167          const char *pcmpgt;
 168          LLVMValueRef args[2];
 169          LLVMValueRef res;
 170          LLVMTypeRef vec_type = lp_build_vec_type(type);
 171
 172          switch (type.width) {
 173          case 8:
 174             pcmpeq = "llvm.x86.sse2.pcmpeq.b";
 175             pcmpgt = "llvm.x86.sse2.pcmpgt.b";
 176             break;
 177          case 16:
 178             pcmpeq = "llvm.x86.sse2.pcmpeq.w";
 179             pcmpgt = "llvm.x86.sse2.pcmpgt.w";
 180             break;
 181          case 32:
 182             pcmpeq = "llvm.x86.sse2.pcmpeq.d";
 183             pcmpgt = "llvm.x86.sse2.pcmpgt.d";
 184             break;
 185          default:
 186             assert(0);
 187             return lp_build_undef(type);
 188          }
 189
 190          /* There are no signed byte and unsigned word/dword comparison
 191           * instructions. So flip the sign bit so that the results match.
 192           */
 193          if(table[func].gt &&
 194             ((type.width == 8 && type.sign) ||
 195              (type.width != 8 && !type.sign))) {
 196             LLVMValueRef msb = lp_build_const_int_vec(type, (unsigned long long)1 << (type.width - 1));
 197             a = LLVMBuildXor(builder, a, msb, "");
 198             b = LLVMBuildXor(builder, b, msb, "");
 199          }
 200
 201          if(table[func].swap) {
 202             args[0] = b;
 203             args[1] = a;
 204          }
 205          else {
 206             args[0] = a;
 207             args[1] = b;
 208          }
 209
 210          if(table[func].eq)
 211             res = lp_build_intrinsic(builder, pcmpeq, vec_type, args, 2);
 212          else if (table[func].gt)
 213             res = lp_build_intrinsic(builder, pcmpgt, vec_type, args, 2);
 214          else
 215             res = LLVMConstNull(vec_type);
 216
 217          if(table[func].not)
 218             res = LLVMBuildNot(builder, res, "");
 219
 220          return res;
 221       }
 222    } /* if (type.width * type.length == 128) */
 223 #endif
 224 #endif /* HAVE_LLVM < 0x0207 */
 225
 226    if(type.floating) {
 227       LLVMRealPredicate op;
 228       switch(func) {
 229       case PIPE_FUNC_NEVER:
 230          op = LLVMRealPredicateFalse;
 231          break;
 232       case PIPE_FUNC_ALWAYS:
 233          op = LLVMRealPredicateTrue;
 234          break;
 235       case PIPE_FUNC_EQUAL:
 236          op = LLVMRealUEQ;
 237          break;
 238       case PIPE_FUNC_NOTEQUAL:
 239          op = LLVMRealUNE;
 240          break;
 241       case PIPE_FUNC_LESS:
 242          op = LLVMRealULT;
 243          break;
 244       case PIPE_FUNC_LEQUAL:
 245          op = LLVMRealULE;
 246          break;
 247       case PIPE_FUNC_GREATER:
 248          op = LLVMRealUGT;
 249          break;
 250       case PIPE_FUNC_GEQUAL:
 251          op = LLVMRealUGE;
 252          break;
 253       default:
 254          assert(0);
 255          return lp_build_undef(type);
 256       }
 257
 258 #if HAVE_LLVM >= 0x0207
 259       cond = LLVMBuildFCmp(builder, op, a, b, "");
 260       res = LLVMBuildSExt(builder, cond, int_vec_type, "");
 261 #else
 262       if (type.length == 1) {
 263          cond = LLVMBuildFCmp(builder, op, a, b, "");
 264          res = LLVMBuildSExt(builder, cond, int_vec_type, "");
 265       }
 266       else {
 267          unsigned i;
 268
 269          res = LLVMGetUndef(int_vec_type);
 270
 271          debug_printf("%s: warning: using slow element-wise float"
 272                       " vector comparison\n", __FUNCTION__);
 273          for (i = 0; i < type.length; ++i) {
 274             LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
 275             cond = LLVMBuildFCmp(builder, op,
 276                                  LLVMBuildExtractElement(builder, a, index, ""),
 277                                  LLVMBuildExtractElement(builder, b, index, ""),
 278                                  "");
 279             cond = LLVMBuildSelect(builder, cond,
 280                                    LLVMConstExtractElement(ones, index),
 281                                    LLVMConstExtractElement(zeros, index),
 282                                    "");
 283             res = LLVMBuildInsertElement(builder, res, cond, index, "");
 284          }
 285       }
 286 #endif
 287    }
 288    else {
 289       LLVMIntPredicate op;
 290       switch(func) {
 291       case PIPE_FUNC_EQUAL:
 292          op = LLVMIntEQ;
 293          break;
 294       case PIPE_FUNC_NOTEQUAL:
 295          op = LLVMIntNE;
 296          break;
 297       case PIPE_FUNC_LESS:
 298          op = type.sign ? LLVMIntSLT : LLVMIntULT;
 299          break;
 300       case PIPE_FUNC_LEQUAL:
 301          op = type.sign ? LLVMIntSLE : LLVMIntULE;
 302          break;
 303       case PIPE_FUNC_GREATER:
 304          op = type.sign ? LLVMIntSGT : LLVMIntUGT;
 305          break;
 306       case PIPE_FUNC_GEQUAL:
 307          op = type.sign ? LLVMIntSGE : LLVMIntUGE;
 308          break;
 309       default:
 310          assert(0);
 311          return lp_build_undef(type);
 312       }
 313
 314 #if HAVE_LLVM >= 0x0207
 315       cond = LLVMBuildICmp(builder, op, a, b, "");
 316       res = LLVMBuildSExt(builder, cond, int_vec_type, "");
 317 #else
 318       if (type.length == 1) {
 319          cond = LLVMBuildICmp(builder, op, a, b, "");
 320          res = LLVMBuildSExt(builder, cond, int_vec_type, "");
 321       }
 322       else {
 323          unsigned i;
 324
 325          res = LLVMGetUndef(int_vec_type);
 326
 327          debug_printf("%s: warning: using slow element-wise int"
 328                       " vector comparison\n", __FUNCTION__);
 329
 330          for(i = 0; i < type.length; ++i) {
 331             LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
 332             cond = LLVMBuildICmp(builder, op,
 333                                  LLVMBuildExtractElement(builder, a, index, ""),
 334                                  LLVMBuildExtractElement(builder, b, index, ""),
 335                                  "");
 336             cond = LLVMBuildSelect(builder, cond,
 337                                    LLVMConstExtractElement(ones, index),
 338                                    LLVMConstExtractElement(zeros, index),
 339                                    "");
 340             res = LLVMBuildInsertElement(builder, res, cond, index, "");
 341          }
 342       }
 343 #endif
 344    }
 345
 346    return res;
 347 }
 348
 349
 350
 351 /**
 352  * Build code to compare two values 'a' and 'b' using the given func.
 353  * \param func  one of PIPE_FUNC_x
 354  * The result values will be 0 for false or ~0 for true.
 355  */
 356 LLVMValueRef
 357 lp_build_cmp(struct lp_build_context *bld,
 358              unsigned func,
 359              LLVMValueRef a,
 360              LLVMValueRef b)
 361 {
 362    return lp_build_compare(bld->builder, bld->type, func, a, b);
 363 }
 364
 365
 366 /**
 367  * Return mask ? a : b;
 368  *
 369  * mask is a bitwise mask, composed of 0 or ~0 for each element.
 370  */
 371 LLVMValueRef
 372 lp_build_select(struct lp_build_context *bld,
 373                 LLVMValueRef mask,
 374                 LLVMValueRef a,
 375                 LLVMValueRef b)
 376 {
 377    struct lp_type type = bld->type;
 378    LLVMValueRef res;
 379
 380    if(a == b)
 381       return a;
 382
 383    if (type.length == 1) {
 384       mask = LLVMBuildTrunc(bld->builder, mask, LLVMInt1Type(), "");
 385       res = LLVMBuildSelect(bld->builder, mask, a, b, "");
 386    }
 387    else {
 388       if(type.floating) {
 389          LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
 390          a = LLVMBuildBitCast(bld->builder, a, int_vec_type, "");
 391          b = LLVMBuildBitCast(bld->builder, b, int_vec_type, "");
 392       }
 393
 394       a = LLVMBuildAnd(bld->builder, a, mask, "");
 395
 396       /* This often gets translated to PANDN, but sometimes the NOT is
 397        * pre-computed and stored in another constant. The best strategy depends
 398        * on available registers, so it is not a big deal -- hopefully LLVM does
 399        * the right decision attending the rest of the program.
 400        */
 401       b = LLVMBuildAnd(bld->builder, b, LLVMBuildNot(bld->builder, mask, ""), "");
 402
 403       res = LLVMBuildOr(bld->builder, a, b, "");
 404
 405       if(type.floating) {
 406          LLVMTypeRef vec_type = lp_build_vec_type(type);
 407          res = LLVMBuildBitCast(bld->builder, res, vec_type, "");
 408       }
 409    }
 410
 411    return res;
 412 }
 413
 414
 415 LLVMValueRef
 416 lp_build_select_aos(struct lp_build_context *bld,
 417                     LLVMValueRef a,
 418                     LLVMValueRef b,
 419                     const boolean cond[4])
 420 {
 421    const struct lp_type type = bld->type;
 422    const unsigned n = type.length;
 423    unsigned i, j;
 424
 425    if(a == b)
 426       return a;
 427    if(cond[0] && cond[1] && cond[2] && cond[3])
 428       return a;
 429    if(!cond[0] && !cond[1] && !cond[2] && !cond[3])
 430       return b;
 431    if(a == bld->undef || b == bld->undef)
 432       return bld->undef;
 433
 434    /*
 435     * There are three major ways of accomplishing this:
 436     * - with a shuffle,
 437     * - with a select,
 438     * - or with a bit mask.
 439     *
 440     * Select isn't supported for vector types yet.
 441     * The flip between these is empirical and might need to be.
 442     */
 443    if (n <= 4) {
 444       /*
 445        * Shuffle.
 446        */
 447       LLVMTypeRef elem_type = LLVMInt32Type();
 448       LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
 449
 450       for(j = 0; j < n; j += 4)
 451          for(i = 0; i < 4; ++i)
 452             shuffles[j + i] = LLVMConstInt(elem_type, (cond[i] ? 0 : n) + j + i, 0);
 453
 454       return LLVMBuildShuffleVector(bld->builder, a, b, LLVMConstVector(shuffles, n), "");
 455    }
 456    else {
 457 #if 0
 458       /* XXX: Unfortunately select of vectors do not work */
 459       /* Use a select */
 460       LLVMTypeRef elem_type = LLVMInt1Type();
 461       LLVMValueRef cond[LP_MAX_VECTOR_LENGTH];
 462
 463       for(j = 0; j < n; j += 4)
 464          for(i = 0; i < 4; ++i)
 465             cond[j + i] = LLVMConstInt(elem_type, cond[i] ? 1 : 0, 0);
 466
 467       return LLVMBuildSelect(bld->builder, LLVMConstVector(cond, n), a, b, "");
 468 #else
 469       LLVMValueRef mask = lp_build_const_mask_aos(type, cond);
 470       return lp_build_select(bld, mask, a, b);
 471 #endif
 472    }
 473 }
 474
 475 LLVMValueRef
 476 lp_build_alloca(struct lp_build_context *bld)
 477 {
 478    const struct lp_type type = bld->type;
 479
 480    if (type.length > 1) { /*vector*/
 481       return LLVMBuildAlloca(bld->builder, lp_build_vec_type(type), "");
 482    } else { /*scalar*/
 483       return LLVMBuildAlloca(bld->builder, lp_build_elem_type(type), "");
 484    }
 485 }