src/gallium/auxiliary/gallivm/lp_bld_init.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 #include "pipe/p_config.h"
  30 #include "pipe/p_compiler.h"
  31 #include "util/u_cpu_detect.h"
  32 #include "util/u_debug.h"
  33 #include "util/u_memory.h"
  34 #include "util/u_simple_list.h"
  35 #include "lp_bld.h"
  36 #include "lp_bld_debug.h"
  37 #include "lp_bld_misc.h"
  38 #include "lp_bld_init.h"
  39
  40 #include <llvm-c/Analysis.h>
  41 #include <llvm-c/Transforms/Scalar.h>
  42 #include <llvm-c/BitWriter.h>
  43
  44
  45 /**
  46  * AVX is supported in:
  47  * - standard JIT from LLVM 3.2 onwards
  48  * - MC-JIT from LLVM 3.1
  49  *   - MC-JIT supports limited OSes (MacOSX and Linux)
  50  * - standard JIT in LLVM 3.1, with backports
  51  */
  52 #if defined(PIPE_ARCH_PPC_64)
  53 #  define USE_MCJIT 1
  54 #  define HAVE_AVX 0
  55 #elif HAVE_LLVM >= 0x0302 || (HAVE_LLVM == 0x0301 && defined(HAVE_JIT_AVX_SUPPORT))
  56 #  define USE_MCJIT 0
  57 #  define HAVE_AVX 1
  58 #elif HAVE_LLVM == 0x0301 && (defined(PIPE_OS_LINUX) || defined(PIPE_OS_APPLE))
  59 #  define USE_MCJIT 1
  60 #  define HAVE_AVX 1
  61 #else
  62 #  define USE_MCJIT 0
  63 #  define HAVE_AVX 0
  64 #endif
  65
  66
  67 #if USE_MCJIT
  68 void LLVMLinkInMCJIT();
  69 #endif
  70
  71
  72 #ifdef DEBUG
  73 unsigned gallivm_debug = 0;
  74
  75 static const struct debug_named_value lp_bld_debug_flags[] = {
  76    { "tgsi",   GALLIVM_DEBUG_TGSI, NULL },
  77    { "ir",     GALLIVM_DEBUG_IR, NULL },
  78    { "asm",    GALLIVM_DEBUG_ASM, NULL },
  79    { "nopt",   GALLIVM_DEBUG_NO_OPT, NULL },
  80    { "perf",   GALLIVM_DEBUG_PERF, NULL },
  81    { "no_brilinear", GALLIVM_DEBUG_NO_BRILINEAR, NULL },
  82    { "gc",     GALLIVM_DEBUG_GC, NULL },
  83    DEBUG_NAMED_VALUE_END
  84 };
  85
  86 DEBUG_GET_ONCE_FLAGS_OPTION(gallivm_debug, "GALLIVM_DEBUG", lp_bld_debug_flags, 0)
  87 #endif
  88
  89
  90 static boolean gallivm_initialized = FALSE;
  91
  92 unsigned lp_native_vector_width;
  93
  94
  95 /*
  96  * Optimization values are:
  97  * - 0: None (-O0)
  98  * - 1: Less (-O1)
  99  * - 2: Default (-O2, -Os)
 100  * - 3: Aggressive (-O3)
 101  *
 102  * See also CodeGenOpt::Level in llvm/Target/TargetMachine.h
 103  */
 104 enum LLVM_CodeGenOpt_Level {
 105 #if HAVE_LLVM >= 0x207
 106    None,        // -O0
 107    Less,        // -O1
 108    Default,     // -O2, -Os
 109    Aggressive   // -O3
 110 #else
 111    Default,
 112    None,
 113    Aggressive
 114 #endif
 115 };
 116
 117
 118 #if HAVE_LLVM <= 0x0206
 119 /**
 120  * LLVM 2.6 permits only one ExecutionEngine to be created.  So use the
 121  * same gallivm state everywhere.
 122  */
 123 static struct gallivm_state *GlobalGallivm = NULL;
 124 #endif
 125
 126
 127 /**
 128  * Create the LLVM (optimization) pass manager and install
 129  * relevant optimization passes.
 130  * \return  TRUE for success, FALSE for failure
 131  */
 132 static boolean
 133 create_pass_manager(struct gallivm_state *gallivm)
 134 {
 135    assert(!gallivm->passmgr);
 136    assert(gallivm->target);
 137
 138    gallivm->passmgr = LLVMCreateFunctionPassManager(gallivm->provider);
 139    if (!gallivm->passmgr)
 140       return FALSE;
 141
 142    LLVMAddTargetData(gallivm->target, gallivm->passmgr);
 143
 144    if ((gallivm_debug & GALLIVM_DEBUG_NO_OPT) == 0) {
 145       /* These are the passes currently listed in llvm-c/Transforms/Scalar.h,
 146        * but there are more on SVN.
 147        * TODO: Add more passes.
 148        */
 149       LLVMAddCFGSimplificationPass(gallivm->passmgr);
 150
 151       if (HAVE_LLVM >= 0x207 && sizeof(void*) == 4) {
 152          /* For LLVM >= 2.7 and 32-bit build, use this order of passes to
 153           * avoid generating bad code.
 154           * Test with piglit glsl-vs-sqrt-zero test.
 155           */
 156          LLVMAddConstantPropagationPass(gallivm->passmgr);
 157          LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
 158       }
 159       else {
 160          LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
 161          LLVMAddConstantPropagationPass(gallivm->passmgr);
 162       }
 163
 164       if (util_cpu_caps.has_sse4_1) {
 165          /* FIXME: There is a bug in this pass, whereby the combination
 166           * of fptosi and sitofp (necessary for trunc/floor/ceil/round
 167           * implementation) somehow becomes invalid code.
 168           */
 169          LLVMAddInstructionCombiningPass(gallivm->passmgr);
 170       }
 171       LLVMAddGVNPass(gallivm->passmgr);
 172    }
 173    else {
 174       /* We need at least this pass to prevent the backends to fail in
 175        * unexpected ways.
 176        */
 177       LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
 178    }
 179
 180    return TRUE;
 181 }
 182
 183
 184 /**
 185  * Free gallivm object's LLVM allocations, but not the gallivm object itself.
 186  */
 187 static void
 188 free_gallivm_state(struct gallivm_state *gallivm)
 189 {
 190 #if HAVE_LLVM >= 0x207 /* XXX or 0x208? */
 191    /* This leads to crashes w/ some versions of LLVM */
 192    LLVMModuleRef mod;
 193    char *error;
 194
 195    if (gallivm->engine && gallivm->provider)
 196       LLVMRemoveModuleProvider(gallivm->engine, gallivm->provider,
 197                                &mod, &error);
 198 #endif
 199
 200    if (gallivm->passmgr) {
 201       LLVMDisposePassManager(gallivm->passmgr);
 202    }
 203
 204 #if 0
 205    /* XXX this seems to crash with all versions of LLVM */
 206    if (gallivm->provider)
 207       LLVMDisposeModuleProvider(gallivm->provider);
 208 #endif
 209
 210    if (HAVE_LLVM >= 0x207 && gallivm->engine) {
 211       /* This will already destroy any associated module */
 212       LLVMDisposeExecutionEngine(gallivm->engine);
 213    } else {
 214       LLVMDisposeModule(gallivm->module);
 215    }
 216
 217 #if !USE_MCJIT
 218    /* Don't free the TargetData, it's owned by the exec engine */
 219 #else
 220    if (gallivm->target) {
 221       LLVMDisposeTargetData(gallivm->target);
 222    }
 223 #endif
 224
 225    /* Never free the LLVM context.
 226     */
 227 #if 0
 228    if (gallivm->context)
 229       LLVMContextDispose(gallivm->context);
 230 #endif
 231
 232    if (gallivm->builder)
 233       LLVMDisposeBuilder(gallivm->builder);
 234
 235    gallivm->engine = NULL;
 236    gallivm->target = NULL;
 237    gallivm->module = NULL;
 238    gallivm->provider = NULL;
 239    gallivm->passmgr = NULL;
 240    gallivm->context = NULL;
 241    gallivm->builder = NULL;
 242 }
 243
 244
 245 static boolean
 246 init_gallivm_engine(struct gallivm_state *gallivm)
 247 {
 248    if (1) {
 249       /* We can only create one LLVMExecutionEngine (w/ LLVM 2.6 anyway) */
 250       enum LLVM_CodeGenOpt_Level optlevel;
 251       char *error = NULL;
 252       int ret;
 253
 254       if (gallivm_debug & GALLIVM_DEBUG_NO_OPT) {
 255          optlevel = None;
 256       }
 257       else {
 258          optlevel = Default;
 259       }
 260
 261 #if USE_MCJIT
 262       ret = lp_build_create_mcjit_compiler_for_module(&gallivm->engine,
 263                                                       gallivm->module,
 264                                                       (unsigned) optlevel,
 265                                                       &error);
 266 #else
 267       ret = LLVMCreateJITCompiler(&gallivm->engine, gallivm->provider,
 268                                   (unsigned) optlevel, &error);
 269 #endif
 270       if (ret) {
 271          _debug_printf("%s\n", error);
 272          LLVMDisposeMessage(error);
 273          goto fail;
 274       }
 275
 276 #if defined(DEBUG) || defined(PROFILE)
 277       lp_register_oprofile_jit_event_listener(gallivm->engine);
 278 #endif
 279    }
 280
 281    LLVMAddModuleProvider(gallivm->engine, gallivm->provider);//new
 282
 283 #if !USE_MCJIT
 284    gallivm->target = LLVMGetExecutionEngineTargetData(gallivm->engine);
 285    if (!gallivm->target)
 286       goto fail;
 287 #else
 288    if (0) {
 289        /*
 290         * Dump the data layout strings.
 291         */
 292
 293        LLVMTargetDataRef target = LLVMGetExecutionEngineTargetData(gallivm->engine);
 294        char *data_layout;
 295        char *engine_data_layout;
 296
 297        data_layout = LLVMCopyStringRepOfTargetData(gallivm->target);
 298        engine_data_layout = LLVMCopyStringRepOfTargetData(target);
 299
 300        if (1) {
 301           debug_printf("module target data = %s\n", data_layout);
 302           debug_printf("engine target data = %s\n", engine_data_layout);
 303        }
 304
 305        free(data_layout);
 306        free(engine_data_layout);
 307    }
 308 #endif
 309
 310    return TRUE;
 311
 312 fail:
 313    return FALSE;
 314 }
 315
 316
 317 /**
 318  * Singleton
 319  *
 320  * We must never free LLVM contexts, because LLVM has several global caches
 321  * which pointing/derived from objects owned by the context, causing false
 322  * memory leaks and false cache hits when these objects are destroyed.
 323  *
 324  * TODO: For thread safety on multi-threaded OpenGL we should use one LLVM
 325  * context per thread, and put them in a pool when threads are destroyed.
 326  */
 327 static LLVMContextRef gallivm_context = NULL;
 328
 329
 330 /**
 331  * Allocate gallivm LLVM objects.
 332  * \return  TRUE for success, FALSE for failure
 333  */
 334 static boolean
 335 init_gallivm_state(struct gallivm_state *gallivm)
 336 {
 337    assert(!gallivm->context);
 338    assert(!gallivm->module);
 339    assert(!gallivm->provider);
 340
 341    lp_build_init();
 342
 343    if (!gallivm_context) {
 344       gallivm_context = LLVMContextCreate();
 345    }
 346    gallivm->context = gallivm_context;
 347    if (!gallivm->context)
 348       goto fail;
 349
 350    gallivm->module = LLVMModuleCreateWithNameInContext("gallivm",
 351                                                        gallivm->context);
 352    if (!gallivm->module)
 353       goto fail;
 354
 355    gallivm->provider =
 356       LLVMCreateModuleProviderForExistingModule(gallivm->module);
 357    if (!gallivm->provider)
 358       goto fail;
 359
 360    gallivm->builder = LLVMCreateBuilderInContext(gallivm->context);
 361    if (!gallivm->builder)
 362       goto fail;
 363
 364    /* FIXME: MC-JIT only allows compiling one module at a time, and it must be
 365     * complete when MC-JIT is created. So defer the MC-JIT engine creation for
 366     * now.
 367     */
 368 #if !USE_MCJIT
 369    if (!init_gallivm_engine(gallivm)) {
 370       goto fail;
 371    }
 372 #else
 373    /*
 374     * MC-JIT engine compiles the module immediately on creation, so we can't
 375     * obtain the target data from it.  Instead we create a target data layout
 376     * from a string.
 377     *
 378     * The produced layout strings are not precisely the same, but should make
 379     * no difference for the kind of optimization passes we run.
 380     *
 381     * For reference this is the layout string on x64:
 382     *
 383     *   e-p:64:64:64-S128-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f16:16:16-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-f128:128:128-n8:16:32:64
 384     *
 385     * See also:
 386     * - http://llvm.org/docs/LangRef.html#datalayout
 387     */
 388
 389    {
 390       const unsigned pointer_size = 8 * sizeof(void *);
 391       char layout[512];
 392       util_snprintf(layout, sizeof layout, "%c-p:%u:%u:%u-i64:64:64-a0:0:%u-s0:%u:%u",
 393 #ifdef PIPE_ARCH_LITTLE_ENDIAN
 394                     'e', // little endian
 395 #else
 396                     'E', // big endian
 397 #endif
 398                     pointer_size, pointer_size, pointer_size, // pointer size, abi alignment, preferred alignment
 399                     pointer_size, // aggregate preferred alignment
 400                     pointer_size, pointer_size); // stack objects abi alignment, preferred alignment
 401
 402       gallivm->target = LLVMCreateTargetData(layout);
 403       if (!gallivm->target) {
 404          return FALSE;
 405       }
 406    }
 407 #endif
 408
 409    if (!create_pass_manager(gallivm))
 410       goto fail;
 411
 412    return TRUE;
 413
 414 fail:
 415    free_gallivm_state(gallivm);
 416    return FALSE;
 417 }
 418
 419
 420 void
 421 lp_build_init(void)
 422 {
 423    if (gallivm_initialized)
 424       return;
 425
 426 #ifdef DEBUG
 427    gallivm_debug = debug_get_option_gallivm_debug();
 428 #endif
 429
 430    lp_set_target_options();
 431
 432 #if USE_MCJIT
 433    LLVMLinkInMCJIT();
 434 #else
 435    LLVMLinkInJIT();
 436 #endif
 437
 438    util_cpu_detect();
 439
 440    /* AMD Bulldozer AVX's throughput is the same as SSE2; and because using
 441     * 8-wide vector needs more floating ops than 4-wide (due to padding), it is
 442     * actually more efficient to use 4-wide vectors on this processor.
 443     *
 444     * See also:
 445     * - http://www.anandtech.com/show/4955/the-bulldozer-review-amd-fx8150-tested/2
 446     */
 447    if (HAVE_AVX &&
 448        util_cpu_caps.has_avx &&
 449        util_cpu_caps.has_intel) {
 450       lp_native_vector_width = 256;
 451    } else {
 452       /* Leave it at 128, even when no SIMD extensions are available.
 453        * Really needs to be a multiple of 128 so can fit 4 floats.
 454        */
 455       lp_native_vector_width = 128;
 456    }
 457
 458    lp_native_vector_width = debug_get_num_option("LP_NATIVE_VECTOR_WIDTH",
 459                                                  lp_native_vector_width);
 460
 461    gallivm_initialized = TRUE;
 462
 463 #if 0
 464    /* For simulating less capable machines */
 465    util_cpu_caps.has_sse3 = 0;
 466    util_cpu_caps.has_ssse3 = 0;
 467    util_cpu_caps.has_sse4_1 = 0;
 468 #endif
 469 }
 470
 471
 472
 473 /**
 474  * Create a new gallivm_state object.
 475  * Note that we return a singleton.
 476  */
 477 struct gallivm_state *
 478 gallivm_create(void)
 479 {
 480    struct gallivm_state *gallivm;
 481
 482 #if HAVE_LLVM <= 0x206
 483    if (GlobalGallivm) {
 484       return GlobalGallivm;
 485    }
 486 #endif
 487
 488    gallivm = CALLOC_STRUCT(gallivm_state);
 489    if (gallivm) {
 490       if (!init_gallivm_state(gallivm)) {
 491          FREE(gallivm);
 492          gallivm = NULL;
 493       }
 494    }
 495
 496 #if HAVE_LLVM <= 0x206
 497    GlobalGallivm = gallivm;
 498 #endif
 499
 500    return gallivm;
 501 }
 502
 503
 504 /**
 505  * Destroy a gallivm_state object.
 506  */
 507 void
 508 gallivm_destroy(struct gallivm_state *gallivm)
 509 {
 510 #if HAVE_LLVM <= 0x0206
 511    /* No-op: don't destroy the singleton */
 512    (void) gallivm;
 513 #else
 514    free_gallivm_state(gallivm);
 515    FREE(gallivm);
 516 #endif
 517 }
 518
 519
 520 /**
 521  * Validate and optimze a function.
 522  */
 523 static void
 524 gallivm_optimize_function(struct gallivm_state *gallivm,
 525                           LLVMValueRef func)
 526 {
 527    if (0) {
 528       debug_printf("optimizing %s...\n", LLVMGetValueName(func));
 529    }
 530
 531    assert(gallivm->passmgr);
 532
 533    /* Apply optimizations to LLVM IR */
 534    LLVMRunFunctionPassManager(gallivm->passmgr, func);
 535
 536    if (0) {
 537       if (gallivm_debug & GALLIVM_DEBUG_IR) {
 538          /* Print the LLVM IR to stderr */
 539          lp_debug_dump_value(func);
 540          debug_printf("\n");
 541       }
 542    }
 543 }
 544
 545
 546 /**
 547  * Validate a function.
 548  */
 549 void
 550 gallivm_verify_function(struct gallivm_state *gallivm,
 551                         LLVMValueRef func)
 552 {
 553    /* Verify the LLVM IR.  If invalid, dump and abort */
 554 #ifdef DEBUG
 555    if (LLVMVerifyFunction(func, LLVMPrintMessageAction)) {
 556       lp_debug_dump_value(func);
 557       assert(0);
 558       return;
 559    }
 560 #endif
 561
 562    gallivm_optimize_function(gallivm, func);
 563
 564    if (gallivm_debug & GALLIVM_DEBUG_IR) {
 565       /* Print the LLVM IR to stderr */
 566       lp_debug_dump_value(func);
 567       debug_printf("\n");
 568    }
 569 }
 570
 571
 572 void
 573 gallivm_compile_module(struct gallivm_state *gallivm)
 574 {
 575 #if HAVE_LLVM > 0x206
 576    assert(!gallivm->compiled);
 577 #endif
 578
 579    /* Dump byte code to a file */
 580    if (0) {
 581       LLVMWriteBitcodeToFile(gallivm->module, "llvmpipe.bc");
 582       debug_printf("llvmpipe.bc written\n");
 583       debug_printf("Invoke as \"llc -o - llvmpipe.bc\"\n");
 584    }
 585
 586 #if USE_MCJIT
 587    assert(!gallivm->engine);
 588    if (!init_gallivm_engine(gallivm)) {
 589       assert(0);
 590    }
 591 #endif
 592    assert(gallivm->engine);
 593
 594    ++gallivm->compiled;
 595 }
 596
 597
 598 func_pointer
 599 gallivm_jit_function(struct gallivm_state *gallivm,
 600                      LLVMValueRef func)
 601 {
 602    void *code;
 603    func_pointer jit_func;
 604
 605    assert(gallivm->compiled);
 606    assert(gallivm->engine);
 607
 608    code = LLVMGetPointerToGlobal(gallivm->engine, func);
 609    assert(code);
 610    jit_func = pointer_to_func(code);
 611
 612    if (gallivm_debug & GALLIVM_DEBUG_ASM) {
 613       lp_disassemble(code);
 614    }
 615
 616    /* Free the function body to save memory */
 617    lp_func_delete_body(func);
 618
 619    return jit_func;
 620 }
 621
 622
 623 /**
 624  * Free the function (and its machine code).
 625  */
 626 void
 627 gallivm_free_function(struct gallivm_state *gallivm,
 628                       LLVMValueRef func,
 629                       const void *code)
 630 {
 631 #if !USE_MCJIT
 632    if (code) {
 633       LLVMFreeMachineCodeForFunction(gallivm->engine, func);
 634    }
 635
 636    LLVMDeleteFunction(func);
 637 #endif
 638 }