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