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 HAVE_LLVM >= 0x0301
 262       ret = lp_build_create_jit_compiler_for_module(&gallivm->engine,
 263                                                     gallivm->module,
 264                                                     (unsigned) optlevel,
 265                                                     USE_MCJIT,
 266                                                     &error);
 267 #else
 268       ret = LLVMCreateJITCompiler(&gallivm->engine, gallivm->provider,
 269                                   (unsigned) optlevel, &error);
 270 #endif
 271       if (ret) {
 272          _debug_printf("%s\n", error);
 273          LLVMDisposeMessage(error);
 274          goto fail;
 275       }
 276
 277 #if defined(DEBUG) || defined(PROFILE)
 278       lp_register_oprofile_jit_event_listener(gallivm->engine);
 279 #endif
 280    }
 281
 282    LLVMAddModuleProvider(gallivm->engine, gallivm->provider);//new
 283
 284 #if !USE_MCJIT
 285    gallivm->target = LLVMGetExecutionEngineTargetData(gallivm->engine);
 286    if (!gallivm->target)
 287       goto fail;
 288 #else
 289    if (0) {
 290        /*
 291         * Dump the data layout strings.
 292         */
 293
 294        LLVMTargetDataRef target = LLVMGetExecutionEngineTargetData(gallivm->engine);
 295        char *data_layout;
 296        char *engine_data_layout;
 297
 298        data_layout = LLVMCopyStringRepOfTargetData(gallivm->target);
 299        engine_data_layout = LLVMCopyStringRepOfTargetData(target);
 300
 301        if (1) {
 302           debug_printf("module target data = %s\n", data_layout);
 303           debug_printf("engine target data = %s\n", engine_data_layout);
 304        }
 305
 306        free(data_layout);
 307        free(engine_data_layout);
 308    }
 309 #endif
 310
 311    return TRUE;
 312
 313 fail:
 314    return FALSE;
 315 }
 316
 317
 318 /**
 319  * Singleton
 320  *
 321  * We must never free LLVM contexts, because LLVM has several global caches
 322  * which pointing/derived from objects owned by the context, causing false
 323  * memory leaks and false cache hits when these objects are destroyed.
 324  *
 325  * TODO: For thread safety on multi-threaded OpenGL we should use one LLVM
 326  * context per thread, and put them in a pool when threads are destroyed.
 327  */
 328 static LLVMContextRef gallivm_context = NULL;
 329
 330
 331 /**
 332  * Allocate gallivm LLVM objects.
 333  * \return  TRUE for success, FALSE for failure
 334  */
 335 static boolean
 336 init_gallivm_state(struct gallivm_state *gallivm)
 337 {
 338    assert(!gallivm->context);
 339    assert(!gallivm->module);
 340    assert(!gallivm->provider);
 341
 342    lp_build_init();
 343
 344    if (!gallivm_context) {
 345       gallivm_context = LLVMContextCreate();
 346    }
 347    gallivm->context = gallivm_context;
 348    if (!gallivm->context)
 349       goto fail;
 350
 351    gallivm->module = LLVMModuleCreateWithNameInContext("gallivm",
 352                                                        gallivm->context);
 353    if (!gallivm->module)
 354       goto fail;
 355
 356    gallivm->provider =
 357       LLVMCreateModuleProviderForExistingModule(gallivm->module);
 358    if (!gallivm->provider)
 359       goto fail;
 360
 361    gallivm->builder = LLVMCreateBuilderInContext(gallivm->context);
 362    if (!gallivm->builder)
 363       goto fail;
 364
 365    /* FIXME: MC-JIT only allows compiling one module at a time, and it must be
 366     * complete when MC-JIT is created. So defer the MC-JIT engine creation for
 367     * now.
 368     */
 369 #if !USE_MCJIT
 370    if (!init_gallivm_engine(gallivm)) {
 371       goto fail;
 372    }
 373 #else
 374    /*
 375     * MC-JIT engine compiles the module immediately on creation, so we can't
 376     * obtain the target data from it.  Instead we create a target data layout
 377     * from a string.
 378     *
 379     * The produced layout strings are not precisely the same, but should make
 380     * no difference for the kind of optimization passes we run.
 381     *
 382     * For reference this is the layout string on x64:
 383     *
 384     *   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
 385     *
 386     * See also:
 387     * - http://llvm.org/docs/LangRef.html#datalayout
 388     */
 389
 390    {
 391       const unsigned pointer_size = 8 * sizeof(void *);
 392       char layout[512];
 393       util_snprintf(layout, sizeof layout, "%c-p:%u:%u:%u-i64:64:64-a0:0:%u-s0:%u:%u",
 394 #ifdef PIPE_ARCH_LITTLE_ENDIAN
 395                     'e', // little endian
 396 #else
 397                     'E', // big endian
 398 #endif
 399                     pointer_size, pointer_size, pointer_size, // pointer size, abi alignment, preferred alignment
 400                     pointer_size, // aggregate preferred alignment
 401                     pointer_size, pointer_size); // stack objects abi alignment, preferred alignment
 402
 403       gallivm->target = LLVMCreateTargetData(layout);
 404       if (!gallivm->target) {
 405          return FALSE;
 406       }
 407    }
 408 #endif
 409
 410    if (!create_pass_manager(gallivm))
 411       goto fail;
 412
 413    return TRUE;
 414
 415 fail:
 416    free_gallivm_state(gallivm);
 417    return FALSE;
 418 }
 419
 420
 421 void
 422 lp_build_init(void)
 423 {
 424    if (gallivm_initialized)
 425       return;
 426
 427 #ifdef DEBUG
 428    gallivm_debug = debug_get_option_gallivm_debug();
 429 #endif
 430
 431    lp_set_target_options();
 432
 433 #if USE_MCJIT
 434    LLVMLinkInMCJIT();
 435 #else
 436    LLVMLinkInJIT();
 437 #endif
 438
 439    util_cpu_detect();
 440
 441    /* AMD Bulldozer AVX's throughput is the same as SSE2; and because using
 442     * 8-wide vector needs more floating ops than 4-wide (due to padding), it is
 443     * actually more efficient to use 4-wide vectors on this processor.
 444     *
 445     * See also:
 446     * - http://www.anandtech.com/show/4955/the-bulldozer-review-amd-fx8150-tested/2
 447     */
 448    if (HAVE_AVX &&
 449        util_cpu_caps.has_avx &&
 450        util_cpu_caps.has_intel) {
 451       lp_native_vector_width = 256;
 452    } else {
 453       /* Leave it at 128, even when no SIMD extensions are available.
 454        * Really needs to be a multiple of 128 so can fit 4 floats.
 455        */
 456       lp_native_vector_width = 128;
 457    }
 458
 459    lp_native_vector_width = debug_get_num_option("LP_NATIVE_VECTOR_WIDTH",
 460                                                  lp_native_vector_width);
 461
 462    if (lp_native_vector_width <= 128) {
 463       /* Hide AVX support, as often LLVM AVX instrinsics are only guarded by
 464        * "util_cpu_caps.has_avx" predicate, and lack the
 465        * "lp_native_vector_width > 128" predicate. And also to ensure a more
 466        * consistent behavior, allowing one to test SSE2 on AVX machines.
 467        */
 468       util_cpu_caps.has_avx = 0;
 469    }
 470
 471 #ifdef PIPE_ARCH_PPC_64
 472    /* Set the NJ bit in VSCR to 0 so denormalized values are handled as
 473     * specified by IEEE standard (PowerISA 2.06 - Section 6.3). This garantees
 474     * that some rounding and half-float to float handling does not round
 475     * incorrectly to 0.
 476     */
 477    if (util_cpu_caps.has_altivec) {
 478       unsigned short mask[] = { 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF,
 479                                 0xFFFF, 0xFFFF, 0xFFFE, 0xFFFF };
 480       __asm (
 481         "mfvscr %%v1\n"
 482         "vand   %0,%%v1,%0\n"
 483         "mtvscr %0"
 484         :
 485         : "r" (*mask)
 486       );
 487    }
 488 #endif
 489
 490    gallivm_initialized = TRUE;
 491
 492 #if 0
 493    /* For simulating less capable machines */
 494    util_cpu_caps.has_sse3 = 0;
 495    util_cpu_caps.has_ssse3 = 0;
 496    util_cpu_caps.has_sse4_1 = 0;
 497    util_cpu_caps.has_avx = 0;
 498 #endif
 499 }
 500
 501
 502
 503 /**
 504  * Create a new gallivm_state object.
 505  * Note that we return a singleton.
 506  */
 507 struct gallivm_state *
 508 gallivm_create(void)
 509 {
 510    struct gallivm_state *gallivm;
 511
 512 #if HAVE_LLVM <= 0x206
 513    if (GlobalGallivm) {
 514       return GlobalGallivm;
 515    }
 516 #endif
 517
 518    gallivm = CALLOC_STRUCT(gallivm_state);
 519    if (gallivm) {
 520       if (!init_gallivm_state(gallivm)) {
 521          FREE(gallivm);
 522          gallivm = NULL;
 523       }
 524    }
 525
 526 #if HAVE_LLVM <= 0x206
 527    GlobalGallivm = gallivm;
 528 #endif
 529
 530    return gallivm;
 531 }
 532
 533
 534 /**
 535  * Destroy a gallivm_state object.
 536  */
 537 void
 538 gallivm_destroy(struct gallivm_state *gallivm)
 539 {
 540 #if HAVE_LLVM <= 0x0206
 541    /* No-op: don't destroy the singleton */
 542    (void) gallivm;
 543 #else
 544    free_gallivm_state(gallivm);
 545    FREE(gallivm);
 546 #endif
 547 }
 548
 549
 550 /**
 551  * Validate and optimze a function.
 552  */
 553 static void
 554 gallivm_optimize_function(struct gallivm_state *gallivm,
 555                           LLVMValueRef func)
 556 {
 557    if (0) {
 558       debug_printf("optimizing %s...\n", LLVMGetValueName(func));
 559    }
 560
 561    assert(gallivm->passmgr);
 562
 563    /* Apply optimizations to LLVM IR */
 564    LLVMRunFunctionPassManager(gallivm->passmgr, func);
 565
 566    if (0) {
 567       if (gallivm_debug & GALLIVM_DEBUG_IR) {
 568          /* Print the LLVM IR to stderr */
 569          lp_debug_dump_value(func);
 570          debug_printf("\n");
 571       }
 572    }
 573 }
 574
 575
 576 /**
 577  * Validate a function.
 578  */
 579 void
 580 gallivm_verify_function(struct gallivm_state *gallivm,
 581                         LLVMValueRef func)
 582 {
 583    /* Verify the LLVM IR.  If invalid, dump and abort */
 584 #ifdef DEBUG
 585    if (LLVMVerifyFunction(func, LLVMPrintMessageAction)) {
 586       lp_debug_dump_value(func);
 587       assert(0);
 588       return;
 589    }
 590 #endif
 591
 592    gallivm_optimize_function(gallivm, func);
 593
 594    if (gallivm_debug & GALLIVM_DEBUG_IR) {
 595       /* Print the LLVM IR to stderr */
 596       lp_debug_dump_value(func);
 597       debug_printf("\n");
 598    }
 599 }
 600
 601
 602 void
 603 gallivm_compile_module(struct gallivm_state *gallivm)
 604 {
 605 #if HAVE_LLVM > 0x206
 606    assert(!gallivm->compiled);
 607 #endif
 608
 609    /* Dump byte code to a file */
 610    if (0) {
 611       LLVMWriteBitcodeToFile(gallivm->module, "llvmpipe.bc");
 612       debug_printf("llvmpipe.bc written\n");
 613       debug_printf("Invoke as \"llc -o - llvmpipe.bc\"\n");
 614    }
 615
 616 #if USE_MCJIT
 617    assert(!gallivm->engine);
 618    if (!init_gallivm_engine(gallivm)) {
 619       assert(0);
 620    }
 621 #endif
 622    assert(gallivm->engine);
 623
 624    ++gallivm->compiled;
 625 }
 626
 627
 628 func_pointer
 629 gallivm_jit_function(struct gallivm_state *gallivm,
 630                      LLVMValueRef func)
 631 {
 632    void *code;
 633    func_pointer jit_func;
 634
 635    assert(gallivm->compiled);
 636    assert(gallivm->engine);
 637
 638    code = LLVMGetPointerToGlobal(gallivm->engine, func);
 639    assert(code);
 640    jit_func = pointer_to_func(code);
 641
 642    if (gallivm_debug & GALLIVM_DEBUG_ASM) {
 643       lp_disassemble(code);
 644    }
 645
 646    /* Free the function body to save memory */
 647    lp_func_delete_body(func);
 648
 649    return jit_func;
 650 }
 651
 652
 653 /**
 654  * Free the function (and its machine code).
 655  */
 656 void
 657 gallivm_free_function(struct gallivm_state *gallivm,
 658                       LLVMValueRef func,
 659                       const void *code)
 660 {
 661 #if !USE_MCJIT
 662    if (code) {
 663       LLVMFreeMachineCodeForFunction(gallivm->engine, func);
 664    }
 665
 666    LLVMDeleteFunction(func);
 667 #endif
 668 }