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
 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    if (lp_native_vector_width <= 128) {
 459       /* Hide AVX support, as often LLVM AVX instrinsics are only guarded by
 460        * "util_cpu_caps.has_avx" predicate, and lack the
 461        * "lp_native_vector_width > 128" predicate. And also to ensure a more
 462        * consistent behavior, allowing one to test SSE2 on AVX machines.
 463        */
 464       util_cpu_caps.has_avx = 0;
 465    }
 466
 467    if (!HAVE_AVX) {
 468       /*
 469        * note these instructions are VEX-only, so can only emit if we use
 470        * avx (don't want to base it on has_avx & has_f16c later as that would
 471        * omit it unnecessarily on amd cpus, see above).
 472        */
 473       util_cpu_caps.has_f16c = 0;
 474    }
 475
 476 #ifdef PIPE_ARCH_PPC_64
 477    /* Set the NJ bit in VSCR to 0 so denormalized values are handled as
 478     * specified by IEEE standard (PowerISA 2.06 - Section 6.3). This garantees
 479     * that some rounding and half-float to float handling does not round
 480     * incorrectly to 0.
 481     */
 482    if (util_cpu_caps.has_altivec) {
 483       unsigned short mask[] = { 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF,
 484                                 0xFFFF, 0xFFFF, 0xFFFE, 0xFFFF };
 485       __asm (
 486         "mfvscr %%v1\n"
 487         "vand   %0,%%v1,%0\n"
 488         "mtvscr %0"
 489         :
 490         : "r" (*mask)
 491       );
 492    }
 493 #endif
 494
 495    gallivm_initialized = TRUE;
 496
 497 #if 0
 498    /* For simulating less capable machines */
 499    util_cpu_caps.has_sse3 = 0;
 500    util_cpu_caps.has_ssse3 = 0;
 501    util_cpu_caps.has_sse4_1 = 0;
 502    util_cpu_caps.has_avx = 0;
 503    util_cpu_caps.has_f16c = 0;
 504 #endif
 505 }
 506
 507
 508
 509 /**
 510  * Create a new gallivm_state object.
 511  * Note that we return a singleton.
 512  */
 513 struct gallivm_state *
 514 gallivm_create(void)
 515 {
 516    struct gallivm_state *gallivm;
 517
 518 #if HAVE_LLVM <= 0x206
 519    if (GlobalGallivm) {
 520       return GlobalGallivm;
 521    }
 522 #endif
 523
 524    gallivm = CALLOC_STRUCT(gallivm_state);
 525    if (gallivm) {
 526       if (!init_gallivm_state(gallivm)) {
 527          FREE(gallivm);
 528          gallivm = NULL;
 529       }
 530    }
 531
 532 #if HAVE_LLVM <= 0x206
 533    GlobalGallivm = gallivm;
 534 #endif
 535
 536    return gallivm;
 537 }
 538
 539
 540 /**
 541  * Destroy a gallivm_state object.
 542  */
 543 void
 544 gallivm_destroy(struct gallivm_state *gallivm)
 545 {
 546 #if HAVE_LLVM <= 0x0206
 547    /* No-op: don't destroy the singleton */
 548    (void) gallivm;
 549 #else
 550    free_gallivm_state(gallivm);
 551    FREE(gallivm);
 552 #endif
 553 }
 554
 555
 556 /**
 557  * Validate and optimze a function.
 558  */
 559 static void
 560 gallivm_optimize_function(struct gallivm_state *gallivm,
 561                           LLVMValueRef func)
 562 {
 563    if (0) {
 564       debug_printf("optimizing %s...\n", LLVMGetValueName(func));
 565    }
 566
 567    assert(gallivm->passmgr);
 568
 569    /* Apply optimizations to LLVM IR */
 570    LLVMRunFunctionPassManager(gallivm->passmgr, func);
 571
 572    if (0) {
 573       if (gallivm_debug & GALLIVM_DEBUG_IR) {
 574          /* Print the LLVM IR to stderr */
 575          lp_debug_dump_value(func);
 576          debug_printf("\n");
 577       }
 578    }
 579 }
 580
 581
 582 /**
 583  * Validate a function.
 584  */
 585 void
 586 gallivm_verify_function(struct gallivm_state *gallivm,
 587                         LLVMValueRef func)
 588 {
 589    /* Verify the LLVM IR.  If invalid, dump and abort */
 590 #ifdef DEBUG
 591    if (LLVMVerifyFunction(func, LLVMPrintMessageAction)) {
 592       lp_debug_dump_value(func);
 593       assert(0);
 594       return;
 595    }
 596 #endif
 597
 598    gallivm_optimize_function(gallivm, func);
 599
 600    if (gallivm_debug & GALLIVM_DEBUG_IR) {
 601       /* Print the LLVM IR to stderr */
 602       lp_debug_dump_value(func);
 603       debug_printf("\n");
 604    }
 605 }
 606
 607
 608 void
 609 gallivm_compile_module(struct gallivm_state *gallivm)
 610 {
 611 #if HAVE_LLVM > 0x206
 612    assert(!gallivm->compiled);
 613 #endif
 614
 615    /* Dump byte code to a file */
 616    if (0) {
 617       LLVMWriteBitcodeToFile(gallivm->module, "llvmpipe.bc");
 618       debug_printf("llvmpipe.bc written\n");
 619       debug_printf("Invoke as \"llc -o - llvmpipe.bc\"\n");
 620    }
 621
 622 #if USE_MCJIT
 623    assert(!gallivm->engine);
 624    if (!init_gallivm_engine(gallivm)) {
 625       assert(0);
 626    }
 627 #endif
 628    assert(gallivm->engine);
 629
 630    ++gallivm->compiled;
 631 }
 632
 633
 634
 635 func_pointer
 636 gallivm_jit_function(struct gallivm_state *gallivm,
 637                      LLVMValueRef func)
 638 {
 639    void *code;
 640    func_pointer jit_func;
 641
 642    assert(gallivm->compiled);
 643    assert(gallivm->engine);
 644
 645    code = LLVMGetPointerToGlobal(gallivm->engine, func);
 646    assert(code);
 647    jit_func = pointer_to_func(code);
 648
 649    if (gallivm_debug & GALLIVM_DEBUG_ASM) {
 650       lp_disassemble(func, code);
 651    }
 652
 653 #if defined(PROFILE)
 654    lp_profile(func, code);
 655 #endif
 656
 657    /* Free the function body to save memory */
 658    lp_func_delete_body(func);
 659
 660    return jit_func;
 661 }
 662
 663
 664 /**
 665  * Free the function (and its machine code).
 666  */
 667 void
 668 gallivm_free_function(struct gallivm_state *gallivm,
 669                       LLVMValueRef func,
 670                       const void *code)
 671 {
 672 #if !USE_MCJIT
 673    if (code) {
 674       LLVMFreeMachineCodeForFunction(gallivm->engine, func);
 675    }
 676
 677    LLVMDeleteFunction(func);
 678 #endif
 679 }