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/simple_list.h"
  35 #include "util/os_time.h"
  36 #include "lp_bld.h"
  37 #include "lp_bld_debug.h"
  38 #include "lp_bld_misc.h"
  39 #include "lp_bld_init.h"
  40
  41 #include <llvm-c/Analysis.h>
  42 #include <llvm-c/Transforms/Scalar.h>
  43 #if HAVE_LLVM >= 0x0700
  44 #include <llvm-c/Transforms/Utils.h>
  45 #endif
  46 #include <llvm-c/BitWriter.h>
  47
  48
  49 /* Only MCJIT is available as of LLVM SVN r216982 */
  50 #if HAVE_LLVM >= 0x0306
  51 #  define USE_MCJIT 1
  52 #elif defined(PIPE_ARCH_PPC_64) || defined(PIPE_ARCH_S390) || defined(PIPE_ARCH_ARM) || defined(PIPE_ARCH_AARCH64)
  53 #  define USE_MCJIT 1
  54 #endif
  55
  56 #if defined(USE_MCJIT)
  57 static const bool use_mcjit = USE_MCJIT;
  58 #else
  59 static bool use_mcjit = FALSE;
  60 #endif
  61
  62 unsigned gallivm_perf = 0;
  63
  64 static const struct debug_named_value lp_bld_perf_flags[] = {
  65    { "no_brilinear", GALLIVM_PERF_NO_BRILINEAR, "disable brilinear optimization" },
  66    { "no_rho_approx", GALLIVM_PERF_NO_RHO_APPROX, "disable rho_approx optimization" },
  67    { "no_quad_lod", GALLIVM_PERF_NO_QUAD_LOD, "disable quad_lod optimization" },
  68    { "nopt",   GALLIVM_PERF_NO_OPT, "disable optimization passes to speed up shader compilation" },
  69    { "no_filter_hacks", GALLIVM_PERF_NO_BRILINEAR | GALLIVM_PERF_NO_RHO_APPROX |
  70      GALLIVM_PERF_NO_QUAD_LOD, "disable filter optimization hacks" },
  71    DEBUG_NAMED_VALUE_END
  72 };
  73
  74 #ifdef DEBUG
  75 unsigned gallivm_debug = 0;
  76
  77 static const struct debug_named_value lp_bld_debug_flags[] = {
  78    { "tgsi",   GALLIVM_DEBUG_TGSI, NULL },
  79    { "ir",     GALLIVM_DEBUG_IR, NULL },
  80    { "asm",    GALLIVM_DEBUG_ASM, NULL },
  81    { "perf",   GALLIVM_DEBUG_PERF, NULL },
  82    { "gc",     GALLIVM_DEBUG_GC, NULL },
  83    { "dumpbc", GALLIVM_DEBUG_DUMP_BC, NULL },
  84    DEBUG_NAMED_VALUE_END
  85 };
  86
  87 DEBUG_GET_ONCE_FLAGS_OPTION(gallivm_debug, "GALLIVM_DEBUG", lp_bld_debug_flags, 0)
  88 #endif
  89
  90
  91 static boolean gallivm_initialized = FALSE;
  92
  93 unsigned lp_native_vector_width;
  94
  95
  96 /*
  97  * Optimization values are:
  98  * - 0: None (-O0)
  99  * - 1: Less (-O1)
 100  * - 2: Default (-O2, -Os)
 101  * - 3: Aggressive (-O3)
 102  *
 103  * See also CodeGenOpt::Level in llvm/Target/TargetMachine.h
 104  */
 105 enum LLVM_CodeGenOpt_Level {
 106    None,        // -O0
 107    Less,        // -O1
 108    Default,     // -O2, -Os
 109    Aggressive   // -O3
 110 };
 111
 112
 113 /**
 114  * Create the LLVM (optimization) pass manager and install
 115  * relevant optimization passes.
 116  * \return  TRUE for success, FALSE for failure
 117  */
 118 static boolean
 119 create_pass_manager(struct gallivm_state *gallivm)
 120 {
 121    assert(!gallivm->passmgr);
 122    assert(gallivm->target);
 123
 124    gallivm->passmgr = LLVMCreateFunctionPassManagerForModule(gallivm->module);
 125    if (!gallivm->passmgr)
 126       return FALSE;
 127    /*
 128     * TODO: some per module pass manager with IPO passes might be helpful -
 129     * the generated texture functions may benefit from inlining if they are
 130     * simple, or constant propagation into them, etc.
 131     */
 132
 133 #if HAVE_LLVM < 0x0309
 134    // Old versions of LLVM get the DataLayout from the pass manager.
 135    LLVMAddTargetData(gallivm->target, gallivm->passmgr);
 136 #endif
 137
 138    {
 139       char *td_str;
 140       // New ones from the Module.
 141       td_str = LLVMCopyStringRepOfTargetData(gallivm->target);
 142       LLVMSetDataLayout(gallivm->module, td_str);
 143       free(td_str);
 144    }
 145
 146    if ((gallivm_perf & GALLIVM_PERF_NO_OPT) == 0) {
 147       /*
 148        * TODO: Evaluate passes some more - keeping in mind
 149        * both quality of generated code and compile times.
 150        */
 151       /*
 152        * NOTE: if you change this, don't forget to change the output
 153        * with GALLIVM_DEBUG_DUMP_BC in gallivm_compile_module.
 154        */
 155       LLVMAddScalarReplAggregatesPass(gallivm->passmgr);
 156       LLVMAddEarlyCSEPass(gallivm->passmgr);
 157       LLVMAddCFGSimplificationPass(gallivm->passmgr);
 158       /*
 159        * FIXME: LICM is potentially quite useful. However, for some
 160        * rather crazy shaders the compile time can reach _hours_ per shader,
 161        * due to licm implying lcssa (since llvm 3.5), which can take forever.
 162        * Even for sane shaders, the cost of licm is rather high (and not just
 163        * due to lcssa, licm itself too), though mostly only in cases when it
 164        * can actually move things, so having to disable it is a pity.
 165        * LLVMAddLICMPass(gallivm->passmgr);
 166        */
 167       LLVMAddReassociatePass(gallivm->passmgr);
 168       LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
 169       LLVMAddConstantPropagationPass(gallivm->passmgr);
 170       LLVMAddInstructionCombiningPass(gallivm->passmgr);
 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 any generated code
 186  * nor the gallivm object itself.
 187  */
 188 void
 189 gallivm_free_ir(struct gallivm_state *gallivm)
 190 {
 191    if (gallivm->passmgr) {
 192       LLVMDisposePassManager(gallivm->passmgr);
 193    }
 194
 195    if (gallivm->engine) {
 196       /* This will already destroy any associated module */
 197       LLVMDisposeExecutionEngine(gallivm->engine);
 198    } else if (gallivm->module) {
 199       LLVMDisposeModule(gallivm->module);
 200    }
 201
 202    FREE(gallivm->module_name);
 203
 204    if (!use_mcjit) {
 205       /* Don't free the TargetData, it's owned by the exec engine */
 206    } else {
 207       if (gallivm->target) {
 208          LLVMDisposeTargetData(gallivm->target);
 209       }
 210    }
 211
 212    if (gallivm->builder)
 213       LLVMDisposeBuilder(gallivm->builder);
 214
 215    /* The LLVMContext should be owned by the parent of gallivm. */
 216
 217    gallivm->engine = NULL;
 218    gallivm->target = NULL;
 219    gallivm->module = NULL;
 220    gallivm->module_name = NULL;
 221    gallivm->passmgr = NULL;
 222    gallivm->context = NULL;
 223    gallivm->builder = NULL;
 224 }
 225
 226
 227 /**
 228  * Free LLVM-generated code.  Should be done AFTER gallivm_free_ir().
 229  */
 230 static void
 231 gallivm_free_code(struct gallivm_state *gallivm)
 232 {
 233    assert(!gallivm->module);
 234    assert(!gallivm->engine);
 235    lp_free_generated_code(gallivm->code);
 236    gallivm->code = NULL;
 237    lp_free_memory_manager(gallivm->memorymgr);
 238    gallivm->memorymgr = NULL;
 239 }
 240
 241
 242 static boolean
 243 init_gallivm_engine(struct gallivm_state *gallivm)
 244 {
 245    if (1) {
 246       enum LLVM_CodeGenOpt_Level optlevel;
 247       char *error = NULL;
 248       int ret;
 249
 250       if (gallivm_perf & GALLIVM_PERF_NO_OPT) {
 251          optlevel = None;
 252       }
 253       else {
 254          optlevel = Default;
 255       }
 256
 257       ret = lp_build_create_jit_compiler_for_module(&gallivm->engine,
 258                                                     &gallivm->code,
 259                                                     gallivm->module,
 260                                                     gallivm->memorymgr,
 261                                                     (unsigned) optlevel,
 262                                                     use_mcjit,
 263                                                     &error);
 264       if (ret) {
 265          _debug_printf("%s\n", error);
 266          LLVMDisposeMessage(error);
 267          goto fail;
 268       }
 269    }
 270
 271    if (!use_mcjit) {
 272       gallivm->target = LLVMGetExecutionEngineTargetData(gallivm->engine);
 273       if (!gallivm->target)
 274          goto fail;
 275    } else {
 276       if (0) {
 277           /*
 278            * Dump the data layout strings.
 279            */
 280
 281           LLVMTargetDataRef target = LLVMGetExecutionEngineTargetData(gallivm->engine);
 282           char *data_layout;
 283           char *engine_data_layout;
 284
 285           data_layout = LLVMCopyStringRepOfTargetData(gallivm->target);
 286           engine_data_layout = LLVMCopyStringRepOfTargetData(target);
 287
 288           if (1) {
 289              debug_printf("module target data = %s\n", data_layout);
 290              debug_printf("engine target data = %s\n", engine_data_layout);
 291           }
 292
 293           free(data_layout);
 294           free(engine_data_layout);
 295       }
 296    }
 297
 298    return TRUE;
 299
 300 fail:
 301    return FALSE;
 302 }
 303
 304
 305 /**
 306  * Allocate gallivm LLVM objects.
 307  * \return  TRUE for success, FALSE for failure
 308  */
 309 static boolean
 310 init_gallivm_state(struct gallivm_state *gallivm, const char *name,
 311                    LLVMContextRef context)
 312 {
 313    assert(!gallivm->context);
 314    assert(!gallivm->module);
 315
 316    if (!lp_build_init())
 317       return FALSE;
 318
 319    gallivm->context = context;
 320
 321    if (!gallivm->context)
 322       goto fail;
 323
 324    gallivm->module_name = NULL;
 325    if (name) {
 326       size_t size = strlen(name) + 1;
 327       gallivm->module_name = MALLOC(size);
 328       if (gallivm->module_name) {
 329          memcpy(gallivm->module_name, name, size);
 330       }
 331    }
 332
 333    gallivm->module = LLVMModuleCreateWithNameInContext(name,
 334                                                        gallivm->context);
 335    if (!gallivm->module)
 336       goto fail;
 337
 338    gallivm->builder = LLVMCreateBuilderInContext(gallivm->context);
 339    if (!gallivm->builder)
 340       goto fail;
 341
 342    gallivm->memorymgr = lp_get_default_memory_manager();
 343    if (!gallivm->memorymgr)
 344       goto fail;
 345
 346    /* FIXME: MC-JIT only allows compiling one module at a time, and it must be
 347     * complete when MC-JIT is created. So defer the MC-JIT engine creation for
 348     * now.
 349     */
 350    if (!use_mcjit) {
 351       if (!init_gallivm_engine(gallivm)) {
 352          goto fail;
 353       }
 354    } else {
 355       /*
 356        * MC-JIT engine compiles the module immediately on creation, so we can't
 357        * obtain the target data from it.  Instead we create a target data layout
 358        * from a string.
 359        *
 360        * The produced layout strings are not precisely the same, but should make
 361        * no difference for the kind of optimization passes we run.
 362        *
 363        * For reference this is the layout string on x64:
 364        *
 365        *   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
 366        *
 367        * See also:
 368        * - http://llvm.org/docs/LangRef.html#datalayout
 369        */
 370
 371       {
 372          const unsigned pointer_size = 8 * sizeof(void *);
 373          char layout[512];
 374          util_snprintf(layout, sizeof layout, "%c-p:%u:%u:%u-i64:64:64-a0:0:%u-s0:%u:%u",
 375 #ifdef PIPE_ARCH_LITTLE_ENDIAN
 376                        'e', // little endian
 377 #else
 378                        'E', // big endian
 379 #endif
 380                        pointer_size, pointer_size, pointer_size, // pointer size, abi alignment, preferred alignment
 381                        pointer_size, // aggregate preferred alignment
 382                        pointer_size, pointer_size); // stack objects abi alignment, preferred alignment
 383
 384          gallivm->target = LLVMCreateTargetData(layout);
 385          if (!gallivm->target) {
 386             return FALSE;
 387          }
 388       }
 389    }
 390
 391    if (!create_pass_manager(gallivm))
 392       goto fail;
 393
 394    return TRUE;
 395
 396 fail:
 397    gallivm_free_ir(gallivm);
 398    gallivm_free_code(gallivm);
 399    return FALSE;
 400 }
 401
 402
 403 boolean
 404 lp_build_init(void)
 405 {
 406    if (gallivm_initialized)
 407       return TRUE;
 408
 409
 410    /* LLVMLinkIn* are no-ops at runtime.  They just ensure the respective
 411     * component is linked at buildtime, which is sufficient for its static
 412     * constructors to be called at load time.
 413     */
 414 #if defined(USE_MCJIT)
 415 #  if USE_MCJIT
 416       LLVMLinkInMCJIT();
 417 #  else
 418       LLVMLinkInJIT();
 419 #  endif
 420 #else
 421    use_mcjit = debug_get_bool_option("GALLIVM_MCJIT", FALSE);
 422    LLVMLinkInJIT();
 423    LLVMLinkInMCJIT();
 424 #endif
 425
 426 #ifdef DEBUG
 427    gallivm_debug = debug_get_option_gallivm_debug();
 428 #endif
 429
 430    gallivm_perf = debug_get_flags_option("GALLIVM_PERF", lp_bld_perf_flags, 0 );
 431
 432    lp_set_target_options();
 433
 434    util_cpu_detect();
 435
 436    /* For simulating less capable machines */
 437 #ifdef DEBUG
 438    if (debug_get_bool_option("LP_FORCE_SSE2", FALSE)) {
 439       assert(util_cpu_caps.has_sse2);
 440       util_cpu_caps.has_sse3 = 0;
 441       util_cpu_caps.has_ssse3 = 0;
 442       util_cpu_caps.has_sse4_1 = 0;
 443       util_cpu_caps.has_sse4_2 = 0;
 444       util_cpu_caps.has_avx = 0;
 445       util_cpu_caps.has_avx2 = 0;
 446       util_cpu_caps.has_f16c = 0;
 447       util_cpu_caps.has_fma = 0;
 448    }
 449 #endif
 450
 451    /* AMD Bulldozer AVX's throughput is the same as SSE2; and because using
 452     * 8-wide vector needs more floating ops than 4-wide (due to padding), it is
 453     * actually more efficient to use 4-wide vectors on this processor.
 454     *
 455     * See also:
 456     * - http://www.anandtech.com/show/4955/the-bulldozer-review-amd-fx8150-tested/2
 457     */
 458    if (util_cpu_caps.has_avx &&
 459        util_cpu_caps.has_intel) {
 460       lp_native_vector_width = 256;
 461    } else {
 462       /* Leave it at 128, even when no SIMD extensions are available.
 463        * Really needs to be a multiple of 128 so can fit 4 floats.
 464        */
 465       lp_native_vector_width = 128;
 466    }
 467
 468    lp_native_vector_width = debug_get_num_option("LP_NATIVE_VECTOR_WIDTH",
 469                                                  lp_native_vector_width);
 470
 471    if (lp_native_vector_width <= 128) {
 472       /* Hide AVX support, as often LLVM AVX intrinsics are only guarded by
 473        * "util_cpu_caps.has_avx" predicate, and lack the
 474        * "lp_native_vector_width > 128" predicate. And also to ensure a more
 475        * consistent behavior, allowing one to test SSE2 on AVX machines.
 476        * XXX: should not play games with util_cpu_caps directly as it might
 477        * get used for other things outside llvm too.
 478        */
 479       util_cpu_caps.has_avx = 0;
 480       util_cpu_caps.has_avx2 = 0;
 481       util_cpu_caps.has_f16c = 0;
 482       util_cpu_caps.has_fma = 0;
 483    }
 484    if (HAVE_LLVM < 0x0304 || !use_mcjit) {
 485       /* AVX2 support has only been tested with LLVM 3.4, and it requires
 486        * MCJIT. */
 487       util_cpu_caps.has_avx2 = 0;
 488    }
 489
 490 #ifdef PIPE_ARCH_PPC_64
 491    /* Set the NJ bit in VSCR to 0 so denormalized values are handled as
 492     * specified by IEEE standard (PowerISA 2.06 - Section 6.3). This guarantees
 493     * that some rounding and half-float to float handling does not round
 494     * incorrectly to 0.
 495     * XXX: should eventually follow same logic on all platforms.
 496     * Right now denorms get explicitly disabled (but elsewhere) for x86,
 497     * whereas ppc64 explicitly enables them...
 498     */
 499    if (util_cpu_caps.has_altivec) {
 500       unsigned short mask[] = { 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF,
 501                                 0xFFFF, 0xFFFF, 0xFFFE, 0xFFFF };
 502       __asm (
 503         "mfvscr %%v1\n"
 504         "vand   %0,%%v1,%0\n"
 505         "mtvscr %0"
 506         :
 507         : "r" (*mask)
 508       );
 509    }
 510 #endif
 511
 512    gallivm_initialized = TRUE;
 513
 514    return TRUE;
 515 }
 516
 517
 518
 519 /**
 520  * Create a new gallivm_state object.
 521  */
 522 struct gallivm_state *
 523 gallivm_create(const char *name, LLVMContextRef context)
 524 {
 525    struct gallivm_state *gallivm;
 526
 527    gallivm = CALLOC_STRUCT(gallivm_state);
 528    if (gallivm) {
 529       if (!init_gallivm_state(gallivm, name, context)) {
 530          FREE(gallivm);
 531          gallivm = NULL;
 532       }
 533    }
 534
 535    return gallivm;
 536 }
 537
 538
 539 /**
 540  * Destroy a gallivm_state object.
 541  */
 542 void
 543 gallivm_destroy(struct gallivm_state *gallivm)
 544 {
 545    gallivm_free_ir(gallivm);
 546    gallivm_free_code(gallivm);
 547    FREE(gallivm);
 548 }
 549
 550
 551 /**
 552  * Validate a function.
 553  * Verification is only done with debug builds.
 554  */
 555 void
 556 gallivm_verify_function(struct gallivm_state *gallivm,
 557                         LLVMValueRef func)
 558 {
 559    /* Verify the LLVM IR.  If invalid, dump and abort */
 560 #ifdef DEBUG
 561    if (LLVMVerifyFunction(func, LLVMPrintMessageAction)) {
 562       lp_debug_dump_value(func);
 563       assert(0);
 564       return;
 565    }
 566 #endif
 567
 568    if (gallivm_debug & GALLIVM_DEBUG_IR) {
 569       /* Print the LLVM IR to stderr */
 570       lp_debug_dump_value(func);
 571       debug_printf("\n");
 572    }
 573 }
 574
 575
 576 /**
 577  * Compile a module.
 578  * This does IR optimization on all functions in the module.
 579  */
 580 void
 581 gallivm_compile_module(struct gallivm_state *gallivm)
 582 {
 583    LLVMValueRef func;
 584    int64_t time_begin = 0;
 585
 586    assert(!gallivm->compiled);
 587
 588    if (gallivm->builder) {
 589       LLVMDisposeBuilder(gallivm->builder);
 590       gallivm->builder = NULL;
 591    }
 592
 593    /* Dump bitcode to a file */
 594    if (gallivm_debug & GALLIVM_DEBUG_DUMP_BC) {
 595       char filename[256];
 596       assert(gallivm->module_name);
 597       util_snprintf(filename, sizeof(filename), "ir_%s.bc", gallivm->module_name);
 598       LLVMWriteBitcodeToFile(gallivm->module, filename);
 599       debug_printf("%s written\n", filename);
 600       debug_printf("Invoke as \"opt %s %s | llc -O%d %s%s\"\n",
 601                    gallivm_debug & GALLIVM_PERF_NO_OPT ? "-mem2reg" :
 602                    "-sroa -early-cse -simplifycfg -reassociate "
 603                    "-mem2reg -constprop -instcombine -gvn",
 604                    filename, gallivm_debug & GALLIVM_PERF_NO_OPT ? 0 : 2,
 605                    (HAVE_LLVM >= 0x0305) ? "[-mcpu=<-mcpu option>] " : "",
 606                    "[-mattr=<-mattr option(s)>]");
 607    }
 608
 609    if (gallivm_debug & GALLIVM_DEBUG_PERF)
 610       time_begin = os_time_get();
 611
 612    /* Run optimization passes */
 613    LLVMInitializeFunctionPassManager(gallivm->passmgr);
 614    func = LLVMGetFirstFunction(gallivm->module);
 615    while (func) {
 616       if (0) {
 617          debug_printf("optimizing func %s...\n", LLVMGetValueName(func));
 618       }
 619
 620    /* Disable frame pointer omission on debug/profile builds */
 621    /* XXX: And workaround http://llvm.org/PR21435 */
 622 #if HAVE_LLVM >= 0x0307 && \
 623     (defined(DEBUG) || defined(PROFILE) || \
 624      defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64))
 625       LLVMAddTargetDependentFunctionAttr(func, "no-frame-pointer-elim", "true");
 626       LLVMAddTargetDependentFunctionAttr(func, "no-frame-pointer-elim-non-leaf", "true");
 627 #endif
 628
 629       LLVMRunFunctionPassManager(gallivm->passmgr, func);
 630       func = LLVMGetNextFunction(func);
 631    }
 632    LLVMFinalizeFunctionPassManager(gallivm->passmgr);
 633
 634    if (gallivm_debug & GALLIVM_DEBUG_PERF) {
 635       int64_t time_end = os_time_get();
 636       int time_msec = (int)((time_end - time_begin) / 1000);
 637       assert(gallivm->module_name);
 638       debug_printf("optimizing module %s took %d msec\n",
 639                    gallivm->module_name, time_msec);
 640    }
 641
 642    if (use_mcjit) {
 643       /* Setting the module's DataLayout to an empty string will cause the
 644        * ExecutionEngine to copy to the DataLayout string from its target
 645        * machine to the module.  As of LLVM 3.8 the module and the execution
 646        * engine are required to have the same DataLayout.
 647        *
 648        * We must make sure we do this after running the optimization passes,
 649        * because those passes need a correct datalayout string.  For example,
 650        * if those optimization passes see an empty datalayout, they will assume
 651        * this is a little endian target and will do optimizations that break big
 652        * endian machines.
 653        *
 654        * TODO: This is just a temporary work-around.  The correct solution is
 655        * for gallivm_init_state() to create a TargetMachine and pull the
 656        * DataLayout from there.  Currently, the TargetMachine used by llvmpipe
 657        * is being implicitly created by the EngineBuilder in
 658        * lp_build_create_jit_compiler_for_module()
 659        */
 660       LLVMSetDataLayout(gallivm->module, "");
 661       assert(!gallivm->engine);
 662       if (!init_gallivm_engine(gallivm)) {
 663          assert(0);
 664       }
 665    }
 666    assert(gallivm->engine);
 667
 668    ++gallivm->compiled;
 669
 670    if (gallivm_debug & GALLIVM_DEBUG_ASM) {
 671       LLVMValueRef llvm_func = LLVMGetFirstFunction(gallivm->module);
 672
 673       while (llvm_func) {
 674          /*
 675           * Need to filter out functions which don't have an implementation,
 676           * such as the intrinsics. May not be sufficient in case of IPO?
 677           * LLVMGetPointerToGlobal() will abort otherwise.
 678           */
 679          if (!LLVMIsDeclaration(llvm_func)) {
 680             void *func_code = LLVMGetPointerToGlobal(gallivm->engine, llvm_func);
 681             lp_disassemble(llvm_func, func_code);
 682          }
 683          llvm_func = LLVMGetNextFunction(llvm_func);
 684       }
 685    }
 686
 687 #if defined(PROFILE)
 688    {
 689       LLVMValueRef llvm_func = LLVMGetFirstFunction(gallivm->module);
 690
 691       while (llvm_func) {
 692          if (!LLVMIsDeclaration(llvm_func)) {
 693             void *func_code = LLVMGetPointerToGlobal(gallivm->engine, llvm_func);
 694             lp_profile(llvm_func, func_code);
 695          }
 696          llvm_func = LLVMGetNextFunction(llvm_func);
 697       }
 698    }
 699 #endif
 700 }
 701
 702
 703
 704 func_pointer
 705 gallivm_jit_function(struct gallivm_state *gallivm,
 706                      LLVMValueRef func)
 707 {
 708    void *code;
 709    func_pointer jit_func;
 710    int64_t time_begin = 0;
 711
 712    assert(gallivm->compiled);
 713    assert(gallivm->engine);
 714
 715    if (gallivm_debug & GALLIVM_DEBUG_PERF)
 716       time_begin = os_time_get();
 717
 718    code = LLVMGetPointerToGlobal(gallivm->engine, func);
 719    assert(code);
 720    jit_func = pointer_to_func(code);
 721
 722    if (gallivm_debug & GALLIVM_DEBUG_PERF) {
 723       int64_t time_end = os_time_get();
 724       int time_msec = (int)(time_end - time_begin) / 1000;
 725       debug_printf("   jitting func %s took %d msec\n",
 726                    LLVMGetValueName(func), time_msec);
 727    }
 728
 729    return jit_func;
 730 }