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 "os/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 #include <llvm-c/BitWriter.h>
  44
  45
  46 /* Only MCJIT is available as of LLVM SVN r216982 */
  47 #if HAVE_LLVM >= 0x0306
  48 #  define USE_MCJIT 1
  49 #elif defined(PIPE_ARCH_PPC_64) || defined(PIPE_ARCH_S390) || defined(PIPE_ARCH_ARM) || defined(PIPE_ARCH_AARCH64)
  50 #  define USE_MCJIT 1
  51 #endif
  52
  53 #if defined(USE_MCJIT)
  54 static const bool use_mcjit = USE_MCJIT;
  55 #else
  56 static bool use_mcjit = FALSE;
  57 #endif
  58
  59
  60 #ifdef DEBUG
  61 unsigned gallivm_debug = 0;
  62
  63 static const struct debug_named_value lp_bld_debug_flags[] = {
  64    { "tgsi",   GALLIVM_DEBUG_TGSI, NULL },
  65    { "ir",     GALLIVM_DEBUG_IR, NULL },
  66    { "asm",    GALLIVM_DEBUG_ASM, NULL },
  67    { "nopt",   GALLIVM_DEBUG_NO_OPT, NULL },
  68    { "perf",   GALLIVM_DEBUG_PERF, NULL },
  69    { "no_brilinear", GALLIVM_DEBUG_NO_BRILINEAR, NULL },
  70    { "no_rho_approx", GALLIVM_DEBUG_NO_RHO_APPROX, NULL },
  71    { "no_quad_lod", GALLIVM_DEBUG_NO_QUAD_LOD, NULL },
  72    { "gc",     GALLIVM_DEBUG_GC, NULL },
  73    { "dumpbc", GALLIVM_DEBUG_DUMP_BC, NULL },
  74    DEBUG_NAMED_VALUE_END
  75 };
  76
  77 DEBUG_GET_ONCE_FLAGS_OPTION(gallivm_debug, "GALLIVM_DEBUG", lp_bld_debug_flags, 0)
  78 #endif
  79
  80
  81 static boolean gallivm_initialized = FALSE;
  82
  83 unsigned lp_native_vector_width;
  84
  85
  86 /*
  87  * Optimization values are:
  88  * - 0: None (-O0)
  89  * - 1: Less (-O1)
  90  * - 2: Default (-O2, -Os)
  91  * - 3: Aggressive (-O3)
  92  *
  93  * See also CodeGenOpt::Level in llvm/Target/TargetMachine.h
  94  */
  95 enum LLVM_CodeGenOpt_Level {
  96    None,        // -O0
  97    Less,        // -O1
  98    Default,     // -O2, -Os
  99    Aggressive   // -O3
 100 };
 101
 102
 103 /**
 104  * Create the LLVM (optimization) pass manager and install
 105  * relevant optimization passes.
 106  * \return  TRUE for success, FALSE for failure
 107  */
 108 static boolean
 109 create_pass_manager(struct gallivm_state *gallivm)
 110 {
 111    assert(!gallivm->passmgr);
 112    assert(gallivm->target);
 113
 114    gallivm->passmgr = LLVMCreateFunctionPassManagerForModule(gallivm->module);
 115    if (!gallivm->passmgr)
 116       return FALSE;
 117    /*
 118     * TODO: some per module pass manager with IPO passes might be helpful -
 119     * the generated texture functions may benefit from inlining if they are
 120     * simple, or constant propagation into them, etc.
 121     */
 122
 123 #if HAVE_LLVM < 0x0309
 124    // Old versions of LLVM get the DataLayout from the pass manager.
 125    LLVMAddTargetData(gallivm->target, gallivm->passmgr);
 126 #endif
 127
 128    /* Setting the module's DataLayout to an empty string will cause the
 129     * ExecutionEngine to copy to the DataLayout string from its target
 130     * machine to the module.  As of LLVM 3.8 the module and the execution
 131     * engine are required to have the same DataLayout.
 132     *
 133     * TODO: This is just a temporary work-around.  The correct solution is
 134     * for gallivm_init_state() to create a TargetMachine and pull the
 135     * DataLayout from there.  Currently, the TargetMachine used by llvmpipe
 136     * is being implicitly created by the EngineBuilder in
 137     * lp_build_create_jit_compiler_for_module()
 138     */
 139
 140 #if HAVE_LLVM < 0x0308
 141    {
 142       char *td_str;
 143       // New ones from the Module.
 144       td_str = LLVMCopyStringRepOfTargetData(gallivm->target);
 145       LLVMSetDataLayout(gallivm->module, td_str);
 146       free(td_str);
 147    }
 148 #else
 149    LLVMSetDataLayout(gallivm->module, "");
 150 #endif
 151
 152    if ((gallivm_debug & GALLIVM_DEBUG_NO_OPT) == 0) {
 153       /* These are the passes currently listed in llvm-c/Transforms/Scalar.h,
 154        * but there are more on SVN.
 155        * TODO: Add more passes.
 156        */
 157       LLVMAddScalarReplAggregatesPass(gallivm->passmgr);
 158       LLVMAddLICMPass(gallivm->passmgr);
 159       LLVMAddCFGSimplificationPass(gallivm->passmgr);
 160       LLVMAddReassociatePass(gallivm->passmgr);
 161       LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
 162       LLVMAddConstantPropagationPass(gallivm->passmgr);
 163       LLVMAddInstructionCombiningPass(gallivm->passmgr);
 164       LLVMAddGVNPass(gallivm->passmgr);
 165    }
 166    else {
 167       /* We need at least this pass to prevent the backends to fail in
 168        * unexpected ways.
 169        */
 170       LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
 171    }
 172
 173    return TRUE;
 174 }
 175
 176
 177 /**
 178  * Free gallivm object's LLVM allocations, but not any generated code
 179  * nor the gallivm object itself.
 180  */
 181 void
 182 gallivm_free_ir(struct gallivm_state *gallivm)
 183 {
 184    if (gallivm->passmgr) {
 185       LLVMDisposePassManager(gallivm->passmgr);
 186    }
 187
 188    if (gallivm->engine) {
 189       /* This will already destroy any associated module */
 190       LLVMDisposeExecutionEngine(gallivm->engine);
 191    } else if (gallivm->module) {
 192       LLVMDisposeModule(gallivm->module);
 193    }
 194
 195    FREE(gallivm->module_name);
 196
 197    if (!use_mcjit) {
 198       /* Don't free the TargetData, it's owned by the exec engine */
 199    } else {
 200       if (gallivm->target) {
 201          LLVMDisposeTargetData(gallivm->target);
 202       }
 203    }
 204
 205    if (gallivm->builder)
 206       LLVMDisposeBuilder(gallivm->builder);
 207
 208    /* The LLVMContext should be owned by the parent of gallivm. */
 209
 210    gallivm->engine = NULL;
 211    gallivm->target = NULL;
 212    gallivm->module = NULL;
 213    gallivm->module_name = NULL;
 214    gallivm->passmgr = NULL;
 215    gallivm->context = NULL;
 216    gallivm->builder = NULL;
 217 }
 218
 219
 220 /**
 221  * Free LLVM-generated code.  Should be done AFTER gallivm_free_ir().
 222  */
 223 static void
 224 gallivm_free_code(struct gallivm_state *gallivm)
 225 {
 226    assert(!gallivm->module);
 227    assert(!gallivm->engine);
 228    lp_free_generated_code(gallivm->code);
 229    gallivm->code = NULL;
 230    lp_free_memory_manager(gallivm->memorymgr);
 231    gallivm->memorymgr = NULL;
 232 }
 233
 234
 235 static boolean
 236 init_gallivm_engine(struct gallivm_state *gallivm)
 237 {
 238    if (1) {
 239       enum LLVM_CodeGenOpt_Level optlevel;
 240       char *error = NULL;
 241       int ret;
 242
 243       if (gallivm_debug & GALLIVM_DEBUG_NO_OPT) {
 244          optlevel = None;
 245       }
 246       else {
 247          optlevel = Default;
 248       }
 249
 250       ret = lp_build_create_jit_compiler_for_module(&gallivm->engine,
 251                                                     &gallivm->code,
 252                                                     gallivm->module,
 253                                                     gallivm->memorymgr,
 254                                                     (unsigned) optlevel,
 255                                                     use_mcjit,
 256                                                     &error);
 257       if (ret) {
 258          _debug_printf("%s\n", error);
 259          LLVMDisposeMessage(error);
 260          goto fail;
 261       }
 262    }
 263
 264    if (!use_mcjit) {
 265       gallivm->target = LLVMGetExecutionEngineTargetData(gallivm->engine);
 266       if (!gallivm->target)
 267          goto fail;
 268    } else {
 269       if (0) {
 270           /*
 271            * Dump the data layout strings.
 272            */
 273
 274           LLVMTargetDataRef target = LLVMGetExecutionEngineTargetData(gallivm->engine);
 275           char *data_layout;
 276           char *engine_data_layout;
 277
 278           data_layout = LLVMCopyStringRepOfTargetData(gallivm->target);
 279           engine_data_layout = LLVMCopyStringRepOfTargetData(target);
 280
 281           if (1) {
 282              debug_printf("module target data = %s\n", data_layout);
 283              debug_printf("engine target data = %s\n", engine_data_layout);
 284           }
 285
 286           free(data_layout);
 287           free(engine_data_layout);
 288       }
 289    }
 290
 291    return TRUE;
 292
 293 fail:
 294    return FALSE;
 295 }
 296
 297
 298 /**
 299  * Allocate gallivm LLVM objects.
 300  * \return  TRUE for success, FALSE for failure
 301  */
 302 static boolean
 303 init_gallivm_state(struct gallivm_state *gallivm, const char *name,
 304                    LLVMContextRef context)
 305 {
 306    assert(!gallivm->context);
 307    assert(!gallivm->module);
 308
 309    if (!lp_build_init())
 310       return FALSE;
 311
 312    gallivm->context = context;
 313
 314    if (!gallivm->context)
 315       goto fail;
 316
 317    gallivm->module_name = NULL;
 318    if (name) {
 319       size_t size = strlen(name) + 1;
 320       gallivm->module_name = MALLOC(size);
 321       if (gallivm->module_name) {
 322          memcpy(gallivm->module_name, name, size);
 323       }
 324    }
 325
 326    gallivm->module = LLVMModuleCreateWithNameInContext(name,
 327                                                        gallivm->context);
 328    if (!gallivm->module)
 329       goto fail;
 330
 331    gallivm->builder = LLVMCreateBuilderInContext(gallivm->context);
 332    if (!gallivm->builder)
 333       goto fail;
 334
 335    gallivm->memorymgr = lp_get_default_memory_manager();
 336    if (!gallivm->memorymgr)
 337       goto fail;
 338
 339    /* FIXME: MC-JIT only allows compiling one module at a time, and it must be
 340     * complete when MC-JIT is created. So defer the MC-JIT engine creation for
 341     * now.
 342     */
 343    if (!use_mcjit) {
 344       if (!init_gallivm_engine(gallivm)) {
 345          goto fail;
 346       }
 347    } else {
 348       /*
 349        * MC-JIT engine compiles the module immediately on creation, so we can't
 350        * obtain the target data from it.  Instead we create a target data layout
 351        * from a string.
 352        *
 353        * The produced layout strings are not precisely the same, but should make
 354        * no difference for the kind of optimization passes we run.
 355        *
 356        * For reference this is the layout string on x64:
 357        *
 358        *   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
 359        *
 360        * See also:
 361        * - http://llvm.org/docs/LangRef.html#datalayout
 362        */
 363
 364       {
 365          const unsigned pointer_size = 8 * sizeof(void *);
 366          char layout[512];
 367          util_snprintf(layout, sizeof layout, "%c-p:%u:%u:%u-i64:64:64-a0:0:%u-s0:%u:%u",
 368 #ifdef PIPE_ARCH_LITTLE_ENDIAN
 369                        'e', // little endian
 370 #else
 371                        'E', // big endian
 372 #endif
 373                        pointer_size, pointer_size, pointer_size, // pointer size, abi alignment, preferred alignment
 374                        pointer_size, // aggregate preferred alignment
 375                        pointer_size, pointer_size); // stack objects abi alignment, preferred alignment
 376
 377          gallivm->target = LLVMCreateTargetData(layout);
 378          if (!gallivm->target) {
 379             return FALSE;
 380          }
 381       }
 382    }
 383
 384    if (!create_pass_manager(gallivm))
 385       goto fail;
 386
 387    return TRUE;
 388
 389 fail:
 390    gallivm_free_ir(gallivm);
 391    gallivm_free_code(gallivm);
 392    return FALSE;
 393 }
 394
 395
 396 boolean
 397 lp_build_init(void)
 398 {
 399    if (gallivm_initialized)
 400       return TRUE;
 401
 402
 403    /* LLVMLinkIn* are no-ops at runtime.  They just ensure the respective
 404     * component is linked at buildtime, which is sufficient for its static
 405     * constructors to be called at load time.
 406     */
 407 #if defined(USE_MCJIT)
 408 #  if USE_MCJIT
 409       LLVMLinkInMCJIT();
 410 #  else
 411       LLVMLinkInJIT();
 412 #  endif
 413 #else
 414    use_mcjit = debug_get_bool_option("GALLIVM_MCJIT", FALSE);
 415    LLVMLinkInJIT();
 416    LLVMLinkInMCJIT();
 417 #endif
 418
 419 #ifdef DEBUG
 420    gallivm_debug = debug_get_option_gallivm_debug();
 421 #endif
 422
 423    lp_set_target_options();
 424
 425    util_cpu_detect();
 426
 427    /* For simulating less capable machines */
 428 #ifdef DEBUG
 429    if (debug_get_bool_option("LP_FORCE_SSE2", FALSE)) {
 430       assert(util_cpu_caps.has_sse2);
 431       util_cpu_caps.has_sse3 = 0;
 432       util_cpu_caps.has_ssse3 = 0;
 433       util_cpu_caps.has_sse4_1 = 0;
 434       util_cpu_caps.has_sse4_2 = 0;
 435       util_cpu_caps.has_avx = 0;
 436       util_cpu_caps.has_avx2 = 0;
 437       util_cpu_caps.has_f16c = 0;
 438       util_cpu_caps.has_fma = 0;
 439    }
 440 #endif
 441
 442    /* AMD Bulldozer AVX's throughput is the same as SSE2; and because using
 443     * 8-wide vector needs more floating ops than 4-wide (due to padding), it is
 444     * actually more efficient to use 4-wide vectors on this processor.
 445     *
 446     * See also:
 447     * - http://www.anandtech.com/show/4955/the-bulldozer-review-amd-fx8150-tested/2
 448     */
 449    if (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 intrinsics 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        * XXX: should not play games with util_cpu_caps directly as it might
 468        * get used for other things outside llvm too.
 469        */
 470       util_cpu_caps.has_avx = 0;
 471       util_cpu_caps.has_avx2 = 0;
 472       util_cpu_caps.has_f16c = 0;
 473       util_cpu_caps.has_fma = 0;
 474    }
 475    if (HAVE_LLVM < 0x0304 || !use_mcjit) {
 476       /* AVX2 support has only been tested with LLVM 3.4, and it requires
 477        * MCJIT. */
 478       util_cpu_caps.has_avx2 = 0;
 479    }
 480
 481 #ifdef PIPE_ARCH_PPC_64
 482    /* Set the NJ bit in VSCR to 0 so denormalized values are handled as
 483     * specified by IEEE standard (PowerISA 2.06 - Section 6.3). This guarantees
 484     * that some rounding and half-float to float handling does not round
 485     * incorrectly to 0.
 486     * XXX: should eventually follow same logic on all platforms.
 487     * Right now denorms get explicitly disabled (but elsewhere) for x86,
 488     * whereas ppc64 explicitly enables them...
 489     */
 490    if (util_cpu_caps.has_altivec) {
 491       unsigned short mask[] = { 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF,
 492                                 0xFFFF, 0xFFFF, 0xFFFE, 0xFFFF };
 493       __asm (
 494         "mfvscr %%v1\n"
 495         "vand   %0,%%v1,%0\n"
 496         "mtvscr %0"
 497         :
 498         : "r" (*mask)
 499       );
 500    }
 501 #endif
 502
 503    gallivm_initialized = TRUE;
 504
 505    return TRUE;
 506 }
 507
 508
 509
 510 /**
 511  * Create a new gallivm_state object.
 512  */
 513 struct gallivm_state *
 514 gallivm_create(const char *name, LLVMContextRef context)
 515 {
 516    struct gallivm_state *gallivm;
 517
 518    gallivm = CALLOC_STRUCT(gallivm_state);
 519    if (gallivm) {
 520       if (!init_gallivm_state(gallivm, name, context)) {
 521          FREE(gallivm);
 522          gallivm = NULL;
 523       }
 524    }
 525
 526    return gallivm;
 527 }
 528
 529
 530 /**
 531  * Destroy a gallivm_state object.
 532  */
 533 void
 534 gallivm_destroy(struct gallivm_state *gallivm)
 535 {
 536    gallivm_free_ir(gallivm);
 537    gallivm_free_code(gallivm);
 538    FREE(gallivm);
 539 }
 540
 541
 542 /**
 543  * Validate a function.
 544  * Verification is only done with debug builds.
 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    if (gallivm_debug & GALLIVM_DEBUG_IR) {
 560       /* Print the LLVM IR to stderr */
 561       lp_debug_dump_value(func);
 562       debug_printf("\n");
 563    }
 564 }
 565
 566
 567 /**
 568  * Compile a module.
 569  * This does IR optimization on all functions in the module.
 570  */
 571 void
 572 gallivm_compile_module(struct gallivm_state *gallivm)
 573 {
 574    LLVMValueRef func;
 575    int64_t time_begin = 0;
 576
 577    assert(!gallivm->compiled);
 578
 579    if (gallivm->builder) {
 580       LLVMDisposeBuilder(gallivm->builder);
 581       gallivm->builder = NULL;
 582    }
 583
 584    if (gallivm_debug & GALLIVM_DEBUG_PERF)
 585       time_begin = os_time_get();
 586
 587    /* Run optimization passes */
 588    LLVMInitializeFunctionPassManager(gallivm->passmgr);
 589    func = LLVMGetFirstFunction(gallivm->module);
 590    while (func) {
 591       if (0) {
 592          debug_printf("optimizing func %s...\n", LLVMGetValueName(func));
 593       }
 594
 595    /* Disable frame pointer omission on debug/profile builds */
 596    /* XXX: And workaround http://llvm.org/PR21435 */
 597 #if HAVE_LLVM >= 0x0307 && \
 598     (defined(DEBUG) || defined(PROFILE) || \
 599      defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64))
 600       LLVMAddTargetDependentFunctionAttr(func, "no-frame-pointer-elim", "true");
 601       LLVMAddTargetDependentFunctionAttr(func, "no-frame-pointer-elim-non-leaf", "true");
 602 #endif
 603
 604       LLVMRunFunctionPassManager(gallivm->passmgr, func);
 605       func = LLVMGetNextFunction(func);
 606    }
 607    LLVMFinalizeFunctionPassManager(gallivm->passmgr);
 608
 609    if (gallivm_debug & GALLIVM_DEBUG_PERF) {
 610       int64_t time_end = os_time_get();
 611       int time_msec = (int)(time_end - time_begin) / 1000;
 612       assert(gallivm->module_name);
 613       debug_printf("optimizing module %s took %d msec\n",
 614                    gallivm->module_name, time_msec);
 615    }
 616
 617    /* Dump byte code to a file */
 618    if (gallivm_debug & GALLIVM_DEBUG_DUMP_BC) {
 619       char filename[256];
 620       assert(gallivm->module_name);
 621       util_snprintf(filename, sizeof(filename), "ir_%s.bc", gallivm->module_name);
 622       LLVMWriteBitcodeToFile(gallivm->module, filename);
 623       debug_printf("%s written\n", filename);
 624       debug_printf("Invoke as \"llc %s%s -o - %s\"\n",
 625                    (HAVE_LLVM >= 0x0305) ? "[-mcpu=<-mcpu option] " : "",
 626                    "[-mattr=<-mattr option(s)>]",
 627                    filename);
 628    }
 629
 630    if (use_mcjit) {
 631       assert(!gallivm->engine);
 632       if (!init_gallivm_engine(gallivm)) {
 633          assert(0);
 634       }
 635    }
 636    assert(gallivm->engine);
 637
 638    ++gallivm->compiled;
 639
 640    if (gallivm_debug & GALLIVM_DEBUG_ASM) {
 641       LLVMValueRef llvm_func = LLVMGetFirstFunction(gallivm->module);
 642
 643       while (llvm_func) {
 644          /*
 645           * Need to filter out functions which don't have an implementation,
 646           * such as the intrinsics. May not be sufficient in case of IPO?
 647           * LLVMGetPointerToGlobal() will abort otherwise.
 648           */
 649          if (!LLVMIsDeclaration(llvm_func)) {
 650             void *func_code = LLVMGetPointerToGlobal(gallivm->engine, llvm_func);
 651             lp_disassemble(llvm_func, func_code);
 652          }
 653          llvm_func = LLVMGetNextFunction(llvm_func);
 654       }
 655    }
 656
 657 #if defined(PROFILE)
 658    {
 659       LLVMValueRef llvm_func = LLVMGetFirstFunction(gallivm->module);
 660
 661       while (llvm_func) {
 662          if (!LLVMIsDeclaration(llvm_func)) {
 663             void *func_code = LLVMGetPointerToGlobal(gallivm->engine, llvm_func);
 664             lp_profile(llvm_func, func_code);
 665          }
 666          llvm_func = LLVMGetNextFunction(llvm_func);
 667       }
 668    }
 669 #endif
 670 }
 671
 672
 673
 674 func_pointer
 675 gallivm_jit_function(struct gallivm_state *gallivm,
 676                      LLVMValueRef func)
 677 {
 678    void *code;
 679    func_pointer jit_func;
 680    int64_t time_begin = 0;
 681
 682    assert(gallivm->compiled);
 683    assert(gallivm->engine);
 684
 685    if (gallivm_debug & GALLIVM_DEBUG_PERF)
 686       time_begin = os_time_get();
 687
 688    code = LLVMGetPointerToGlobal(gallivm->engine, func);
 689    assert(code);
 690    jit_func = pointer_to_func(code);
 691
 692    if (gallivm_debug & GALLIVM_DEBUG_PERF) {
 693       int64_t time_end = os_time_get();
 694       int time_msec = (int)(time_end - time_begin) / 1000;
 695       debug_printf("   jitting func %s took %d msec\n",
 696                    LLVMGetValueName(func), time_msec);
 697    }
 698
 699    return jit_func;
 700 }