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
  63 #ifdef DEBUG
  64 unsigned gallivm_debug = 0;
  65
  66 static const struct debug_named_value lp_bld_debug_flags[] = {
  67    { "tgsi",   GALLIVM_DEBUG_TGSI, NULL },
  68    { "ir",     GALLIVM_DEBUG_IR, NULL },
  69    { "asm",    GALLIVM_DEBUG_ASM, NULL },
  70    { "nopt",   GALLIVM_DEBUG_NO_OPT, NULL },
  71    { "perf",   GALLIVM_DEBUG_PERF, NULL },
  72    { "no_brilinear", GALLIVM_DEBUG_NO_BRILINEAR, NULL },
  73    { "no_rho_approx", GALLIVM_DEBUG_NO_RHO_APPROX, NULL },
  74    { "no_quad_lod", GALLIVM_DEBUG_NO_QUAD_LOD, NULL },
  75    { "gc",     GALLIVM_DEBUG_GC, NULL },
  76    { "dumpbc", GALLIVM_DEBUG_DUMP_BC, NULL },
  77    DEBUG_NAMED_VALUE_END
  78 };
  79
  80 DEBUG_GET_ONCE_FLAGS_OPTION(gallivm_debug, "GALLIVM_DEBUG", lp_bld_debug_flags, 0)
  81 #endif
  82
  83
  84 static boolean gallivm_initialized = FALSE;
  85
  86 unsigned lp_native_vector_width;
  87
  88
  89 /*
  90  * Optimization values are:
  91  * - 0: None (-O0)
  92  * - 1: Less (-O1)
  93  * - 2: Default (-O2, -Os)
  94  * - 3: Aggressive (-O3)
  95  *
  96  * See also CodeGenOpt::Level in llvm/Target/TargetMachine.h
  97  */
  98 enum LLVM_CodeGenOpt_Level {
  99    None,        // -O0
 100    Less,        // -O1
 101    Default,     // -O2, -Os
 102    Aggressive   // -O3
 103 };
 104
 105
 106 /**
 107  * Create the LLVM (optimization) pass manager and install
 108  * relevant optimization passes.
 109  * \return  TRUE for success, FALSE for failure
 110  */
 111 static boolean
 112 create_pass_manager(struct gallivm_state *gallivm)
 113 {
 114    assert(!gallivm->passmgr);
 115    assert(gallivm->target);
 116
 117    gallivm->passmgr = LLVMCreateFunctionPassManagerForModule(gallivm->module);
 118    if (!gallivm->passmgr)
 119       return FALSE;
 120    /*
 121     * TODO: some per module pass manager with IPO passes might be helpful -
 122     * the generated texture functions may benefit from inlining if they are
 123     * simple, or constant propagation into them, etc.
 124     */
 125
 126 #if HAVE_LLVM < 0x0309
 127    // Old versions of LLVM get the DataLayout from the pass manager.
 128    LLVMAddTargetData(gallivm->target, gallivm->passmgr);
 129 #endif
 130
 131    {
 132       char *td_str;
 133       // New ones from the Module.
 134       td_str = LLVMCopyStringRepOfTargetData(gallivm->target);
 135       LLVMSetDataLayout(gallivm->module, td_str);
 136       free(td_str);
 137    }
 138
 139    if ((gallivm_debug & GALLIVM_DEBUG_NO_OPT) == 0) {
 140       /* These are the passes currently listed in llvm-c/Transforms/Scalar.h,
 141        * but there are more on SVN.
 142        * TODO: Add more passes.
 143        */
 144       LLVMAddScalarReplAggregatesPass(gallivm->passmgr);
 145       LLVMAddLICMPass(gallivm->passmgr);
 146       LLVMAddCFGSimplificationPass(gallivm->passmgr);
 147       LLVMAddReassociatePass(gallivm->passmgr);
 148       LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
 149       LLVMAddConstantPropagationPass(gallivm->passmgr);
 150       LLVMAddInstructionCombiningPass(gallivm->passmgr);
 151       LLVMAddGVNPass(gallivm->passmgr);
 152    }
 153    else {
 154       /* We need at least this pass to prevent the backends to fail in
 155        * unexpected ways.
 156        */
 157       LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
 158    }
 159
 160    return TRUE;
 161 }
 162
 163
 164 /**
 165  * Free gallivm object's LLVM allocations, but not any generated code
 166  * nor the gallivm object itself.
 167  */
 168 void
 169 gallivm_free_ir(struct gallivm_state *gallivm)
 170 {
 171    if (gallivm->passmgr) {
 172       LLVMDisposePassManager(gallivm->passmgr);
 173    }
 174
 175    if (gallivm->engine) {
 176       /* This will already destroy any associated module */
 177       LLVMDisposeExecutionEngine(gallivm->engine);
 178    } else if (gallivm->module) {
 179       LLVMDisposeModule(gallivm->module);
 180    }
 181
 182    FREE(gallivm->module_name);
 183
 184    if (!use_mcjit) {
 185       /* Don't free the TargetData, it's owned by the exec engine */
 186    } else {
 187       if (gallivm->target) {
 188          LLVMDisposeTargetData(gallivm->target);
 189       }
 190    }
 191
 192    if (gallivm->builder)
 193       LLVMDisposeBuilder(gallivm->builder);
 194
 195    /* The LLVMContext should be owned by the parent of gallivm. */
 196
 197    gallivm->engine = NULL;
 198    gallivm->target = NULL;
 199    gallivm->module = NULL;
 200    gallivm->module_name = NULL;
 201    gallivm->passmgr = NULL;
 202    gallivm->context = NULL;
 203    gallivm->builder = NULL;
 204 }
 205
 206
 207 /**
 208  * Free LLVM-generated code.  Should be done AFTER gallivm_free_ir().
 209  */
 210 static void
 211 gallivm_free_code(struct gallivm_state *gallivm)
 212 {
 213    assert(!gallivm->module);
 214    assert(!gallivm->engine);
 215    lp_free_generated_code(gallivm->code);
 216    gallivm->code = NULL;
 217    lp_free_memory_manager(gallivm->memorymgr);
 218    gallivm->memorymgr = NULL;
 219 }
 220
 221
 222 static boolean
 223 init_gallivm_engine(struct gallivm_state *gallivm)
 224 {
 225    if (1) {
 226       enum LLVM_CodeGenOpt_Level optlevel;
 227       char *error = NULL;
 228       int ret;
 229
 230       if (gallivm_debug & GALLIVM_DEBUG_NO_OPT) {
 231          optlevel = None;
 232       }
 233       else {
 234          optlevel = Default;
 235       }
 236
 237       ret = lp_build_create_jit_compiler_for_module(&gallivm->engine,
 238                                                     &gallivm->code,
 239                                                     gallivm->module,
 240                                                     gallivm->memorymgr,
 241                                                     (unsigned) optlevel,
 242                                                     use_mcjit,
 243                                                     &error);
 244       if (ret) {
 245          _debug_printf("%s\n", error);
 246          LLVMDisposeMessage(error);
 247          goto fail;
 248       }
 249    }
 250
 251    if (!use_mcjit) {
 252       gallivm->target = LLVMGetExecutionEngineTargetData(gallivm->engine);
 253       if (!gallivm->target)
 254          goto fail;
 255    } else {
 256       if (0) {
 257           /*
 258            * Dump the data layout strings.
 259            */
 260
 261           LLVMTargetDataRef target = LLVMGetExecutionEngineTargetData(gallivm->engine);
 262           char *data_layout;
 263           char *engine_data_layout;
 264
 265           data_layout = LLVMCopyStringRepOfTargetData(gallivm->target);
 266           engine_data_layout = LLVMCopyStringRepOfTargetData(target);
 267
 268           if (1) {
 269              debug_printf("module target data = %s\n", data_layout);
 270              debug_printf("engine target data = %s\n", engine_data_layout);
 271           }
 272
 273           free(data_layout);
 274           free(engine_data_layout);
 275       }
 276    }
 277
 278    return TRUE;
 279
 280 fail:
 281    return FALSE;
 282 }
 283
 284
 285 /**
 286  * Allocate gallivm LLVM objects.
 287  * \return  TRUE for success, FALSE for failure
 288  */
 289 static boolean
 290 init_gallivm_state(struct gallivm_state *gallivm, const char *name,
 291                    LLVMContextRef context)
 292 {
 293    assert(!gallivm->context);
 294    assert(!gallivm->module);
 295
 296    if (!lp_build_init())
 297       return FALSE;
 298
 299    gallivm->context = context;
 300
 301    if (!gallivm->context)
 302       goto fail;
 303
 304    gallivm->module_name = NULL;
 305    if (name) {
 306       size_t size = strlen(name) + 1;
 307       gallivm->module_name = MALLOC(size);
 308       if (gallivm->module_name) {
 309          memcpy(gallivm->module_name, name, size);
 310       }
 311    }
 312
 313    gallivm->module = LLVMModuleCreateWithNameInContext(name,
 314                                                        gallivm->context);
 315    if (!gallivm->module)
 316       goto fail;
 317
 318    gallivm->builder = LLVMCreateBuilderInContext(gallivm->context);
 319    if (!gallivm->builder)
 320       goto fail;
 321
 322    gallivm->memorymgr = lp_get_default_memory_manager();
 323    if (!gallivm->memorymgr)
 324       goto fail;
 325
 326    /* FIXME: MC-JIT only allows compiling one module at a time, and it must be
 327     * complete when MC-JIT is created. So defer the MC-JIT engine creation for
 328     * now.
 329     */
 330    if (!use_mcjit) {
 331       if (!init_gallivm_engine(gallivm)) {
 332          goto fail;
 333       }
 334    } else {
 335       /*
 336        * MC-JIT engine compiles the module immediately on creation, so we can't
 337        * obtain the target data from it.  Instead we create a target data layout
 338        * from a string.
 339        *
 340        * The produced layout strings are not precisely the same, but should make
 341        * no difference for the kind of optimization passes we run.
 342        *
 343        * For reference this is the layout string on x64:
 344        *
 345        *   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
 346        *
 347        * See also:
 348        * - http://llvm.org/docs/LangRef.html#datalayout
 349        */
 350
 351       {
 352          const unsigned pointer_size = 8 * sizeof(void *);
 353          char layout[512];
 354          util_snprintf(layout, sizeof layout, "%c-p:%u:%u:%u-i64:64:64-a0:0:%u-s0:%u:%u",
 355 #ifdef PIPE_ARCH_LITTLE_ENDIAN
 356                        'e', // little endian
 357 #else
 358                        'E', // big endian
 359 #endif
 360                        pointer_size, pointer_size, pointer_size, // pointer size, abi alignment, preferred alignment
 361                        pointer_size, // aggregate preferred alignment
 362                        pointer_size, pointer_size); // stack objects abi alignment, preferred alignment
 363
 364          gallivm->target = LLVMCreateTargetData(layout);
 365          if (!gallivm->target) {
 366             return FALSE;
 367          }
 368       }
 369    }
 370
 371    if (!create_pass_manager(gallivm))
 372       goto fail;
 373
 374    return TRUE;
 375
 376 fail:
 377    gallivm_free_ir(gallivm);
 378    gallivm_free_code(gallivm);
 379    return FALSE;
 380 }
 381
 382
 383 boolean
 384 lp_build_init(void)
 385 {
 386    if (gallivm_initialized)
 387       return TRUE;
 388
 389
 390    /* LLVMLinkIn* are no-ops at runtime.  They just ensure the respective
 391     * component is linked at buildtime, which is sufficient for its static
 392     * constructors to be called at load time.
 393     */
 394 #if defined(USE_MCJIT)
 395 #  if USE_MCJIT
 396       LLVMLinkInMCJIT();
 397 #  else
 398       LLVMLinkInJIT();
 399 #  endif
 400 #else
 401    use_mcjit = debug_get_bool_option("GALLIVM_MCJIT", FALSE);
 402    LLVMLinkInJIT();
 403    LLVMLinkInMCJIT();
 404 #endif
 405
 406 #ifdef DEBUG
 407    gallivm_debug = debug_get_option_gallivm_debug();
 408 #endif
 409
 410    lp_set_target_options();
 411
 412    util_cpu_detect();
 413
 414    /* For simulating less capable machines */
 415 #ifdef DEBUG
 416    if (debug_get_bool_option("LP_FORCE_SSE2", FALSE)) {
 417       assert(util_cpu_caps.has_sse2);
 418       util_cpu_caps.has_sse3 = 0;
 419       util_cpu_caps.has_ssse3 = 0;
 420       util_cpu_caps.has_sse4_1 = 0;
 421       util_cpu_caps.has_sse4_2 = 0;
 422       util_cpu_caps.has_avx = 0;
 423       util_cpu_caps.has_avx2 = 0;
 424       util_cpu_caps.has_f16c = 0;
 425       util_cpu_caps.has_fma = 0;
 426    }
 427 #endif
 428
 429    /* AMD Bulldozer AVX's throughput is the same as SSE2; and because using
 430     * 8-wide vector needs more floating ops than 4-wide (due to padding), it is
 431     * actually more efficient to use 4-wide vectors on this processor.
 432     *
 433     * See also:
 434     * - http://www.anandtech.com/show/4955/the-bulldozer-review-amd-fx8150-tested/2
 435     */
 436    if (util_cpu_caps.has_avx &&
 437        util_cpu_caps.has_intel) {
 438       lp_native_vector_width = 256;
 439    } else {
 440       /* Leave it at 128, even when no SIMD extensions are available.
 441        * Really needs to be a multiple of 128 so can fit 4 floats.
 442        */
 443       lp_native_vector_width = 128;
 444    }
 445
 446    lp_native_vector_width = debug_get_num_option("LP_NATIVE_VECTOR_WIDTH",
 447                                                  lp_native_vector_width);
 448
 449    if (lp_native_vector_width <= 128) {
 450       /* Hide AVX support, as often LLVM AVX intrinsics are only guarded by
 451        * "util_cpu_caps.has_avx" predicate, and lack the
 452        * "lp_native_vector_width > 128" predicate. And also to ensure a more
 453        * consistent behavior, allowing one to test SSE2 on AVX machines.
 454        * XXX: should not play games with util_cpu_caps directly as it might
 455        * get used for other things outside llvm too.
 456        */
 457       util_cpu_caps.has_avx = 0;
 458       util_cpu_caps.has_avx2 = 0;
 459       util_cpu_caps.has_f16c = 0;
 460       util_cpu_caps.has_fma = 0;
 461    }
 462    if (HAVE_LLVM < 0x0304 || !use_mcjit) {
 463       /* AVX2 support has only been tested with LLVM 3.4, and it requires
 464        * MCJIT. */
 465       util_cpu_caps.has_avx2 = 0;
 466    }
 467
 468 #ifdef PIPE_ARCH_PPC_64
 469    /* Set the NJ bit in VSCR to 0 so denormalized values are handled as
 470     * specified by IEEE standard (PowerISA 2.06 - Section 6.3). This guarantees
 471     * that some rounding and half-float to float handling does not round
 472     * incorrectly to 0.
 473     * XXX: should eventually follow same logic on all platforms.
 474     * Right now denorms get explicitly disabled (but elsewhere) for x86,
 475     * whereas ppc64 explicitly enables them...
 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    return TRUE;
 493 }
 494
 495
 496
 497 /**
 498  * Create a new gallivm_state object.
 499  */
 500 struct gallivm_state *
 501 gallivm_create(const char *name, LLVMContextRef context)
 502 {
 503    struct gallivm_state *gallivm;
 504
 505    gallivm = CALLOC_STRUCT(gallivm_state);
 506    if (gallivm) {
 507       if (!init_gallivm_state(gallivm, name, context)) {
 508          FREE(gallivm);
 509          gallivm = NULL;
 510       }
 511    }
 512
 513    return gallivm;
 514 }
 515
 516
 517 /**
 518  * Destroy a gallivm_state object.
 519  */
 520 void
 521 gallivm_destroy(struct gallivm_state *gallivm)
 522 {
 523    gallivm_free_ir(gallivm);
 524    gallivm_free_code(gallivm);
 525    FREE(gallivm);
 526 }
 527
 528
 529 /**
 530  * Validate a function.
 531  * Verification is only done with debug builds.
 532  */
 533 void
 534 gallivm_verify_function(struct gallivm_state *gallivm,
 535                         LLVMValueRef func)
 536 {
 537    /* Verify the LLVM IR.  If invalid, dump and abort */
 538 #ifdef DEBUG
 539    if (LLVMVerifyFunction(func, LLVMPrintMessageAction)) {
 540       lp_debug_dump_value(func);
 541       assert(0);
 542       return;
 543    }
 544 #endif
 545
 546    if (gallivm_debug & GALLIVM_DEBUG_IR) {
 547       /* Print the LLVM IR to stderr */
 548       lp_debug_dump_value(func);
 549       debug_printf("\n");
 550    }
 551 }
 552
 553
 554 /**
 555  * Compile a module.
 556  * This does IR optimization on all functions in the module.
 557  */
 558 void
 559 gallivm_compile_module(struct gallivm_state *gallivm)
 560 {
 561    LLVMValueRef func;
 562    int64_t time_begin = 0;
 563
 564    assert(!gallivm->compiled);
 565
 566    if (gallivm->builder) {
 567       LLVMDisposeBuilder(gallivm->builder);
 568       gallivm->builder = NULL;
 569    }
 570
 571    if (gallivm_debug & GALLIVM_DEBUG_PERF)
 572       time_begin = os_time_get();
 573
 574    /* Run optimization passes */
 575    LLVMInitializeFunctionPassManager(gallivm->passmgr);
 576    func = LLVMGetFirstFunction(gallivm->module);
 577    while (func) {
 578       if (0) {
 579          debug_printf("optimizing func %s...\n", LLVMGetValueName(func));
 580       }
 581
 582    /* Disable frame pointer omission on debug/profile builds */
 583    /* XXX: And workaround http://llvm.org/PR21435 */
 584 #if HAVE_LLVM >= 0x0307 && \
 585     (defined(DEBUG) || defined(PROFILE) || \
 586      defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64))
 587       LLVMAddTargetDependentFunctionAttr(func, "no-frame-pointer-elim", "true");
 588       LLVMAddTargetDependentFunctionAttr(func, "no-frame-pointer-elim-non-leaf", "true");
 589 #endif
 590
 591       LLVMRunFunctionPassManager(gallivm->passmgr, func);
 592       func = LLVMGetNextFunction(func);
 593    }
 594    LLVMFinalizeFunctionPassManager(gallivm->passmgr);
 595
 596    if (gallivm_debug & GALLIVM_DEBUG_PERF) {
 597       int64_t time_end = os_time_get();
 598       int time_msec = (int)(time_end - time_begin) / 1000;
 599       assert(gallivm->module_name);
 600       debug_printf("optimizing module %s took %d msec\n",
 601                    gallivm->module_name, time_msec);
 602    }
 603
 604    /* Dump byte code to a file */
 605    if (gallivm_debug & GALLIVM_DEBUG_DUMP_BC) {
 606       char filename[256];
 607       assert(gallivm->module_name);
 608       util_snprintf(filename, sizeof(filename), "ir_%s.bc", gallivm->module_name);
 609       LLVMWriteBitcodeToFile(gallivm->module, filename);
 610       debug_printf("%s written\n", filename);
 611       debug_printf("Invoke as \"llc %s%s -o - %s\"\n",
 612                    (HAVE_LLVM >= 0x0305) ? "[-mcpu=<-mcpu option>] " : "",
 613                    "[-mattr=<-mattr option(s)>]",
 614                    filename);
 615    }
 616
 617    if (use_mcjit) {
 618       /* Setting the module's DataLayout to an empty string will cause the
 619        * ExecutionEngine to copy to the DataLayout string from its target
 620        * machine to the module.  As of LLVM 3.8 the module and the execution
 621        * engine are required to have the same DataLayout.
 622        *
 623        * We must make sure we do this after running the optimization passes,
 624        * because those passes need a correct datalayout string.  For example,
 625        * if those optimization passes see an empty datalayout, they will assume
 626        * this is a little endian target and will do optimizations that break big
 627        * endian machines.
 628        *
 629        * TODO: This is just a temporary work-around.  The correct solution is
 630        * for gallivm_init_state() to create a TargetMachine and pull the
 631        * DataLayout from there.  Currently, the TargetMachine used by llvmpipe
 632        * is being implicitly created by the EngineBuilder in
 633        * lp_build_create_jit_compiler_for_module()
 634        */
 635       LLVMSetDataLayout(gallivm->module, "");
 636       assert(!gallivm->engine);
 637       if (!init_gallivm_engine(gallivm)) {
 638          assert(0);
 639       }
 640    }
 641    assert(gallivm->engine);
 642
 643    ++gallivm->compiled;
 644
 645    if (gallivm_debug & GALLIVM_DEBUG_ASM) {
 646       LLVMValueRef llvm_func = LLVMGetFirstFunction(gallivm->module);
 647
 648       while (llvm_func) {
 649          /*
 650           * Need to filter out functions which don't have an implementation,
 651           * such as the intrinsics. May not be sufficient in case of IPO?
 652           * LLVMGetPointerToGlobal() will abort otherwise.
 653           */
 654          if (!LLVMIsDeclaration(llvm_func)) {
 655             void *func_code = LLVMGetPointerToGlobal(gallivm->engine, llvm_func);
 656             lp_disassemble(llvm_func, func_code);
 657          }
 658          llvm_func = LLVMGetNextFunction(llvm_func);
 659       }
 660    }
 661
 662 #if defined(PROFILE)
 663    {
 664       LLVMValueRef llvm_func = LLVMGetFirstFunction(gallivm->module);
 665
 666       while (llvm_func) {
 667          if (!LLVMIsDeclaration(llvm_func)) {
 668             void *func_code = LLVMGetPointerToGlobal(gallivm->engine, llvm_func);
 669             lp_profile(llvm_func, func_code);
 670          }
 671          llvm_func = LLVMGetNextFunction(llvm_func);
 672       }
 673    }
 674 #endif
 675 }
 676
 677
 678
 679 func_pointer
 680 gallivm_jit_function(struct gallivm_state *gallivm,
 681                      LLVMValueRef func)
 682 {
 683    void *code;
 684    func_pointer jit_func;
 685    int64_t time_begin = 0;
 686
 687    assert(gallivm->compiled);
 688    assert(gallivm->engine);
 689
 690    if (gallivm_debug & GALLIVM_DEBUG_PERF)
 691       time_begin = os_time_get();
 692
 693    code = LLVMGetPointerToGlobal(gallivm->engine, func);
 694    assert(code);
 695    jit_func = pointer_to_func(code);
 696
 697    if (gallivm_debug & GALLIVM_DEBUG_PERF) {
 698       int64_t time_end = os_time_get();
 699       int time_msec = (int)(time_end - time_begin) / 1000;
 700       debug_printf("   jitting func %s took %d msec\n",
 701                    LLVMGetValueName(func), time_msec);
 702    }
 703
 704    return jit_func;
 705 }