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       /*
 141        * TODO: Evaluate passes some more - keeping in mind
 142        * both quality of generated code and compile times.
 143        */
 144       LLVMAddScalarReplAggregatesPass(gallivm->passmgr);
 145       LLVMAddEarlyCSEPass(gallivm->passmgr);
 146       LLVMAddCFGSimplificationPass(gallivm->passmgr);
 147       LLVMAddLICMPass(gallivm->passmgr);
 148       LLVMAddReassociatePass(gallivm->passmgr);
 149       LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
 150       LLVMAddConstantPropagationPass(gallivm->passmgr);
 151       LLVMAddInstructionCombiningPass(gallivm->passmgr);
 152       LLVMAddGVNPass(gallivm->passmgr);
 153    }
 154    else {
 155       /* We need at least this pass to prevent the backends to fail in
 156        * unexpected ways.
 157        */
 158       LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
 159    }
 160
 161    return TRUE;
 162 }
 163
 164
 165 /**
 166  * Free gallivm object's LLVM allocations, but not any generated code
 167  * nor the gallivm object itself.
 168  */
 169 void
 170 gallivm_free_ir(struct gallivm_state *gallivm)
 171 {
 172    if (gallivm->passmgr) {
 173       LLVMDisposePassManager(gallivm->passmgr);
 174    }
 175
 176    if (gallivm->engine) {
 177       /* This will already destroy any associated module */
 178       LLVMDisposeExecutionEngine(gallivm->engine);
 179    } else if (gallivm->module) {
 180       LLVMDisposeModule(gallivm->module);
 181    }
 182
 183    FREE(gallivm->module_name);
 184
 185    if (!use_mcjit) {
 186       /* Don't free the TargetData, it's owned by the exec engine */
 187    } else {
 188       if (gallivm->target) {
 189          LLVMDisposeTargetData(gallivm->target);
 190       }
 191    }
 192
 193    if (gallivm->builder)
 194       LLVMDisposeBuilder(gallivm->builder);
 195
 196    /* The LLVMContext should be owned by the parent of gallivm. */
 197
 198    gallivm->engine = NULL;
 199    gallivm->target = NULL;
 200    gallivm->module = NULL;
 201    gallivm->module_name = NULL;
 202    gallivm->passmgr = NULL;
 203    gallivm->context = NULL;
 204    gallivm->builder = NULL;
 205 }
 206
 207
 208 /**
 209  * Free LLVM-generated code.  Should be done AFTER gallivm_free_ir().
 210  */
 211 static void
 212 gallivm_free_code(struct gallivm_state *gallivm)
 213 {
 214    assert(!gallivm->module);
 215    assert(!gallivm->engine);
 216    lp_free_generated_code(gallivm->code);
 217    gallivm->code = NULL;
 218    lp_free_memory_manager(gallivm->memorymgr);
 219    gallivm->memorymgr = NULL;
 220 }
 221
 222
 223 static boolean
 224 init_gallivm_engine(struct gallivm_state *gallivm)
 225 {
 226    if (1) {
 227       enum LLVM_CodeGenOpt_Level optlevel;
 228       char *error = NULL;
 229       int ret;
 230
 231       if (gallivm_debug & GALLIVM_DEBUG_NO_OPT) {
 232          optlevel = None;
 233       }
 234       else {
 235          optlevel = Default;
 236       }
 237
 238       ret = lp_build_create_jit_compiler_for_module(&gallivm->engine,
 239                                                     &gallivm->code,
 240                                                     gallivm->module,
 241                                                     gallivm->memorymgr,
 242                                                     (unsigned) optlevel,
 243                                                     use_mcjit,
 244                                                     &error);
 245       if (ret) {
 246          _debug_printf("%s\n", error);
 247          LLVMDisposeMessage(error);
 248          goto fail;
 249       }
 250    }
 251
 252    if (!use_mcjit) {
 253       gallivm->target = LLVMGetExecutionEngineTargetData(gallivm->engine);
 254       if (!gallivm->target)
 255          goto fail;
 256    } else {
 257       if (0) {
 258           /*
 259            * Dump the data layout strings.
 260            */
 261
 262           LLVMTargetDataRef target = LLVMGetExecutionEngineTargetData(gallivm->engine);
 263           char *data_layout;
 264           char *engine_data_layout;
 265
 266           data_layout = LLVMCopyStringRepOfTargetData(gallivm->target);
 267           engine_data_layout = LLVMCopyStringRepOfTargetData(target);
 268
 269           if (1) {
 270              debug_printf("module target data = %s\n", data_layout);
 271              debug_printf("engine target data = %s\n", engine_data_layout);
 272           }
 273
 274           free(data_layout);
 275           free(engine_data_layout);
 276       }
 277    }
 278
 279    return TRUE;
 280
 281 fail:
 282    return FALSE;
 283 }
 284
 285
 286 /**
 287  * Allocate gallivm LLVM objects.
 288  * \return  TRUE for success, FALSE for failure
 289  */
 290 static boolean
 291 init_gallivm_state(struct gallivm_state *gallivm, const char *name,
 292                    LLVMContextRef context)
 293 {
 294    assert(!gallivm->context);
 295    assert(!gallivm->module);
 296
 297    if (!lp_build_init())
 298       return FALSE;
 299
 300    gallivm->context = context;
 301
 302    if (!gallivm->context)
 303       goto fail;
 304
 305    gallivm->module_name = NULL;
 306    if (name) {
 307       size_t size = strlen(name) + 1;
 308       gallivm->module_name = MALLOC(size);
 309       if (gallivm->module_name) {
 310          memcpy(gallivm->module_name, name, size);
 311       }
 312    }
 313
 314    gallivm->module = LLVMModuleCreateWithNameInContext(name,
 315                                                        gallivm->context);
 316    if (!gallivm->module)
 317       goto fail;
 318
 319    gallivm->builder = LLVMCreateBuilderInContext(gallivm->context);
 320    if (!gallivm->builder)
 321       goto fail;
 322
 323    gallivm->memorymgr = lp_get_default_memory_manager();
 324    if (!gallivm->memorymgr)
 325       goto fail;
 326
 327    /* FIXME: MC-JIT only allows compiling one module at a time, and it must be
 328     * complete when MC-JIT is created. So defer the MC-JIT engine creation for
 329     * now.
 330     */
 331    if (!use_mcjit) {
 332       if (!init_gallivm_engine(gallivm)) {
 333          goto fail;
 334       }
 335    } else {
 336       /*
 337        * MC-JIT engine compiles the module immediately on creation, so we can't
 338        * obtain the target data from it.  Instead we create a target data layout
 339        * from a string.
 340        *
 341        * The produced layout strings are not precisely the same, but should make
 342        * no difference for the kind of optimization passes we run.
 343        *
 344        * For reference this is the layout string on x64:
 345        *
 346        *   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
 347        *
 348        * See also:
 349        * - http://llvm.org/docs/LangRef.html#datalayout
 350        */
 351
 352       {
 353          const unsigned pointer_size = 8 * sizeof(void *);
 354          char layout[512];
 355          util_snprintf(layout, sizeof layout, "%c-p:%u:%u:%u-i64:64:64-a0:0:%u-s0:%u:%u",
 356 #ifdef PIPE_ARCH_LITTLE_ENDIAN
 357                        'e', // little endian
 358 #else
 359                        'E', // big endian
 360 #endif
 361                        pointer_size, pointer_size, pointer_size, // pointer size, abi alignment, preferred alignment
 362                        pointer_size, // aggregate preferred alignment
 363                        pointer_size, pointer_size); // stack objects abi alignment, preferred alignment
 364
 365          gallivm->target = LLVMCreateTargetData(layout);
 366          if (!gallivm->target) {
 367             return FALSE;
 368          }
 369       }
 370    }
 371
 372    if (!create_pass_manager(gallivm))
 373       goto fail;
 374
 375    return TRUE;
 376
 377 fail:
 378    gallivm_free_ir(gallivm);
 379    gallivm_free_code(gallivm);
 380    return FALSE;
 381 }
 382
 383
 384 boolean
 385 lp_build_init(void)
 386 {
 387    if (gallivm_initialized)
 388       return TRUE;
 389
 390
 391    /* LLVMLinkIn* are no-ops at runtime.  They just ensure the respective
 392     * component is linked at buildtime, which is sufficient for its static
 393     * constructors to be called at load time.
 394     */
 395 #if defined(USE_MCJIT)
 396 #  if USE_MCJIT
 397       LLVMLinkInMCJIT();
 398 #  else
 399       LLVMLinkInJIT();
 400 #  endif
 401 #else
 402    use_mcjit = debug_get_bool_option("GALLIVM_MCJIT", FALSE);
 403    LLVMLinkInJIT();
 404    LLVMLinkInMCJIT();
 405 #endif
 406
 407 #ifdef DEBUG
 408    gallivm_debug = debug_get_option_gallivm_debug();
 409 #endif
 410
 411    lp_set_target_options();
 412
 413    util_cpu_detect();
 414
 415    /* For simulating less capable machines */
 416 #ifdef DEBUG
 417    if (debug_get_bool_option("LP_FORCE_SSE2", FALSE)) {
 418       assert(util_cpu_caps.has_sse2);
 419       util_cpu_caps.has_sse3 = 0;
 420       util_cpu_caps.has_ssse3 = 0;
 421       util_cpu_caps.has_sse4_1 = 0;
 422       util_cpu_caps.has_sse4_2 = 0;
 423       util_cpu_caps.has_avx = 0;
 424       util_cpu_caps.has_avx2 = 0;
 425       util_cpu_caps.has_f16c = 0;
 426       util_cpu_caps.has_fma = 0;
 427    }
 428 #endif
 429
 430    /* AMD Bulldozer AVX's throughput is the same as SSE2; and because using
 431     * 8-wide vector needs more floating ops than 4-wide (due to padding), it is
 432     * actually more efficient to use 4-wide vectors on this processor.
 433     *
 434     * See also:
 435     * - http://www.anandtech.com/show/4955/the-bulldozer-review-amd-fx8150-tested/2
 436     */
 437    if (util_cpu_caps.has_avx &&
 438        util_cpu_caps.has_intel) {
 439       lp_native_vector_width = 256;
 440    } else {
 441       /* Leave it at 128, even when no SIMD extensions are available.
 442        * Really needs to be a multiple of 128 so can fit 4 floats.
 443        */
 444       lp_native_vector_width = 128;
 445    }
 446
 447    lp_native_vector_width = debug_get_num_option("LP_NATIVE_VECTOR_WIDTH",
 448                                                  lp_native_vector_width);
 449
 450    if (lp_native_vector_width <= 128) {
 451       /* Hide AVX support, as often LLVM AVX intrinsics are only guarded by
 452        * "util_cpu_caps.has_avx" predicate, and lack the
 453        * "lp_native_vector_width > 128" predicate. And also to ensure a more
 454        * consistent behavior, allowing one to test SSE2 on AVX machines.
 455        * XXX: should not play games with util_cpu_caps directly as it might
 456        * get used for other things outside llvm too.
 457        */
 458       util_cpu_caps.has_avx = 0;
 459       util_cpu_caps.has_avx2 = 0;
 460       util_cpu_caps.has_f16c = 0;
 461       util_cpu_caps.has_fma = 0;
 462    }
 463    if (HAVE_LLVM < 0x0304 || !use_mcjit) {
 464       /* AVX2 support has only been tested with LLVM 3.4, and it requires
 465        * MCJIT. */
 466       util_cpu_caps.has_avx2 = 0;
 467    }
 468
 469 #ifdef PIPE_ARCH_PPC_64
 470    /* Set the NJ bit in VSCR to 0 so denormalized values are handled as
 471     * specified by IEEE standard (PowerISA 2.06 - Section 6.3). This guarantees
 472     * that some rounding and half-float to float handling does not round
 473     * incorrectly to 0.
 474     * XXX: should eventually follow same logic on all platforms.
 475     * Right now denorms get explicitly disabled (but elsewhere) for x86,
 476     * whereas ppc64 explicitly enables them...
 477     */
 478    if (util_cpu_caps.has_altivec) {
 479       unsigned short mask[] = { 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF,
 480                                 0xFFFF, 0xFFFF, 0xFFFE, 0xFFFF };
 481       __asm (
 482         "mfvscr %%v1\n"
 483         "vand   %0,%%v1,%0\n"
 484         "mtvscr %0"
 485         :
 486         : "r" (*mask)
 487       );
 488    }
 489 #endif
 490
 491    gallivm_initialized = TRUE;
 492
 493    return TRUE;
 494 }
 495
 496
 497
 498 /**
 499  * Create a new gallivm_state object.
 500  */
 501 struct gallivm_state *
 502 gallivm_create(const char *name, LLVMContextRef context)
 503 {
 504    struct gallivm_state *gallivm;
 505
 506    gallivm = CALLOC_STRUCT(gallivm_state);
 507    if (gallivm) {
 508       if (!init_gallivm_state(gallivm, name, context)) {
 509          FREE(gallivm);
 510          gallivm = NULL;
 511       }
 512    }
 513
 514    return gallivm;
 515 }
 516
 517
 518 /**
 519  * Destroy a gallivm_state object.
 520  */
 521 void
 522 gallivm_destroy(struct gallivm_state *gallivm)
 523 {
 524    gallivm_free_ir(gallivm);
 525    gallivm_free_code(gallivm);
 526    FREE(gallivm);
 527 }
 528
 529
 530 /**
 531  * Validate a function.
 532  * Verification is only done with debug builds.
 533  */
 534 void
 535 gallivm_verify_function(struct gallivm_state *gallivm,
 536                         LLVMValueRef func)
 537 {
 538    /* Verify the LLVM IR.  If invalid, dump and abort */
 539 #ifdef DEBUG
 540    if (LLVMVerifyFunction(func, LLVMPrintMessageAction)) {
 541       lp_debug_dump_value(func);
 542       assert(0);
 543       return;
 544    }
 545 #endif
 546
 547    if (gallivm_debug & GALLIVM_DEBUG_IR) {
 548       /* Print the LLVM IR to stderr */
 549       lp_debug_dump_value(func);
 550       debug_printf("\n");
 551    }
 552 }
 553
 554
 555 /**
 556  * Compile a module.
 557  * This does IR optimization on all functions in the module.
 558  */
 559 void
 560 gallivm_compile_module(struct gallivm_state *gallivm)
 561 {
 562    LLVMValueRef func;
 563    int64_t time_begin = 0;
 564
 565    assert(!gallivm->compiled);
 566
 567    if (gallivm->builder) {
 568       LLVMDisposeBuilder(gallivm->builder);
 569       gallivm->builder = NULL;
 570    }
 571
 572    if (gallivm_debug & GALLIVM_DEBUG_PERF)
 573       time_begin = os_time_get();
 574
 575    /* Run optimization passes */
 576    LLVMInitializeFunctionPassManager(gallivm->passmgr);
 577    func = LLVMGetFirstFunction(gallivm->module);
 578    while (func) {
 579       if (0) {
 580          debug_printf("optimizing func %s...\n", LLVMGetValueName(func));
 581       }
 582
 583    /* Disable frame pointer omission on debug/profile builds */
 584    /* XXX: And workaround http://llvm.org/PR21435 */
 585 #if HAVE_LLVM >= 0x0307 && \
 586     (defined(DEBUG) || defined(PROFILE) || \
 587      defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64))
 588       LLVMAddTargetDependentFunctionAttr(func, "no-frame-pointer-elim", "true");
 589       LLVMAddTargetDependentFunctionAttr(func, "no-frame-pointer-elim-non-leaf", "true");
 590 #endif
 591
 592       LLVMRunFunctionPassManager(gallivm->passmgr, func);
 593       func = LLVMGetNextFunction(func);
 594    }
 595    LLVMFinalizeFunctionPassManager(gallivm->passmgr);
 596
 597    if (gallivm_debug & GALLIVM_DEBUG_PERF) {
 598       int64_t time_end = os_time_get();
 599       int time_msec = (int)(time_end - time_begin) / 1000;
 600       assert(gallivm->module_name);
 601       debug_printf("optimizing module %s took %d msec\n",
 602                    gallivm->module_name, time_msec);
 603    }
 604
 605    /* Dump byte code to a file */
 606    if (gallivm_debug & GALLIVM_DEBUG_DUMP_BC) {
 607       char filename[256];
 608       assert(gallivm->module_name);
 609       util_snprintf(filename, sizeof(filename), "ir_%s.bc", gallivm->module_name);
 610       LLVMWriteBitcodeToFile(gallivm->module, filename);
 611       debug_printf("%s written\n", filename);
 612       debug_printf("Invoke as \"llc %s%s -o - %s\"\n",
 613                    (HAVE_LLVM >= 0x0305) ? "[-mcpu=<-mcpu option>] " : "",
 614                    "[-mattr=<-mattr option(s)>]",
 615                    filename);
 616    }
 617
 618    if (use_mcjit) {
 619       /* Setting the module's DataLayout to an empty string will cause the
 620        * ExecutionEngine to copy to the DataLayout string from its target
 621        * machine to the module.  As of LLVM 3.8 the module and the execution
 622        * engine are required to have the same DataLayout.
 623        *
 624        * We must make sure we do this after running the optimization passes,
 625        * because those passes need a correct datalayout string.  For example,
 626        * if those optimization passes see an empty datalayout, they will assume
 627        * this is a little endian target and will do optimizations that break big
 628        * endian machines.
 629        *
 630        * TODO: This is just a temporary work-around.  The correct solution is
 631        * for gallivm_init_state() to create a TargetMachine and pull the
 632        * DataLayout from there.  Currently, the TargetMachine used by llvmpipe
 633        * is being implicitly created by the EngineBuilder in
 634        * lp_build_create_jit_compiler_for_module()
 635        */
 636       LLVMSetDataLayout(gallivm->module, "");
 637       assert(!gallivm->engine);
 638       if (!init_gallivm_engine(gallivm)) {
 639          assert(0);
 640       }
 641    }
 642    assert(gallivm->engine);
 643
 644    ++gallivm->compiled;
 645
 646    if (gallivm_debug & GALLIVM_DEBUG_ASM) {
 647       LLVMValueRef llvm_func = LLVMGetFirstFunction(gallivm->module);
 648
 649       while (llvm_func) {
 650          /*
 651           * Need to filter out functions which don't have an implementation,
 652           * such as the intrinsics. May not be sufficient in case of IPO?
 653           * LLVMGetPointerToGlobal() will abort otherwise.
 654           */
 655          if (!LLVMIsDeclaration(llvm_func)) {
 656             void *func_code = LLVMGetPointerToGlobal(gallivm->engine, llvm_func);
 657             lp_disassemble(llvm_func, func_code);
 658          }
 659          llvm_func = LLVMGetNextFunction(llvm_func);
 660       }
 661    }
 662
 663 #if defined(PROFILE)
 664    {
 665       LLVMValueRef llvm_func = LLVMGetFirstFunction(gallivm->module);
 666
 667       while (llvm_func) {
 668          if (!LLVMIsDeclaration(llvm_func)) {
 669             void *func_code = LLVMGetPointerToGlobal(gallivm->engine, llvm_func);
 670             lp_profile(llvm_func, func_code);
 671          }
 672          llvm_func = LLVMGetNextFunction(llvm_func);
 673       }
 674    }
 675 #endif
 676 }
 677
 678
 679
 680 func_pointer
 681 gallivm_jit_function(struct gallivm_state *gallivm,
 682                      LLVMValueRef func)
 683 {
 684    void *code;
 685    func_pointer jit_func;
 686    int64_t time_begin = 0;
 687
 688    assert(gallivm->compiled);
 689    assert(gallivm->engine);
 690
 691    if (gallivm_debug & GALLIVM_DEBUG_PERF)
 692       time_begin = os_time_get();
 693
 694    code = LLVMGetPointerToGlobal(gallivm->engine, func);
 695    assert(code);
 696    jit_func = pointer_to_func(code);
 697
 698    if (gallivm_debug & GALLIVM_DEBUG_PERF) {
 699       int64_t time_end = os_time_get();
 700       int time_msec = (int)(time_end - time_begin) / 1000;
 701       debug_printf("   jitting func %s took %d msec\n",
 702                    LLVMGetValueName(func), time_msec);
 703    }
 704
 705    return jit_func;
 706 }