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/Config/llvm-config.h>
  42 #include <llvm-c/Analysis.h>
  43 #include <llvm-c/Transforms/Scalar.h>
  44 #if LLVM_VERSION_MAJOR >= 7
  45 #include <llvm-c/Transforms/Utils.h>
  46 #endif
  47 #include <llvm-c/BitWriter.h>
  48 #if GALLIVM_HAVE_CORO
  49 #if LLVM_VERSION_MAJOR <= 8 && (defined(PIPE_ARCH_AARCH64) || defined (PIPE_ARCH_ARM) || defined(PIPE_ARCH_S390))
  50 #include <llvm-c/Transforms/IPO.h>
  51 #endif
  52 #include <llvm-c/Transforms/Coroutines.h>
  53 #endif
  54
  55 unsigned gallivm_perf = 0;
  56
  57 static const struct debug_named_value lp_bld_perf_flags[] = {
  58    { "no_brilinear", GALLIVM_PERF_NO_BRILINEAR, "disable brilinear optimization" },
  59    { "no_rho_approx", GALLIVM_PERF_NO_RHO_APPROX, "disable rho_approx optimization" },
  60    { "no_quad_lod", GALLIVM_PERF_NO_QUAD_LOD, "disable quad_lod optimization" },
  61    { "no_aos_sampling", GALLIVM_PERF_NO_AOS_SAMPLING, "disable aos sampling optimization" },
  62    { "nopt",   GALLIVM_PERF_NO_OPT, "disable optimization passes to speed up shader compilation" },
  63    { "no_filter_hacks", GALLIVM_PERF_NO_BRILINEAR | GALLIVM_PERF_NO_RHO_APPROX |
  64      GALLIVM_PERF_NO_QUAD_LOD, "disable filter optimization hacks" },
  65    DEBUG_NAMED_VALUE_END
  66 };
  67
  68 #ifdef DEBUG
  69 unsigned gallivm_debug = 0;
  70
  71 static const struct debug_named_value lp_bld_debug_flags[] = {
  72    { "tgsi",   GALLIVM_DEBUG_TGSI, NULL },
  73    { "ir",     GALLIVM_DEBUG_IR, NULL },
  74    { "asm",    GALLIVM_DEBUG_ASM, NULL },
  75    { "perf",   GALLIVM_DEBUG_PERF, NULL },
  76    { "gc",     GALLIVM_DEBUG_GC, NULL },
  77    { "dumpbc", GALLIVM_DEBUG_DUMP_BC, NULL },
  78    DEBUG_NAMED_VALUE_END
  79 };
  80
  81 DEBUG_GET_ONCE_FLAGS_OPTION(gallivm_debug, "GALLIVM_DEBUG", lp_bld_debug_flags, 0)
  82 #endif
  83
  84
  85 static boolean gallivm_initialized = FALSE;
  86
  87 unsigned lp_native_vector_width;
  88
  89
  90 /*
  91  * Optimization values are:
  92  * - 0: None (-O0)
  93  * - 1: Less (-O1)
  94  * - 2: Default (-O2, -Os)
  95  * - 3: Aggressive (-O3)
  96  *
  97  * See also CodeGenOpt::Level in llvm/Target/TargetMachine.h
  98  */
  99 enum LLVM_CodeGenOpt_Level {
 100    None,        // -O0
 101    Less,        // -O1
 102    Default,     // -O2, -Os
 103    Aggressive   // -O3
 104 };
 105
 106
 107 /**
 108  * Create the LLVM (optimization) pass manager and install
 109  * relevant optimization passes.
 110  * \return  TRUE for success, FALSE for failure
 111  */
 112 static boolean
 113 create_pass_manager(struct gallivm_state *gallivm)
 114 {
 115    assert(!gallivm->passmgr);
 116    assert(gallivm->target);
 117
 118    gallivm->passmgr = LLVMCreateFunctionPassManagerForModule(gallivm->module);
 119    if (!gallivm->passmgr)
 120       return FALSE;
 121
 122 #if GALLIVM_HAVE_CORO
 123    gallivm->cgpassmgr = LLVMCreatePassManager();
 124 #endif
 125    /*
 126     * TODO: some per module pass manager with IPO passes might be helpful -
 127     * the generated texture functions may benefit from inlining if they are
 128     * simple, or constant propagation into them, etc.
 129     */
 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_HAVE_CORO
 140 #if LLVM_VERSION_MAJOR <= 8 && (defined(PIPE_ARCH_AARCH64) || defined (PIPE_ARCH_ARM) || defined(PIPE_ARCH_S390))
 141    LLVMAddArgumentPromotionPass(gallivm->cgpassmgr);
 142    LLVMAddFunctionAttrsPass(gallivm->cgpassmgr);
 143 #endif
 144    LLVMAddCoroEarlyPass(gallivm->cgpassmgr);
 145    LLVMAddCoroSplitPass(gallivm->cgpassmgr);
 146    LLVMAddCoroElidePass(gallivm->cgpassmgr);
 147 #endif
 148
 149    if ((gallivm_perf & GALLIVM_PERF_NO_OPT) == 0) {
 150       /*
 151        * TODO: Evaluate passes some more - keeping in mind
 152        * both quality of generated code and compile times.
 153        */
 154       /*
 155        * NOTE: if you change this, don't forget to change the output
 156        * with GALLIVM_DEBUG_DUMP_BC in gallivm_compile_module.
 157        */
 158       LLVMAddScalarReplAggregatesPass(gallivm->passmgr);
 159       LLVMAddEarlyCSEPass(gallivm->passmgr);
 160       LLVMAddCFGSimplificationPass(gallivm->passmgr);
 161       /*
 162        * FIXME: LICM is potentially quite useful. However, for some
 163        * rather crazy shaders the compile time can reach _hours_ per shader,
 164        * due to licm implying lcssa (since llvm 3.5), which can take forever.
 165        * Even for sane shaders, the cost of licm is rather high (and not just
 166        * due to lcssa, licm itself too), though mostly only in cases when it
 167        * can actually move things, so having to disable it is a pity.
 168        * LLVMAddLICMPass(gallivm->passmgr);
 169        */
 170       LLVMAddReassociatePass(gallivm->passmgr);
 171       LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
 172 #if LLVM_VERSION_MAJOR <= 11
 173       LLVMAddConstantPropagationPass(gallivm->passmgr);
 174 #else
 175       LLVMAddInstructionSimplifyPass(gallivm->passmgr);
 176 #endif
 177       LLVMAddInstructionCombiningPass(gallivm->passmgr);
 178       LLVMAddGVNPass(gallivm->passmgr);
 179    }
 180    else {
 181       /* We need at least this pass to prevent the backends to fail in
 182        * unexpected ways.
 183        */
 184       LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
 185    }
 186 #if GALLIVM_HAVE_CORO
 187    LLVMAddCoroCleanupPass(gallivm->passmgr);
 188 #endif
 189
 190    return TRUE;
 191 }
 192
 193
 194 /**
 195  * Free gallivm object's LLVM allocations, but not any generated code
 196  * nor the gallivm object itself.
 197  */
 198 void
 199 gallivm_free_ir(struct gallivm_state *gallivm)
 200 {
 201    if (gallivm->passmgr) {
 202       LLVMDisposePassManager(gallivm->passmgr);
 203    }
 204
 205 #if GALLIVM_HAVE_CORO
 206    if (gallivm->cgpassmgr) {
 207       LLVMDisposePassManager(gallivm->cgpassmgr);
 208    }
 209 #endif
 210
 211    if (gallivm->engine) {
 212       /* This will already destroy any associated module */
 213       LLVMDisposeExecutionEngine(gallivm->engine);
 214    } else if (gallivm->module) {
 215       LLVMDisposeModule(gallivm->module);
 216    }
 217
 218    if (gallivm->cache) {
 219       lp_free_objcache(gallivm->cache->jit_obj_cache);
 220       free(gallivm->cache->data);
 221    }
 222    FREE(gallivm->module_name);
 223
 224    if (gallivm->target) {
 225       LLVMDisposeTargetData(gallivm->target);
 226    }
 227
 228    if (gallivm->builder)
 229       LLVMDisposeBuilder(gallivm->builder);
 230
 231    /* The LLVMContext should be owned by the parent of gallivm. */
 232
 233    gallivm->engine = NULL;
 234    gallivm->target = NULL;
 235    gallivm->module = NULL;
 236    gallivm->module_name = NULL;
 237    gallivm->cgpassmgr = NULL;
 238    gallivm->passmgr = NULL;
 239    gallivm->context = NULL;
 240    gallivm->builder = NULL;
 241    gallivm->cache = NULL;
 242 }
 243
 244
 245 /**
 246  * Free LLVM-generated code.  Should be done AFTER gallivm_free_ir().
 247  */
 248 static void
 249 gallivm_free_code(struct gallivm_state *gallivm)
 250 {
 251    assert(!gallivm->module);
 252    assert(!gallivm->engine);
 253    lp_free_generated_code(gallivm->code);
 254    gallivm->code = NULL;
 255    lp_free_memory_manager(gallivm->memorymgr);
 256    gallivm->memorymgr = NULL;
 257 }
 258
 259
 260 static boolean
 261 init_gallivm_engine(struct gallivm_state *gallivm)
 262 {
 263    if (1) {
 264       enum LLVM_CodeGenOpt_Level optlevel;
 265       char *error = NULL;
 266       int ret;
 267
 268       if (gallivm_perf & GALLIVM_PERF_NO_OPT) {
 269          optlevel = None;
 270       }
 271       else {
 272          optlevel = Default;
 273       }
 274
 275       ret = lp_build_create_jit_compiler_for_module(&gallivm->engine,
 276                                                     &gallivm->code,
 277                                                     gallivm->cache,
 278                                                     gallivm->module,
 279                                                     gallivm->memorymgr,
 280                                                     (unsigned) optlevel,
 281                                                     &error);
 282       if (ret) {
 283          _debug_printf("%s\n", error);
 284          LLVMDisposeMessage(error);
 285          goto fail;
 286       }
 287    }
 288
 289    if (0) {
 290        /*
 291         * Dump the data layout strings.
 292         */
 293
 294        LLVMTargetDataRef target = LLVMGetExecutionEngineTargetData(gallivm->engine);
 295        char *data_layout;
 296        char *engine_data_layout;
 297
 298        data_layout = LLVMCopyStringRepOfTargetData(gallivm->target);
 299        engine_data_layout = LLVMCopyStringRepOfTargetData(target);
 300
 301        if (1) {
 302           debug_printf("module target data = %s\n", data_layout);
 303           debug_printf("engine target data = %s\n", engine_data_layout);
 304        }
 305
 306        free(data_layout);
 307        free(engine_data_layout);
 308    }
 309
 310    return TRUE;
 311
 312 fail:
 313    return FALSE;
 314 }
 315
 316
 317 /**
 318  * Allocate gallivm LLVM objects.
 319  * \return  TRUE for success, FALSE for failure
 320  */
 321 static boolean
 322 init_gallivm_state(struct gallivm_state *gallivm, const char *name,
 323                    LLVMContextRef context, struct lp_cached_code *cache)
 324 {
 325    assert(!gallivm->context);
 326    assert(!gallivm->module);
 327
 328    if (!lp_build_init())
 329       return FALSE;
 330
 331    gallivm->context = context;
 332    gallivm->cache = cache;
 333    if (!gallivm->context)
 334       goto fail;
 335
 336    gallivm->module_name = NULL;
 337    if (name) {
 338       size_t size = strlen(name) + 1;
 339       gallivm->module_name = MALLOC(size);
 340       if (gallivm->module_name) {
 341          memcpy(gallivm->module_name, name, size);
 342       }
 343    }
 344
 345    gallivm->module = LLVMModuleCreateWithNameInContext(name,
 346                                                        gallivm->context);
 347    if (!gallivm->module)
 348       goto fail;
 349
 350    gallivm->builder = LLVMCreateBuilderInContext(gallivm->context);
 351    if (!gallivm->builder)
 352       goto fail;
 353
 354    gallivm->memorymgr = lp_get_default_memory_manager();
 355    if (!gallivm->memorymgr)
 356       goto fail;
 357
 358    /* FIXME: MC-JIT only allows compiling one module at a time, and it must be
 359     * complete when MC-JIT is created. So defer the MC-JIT engine creation for
 360     * now.
 361     */
 362
 363    /*
 364     * MC-JIT engine compiles the module immediately on creation, so we can't
 365     * obtain the target data from it.  Instead we create a target data layout
 366     * from a string.
 367     *
 368     * The produced layout strings are not precisely the same, but should make
 369     * no difference for the kind of optimization passes we run.
 370     *
 371     * For reference this is the layout string on x64:
 372     *
 373     *   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
 374     *
 375     * See also:
 376     * - http://llvm.org/docs/LangRef.html#datalayout
 377     */
 378
 379    {
 380       const unsigned pointer_size = 8 * sizeof(void *);
 381       char layout[512];
 382       snprintf(layout, sizeof layout, "%c-p:%u:%u:%u-i64:64:64-a0:0:%u-s0:%u:%u",
 383 #if UTIL_ARCH_LITTLE_ENDIAN
 384                     'e', // little endian
 385 #else
 386                     'E', // big endian
 387 #endif
 388                     pointer_size, pointer_size, pointer_size, // pointer size, abi alignment, preferred alignment
 389                     pointer_size, // aggregate preferred alignment
 390                     pointer_size, pointer_size); // stack objects abi alignment, preferred alignment
 391
 392       gallivm->target = LLVMCreateTargetData(layout);
 393       if (!gallivm->target) {
 394          return FALSE;
 395       }
 396    }
 397
 398    if (!create_pass_manager(gallivm))
 399       goto fail;
 400
 401    return TRUE;
 402
 403 fail:
 404    gallivm_free_ir(gallivm);
 405    gallivm_free_code(gallivm);
 406    return FALSE;
 407 }
 408
 409
 410 boolean
 411 lp_build_init(void)
 412 {
 413    if (gallivm_initialized)
 414       return TRUE;
 415
 416
 417    /* LLVMLinkIn* are no-ops at runtime.  They just ensure the respective
 418     * component is linked at buildtime, which is sufficient for its static
 419     * constructors to be called at load time.
 420     */
 421    LLVMLinkInMCJIT();
 422
 423 #ifdef DEBUG
 424    gallivm_debug = debug_get_option_gallivm_debug();
 425 #endif
 426
 427    gallivm_perf = debug_get_flags_option("GALLIVM_PERF", lp_bld_perf_flags, 0 );
 428
 429    lp_set_target_options();
 430
 431    util_cpu_detect();
 432
 433    /* For simulating less capable machines */
 434 #ifdef DEBUG
 435    if (debug_get_bool_option("LP_FORCE_SSE2", FALSE)) {
 436       assert(util_cpu_caps.has_sse2);
 437       util_cpu_caps.has_sse3 = 0;
 438       util_cpu_caps.has_ssse3 = 0;
 439       util_cpu_caps.has_sse4_1 = 0;
 440       util_cpu_caps.has_sse4_2 = 0;
 441       util_cpu_caps.has_avx = 0;
 442       util_cpu_caps.has_avx2 = 0;
 443       util_cpu_caps.has_f16c = 0;
 444       util_cpu_caps.has_fma = 0;
 445    }
 446 #endif
 447
 448    if (util_cpu_caps.has_avx2 || util_cpu_caps.has_avx) {
 449       lp_native_vector_width = 256;
 450    } else {
 451       /* Leave it at 128, even when no SIMD extensions are available.
 452        * Really needs to be a multiple of 128 so can fit 4 floats.
 453        */
 454       lp_native_vector_width = 128;
 455    }
 456
 457    lp_native_vector_width = debug_get_num_option("LP_NATIVE_VECTOR_WIDTH",
 458                                                  lp_native_vector_width);
 459
 460 #if LLVM_VERSION_MAJOR < 4
 461    if (lp_native_vector_width <= 128) {
 462       /* Hide AVX support, as often LLVM AVX intrinsics are only guarded by
 463        * "util_cpu_caps.has_avx" predicate, and lack the
 464        * "lp_native_vector_width > 128" predicate. And also to ensure a more
 465        * consistent behavior, allowing one to test SSE2 on AVX machines.
 466        * XXX: should not play games with util_cpu_caps directly as it might
 467        * get used for other things outside llvm too.
 468        */
 469       util_cpu_caps.has_avx = 0;
 470       util_cpu_caps.has_avx2 = 0;
 471       util_cpu_caps.has_f16c = 0;
 472       util_cpu_caps.has_fma = 0;
 473    }
 474 #endif
 475
 476 #ifdef PIPE_ARCH_PPC_64
 477    /* Set the NJ bit in VSCR to 0 so denormalized values are handled as
 478     * specified by IEEE standard (PowerISA 2.06 - Section 6.3). This guarantees
 479     * that some rounding and half-float to float handling does not round
 480     * incorrectly to 0.
 481     * XXX: should eventually follow same logic on all platforms.
 482     * Right now denorms get explicitly disabled (but elsewhere) for x86,
 483     * whereas ppc64 explicitly enables them...
 484     */
 485    if (util_cpu_caps.has_altivec) {
 486       unsigned short mask[] = { 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF,
 487                                 0xFFFF, 0xFFFF, 0xFFFE, 0xFFFF };
 488       __asm (
 489         "mfvscr %%v1\n"
 490         "vand   %0,%%v1,%0\n"
 491         "mtvscr %0"
 492         :
 493         : "r" (*mask)
 494       );
 495    }
 496 #endif
 497
 498    gallivm_initialized = TRUE;
 499
 500    return TRUE;
 501 }
 502
 503
 504
 505 /**
 506  * Create a new gallivm_state object.
 507  */
 508 struct gallivm_state *
 509 gallivm_create(const char *name, LLVMContextRef context,
 510                struct lp_cached_code *cache)
 511 {
 512    struct gallivm_state *gallivm;
 513
 514    gallivm = CALLOC_STRUCT(gallivm_state);
 515    if (gallivm) {
 516       if (!init_gallivm_state(gallivm, name, context, cache)) {
 517          FREE(gallivm);
 518          gallivm = NULL;
 519       }
 520    }
 521
 522    assert(gallivm != NULL);
 523    return gallivm;
 524 }
 525
 526
 527 /**
 528  * Destroy a gallivm_state object.
 529  */
 530 void
 531 gallivm_destroy(struct gallivm_state *gallivm)
 532 {
 533    gallivm_free_ir(gallivm);
 534    gallivm_free_code(gallivm);
 535    FREE(gallivm);
 536 }
 537
 538
 539 /**
 540  * Validate a function.
 541  * Verification is only done with debug builds.
 542  */
 543 void
 544 gallivm_verify_function(struct gallivm_state *gallivm,
 545                         LLVMValueRef func)
 546 {
 547    /* Verify the LLVM IR.  If invalid, dump and abort */
 548 #ifdef DEBUG
 549    if (LLVMVerifyFunction(func, LLVMPrintMessageAction)) {
 550       lp_debug_dump_value(func);
 551       assert(0);
 552       return;
 553    }
 554 #endif
 555
 556    if (gallivm_debug & GALLIVM_DEBUG_IR) {
 557       /* Print the LLVM IR to stderr */
 558       lp_debug_dump_value(func);
 559       debug_printf("\n");
 560    }
 561 }
 562
 563
 564 /**
 565  * Compile a module.
 566  * This does IR optimization on all functions in the module.
 567  */
 568 void
 569 gallivm_compile_module(struct gallivm_state *gallivm)
 570 {
 571    LLVMValueRef func;
 572    int64_t time_begin = 0;
 573
 574    assert(!gallivm->compiled);
 575
 576    if (gallivm->builder) {
 577       LLVMDisposeBuilder(gallivm->builder);
 578       gallivm->builder = NULL;
 579    }
 580
 581    if (gallivm->cache && gallivm->cache->data_size) {
 582       goto skip_cached;
 583    }
 584
 585    /* Dump bitcode to a file */
 586    if (gallivm_debug & GALLIVM_DEBUG_DUMP_BC) {
 587       char filename[256];
 588       assert(gallivm->module_name);
 589       snprintf(filename, sizeof(filename), "ir_%s.bc", gallivm->module_name);
 590       LLVMWriteBitcodeToFile(gallivm->module, filename);
 591       debug_printf("%s written\n", filename);
 592       debug_printf("Invoke as \"opt %s %s | llc -O%d %s%s\"\n",
 593                    gallivm_debug & GALLIVM_PERF_NO_OPT ? "-mem2reg" :
 594                    "-sroa -early-cse -simplifycfg -reassociate "
 595                    "-mem2reg -constprop -instcombine -gvn",
 596                    filename, gallivm_debug & GALLIVM_PERF_NO_OPT ? 0 : 2,
 597                    "[-mcpu=<-mcpu option>] ",
 598                    "[-mattr=<-mattr option(s)>]");
 599    }
 600
 601    if (gallivm_debug & GALLIVM_DEBUG_PERF)
 602       time_begin = os_time_get();
 603
 604 #if GALLIVM_HAVE_CORO
 605    LLVMRunPassManager(gallivm->cgpassmgr, gallivm->module);
 606 #endif
 607    /* Run optimization passes */
 608    LLVMInitializeFunctionPassManager(gallivm->passmgr);
 609    func = LLVMGetFirstFunction(gallivm->module);
 610    while (func) {
 611       if (0) {
 612          debug_printf("optimizing func %s...\n", LLVMGetValueName(func));
 613       }
 614
 615    /* Disable frame pointer omission on debug/profile builds */
 616    /* XXX: And workaround http://llvm.org/PR21435 */
 617 #if defined(DEBUG) || defined(PROFILE) || defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
 618       LLVMAddTargetDependentFunctionAttr(func, "no-frame-pointer-elim", "true");
 619       LLVMAddTargetDependentFunctionAttr(func, "no-frame-pointer-elim-non-leaf", "true");
 620 #endif
 621
 622       LLVMRunFunctionPassManager(gallivm->passmgr, func);
 623       func = LLVMGetNextFunction(func);
 624    }
 625    LLVMFinalizeFunctionPassManager(gallivm->passmgr);
 626
 627    if (gallivm_debug & GALLIVM_DEBUG_PERF) {
 628       int64_t time_end = os_time_get();
 629       int time_msec = (int)((time_end - time_begin) / 1000);
 630       assert(gallivm->module_name);
 631       debug_printf("optimizing module %s took %d msec\n",
 632                    gallivm->module_name, time_msec);
 633    }
 634
 635    /* Setting the module's DataLayout to an empty string will cause the
 636     * ExecutionEngine to copy to the DataLayout string from its target machine
 637     * to the module.  As of LLVM 3.8 the module and the execution engine are
 638     * required to have the same DataLayout.
 639     *
 640     * We must make sure we do this after running the optimization passes,
 641     * because those passes need a correct datalayout string.  For example, if
 642     * those optimization passes see an empty datalayout, they will assume this
 643     * is a little endian target and will do optimizations that break big endian
 644     * machines.
 645     *
 646     * TODO: This is just a temporary work-around.  The correct solution is for
 647     * gallivm_init_state() to create a TargetMachine and pull the DataLayout
 648     * from there.  Currently, the TargetMachine used by llvmpipe is being
 649     * implicitly created by the EngineBuilder in
 650     * lp_build_create_jit_compiler_for_module()
 651     */
 652  skip_cached:
 653    LLVMSetDataLayout(gallivm->module, "");
 654    assert(!gallivm->engine);
 655    if (!init_gallivm_engine(gallivm)) {
 656       assert(0);
 657    }
 658    assert(gallivm->engine);
 659
 660    ++gallivm->compiled;
 661
 662    if (gallivm->debug_printf_hook)
 663       LLVMAddGlobalMapping(gallivm->engine, gallivm->debug_printf_hook, debug_printf);
 664
 665    if (gallivm_debug & GALLIVM_DEBUG_ASM) {
 666       LLVMValueRef llvm_func = LLVMGetFirstFunction(gallivm->module);
 667
 668       while (llvm_func) {
 669          /*
 670           * Need to filter out functions which don't have an implementation,
 671           * such as the intrinsics. May not be sufficient in case of IPO?
 672           * LLVMGetPointerToGlobal() will abort otherwise.
 673           */
 674          if (!LLVMIsDeclaration(llvm_func)) {
 675             void *func_code = LLVMGetPointerToGlobal(gallivm->engine, llvm_func);
 676             lp_disassemble(llvm_func, func_code);
 677          }
 678          llvm_func = LLVMGetNextFunction(llvm_func);
 679       }
 680    }
 681
 682 #if defined(PROFILE)
 683    {
 684       LLVMValueRef llvm_func = LLVMGetFirstFunction(gallivm->module);
 685
 686       while (llvm_func) {
 687          if (!LLVMIsDeclaration(llvm_func)) {
 688             void *func_code = LLVMGetPointerToGlobal(gallivm->engine, llvm_func);
 689             lp_profile(llvm_func, func_code);
 690          }
 691          llvm_func = LLVMGetNextFunction(llvm_func);
 692       }
 693    }
 694 #endif
 695 }
 696
 697
 698
 699 func_pointer
 700 gallivm_jit_function(struct gallivm_state *gallivm,
 701                      LLVMValueRef func)
 702 {
 703    void *code;
 704    func_pointer jit_func;
 705    int64_t time_begin = 0;
 706
 707    assert(gallivm->compiled);
 708    assert(gallivm->engine);
 709
 710    if (gallivm_debug & GALLIVM_DEBUG_PERF)
 711       time_begin = os_time_get();
 712
 713    code = LLVMGetPointerToGlobal(gallivm->engine, func);
 714    assert(code);
 715    jit_func = pointer_to_func(code);
 716
 717    if (gallivm_debug & GALLIVM_DEBUG_PERF) {
 718       int64_t time_end = os_time_get();
 719       int time_msec = (int)(time_end - time_begin) / 1000;
 720       debug_printf("   jitting func %s took %d msec\n",
 721                    LLVMGetValueName(func), time_msec);
 722    }
 723
 724    return jit_func;
 725 }