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