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)
  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)
 141    LLVMAddFunctionAttrsPass(gallivm->cgpassmgr);
 142 #endif
 143    LLVMAddCoroEarlyPass(gallivm->cgpassmgr);
 144    LLVMAddCoroSplitPass(gallivm->cgpassmgr);
 145    LLVMAddCoroElidePass(gallivm->cgpassmgr);
 146 #endif
 147
 148    if ((gallivm_perf & GALLIVM_PERF_NO_OPT) == 0) {
 149       /*
 150        * TODO: Evaluate passes some more - keeping in mind
 151        * both quality of generated code and compile times.
 152        */
 153       /*
 154        * NOTE: if you change this, don't forget to change the output
 155        * with GALLIVM_DEBUG_DUMP_BC in gallivm_compile_module.
 156        */
 157       LLVMAddScalarReplAggregatesPass(gallivm->passmgr);
 158       LLVMAddEarlyCSEPass(gallivm->passmgr);
 159       LLVMAddCFGSimplificationPass(gallivm->passmgr);
 160       /*
 161        * FIXME: LICM is potentially quite useful. However, for some
 162        * rather crazy shaders the compile time can reach _hours_ per shader,
 163        * due to licm implying lcssa (since llvm 3.5), which can take forever.
 164        * Even for sane shaders, the cost of licm is rather high (and not just
 165        * due to lcssa, licm itself too), though mostly only in cases when it
 166        * can actually move things, so having to disable it is a pity.
 167        * LLVMAddLICMPass(gallivm->passmgr);
 168        */
 169       LLVMAddReassociatePass(gallivm->passmgr);
 170       LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
 171       LLVMAddConstantPropagationPass(gallivm->passmgr);
 172       LLVMAddInstructionCombiningPass(gallivm->passmgr);
 173       LLVMAddGVNPass(gallivm->passmgr);
 174 #if GALLIVM_HAVE_CORO
 175       LLVMAddCoroCleanupPass(gallivm->passmgr);
 176 #endif
 177    }
 178    else {
 179       /* We need at least this pass to prevent the backends to fail in
 180        * unexpected ways.
 181        */
 182       LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
 183    }
 184
 185    return TRUE;
 186 }
 187
 188
 189 /**
 190  * Free gallivm object's LLVM allocations, but not any generated code
 191  * nor the gallivm object itself.
 192  */
 193 void
 194 gallivm_free_ir(struct gallivm_state *gallivm)
 195 {
 196    if (gallivm->passmgr) {
 197       LLVMDisposePassManager(gallivm->passmgr);
 198    }
 199
 200 #if GALLIVM_HAVE_CORO
 201    if (gallivm->cgpassmgr) {
 202       LLVMDisposePassManager(gallivm->cgpassmgr);
 203    }
 204 #endif
 205
 206    if (gallivm->engine) {
 207       /* This will already destroy any associated module */
 208       LLVMDisposeExecutionEngine(gallivm->engine);
 209    } else if (gallivm->module) {
 210       LLVMDisposeModule(gallivm->module);
 211    }
 212
 213    FREE(gallivm->module_name);
 214
 215    if (gallivm->target) {
 216       LLVMDisposeTargetData(gallivm->target);
 217    }
 218
 219    if (gallivm->builder)
 220       LLVMDisposeBuilder(gallivm->builder);
 221
 222    /* The LLVMContext should be owned by the parent of gallivm. */
 223
 224    gallivm->engine = NULL;
 225    gallivm->target = NULL;
 226    gallivm->module = NULL;
 227    gallivm->module_name = NULL;
 228    gallivm->cgpassmgr = NULL;
 229    gallivm->passmgr = NULL;
 230    gallivm->context = NULL;
 231    gallivm->builder = NULL;
 232 }
 233
 234
 235 /**
 236  * Free LLVM-generated code.  Should be done AFTER gallivm_free_ir().
 237  */
 238 static void
 239 gallivm_free_code(struct gallivm_state *gallivm)
 240 {
 241    assert(!gallivm->module);
 242    assert(!gallivm->engine);
 243    lp_free_generated_code(gallivm->code);
 244    gallivm->code = NULL;
 245    lp_free_memory_manager(gallivm->memorymgr);
 246    gallivm->memorymgr = NULL;
 247 }
 248
 249
 250 static boolean
 251 init_gallivm_engine(struct gallivm_state *gallivm)
 252 {
 253    if (1) {
 254       enum LLVM_CodeGenOpt_Level optlevel;
 255       char *error = NULL;
 256       int ret;
 257
 258       if (gallivm_perf & GALLIVM_PERF_NO_OPT) {
 259          optlevel = None;
 260       }
 261       else {
 262          optlevel = Default;
 263       }
 264
 265       ret = lp_build_create_jit_compiler_for_module(&gallivm->engine,
 266                                                     &gallivm->code,
 267                                                     gallivm->module,
 268                                                     gallivm->memorymgr,
 269                                                     (unsigned) optlevel,
 270                                                     &error);
 271       if (ret) {
 272          _debug_printf("%s\n", error);
 273          LLVMDisposeMessage(error);
 274          goto fail;
 275       }
 276    }
 277
 278    if (0) {
 279        /*
 280         * Dump the data layout strings.
 281         */
 282
 283        LLVMTargetDataRef target = LLVMGetExecutionEngineTargetData(gallivm->engine);
 284        char *data_layout;
 285        char *engine_data_layout;
 286
 287        data_layout = LLVMCopyStringRepOfTargetData(gallivm->target);
 288        engine_data_layout = LLVMCopyStringRepOfTargetData(target);
 289
 290        if (1) {
 291           debug_printf("module target data = %s\n", data_layout);
 292           debug_printf("engine target data = %s\n", engine_data_layout);
 293        }
 294
 295        free(data_layout);
 296        free(engine_data_layout);
 297    }
 298
 299    return TRUE;
 300
 301 fail:
 302    return FALSE;
 303 }
 304
 305
 306 /**
 307  * Allocate gallivm LLVM objects.
 308  * \return  TRUE for success, FALSE for failure
 309  */
 310 static boolean
 311 init_gallivm_state(struct gallivm_state *gallivm, const char *name,
 312                    LLVMContextRef context)
 313 {
 314    assert(!gallivm->context);
 315    assert(!gallivm->module);
 316
 317    if (!lp_build_init())
 318       return FALSE;
 319
 320    gallivm->context = context;
 321
 322    if (!gallivm->context)
 323       goto fail;
 324
 325    gallivm->module_name = NULL;
 326    if (name) {
 327       size_t size = strlen(name) + 1;
 328       gallivm->module_name = MALLOC(size);
 329       if (gallivm->module_name) {
 330          memcpy(gallivm->module_name, name, size);
 331       }
 332    }
 333
 334    gallivm->module = LLVMModuleCreateWithNameInContext(name,
 335                                                        gallivm->context);
 336    if (!gallivm->module)
 337       goto fail;
 338
 339    gallivm->builder = LLVMCreateBuilderInContext(gallivm->context);
 340    if (!gallivm->builder)
 341       goto fail;
 342
 343    gallivm->memorymgr = lp_get_default_memory_manager();
 344    if (!gallivm->memorymgr)
 345       goto fail;
 346
 347    /* FIXME: MC-JIT only allows compiling one module at a time, and it must be
 348     * complete when MC-JIT is created. So defer the MC-JIT engine creation for
 349     * now.
 350     */
 351
 352    /*
 353     * MC-JIT engine compiles the module immediately on creation, so we can't
 354     * obtain the target data from it.  Instead we create a target data layout
 355     * from a string.
 356     *
 357     * The produced layout strings are not precisely the same, but should make
 358     * no difference for the kind of optimization passes we run.
 359     *
 360     * For reference this is the layout string on x64:
 361     *
 362     *   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
 363     *
 364     * See also:
 365     * - http://llvm.org/docs/LangRef.html#datalayout
 366     */
 367
 368    {
 369       const unsigned pointer_size = 8 * sizeof(void *);
 370       char layout[512];
 371       snprintf(layout, sizeof layout, "%c-p:%u:%u:%u-i64:64:64-a0:0:%u-s0:%u:%u",
 372 #if UTIL_ARCH_LITTLE_ENDIAN
 373                     'e', // little endian
 374 #else
 375                     'E', // big endian
 376 #endif
 377                     pointer_size, pointer_size, pointer_size, // pointer size, abi alignment, preferred alignment
 378                     pointer_size, // aggregate preferred alignment
 379                     pointer_size, pointer_size); // stack objects abi alignment, preferred alignment
 380
 381       gallivm->target = LLVMCreateTargetData(layout);
 382       if (!gallivm->target) {
 383          return FALSE;
 384       }
 385    }
 386
 387    if (!create_pass_manager(gallivm))
 388       goto fail;
 389
 390    return TRUE;
 391
 392 fail:
 393    gallivm_free_ir(gallivm);
 394    gallivm_free_code(gallivm);
 395    return FALSE;
 396 }
 397
 398
 399 boolean
 400 lp_build_init(void)
 401 {
 402    if (gallivm_initialized)
 403       return TRUE;
 404
 405
 406    /* LLVMLinkIn* are no-ops at runtime.  They just ensure the respective
 407     * component is linked at buildtime, which is sufficient for its static
 408     * constructors to be called at load time.
 409     */
 410    LLVMLinkInMCJIT();
 411
 412 #ifdef DEBUG
 413    gallivm_debug = debug_get_option_gallivm_debug();
 414 #endif
 415
 416    gallivm_perf = debug_get_flags_option("GALLIVM_PERF", lp_bld_perf_flags, 0 );
 417
 418    lp_set_target_options();
 419
 420    util_cpu_detect();
 421
 422    /* For simulating less capable machines */
 423 #ifdef DEBUG
 424    if (debug_get_bool_option("LP_FORCE_SSE2", FALSE)) {
 425       assert(util_cpu_caps.has_sse2);
 426       util_cpu_caps.has_sse3 = 0;
 427       util_cpu_caps.has_ssse3 = 0;
 428       util_cpu_caps.has_sse4_1 = 0;
 429       util_cpu_caps.has_sse4_2 = 0;
 430       util_cpu_caps.has_avx = 0;
 431       util_cpu_caps.has_avx2 = 0;
 432       util_cpu_caps.has_f16c = 0;
 433       util_cpu_caps.has_fma = 0;
 434    }
 435 #endif
 436
 437    /* AMD Bulldozer AVX's throughput is the same as SSE2; and because using
 438     * 8-wide vector needs more floating ops than 4-wide (due to padding), it is
 439     * actually more efficient to use 4-wide vectors on this processor.
 440     *
 441     * See also:
 442     * - http://www.anandtech.com/show/4955/the-bulldozer-review-amd-fx8150-tested/2
 443     */
 444    if (util_cpu_caps.has_avx &&
 445        util_cpu_caps.has_intel) {
 446       lp_native_vector_width = 256;
 447    } else {
 448       /* Leave it at 128, even when no SIMD extensions are available.
 449        * Really needs to be a multiple of 128 so can fit 4 floats.
 450        */
 451       lp_native_vector_width = 128;
 452    }
 453
 454    lp_native_vector_width = debug_get_num_option("LP_NATIVE_VECTOR_WIDTH",
 455                                                  lp_native_vector_width);
 456
 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
 471 #ifdef PIPE_ARCH_PPC_64
 472    /* Set the NJ bit in VSCR to 0 so denormalized values are handled as
 473     * specified by IEEE standard (PowerISA 2.06 - Section 6.3). This guarantees
 474     * that some rounding and half-float to float handling does not round
 475     * incorrectly to 0.
 476     * XXX: should eventually follow same logic on all platforms.
 477     * Right now denorms get explicitly disabled (but elsewhere) for x86,
 478     * whereas ppc64 explicitly enables them...
 479     */
 480    if (util_cpu_caps.has_altivec) {
 481       unsigned short mask[] = { 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF,
 482                                 0xFFFF, 0xFFFF, 0xFFFE, 0xFFFF };
 483       __asm (
 484         "mfvscr %%v1\n"
 485         "vand   %0,%%v1,%0\n"
 486         "mtvscr %0"
 487         :
 488         : "r" (*mask)
 489       );
 490    }
 491 #endif
 492
 493    gallivm_initialized = TRUE;
 494
 495    return TRUE;
 496 }
 497
 498
 499
 500 /**
 501  * Create a new gallivm_state object.
 502  */
 503 struct gallivm_state *
 504 gallivm_create(const char *name, LLVMContextRef context)
 505 {
 506    struct gallivm_state *gallivm;
 507
 508    gallivm = CALLOC_STRUCT(gallivm_state);
 509    if (gallivm) {
 510       if (!init_gallivm_state(gallivm, name, context)) {
 511          FREE(gallivm);
 512          gallivm = NULL;
 513       }
 514    }
 515
 516    return gallivm;
 517 }
 518
 519
 520 /**
 521  * Destroy a gallivm_state object.
 522  */
 523 void
 524 gallivm_destroy(struct gallivm_state *gallivm)
 525 {
 526    gallivm_free_ir(gallivm);
 527    gallivm_free_code(gallivm);
 528    FREE(gallivm);
 529 }
 530
 531
 532 /**
 533  * Validate a function.
 534  * Verification is only done with debug builds.
 535  */
 536 void
 537 gallivm_verify_function(struct gallivm_state *gallivm,
 538                         LLVMValueRef func)
 539 {
 540    /* Verify the LLVM IR.  If invalid, dump and abort */
 541 #ifdef DEBUG
 542    if (LLVMVerifyFunction(func, LLVMPrintMessageAction)) {
 543       lp_debug_dump_value(func);
 544       assert(0);
 545       return;
 546    }
 547 #endif
 548
 549    if (gallivm_debug & GALLIVM_DEBUG_IR) {
 550       /* Print the LLVM IR to stderr */
 551       lp_debug_dump_value(func);
 552       debug_printf("\n");
 553    }
 554 }
 555
 556
 557 /**
 558  * Compile a module.
 559  * This does IR optimization on all functions in the module.
 560  */
 561 void
 562 gallivm_compile_module(struct gallivm_state *gallivm)
 563 {
 564    LLVMValueRef func;
 565    int64_t time_begin = 0;
 566
 567    assert(!gallivm->compiled);
 568
 569    if (gallivm->builder) {
 570       LLVMDisposeBuilder(gallivm->builder);
 571       gallivm->builder = NULL;
 572    }
 573
 574    /* Dump bitcode to a file */
 575    if (gallivm_debug & GALLIVM_DEBUG_DUMP_BC) {
 576       char filename[256];
 577       assert(gallivm->module_name);
 578       snprintf(filename, sizeof(filename), "ir_%s.bc", gallivm->module_name);
 579       LLVMWriteBitcodeToFile(gallivm->module, filename);
 580       debug_printf("%s written\n", filename);
 581       debug_printf("Invoke as \"opt %s %s | llc -O%d %s%s\"\n",
 582                    gallivm_debug & GALLIVM_PERF_NO_OPT ? "-mem2reg" :
 583                    "-sroa -early-cse -simplifycfg -reassociate "
 584                    "-mem2reg -constprop -instcombine -gvn",
 585                    filename, gallivm_debug & GALLIVM_PERF_NO_OPT ? 0 : 2,
 586                    "[-mcpu=<-mcpu option>] ",
 587                    "[-mattr=<-mattr option(s)>]");
 588    }
 589
 590    if (gallivm_debug & GALLIVM_DEBUG_PERF)
 591       time_begin = os_time_get();
 592
 593 #if GALLIVM_HAVE_CORO
 594    LLVMRunPassManager(gallivm->cgpassmgr, gallivm->module);
 595 #endif
 596    /* Run optimization passes */
 597    LLVMInitializeFunctionPassManager(gallivm->passmgr);
 598    func = LLVMGetFirstFunction(gallivm->module);
 599    while (func) {
 600       if (0) {
 601          debug_printf("optimizing func %s...\n", LLVMGetValueName(func));
 602       }
 603
 604    /* Disable frame pointer omission on debug/profile builds */
 605    /* XXX: And workaround http://llvm.org/PR21435 */
 606 #if defined(DEBUG) || defined(PROFILE) || defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
 607       LLVMAddTargetDependentFunctionAttr(func, "no-frame-pointer-elim", "true");
 608       LLVMAddTargetDependentFunctionAttr(func, "no-frame-pointer-elim-non-leaf", "true");
 609 #endif
 610
 611       LLVMRunFunctionPassManager(gallivm->passmgr, func);
 612       func = LLVMGetNextFunction(func);
 613    }
 614    LLVMFinalizeFunctionPassManager(gallivm->passmgr);
 615
 616    if (gallivm_debug & GALLIVM_DEBUG_PERF) {
 617       int64_t time_end = os_time_get();
 618       int time_msec = (int)((time_end - time_begin) / 1000);
 619       assert(gallivm->module_name);
 620       debug_printf("optimizing module %s took %d msec\n",
 621                    gallivm->module_name, time_msec);
 622    }
 623
 624    /* Setting the module's DataLayout to an empty string will cause the
 625     * ExecutionEngine to copy to the DataLayout string from its target machine
 626     * to the module.  As of LLVM 3.8 the module and the execution engine are
 627     * required to have the same DataLayout.
 628     *
 629     * We must make sure we do this after running the optimization passes,
 630     * because those passes need a correct datalayout string.  For example, if
 631     * those optimization passes see an empty datalayout, they will assume this
 632     * is a little endian target and will do optimizations that break big endian
 633     * machines.
 634     *
 635     * TODO: This is just a temporary work-around.  The correct solution is for
 636     * gallivm_init_state() to create a TargetMachine and pull the DataLayout
 637     * from there.  Currently, the TargetMachine used by llvmpipe is being
 638     * implicitly created by the EngineBuilder in
 639     * lp_build_create_jit_compiler_for_module()
 640     */
 641    LLVMSetDataLayout(gallivm->module, "");
 642    assert(!gallivm->engine);
 643    if (!init_gallivm_engine(gallivm)) {
 644       assert(0);
 645    }
 646    assert(gallivm->engine);
 647
 648    ++gallivm->compiled;
 649
 650    if (gallivm_debug & GALLIVM_DEBUG_ASM) {
 651       LLVMValueRef llvm_func = LLVMGetFirstFunction(gallivm->module);
 652
 653       while (llvm_func) {
 654          /*
 655           * Need to filter out functions which don't have an implementation,
 656           * such as the intrinsics. May not be sufficient in case of IPO?
 657           * LLVMGetPointerToGlobal() will abort otherwise.
 658           */
 659          if (!LLVMIsDeclaration(llvm_func)) {
 660             void *func_code = LLVMGetPointerToGlobal(gallivm->engine, llvm_func);
 661             lp_disassemble(llvm_func, func_code);
 662          }
 663          llvm_func = LLVMGetNextFunction(llvm_func);
 664       }
 665    }
 666
 667 #if defined(PROFILE)
 668    {
 669       LLVMValueRef llvm_func = LLVMGetFirstFunction(gallivm->module);
 670
 671       while (llvm_func) {
 672          if (!LLVMIsDeclaration(llvm_func)) {
 673             void *func_code = LLVMGetPointerToGlobal(gallivm->engine, llvm_func);
 674             lp_profile(llvm_func, func_code);
 675          }
 676          llvm_func = LLVMGetNextFunction(llvm_func);
 677       }
 678    }
 679 #endif
 680 }
 681
 682
 683
 684 func_pointer
 685 gallivm_jit_function(struct gallivm_state *gallivm,
 686                      LLVMValueRef func)
 687 {
 688    void *code;
 689    func_pointer jit_func;
 690    int64_t time_begin = 0;
 691
 692    assert(gallivm->compiled);
 693    assert(gallivm->engine);
 694
 695    if (gallivm_debug & GALLIVM_DEBUG_PERF)
 696       time_begin = os_time_get();
 697
 698    code = LLVMGetPointerToGlobal(gallivm->engine, func);
 699    assert(code);
 700    jit_func = pointer_to_func(code);
 701
 702    if (gallivm_debug & GALLIVM_DEBUG_PERF) {
 703       int64_t time_end = os_time_get();
 704       int time_msec = (int)(time_end - time_begin) / 1000;
 705       debug_printf("   jitting func %s took %d msec\n",
 706                    LLVMGetValueName(func), time_msec);
 707    }
 708
 709    return jit_func;
 710 }