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 #include <llvm-c/Transforms/Coroutines.h>
  50 #endif
  51
  52 unsigned gallivm_perf = 0;
  53
  54 static const struct debug_named_value lp_bld_perf_flags[] = {
  55    { "no_brilinear", GALLIVM_PERF_NO_BRILINEAR, "disable brilinear optimization" },
  56    { "no_rho_approx", GALLIVM_PERF_NO_RHO_APPROX, "disable rho_approx optimization" },
  57    { "no_quad_lod", GALLIVM_PERF_NO_QUAD_LOD, "disable quad_lod optimization" },
  58    { "no_aos_sampling", GALLIVM_PERF_NO_AOS_SAMPLING, "disable aos sampling optimization" },
  59    { "nopt",   GALLIVM_PERF_NO_OPT, "disable optimization passes to speed up shader compilation" },
  60    { "no_filter_hacks", GALLIVM_PERF_NO_BRILINEAR | GALLIVM_PERF_NO_RHO_APPROX |
  61      GALLIVM_PERF_NO_QUAD_LOD, "disable filter optimization hacks" },
  62    DEBUG_NAMED_VALUE_END
  63 };
  64
  65 #ifdef DEBUG
  66 unsigned gallivm_debug = 0;
  67
  68 static const struct debug_named_value lp_bld_debug_flags[] = {
  69    { "tgsi",   GALLIVM_DEBUG_TGSI, NULL },
  70    { "ir",     GALLIVM_DEBUG_IR, NULL },
  71    { "asm",    GALLIVM_DEBUG_ASM, NULL },
  72    { "perf",   GALLIVM_DEBUG_PERF, NULL },
  73    { "gc",     GALLIVM_DEBUG_GC, NULL },
  74    { "dumpbc", GALLIVM_DEBUG_DUMP_BC, NULL },
  75    DEBUG_NAMED_VALUE_END
  76 };
  77
  78 DEBUG_GET_ONCE_FLAGS_OPTION(gallivm_debug, "GALLIVM_DEBUG", lp_bld_debug_flags, 0)
  79 #endif
  80
  81
  82 static boolean gallivm_initialized = FALSE;
  83
  84 unsigned lp_native_vector_width;
  85
  86
  87 /*
  88  * Optimization values are:
  89  * - 0: None (-O0)
  90  * - 1: Less (-O1)
  91  * - 2: Default (-O2, -Os)
  92  * - 3: Aggressive (-O3)
  93  *
  94  * See also CodeGenOpt::Level in llvm/Target/TargetMachine.h
  95  */
  96 enum LLVM_CodeGenOpt_Level {
  97    None,        // -O0
  98    Less,        // -O1
  99    Default,     // -O2, -Os
 100    Aggressive   // -O3
 101 };
 102
 103
 104 /**
 105  * Create the LLVM (optimization) pass manager and install
 106  * relevant optimization passes.
 107  * \return  TRUE for success, FALSE for failure
 108  */
 109 static boolean
 110 create_pass_manager(struct gallivm_state *gallivm)
 111 {
 112    assert(!gallivm->passmgr);
 113    assert(gallivm->target);
 114
 115    gallivm->passmgr = LLVMCreateFunctionPassManagerForModule(gallivm->module);
 116    if (!gallivm->passmgr)
 117       return FALSE;
 118
 119 #if GALLIVM_HAVE_CORO
 120    gallivm->cgpassmgr = LLVMCreatePassManager();
 121 #endif
 122    /*
 123     * TODO: some per module pass manager with IPO passes might be helpful -
 124     * the generated texture functions may benefit from inlining if they are
 125     * simple, or constant propagation into them, etc.
 126     */
 127
 128 #if LLVM_VERSION_MAJOR == 3 && LLVM_VERSION_MINOR < 9
 129    // Old versions of LLVM get the DataLayout from the pass manager.
 130    LLVMAddTargetData(gallivm->target, gallivm->passmgr);
 131 #endif
 132
 133    {
 134       char *td_str;
 135       // New ones from the Module.
 136       td_str = LLVMCopyStringRepOfTargetData(gallivm->target);
 137       LLVMSetDataLayout(gallivm->module, td_str);
 138       free(td_str);
 139    }
 140
 141 #if GALLIVM_HAVE_CORO
 142    LLVMAddCoroEarlyPass(gallivm->cgpassmgr);
 143    LLVMAddCoroSplitPass(gallivm->cgpassmgr);
 144    LLVMAddCoroElidePass(gallivm->cgpassmgr);
 145 #endif
 146
 147    if ((gallivm_perf & GALLIVM_PERF_NO_OPT) == 0) {
 148       /*
 149        * TODO: Evaluate passes some more - keeping in mind
 150        * both quality of generated code and compile times.
 151        */
 152       /*
 153        * NOTE: if you change this, don't forget to change the output
 154        * with GALLIVM_DEBUG_DUMP_BC in gallivm_compile_module.
 155        */
 156       LLVMAddScalarReplAggregatesPass(gallivm->passmgr);
 157       LLVMAddEarlyCSEPass(gallivm->passmgr);
 158       LLVMAddCFGSimplificationPass(gallivm->passmgr);
 159       /*
 160        * FIXME: LICM is potentially quite useful. However, for some
 161        * rather crazy shaders the compile time can reach _hours_ per shader,
 162        * due to licm implying lcssa (since llvm 3.5), which can take forever.
 163        * Even for sane shaders, the cost of licm is rather high (and not just
 164        * due to lcssa, licm itself too), though mostly only in cases when it
 165        * can actually move things, so having to disable it is a pity.
 166        * LLVMAddLICMPass(gallivm->passmgr);
 167        */
 168       LLVMAddReassociatePass(gallivm->passmgr);
 169       LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
 170       LLVMAddConstantPropagationPass(gallivm->passmgr);
 171       LLVMAddInstructionCombiningPass(gallivm->passmgr);
 172       LLVMAddGVNPass(gallivm->passmgr);
 173 #if GALLIVM_HAVE_CORO
 174       LLVMAddCoroCleanupPass(gallivm->passmgr);
 175 #endif
 176    }
 177    else {
 178       /* We need at least this pass to prevent the backends to fail in
 179        * unexpected ways.
 180        */
 181       LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
 182    }
 183
 184    return TRUE;
 185 }
 186
 187
 188 /**
 189  * Free gallivm object's LLVM allocations, but not any generated code
 190  * nor the gallivm object itself.
 191  */
 192 void
 193 gallivm_free_ir(struct gallivm_state *gallivm)
 194 {
 195    if (gallivm->passmgr) {
 196       LLVMDisposePassManager(gallivm->passmgr);
 197    }
 198
 199 #if GALLIVM_HAVE_CORO
 200    if (gallivm->cgpassmgr) {
 201       LLVMDisposePassManager(gallivm->cgpassmgr);
 202    }
 203 #endif
 204
 205    if (gallivm->engine) {
 206       /* This will already destroy any associated module */
 207       LLVMDisposeExecutionEngine(gallivm->engine);
 208    } else if (gallivm->module) {
 209       LLVMDisposeModule(gallivm->module);
 210    }
 211
 212    FREE(gallivm->module_name);
 213
 214    if (gallivm->target) {
 215       LLVMDisposeTargetData(gallivm->target);
 216    }
 217
 218    if (gallivm->builder)
 219       LLVMDisposeBuilder(gallivm->builder);
 220
 221    /* The LLVMContext should be owned by the parent of gallivm. */
 222
 223    gallivm->engine = NULL;
 224    gallivm->target = NULL;
 225    gallivm->module = NULL;
 226    gallivm->module_name = NULL;
 227    gallivm->cgpassmgr = NULL;
 228    gallivm->passmgr = NULL;
 229    gallivm->context = NULL;
 230    gallivm->builder = NULL;
 231 }
 232
 233
 234 /**
 235  * Free LLVM-generated code.  Should be done AFTER gallivm_free_ir().
 236  */
 237 static void
 238 gallivm_free_code(struct gallivm_state *gallivm)
 239 {
 240    assert(!gallivm->module);
 241    assert(!gallivm->engine);
 242    lp_free_generated_code(gallivm->code);
 243    gallivm->code = NULL;
 244    lp_free_memory_manager(gallivm->memorymgr);
 245    gallivm->memorymgr = NULL;
 246 }
 247
 248
 249 static boolean
 250 init_gallivm_engine(struct gallivm_state *gallivm)
 251 {
 252    if (1) {
 253       enum LLVM_CodeGenOpt_Level optlevel;
 254       char *error = NULL;
 255       int ret;
 256
 257       if (gallivm_perf & GALLIVM_PERF_NO_OPT) {
 258          optlevel = None;
 259       }
 260       else {
 261          optlevel = Default;
 262       }
 263
 264       ret = lp_build_create_jit_compiler_for_module(&gallivm->engine,
 265                                                     &gallivm->code,
 266                                                     gallivm->module,
 267                                                     gallivm->memorymgr,
 268                                                     (unsigned) optlevel,
 269                                                     &error);
 270       if (ret) {
 271          _debug_printf("%s\n", error);
 272          LLVMDisposeMessage(error);
 273          goto fail;
 274       }
 275    }
 276
 277    if (0) {
 278        /*
 279         * Dump the data layout strings.
 280         */
 281
 282        LLVMTargetDataRef target = LLVMGetExecutionEngineTargetData(gallivm->engine);
 283        char *data_layout;
 284        char *engine_data_layout;
 285
 286        data_layout = LLVMCopyStringRepOfTargetData(gallivm->target);
 287        engine_data_layout = LLVMCopyStringRepOfTargetData(target);
 288
 289        if (1) {
 290           debug_printf("module target data = %s\n", data_layout);
 291           debug_printf("engine target data = %s\n", engine_data_layout);
 292        }
 293
 294        free(data_layout);
 295        free(engine_data_layout);
 296    }
 297
 298    return TRUE;
 299
 300 fail:
 301    return FALSE;
 302 }
 303
 304
 305 /**
 306  * Allocate gallivm LLVM objects.
 307  * \return  TRUE for success, FALSE for failure
 308  */
 309 static boolean
 310 init_gallivm_state(struct gallivm_state *gallivm, const char *name,
 311                    LLVMContextRef context)
 312 {
 313    assert(!gallivm->context);
 314    assert(!gallivm->module);
 315
 316    if (!lp_build_init())
 317       return FALSE;
 318
 319    gallivm->context = context;
 320
 321    if (!gallivm->context)
 322       goto fail;
 323
 324    gallivm->module_name = NULL;
 325    if (name) {
 326       size_t size = strlen(name) + 1;
 327       gallivm->module_name = MALLOC(size);
 328       if (gallivm->module_name) {
 329          memcpy(gallivm->module_name, name, size);
 330       }
 331    }
 332
 333    gallivm->module = LLVMModuleCreateWithNameInContext(name,
 334                                                        gallivm->context);
 335    if (!gallivm->module)
 336       goto fail;
 337
 338    gallivm->builder = LLVMCreateBuilderInContext(gallivm->context);
 339    if (!gallivm->builder)
 340       goto fail;
 341
 342    gallivm->memorymgr = lp_get_default_memory_manager();
 343    if (!gallivm->memorymgr)
 344       goto fail;
 345
 346    /* FIXME: MC-JIT only allows compiling one module at a time, and it must be
 347     * complete when MC-JIT is created. So defer the MC-JIT engine creation for
 348     * now.
 349     */
 350
 351    /*
 352     * MC-JIT engine compiles the module immediately on creation, so we can't
 353     * obtain the target data from it.  Instead we create a target data layout
 354     * from a string.
 355     *
 356     * The produced layout strings are not precisely the same, but should make
 357     * no difference for the kind of optimization passes we run.
 358     *
 359     * For reference this is the layout string on x64:
 360     *
 361     *   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
 362     *
 363     * See also:
 364     * - http://llvm.org/docs/LangRef.html#datalayout
 365     */
 366
 367    {
 368       const unsigned pointer_size = 8 * sizeof(void *);
 369       char layout[512];
 370       snprintf(layout, sizeof layout, "%c-p:%u:%u:%u-i64:64:64-a0:0:%u-s0:%u:%u",
 371 #ifdef PIPE_ARCH_LITTLE_ENDIAN
 372                     'e', // little endian
 373 #else
 374                     'E', // big endian
 375 #endif
 376                     pointer_size, pointer_size, pointer_size, // pointer size, abi alignment, preferred alignment
 377                     pointer_size, // aggregate preferred alignment
 378                     pointer_size, pointer_size); // stack objects abi alignment, preferred alignment
 379
 380       gallivm->target = LLVMCreateTargetData(layout);
 381       if (!gallivm->target) {
 382          return FALSE;
 383       }
 384    }
 385
 386    if (!create_pass_manager(gallivm))
 387       goto fail;
 388
 389    return TRUE;
 390
 391 fail:
 392    gallivm_free_ir(gallivm);
 393    gallivm_free_code(gallivm);
 394    return FALSE;
 395 }
 396
 397
 398 boolean
 399 lp_build_init(void)
 400 {
 401    if (gallivm_initialized)
 402       return TRUE;
 403
 404
 405    /* LLVMLinkIn* are no-ops at runtime.  They just ensure the respective
 406     * component is linked at buildtime, which is sufficient for its static
 407     * constructors to be called at load time.
 408     */
 409    LLVMLinkInMCJIT();
 410
 411 #ifdef DEBUG
 412    gallivm_debug = debug_get_option_gallivm_debug();
 413 #endif
 414
 415    gallivm_perf = debug_get_flags_option("GALLIVM_PERF", lp_bld_perf_flags, 0 );
 416
 417    lp_set_target_options();
 418
 419    util_cpu_detect();
 420
 421    /* For simulating less capable machines */
 422 #ifdef DEBUG
 423    if (debug_get_bool_option("LP_FORCE_SSE2", FALSE)) {
 424       assert(util_cpu_caps.has_sse2);
 425       util_cpu_caps.has_sse3 = 0;
 426       util_cpu_caps.has_ssse3 = 0;
 427       util_cpu_caps.has_sse4_1 = 0;
 428       util_cpu_caps.has_sse4_2 = 0;
 429       util_cpu_caps.has_avx = 0;
 430       util_cpu_caps.has_avx2 = 0;
 431       util_cpu_caps.has_f16c = 0;
 432       util_cpu_caps.has_fma = 0;
 433    }
 434 #endif
 435
 436    /* AMD Bulldozer AVX's throughput is the same as SSE2; and because using
 437     * 8-wide vector needs more floating ops than 4-wide (due to padding), it is
 438     * actually more efficient to use 4-wide vectors on this processor.
 439     *
 440     * See also:
 441     * - http://www.anandtech.com/show/4955/the-bulldozer-review-amd-fx8150-tested/2
 442     */
 443    if (util_cpu_caps.has_avx &&
 444        util_cpu_caps.has_intel) {
 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 (lp_native_vector_width <= 128) {
 457       /* Hide AVX support, as often LLVM AVX intrinsics are only guarded by
 458        * "util_cpu_caps.has_avx" predicate, and lack the
 459        * "lp_native_vector_width > 128" predicate. And also to ensure a more
 460        * consistent behavior, allowing one to test SSE2 on AVX machines.
 461        * XXX: should not play games with util_cpu_caps directly as it might
 462        * get used for other things outside llvm too.
 463        */
 464       util_cpu_caps.has_avx = 0;
 465       util_cpu_caps.has_avx2 = 0;
 466       util_cpu_caps.has_f16c = 0;
 467       util_cpu_caps.has_fma = 0;
 468    }
 469
 470 #ifdef PIPE_ARCH_PPC_64
 471    /* Set the NJ bit in VSCR to 0 so denormalized values are handled as
 472     * specified by IEEE standard (PowerISA 2.06 - Section 6.3). This guarantees
 473     * that some rounding and half-float to float handling does not round
 474     * incorrectly to 0.
 475     * XXX: should eventually follow same logic on all platforms.
 476     * Right now denorms get explicitly disabled (but elsewhere) for x86,
 477     * whereas ppc64 explicitly enables them...
 478     */
 479    if (util_cpu_caps.has_altivec) {
 480       unsigned short mask[] = { 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF,
 481                                 0xFFFF, 0xFFFF, 0xFFFE, 0xFFFF };
 482       __asm (
 483         "mfvscr %%v1\n"
 484         "vand   %0,%%v1,%0\n"
 485         "mtvscr %0"
 486         :
 487         : "r" (*mask)
 488       );
 489    }
 490 #endif
 491
 492    gallivm_initialized = TRUE;
 493
 494    return TRUE;
 495 }
 496
 497
 498
 499 /**
 500  * Create a new gallivm_state object.
 501  */
 502 struct gallivm_state *
 503 gallivm_create(const char *name, LLVMContextRef context)
 504 {
 505    struct gallivm_state *gallivm;
 506
 507    gallivm = CALLOC_STRUCT(gallivm_state);
 508    if (gallivm) {
 509       if (!init_gallivm_state(gallivm, name, context)) {
 510          FREE(gallivm);
 511          gallivm = NULL;
 512       }
 513    }
 514
 515    return gallivm;
 516 }
 517
 518
 519 /**
 520  * Destroy a gallivm_state object.
 521  */
 522 void
 523 gallivm_destroy(struct gallivm_state *gallivm)
 524 {
 525    gallivm_free_ir(gallivm);
 526    gallivm_free_code(gallivm);
 527    FREE(gallivm);
 528 }
 529
 530
 531 /**
 532  * Validate a function.
 533  * Verification is only done with debug builds.
 534  */
 535 void
 536 gallivm_verify_function(struct gallivm_state *gallivm,
 537                         LLVMValueRef func)
 538 {
 539    /* Verify the LLVM IR.  If invalid, dump and abort */
 540 #ifdef DEBUG
 541    if (LLVMVerifyFunction(func, LLVMPrintMessageAction)) {
 542       lp_debug_dump_value(func);
 543       assert(0);
 544       return;
 545    }
 546 #endif
 547
 548    if (gallivm_debug & GALLIVM_DEBUG_IR) {
 549       /* Print the LLVM IR to stderr */
 550       lp_debug_dump_value(func);
 551       debug_printf("\n");
 552    }
 553 }
 554
 555
 556 /**
 557  * Compile a module.
 558  * This does IR optimization on all functions in the module.
 559  */
 560 void
 561 gallivm_compile_module(struct gallivm_state *gallivm)
 562 {
 563    LLVMValueRef func;
 564    int64_t time_begin = 0;
 565
 566    assert(!gallivm->compiled);
 567
 568    if (gallivm->builder) {
 569       LLVMDisposeBuilder(gallivm->builder);
 570       gallivm->builder = NULL;
 571    }
 572
 573    /* Dump bitcode to a file */
 574    if (gallivm_debug & GALLIVM_DEBUG_DUMP_BC) {
 575       char filename[256];
 576       assert(gallivm->module_name);
 577       snprintf(filename, sizeof(filename), "ir_%s.bc", gallivm->module_name);
 578       LLVMWriteBitcodeToFile(gallivm->module, filename);
 579       debug_printf("%s written\n", filename);
 580       debug_printf("Invoke as \"opt %s %s | llc -O%d %s%s\"\n",
 581                    gallivm_debug & GALLIVM_PERF_NO_OPT ? "-mem2reg" :
 582                    "-sroa -early-cse -simplifycfg -reassociate "
 583                    "-mem2reg -constprop -instcombine -gvn",
 584                    filename, gallivm_debug & GALLIVM_PERF_NO_OPT ? 0 : 2,
 585                    "[-mcpu=<-mcpu option>] ",
 586                    "[-mattr=<-mattr option(s)>]");
 587    }
 588
 589    if (gallivm_debug & GALLIVM_DEBUG_PERF)
 590       time_begin = os_time_get();
 591
 592 #if GALLIVM_HAVE_CORO
 593    LLVMRunPassManager(gallivm->cgpassmgr, gallivm->module);
 594 #endif
 595    /* Run optimization passes */
 596    LLVMInitializeFunctionPassManager(gallivm->passmgr);
 597    func = LLVMGetFirstFunction(gallivm->module);
 598    while (func) {
 599       if (0) {
 600          debug_printf("optimizing func %s...\n", LLVMGetValueName(func));
 601       }
 602
 603    /* Disable frame pointer omission on debug/profile builds */
 604    /* XXX: And workaround http://llvm.org/PR21435 */
 605 #if defined(DEBUG) || defined(PROFILE) || defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
 606       LLVMAddTargetDependentFunctionAttr(func, "no-frame-pointer-elim", "true");
 607       LLVMAddTargetDependentFunctionAttr(func, "no-frame-pointer-elim-non-leaf", "true");
 608 #endif
 609
 610       LLVMRunFunctionPassManager(gallivm->passmgr, func);
 611       func = LLVMGetNextFunction(func);
 612    }
 613    LLVMFinalizeFunctionPassManager(gallivm->passmgr);
 614
 615    if (gallivm_debug & GALLIVM_DEBUG_PERF) {
 616       int64_t time_end = os_time_get();
 617       int time_msec = (int)((time_end - time_begin) / 1000);
 618       assert(gallivm->module_name);
 619       debug_printf("optimizing module %s took %d msec\n",
 620                    gallivm->module_name, time_msec);
 621    }
 622
 623    /* Setting the module's DataLayout to an empty string will cause the
 624     * ExecutionEngine to copy to the DataLayout string from its target machine
 625     * to the module.  As of LLVM 3.8 the module and the execution engine are
 626     * required to have the same DataLayout.
 627     *
 628     * We must make sure we do this after running the optimization passes,
 629     * because those passes need a correct datalayout string.  For example, if
 630     * those optimization passes see an empty datalayout, they will assume this
 631     * is a little endian target and will do optimizations that break big endian
 632     * machines.
 633     *
 634     * TODO: This is just a temporary work-around.  The correct solution is for
 635     * gallivm_init_state() to create a TargetMachine and pull the DataLayout
 636     * from there.  Currently, the TargetMachine used by llvmpipe is being
 637     * implicitly created by the EngineBuilder in
 638     * lp_build_create_jit_compiler_for_module()
 639     */
 640    LLVMSetDataLayout(gallivm->module, "");
 641    assert(!gallivm->engine);
 642    if (!init_gallivm_engine(gallivm)) {
 643       assert(0);
 644    }
 645    assert(gallivm->engine);
 646
 647    ++gallivm->compiled;
 648
 649    if (gallivm_debug & GALLIVM_DEBUG_ASM) {
 650       LLVMValueRef llvm_func = LLVMGetFirstFunction(gallivm->module);
 651
 652       while (llvm_func) {
 653          /*
 654           * Need to filter out functions which don't have an implementation,
 655           * such as the intrinsics. May not be sufficient in case of IPO?
 656           * LLVMGetPointerToGlobal() will abort otherwise.
 657           */
 658          if (!LLVMIsDeclaration(llvm_func)) {
 659             void *func_code = LLVMGetPointerToGlobal(gallivm->engine, llvm_func);
 660             lp_disassemble(llvm_func, func_code);
 661          }
 662          llvm_func = LLVMGetNextFunction(llvm_func);
 663       }
 664    }
 665
 666 #if defined(PROFILE)
 667    {
 668       LLVMValueRef llvm_func = LLVMGetFirstFunction(gallivm->module);
 669
 670       while (llvm_func) {
 671          if (!LLVMIsDeclaration(llvm_func)) {
 672             void *func_code = LLVMGetPointerToGlobal(gallivm->engine, llvm_func);
 673             lp_profile(llvm_func, func_code);
 674          }
 675          llvm_func = LLVMGetNextFunction(llvm_func);
 676       }
 677    }
 678 #endif
 679 }
 680
 681
 682
 683 func_pointer
 684 gallivm_jit_function(struct gallivm_state *gallivm,
 685                      LLVMValueRef func)
 686 {
 687    void *code;
 688    func_pointer jit_func;
 689    int64_t time_begin = 0;
 690
 691    assert(gallivm->compiled);
 692    assert(gallivm->engine);
 693
 694    if (gallivm_debug & GALLIVM_DEBUG_PERF)
 695       time_begin = os_time_get();
 696
 697    code = LLVMGetPointerToGlobal(gallivm->engine, func);
 698    assert(code);
 699    jit_func = pointer_to_func(code);
 700
 701    if (gallivm_debug & GALLIVM_DEBUG_PERF) {
 702       int64_t time_end = os_time_get();
 703       int time_msec = (int)(time_end - time_begin) / 1000;
 704       debug_printf("   jitting func %s took %d msec\n",
 705                    LLVMGetValueName(func), time_msec);
 706    }
 707
 708    return jit_func;
 709 }