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    {
 129       char *td_str;
 130       // New ones from the Module.
 131       td_str = LLVMCopyStringRepOfTargetData(gallivm->target);
 132       LLVMSetDataLayout(gallivm->module, td_str);
 133       free(td_str);
 134    }
 135
 136 #if GALLIVM_HAVE_CORO
 137    LLVMAddCoroEarlyPass(gallivm->cgpassmgr);
 138    LLVMAddCoroSplitPass(gallivm->cgpassmgr);
 139    LLVMAddCoroElidePass(gallivm->cgpassmgr);
 140 #endif
 141
 142    if ((gallivm_perf & GALLIVM_PERF_NO_OPT) == 0) {
 143       /*
 144        * TODO: Evaluate passes some more - keeping in mind
 145        * both quality of generated code and compile times.
 146        */
 147       /*
 148        * NOTE: if you change this, don't forget to change the output
 149        * with GALLIVM_DEBUG_DUMP_BC in gallivm_compile_module.
 150        */
 151       LLVMAddScalarReplAggregatesPass(gallivm->passmgr);
 152       LLVMAddEarlyCSEPass(gallivm->passmgr);
 153       LLVMAddCFGSimplificationPass(gallivm->passmgr);
 154       /*
 155        * FIXME: LICM is potentially quite useful. However, for some
 156        * rather crazy shaders the compile time can reach _hours_ per shader,
 157        * due to licm implying lcssa (since llvm 3.5), which can take forever.
 158        * Even for sane shaders, the cost of licm is rather high (and not just
 159        * due to lcssa, licm itself too), though mostly only in cases when it
 160        * can actually move things, so having to disable it is a pity.
 161        * LLVMAddLICMPass(gallivm->passmgr);
 162        */
 163       LLVMAddReassociatePass(gallivm->passmgr);
 164       LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
 165       LLVMAddConstantPropagationPass(gallivm->passmgr);
 166       LLVMAddInstructionCombiningPass(gallivm->passmgr);
 167       LLVMAddGVNPass(gallivm->passmgr);
 168 #if GALLIVM_HAVE_CORO
 169       LLVMAddCoroCleanupPass(gallivm->passmgr);
 170 #endif
 171    }
 172    else {
 173       /* We need at least this pass to prevent the backends to fail in
 174        * unexpected ways.
 175        */
 176       LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
 177    }
 178
 179    return TRUE;
 180 }
 181
 182
 183 /**
 184  * Free gallivm object's LLVM allocations, but not any generated code
 185  * nor the gallivm object itself.
 186  */
 187 void
 188 gallivm_free_ir(struct gallivm_state *gallivm)
 189 {
 190    if (gallivm->passmgr) {
 191       LLVMDisposePassManager(gallivm->passmgr);
 192    }
 193
 194 #if GALLIVM_HAVE_CORO
 195    if (gallivm->cgpassmgr) {
 196       LLVMDisposePassManager(gallivm->cgpassmgr);
 197    }
 198 #endif
 199
 200    if (gallivm->engine) {
 201       /* This will already destroy any associated module */
 202       LLVMDisposeExecutionEngine(gallivm->engine);
 203    } else if (gallivm->module) {
 204       LLVMDisposeModule(gallivm->module);
 205    }
 206
 207    FREE(gallivm->module_name);
 208
 209    if (gallivm->target) {
 210       LLVMDisposeTargetData(gallivm->target);
 211    }
 212
 213    if (gallivm->builder)
 214       LLVMDisposeBuilder(gallivm->builder);
 215
 216    /* The LLVMContext should be owned by the parent of gallivm. */
 217
 218    gallivm->engine = NULL;
 219    gallivm->target = NULL;
 220    gallivm->module = NULL;
 221    gallivm->module_name = NULL;
 222    gallivm->cgpassmgr = NULL;
 223    gallivm->passmgr = NULL;
 224    gallivm->context = NULL;
 225    gallivm->builder = NULL;
 226 }
 227
 228
 229 /**
 230  * Free LLVM-generated code.  Should be done AFTER gallivm_free_ir().
 231  */
 232 static void
 233 gallivm_free_code(struct gallivm_state *gallivm)
 234 {
 235    assert(!gallivm->module);
 236    assert(!gallivm->engine);
 237    lp_free_generated_code(gallivm->code);
 238    gallivm->code = NULL;
 239    lp_free_memory_manager(gallivm->memorymgr);
 240    gallivm->memorymgr = NULL;
 241 }
 242
 243
 244 static boolean
 245 init_gallivm_engine(struct gallivm_state *gallivm)
 246 {
 247    if (1) {
 248       enum LLVM_CodeGenOpt_Level optlevel;
 249       char *error = NULL;
 250       int ret;
 251
 252       if (gallivm_perf & GALLIVM_PERF_NO_OPT) {
 253          optlevel = None;
 254       }
 255       else {
 256          optlevel = Default;
 257       }
 258
 259       ret = lp_build_create_jit_compiler_for_module(&gallivm->engine,
 260                                                     &gallivm->code,
 261                                                     gallivm->module,
 262                                                     gallivm->memorymgr,
 263                                                     (unsigned) optlevel,
 264                                                     &error);
 265       if (ret) {
 266          _debug_printf("%s\n", error);
 267          LLVMDisposeMessage(error);
 268          goto fail;
 269       }
 270    }
 271
 272    if (0) {
 273        /*
 274         * Dump the data layout strings.
 275         */
 276
 277        LLVMTargetDataRef target = LLVMGetExecutionEngineTargetData(gallivm->engine);
 278        char *data_layout;
 279        char *engine_data_layout;
 280
 281        data_layout = LLVMCopyStringRepOfTargetData(gallivm->target);
 282        engine_data_layout = LLVMCopyStringRepOfTargetData(target);
 283
 284        if (1) {
 285           debug_printf("module target data = %s\n", data_layout);
 286           debug_printf("engine target data = %s\n", engine_data_layout);
 287        }
 288
 289        free(data_layout);
 290        free(engine_data_layout);
 291    }
 292
 293    return TRUE;
 294
 295 fail:
 296    return FALSE;
 297 }
 298
 299
 300 /**
 301  * Allocate gallivm LLVM objects.
 302  * \return  TRUE for success, FALSE for failure
 303  */
 304 static boolean
 305 init_gallivm_state(struct gallivm_state *gallivm, const char *name,
 306                    LLVMContextRef context)
 307 {
 308    assert(!gallivm->context);
 309    assert(!gallivm->module);
 310
 311    if (!lp_build_init())
 312       return FALSE;
 313
 314    gallivm->context = context;
 315
 316    if (!gallivm->context)
 317       goto fail;
 318
 319    gallivm->module_name = NULL;
 320    if (name) {
 321       size_t size = strlen(name) + 1;
 322       gallivm->module_name = MALLOC(size);
 323       if (gallivm->module_name) {
 324          memcpy(gallivm->module_name, name, size);
 325       }
 326    }
 327
 328    gallivm->module = LLVMModuleCreateWithNameInContext(name,
 329                                                        gallivm->context);
 330    if (!gallivm->module)
 331       goto fail;
 332
 333    gallivm->builder = LLVMCreateBuilderInContext(gallivm->context);
 334    if (!gallivm->builder)
 335       goto fail;
 336
 337    gallivm->memorymgr = lp_get_default_memory_manager();
 338    if (!gallivm->memorymgr)
 339       goto fail;
 340
 341    /* FIXME: MC-JIT only allows compiling one module at a time, and it must be
 342     * complete when MC-JIT is created. So defer the MC-JIT engine creation for
 343     * now.
 344     */
 345
 346    /*
 347     * MC-JIT engine compiles the module immediately on creation, so we can't
 348     * obtain the target data from it.  Instead we create a target data layout
 349     * from a string.
 350     *
 351     * The produced layout strings are not precisely the same, but should make
 352     * no difference for the kind of optimization passes we run.
 353     *
 354     * For reference this is the layout string on x64:
 355     *
 356     *   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
 357     *
 358     * See also:
 359     * - http://llvm.org/docs/LangRef.html#datalayout
 360     */
 361
 362    {
 363       const unsigned pointer_size = 8 * sizeof(void *);
 364       char layout[512];
 365       snprintf(layout, sizeof layout, "%c-p:%u:%u:%u-i64:64:64-a0:0:%u-s0:%u:%u",
 366 #ifdef PIPE_ARCH_LITTLE_ENDIAN
 367                     'e', // little endian
 368 #else
 369                     'E', // big endian
 370 #endif
 371                     pointer_size, pointer_size, pointer_size, // pointer size, abi alignment, preferred alignment
 372                     pointer_size, // aggregate preferred alignment
 373                     pointer_size, pointer_size); // stack objects abi alignment, preferred alignment
 374
 375       gallivm->target = LLVMCreateTargetData(layout);
 376       if (!gallivm->target) {
 377          return FALSE;
 378       }
 379    }
 380
 381    if (!create_pass_manager(gallivm))
 382       goto fail;
 383
 384    return TRUE;
 385
 386 fail:
 387    gallivm_free_ir(gallivm);
 388    gallivm_free_code(gallivm);
 389    return FALSE;
 390 }
 391
 392
 393 boolean
 394 lp_build_init(void)
 395 {
 396    if (gallivm_initialized)
 397       return TRUE;
 398
 399
 400    /* LLVMLinkIn* are no-ops at runtime.  They just ensure the respective
 401     * component is linked at buildtime, which is sufficient for its static
 402     * constructors to be called at load time.
 403     */
 404    LLVMLinkInMCJIT();
 405
 406 #ifdef DEBUG
 407    gallivm_debug = debug_get_option_gallivm_debug();
 408 #endif
 409
 410    gallivm_perf = debug_get_flags_option("GALLIVM_PERF", lp_bld_perf_flags, 0 );
 411
 412    lp_set_target_options();
 413
 414    util_cpu_detect();
 415
 416    /* For simulating less capable machines */
 417 #ifdef DEBUG
 418    if (debug_get_bool_option("LP_FORCE_SSE2", FALSE)) {
 419       assert(util_cpu_caps.has_sse2);
 420       util_cpu_caps.has_sse3 = 0;
 421       util_cpu_caps.has_ssse3 = 0;
 422       util_cpu_caps.has_sse4_1 = 0;
 423       util_cpu_caps.has_sse4_2 = 0;
 424       util_cpu_caps.has_avx = 0;
 425       util_cpu_caps.has_avx2 = 0;
 426       util_cpu_caps.has_f16c = 0;
 427       util_cpu_caps.has_fma = 0;
 428    }
 429 #endif
 430
 431    /* AMD Bulldozer AVX's throughput is the same as SSE2; and because using
 432     * 8-wide vector needs more floating ops than 4-wide (due to padding), it is
 433     * actually more efficient to use 4-wide vectors on this processor.
 434     *
 435     * See also:
 436     * - http://www.anandtech.com/show/4955/the-bulldozer-review-amd-fx8150-tested/2
 437     */
 438    if (util_cpu_caps.has_avx &&
 439        util_cpu_caps.has_intel) {
 440       lp_native_vector_width = 256;
 441    } else {
 442       /* Leave it at 128, even when no SIMD extensions are available.
 443        * Really needs to be a multiple of 128 so can fit 4 floats.
 444        */
 445       lp_native_vector_width = 128;
 446    }
 447
 448    lp_native_vector_width = debug_get_num_option("LP_NATIVE_VECTOR_WIDTH",
 449                                                  lp_native_vector_width);
 450
 451    if (lp_native_vector_width <= 128) {
 452       /* Hide AVX support, as often LLVM AVX intrinsics are only guarded by
 453        * "util_cpu_caps.has_avx" predicate, and lack the
 454        * "lp_native_vector_width > 128" predicate. And also to ensure a more
 455        * consistent behavior, allowing one to test SSE2 on AVX machines.
 456        * XXX: should not play games with util_cpu_caps directly as it might
 457        * get used for other things outside llvm too.
 458        */
 459       util_cpu_caps.has_avx = 0;
 460       util_cpu_caps.has_avx2 = 0;
 461       util_cpu_caps.has_f16c = 0;
 462       util_cpu_caps.has_fma = 0;
 463    }
 464
 465 #ifdef PIPE_ARCH_PPC_64
 466    /* Set the NJ bit in VSCR to 0 so denormalized values are handled as
 467     * specified by IEEE standard (PowerISA 2.06 - Section 6.3). This guarantees
 468     * that some rounding and half-float to float handling does not round
 469     * incorrectly to 0.
 470     * XXX: should eventually follow same logic on all platforms.
 471     * Right now denorms get explicitly disabled (but elsewhere) for x86,
 472     * whereas ppc64 explicitly enables them...
 473     */
 474    if (util_cpu_caps.has_altivec) {
 475       unsigned short mask[] = { 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF,
 476                                 0xFFFF, 0xFFFF, 0xFFFE, 0xFFFF };
 477       __asm (
 478         "mfvscr %%v1\n"
 479         "vand   %0,%%v1,%0\n"
 480         "mtvscr %0"
 481         :
 482         : "r" (*mask)
 483       );
 484    }
 485 #endif
 486
 487    gallivm_initialized = TRUE;
 488
 489    return TRUE;
 490 }
 491
 492
 493
 494 /**
 495  * Create a new gallivm_state object.
 496  */
 497 struct gallivm_state *
 498 gallivm_create(const char *name, LLVMContextRef context)
 499 {
 500    struct gallivm_state *gallivm;
 501
 502    gallivm = CALLOC_STRUCT(gallivm_state);
 503    if (gallivm) {
 504       if (!init_gallivm_state(gallivm, name, context)) {
 505          FREE(gallivm);
 506          gallivm = NULL;
 507       }
 508    }
 509
 510    return gallivm;
 511 }
 512
 513
 514 /**
 515  * Destroy a gallivm_state object.
 516  */
 517 void
 518 gallivm_destroy(struct gallivm_state *gallivm)
 519 {
 520    gallivm_free_ir(gallivm);
 521    gallivm_free_code(gallivm);
 522    FREE(gallivm);
 523 }
 524
 525
 526 /**
 527  * Validate a function.
 528  * Verification is only done with debug builds.
 529  */
 530 void
 531 gallivm_verify_function(struct gallivm_state *gallivm,
 532                         LLVMValueRef func)
 533 {
 534    /* Verify the LLVM IR.  If invalid, dump and abort */
 535 #ifdef DEBUG
 536    if (LLVMVerifyFunction(func, LLVMPrintMessageAction)) {
 537       lp_debug_dump_value(func);
 538       assert(0);
 539       return;
 540    }
 541 #endif
 542
 543    if (gallivm_debug & GALLIVM_DEBUG_IR) {
 544       /* Print the LLVM IR to stderr */
 545       lp_debug_dump_value(func);
 546       debug_printf("\n");
 547    }
 548 }
 549
 550
 551 /**
 552  * Compile a module.
 553  * This does IR optimization on all functions in the module.
 554  */
 555 void
 556 gallivm_compile_module(struct gallivm_state *gallivm)
 557 {
 558    LLVMValueRef func;
 559    int64_t time_begin = 0;
 560
 561    assert(!gallivm->compiled);
 562
 563    if (gallivm->builder) {
 564       LLVMDisposeBuilder(gallivm->builder);
 565       gallivm->builder = NULL;
 566    }
 567
 568    /* Dump bitcode to a file */
 569    if (gallivm_debug & GALLIVM_DEBUG_DUMP_BC) {
 570       char filename[256];
 571       assert(gallivm->module_name);
 572       snprintf(filename, sizeof(filename), "ir_%s.bc", gallivm->module_name);
 573       LLVMWriteBitcodeToFile(gallivm->module, filename);
 574       debug_printf("%s written\n", filename);
 575       debug_printf("Invoke as \"opt %s %s | llc -O%d %s%s\"\n",
 576                    gallivm_debug & GALLIVM_PERF_NO_OPT ? "-mem2reg" :
 577                    "-sroa -early-cse -simplifycfg -reassociate "
 578                    "-mem2reg -constprop -instcombine -gvn",
 579                    filename, gallivm_debug & GALLIVM_PERF_NO_OPT ? 0 : 2,
 580                    "[-mcpu=<-mcpu option>] ",
 581                    "[-mattr=<-mattr option(s)>]");
 582    }
 583
 584    if (gallivm_debug & GALLIVM_DEBUG_PERF)
 585       time_begin = os_time_get();
 586
 587 #if GALLIVM_HAVE_CORO
 588    LLVMRunPassManager(gallivm->cgpassmgr, gallivm->module);
 589 #endif
 590    /* Run optimization passes */
 591    LLVMInitializeFunctionPassManager(gallivm->passmgr);
 592    func = LLVMGetFirstFunction(gallivm->module);
 593    while (func) {
 594       if (0) {
 595          debug_printf("optimizing func %s...\n", LLVMGetValueName(func));
 596       }
 597
 598    /* Disable frame pointer omission on debug/profile builds */
 599    /* XXX: And workaround http://llvm.org/PR21435 */
 600 #if defined(DEBUG) || defined(PROFILE) || defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
 601       LLVMAddTargetDependentFunctionAttr(func, "no-frame-pointer-elim", "true");
 602       LLVMAddTargetDependentFunctionAttr(func, "no-frame-pointer-elim-non-leaf", "true");
 603 #endif
 604
 605       LLVMRunFunctionPassManager(gallivm->passmgr, func);
 606       func = LLVMGetNextFunction(func);
 607    }
 608    LLVMFinalizeFunctionPassManager(gallivm->passmgr);
 609
 610    if (gallivm_debug & GALLIVM_DEBUG_PERF) {
 611       int64_t time_end = os_time_get();
 612       int time_msec = (int)((time_end - time_begin) / 1000);
 613       assert(gallivm->module_name);
 614       debug_printf("optimizing module %s took %d msec\n",
 615                    gallivm->module_name, time_msec);
 616    }
 617
 618    /* Setting the module's DataLayout to an empty string will cause the
 619     * ExecutionEngine to copy to the DataLayout string from its target machine
 620     * to the module.  As of LLVM 3.8 the module and the execution engine are
 621     * required to have the same DataLayout.
 622     *
 623     * We must make sure we do this after running the optimization passes,
 624     * because those passes need a correct datalayout string.  For example, if
 625     * those optimization passes see an empty datalayout, they will assume this
 626     * is a little endian target and will do optimizations that break big endian
 627     * machines.
 628     *
 629     * TODO: This is just a temporary work-around.  The correct solution is for
 630     * gallivm_init_state() to create a TargetMachine and pull the DataLayout
 631     * from there.  Currently, the TargetMachine used by llvmpipe is being
 632     * implicitly created by the EngineBuilder in
 633     * lp_build_create_jit_compiler_for_module()
 634     */
 635    LLVMSetDataLayout(gallivm->module, "");
 636    assert(!gallivm->engine);
 637    if (!init_gallivm_engine(gallivm)) {
 638       assert(0);
 639    }
 640    assert(gallivm->engine);
 641
 642    ++gallivm->compiled;
 643
 644    if (gallivm_debug & GALLIVM_DEBUG_ASM) {
 645       LLVMValueRef llvm_func = LLVMGetFirstFunction(gallivm->module);
 646
 647       while (llvm_func) {
 648          /*
 649           * Need to filter out functions which don't have an implementation,
 650           * such as the intrinsics. May not be sufficient in case of IPO?
 651           * LLVMGetPointerToGlobal() will abort otherwise.
 652           */
 653          if (!LLVMIsDeclaration(llvm_func)) {
 654             void *func_code = LLVMGetPointerToGlobal(gallivm->engine, llvm_func);
 655             lp_disassemble(llvm_func, func_code);
 656          }
 657          llvm_func = LLVMGetNextFunction(llvm_func);
 658       }
 659    }
 660
 661 #if defined(PROFILE)
 662    {
 663       LLVMValueRef llvm_func = LLVMGetFirstFunction(gallivm->module);
 664
 665       while (llvm_func) {
 666          if (!LLVMIsDeclaration(llvm_func)) {
 667             void *func_code = LLVMGetPointerToGlobal(gallivm->engine, llvm_func);
 668             lp_profile(llvm_func, func_code);
 669          }
 670          llvm_func = LLVMGetNextFunction(llvm_func);
 671       }
 672    }
 673 #endif
 674 }
 675
 676
 677
 678 func_pointer
 679 gallivm_jit_function(struct gallivm_state *gallivm,
 680                      LLVMValueRef func)
 681 {
 682    void *code;
 683    func_pointer jit_func;
 684    int64_t time_begin = 0;
 685
 686    assert(gallivm->compiled);
 687    assert(gallivm->engine);
 688
 689    if (gallivm_debug & GALLIVM_DEBUG_PERF)
 690       time_begin = os_time_get();
 691
 692    code = LLVMGetPointerToGlobal(gallivm->engine, func);
 693    assert(code);
 694    jit_func = pointer_to_func(code);
 695
 696    if (gallivm_debug & GALLIVM_DEBUG_PERF) {
 697       int64_t time_end = os_time_get();
 698       int time_msec = (int)(time_end - time_begin) / 1000;
 699       debug_printf("   jitting func %s took %d msec\n",
 700                    LLVMGetValueName(func), time_msec);
 701    }
 702
 703    return jit_func;
 704 }