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 /* Only MCJIT is available as of LLVM SVN r216982 */
  53 #if LLVM_VERSION_MAJOR > 3 || (LLVM_VERSION_MAJOR == 3 && LLVM_VERSION_MINOR >= 6)
  54 #  define USE_MCJIT 1
  55 #elif defined(PIPE_ARCH_PPC_64) || defined(PIPE_ARCH_S390) || defined(PIPE_ARCH_ARM) || defined(PIPE_ARCH_AARCH64)
  56 #  define USE_MCJIT 1
  57 #endif
  58
  59 #if defined(USE_MCJIT)
  60 static const bool use_mcjit = USE_MCJIT;
  61 #else
  62 static bool use_mcjit = FALSE;
  63 #endif
  64
  65 unsigned gallivm_perf = 0;
  66
  67 static const struct debug_named_value lp_bld_perf_flags[] = {
  68    { "no_brilinear", GALLIVM_PERF_NO_BRILINEAR, "disable brilinear optimization" },
  69    { "no_rho_approx", GALLIVM_PERF_NO_RHO_APPROX, "disable rho_approx optimization" },
  70    { "no_quad_lod", GALLIVM_PERF_NO_QUAD_LOD, "disable quad_lod optimization" },
  71    { "no_aos_sampling", GALLIVM_PERF_NO_AOS_SAMPLING, "disable aos sampling optimization" },
  72    { "nopt",   GALLIVM_PERF_NO_OPT, "disable optimization passes to speed up shader compilation" },
  73    { "no_filter_hacks", GALLIVM_PERF_NO_BRILINEAR | GALLIVM_PERF_NO_RHO_APPROX |
  74      GALLIVM_PERF_NO_QUAD_LOD, "disable filter optimization hacks" },
  75    DEBUG_NAMED_VALUE_END
  76 };
  77
  78 #ifdef DEBUG
  79 unsigned gallivm_debug = 0;
  80
  81 static const struct debug_named_value lp_bld_debug_flags[] = {
  82    { "tgsi",   GALLIVM_DEBUG_TGSI, NULL },
  83    { "ir",     GALLIVM_DEBUG_IR, NULL },
  84    { "asm",    GALLIVM_DEBUG_ASM, NULL },
  85    { "perf",   GALLIVM_DEBUG_PERF, NULL },
  86    { "gc",     GALLIVM_DEBUG_GC, NULL },
  87    { "dumpbc", GALLIVM_DEBUG_DUMP_BC, NULL },
  88    DEBUG_NAMED_VALUE_END
  89 };
  90
  91 DEBUG_GET_ONCE_FLAGS_OPTION(gallivm_debug, "GALLIVM_DEBUG", lp_bld_debug_flags, 0)
  92 #endif
  93
  94
  95 static boolean gallivm_initialized = FALSE;
  96
  97 unsigned lp_native_vector_width;
  98
  99
 100 /*
 101  * Optimization values are:
 102  * - 0: None (-O0)
 103  * - 1: Less (-O1)
 104  * - 2: Default (-O2, -Os)
 105  * - 3: Aggressive (-O3)
 106  *
 107  * See also CodeGenOpt::Level in llvm/Target/TargetMachine.h
 108  */
 109 enum LLVM_CodeGenOpt_Level {
 110    None,        // -O0
 111    Less,        // -O1
 112    Default,     // -O2, -Os
 113    Aggressive   // -O3
 114 };
 115
 116
 117 /**
 118  * Create the LLVM (optimization) pass manager and install
 119  * relevant optimization passes.
 120  * \return  TRUE for success, FALSE for failure
 121  */
 122 static boolean
 123 create_pass_manager(struct gallivm_state *gallivm)
 124 {
 125    assert(!gallivm->passmgr);
 126    assert(gallivm->target);
 127
 128    gallivm->passmgr = LLVMCreateFunctionPassManagerForModule(gallivm->module);
 129    if (!gallivm->passmgr)
 130       return FALSE;
 131
 132 #if GALLIVM_HAVE_CORO
 133    gallivm->cgpassmgr = LLVMCreatePassManager();
 134 #endif
 135    /*
 136     * TODO: some per module pass manager with IPO passes might be helpful -
 137     * the generated texture functions may benefit from inlining if they are
 138     * simple, or constant propagation into them, etc.
 139     */
 140
 141 #if LLVM_VERSION_MAJOR == 3 && LLVM_VERSION_MINOR < 9
 142    // Old versions of LLVM get the DataLayout from the pass manager.
 143    LLVMAddTargetData(gallivm->target, gallivm->passmgr);
 144 #endif
 145
 146    {
 147       char *td_str;
 148       // New ones from the Module.
 149       td_str = LLVMCopyStringRepOfTargetData(gallivm->target);
 150       LLVMSetDataLayout(gallivm->module, td_str);
 151       free(td_str);
 152    }
 153
 154 #if GALLIVM_HAVE_CORO
 155    LLVMAddCoroEarlyPass(gallivm->cgpassmgr);
 156    LLVMAddCoroSplitPass(gallivm->cgpassmgr);
 157    LLVMAddCoroElidePass(gallivm->cgpassmgr);
 158 #endif
 159
 160    if ((gallivm_perf & GALLIVM_PERF_NO_OPT) == 0) {
 161       /*
 162        * TODO: Evaluate passes some more - keeping in mind
 163        * both quality of generated code and compile times.
 164        */
 165       /*
 166        * NOTE: if you change this, don't forget to change the output
 167        * with GALLIVM_DEBUG_DUMP_BC in gallivm_compile_module.
 168        */
 169       LLVMAddScalarReplAggregatesPass(gallivm->passmgr);
 170       LLVMAddEarlyCSEPass(gallivm->passmgr);
 171       LLVMAddCFGSimplificationPass(gallivm->passmgr);
 172       /*
 173        * FIXME: LICM is potentially quite useful. However, for some
 174        * rather crazy shaders the compile time can reach _hours_ per shader,
 175        * due to licm implying lcssa (since llvm 3.5), which can take forever.
 176        * Even for sane shaders, the cost of licm is rather high (and not just
 177        * due to lcssa, licm itself too), though mostly only in cases when it
 178        * can actually move things, so having to disable it is a pity.
 179        * LLVMAddLICMPass(gallivm->passmgr);
 180        */
 181       LLVMAddReassociatePass(gallivm->passmgr);
 182       LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
 183       LLVMAddConstantPropagationPass(gallivm->passmgr);
 184       LLVMAddInstructionCombiningPass(gallivm->passmgr);
 185       LLVMAddGVNPass(gallivm->passmgr);
 186 #if GALLIVM_HAVE_CORO
 187       LLVMAddCoroCleanupPass(gallivm->passmgr);
 188 #endif
 189    }
 190    else {
 191       /* We need at least this pass to prevent the backends to fail in
 192        * unexpected ways.
 193        */
 194       LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
 195    }
 196
 197    return TRUE;
 198 }
 199
 200
 201 /**
 202  * Free gallivm object's LLVM allocations, but not any generated code
 203  * nor the gallivm object itself.
 204  */
 205 void
 206 gallivm_free_ir(struct gallivm_state *gallivm)
 207 {
 208    if (gallivm->passmgr) {
 209       LLVMDisposePassManager(gallivm->passmgr);
 210    }
 211
 212 #if GALLIVM_HAVE_CORO
 213    if (gallivm->cgpassmgr) {
 214       LLVMDisposePassManager(gallivm->cgpassmgr);
 215    }
 216 #endif
 217
 218    if (gallivm->engine) {
 219       /* This will already destroy any associated module */
 220       LLVMDisposeExecutionEngine(gallivm->engine);
 221    } else if (gallivm->module) {
 222       LLVMDisposeModule(gallivm->module);
 223    }
 224
 225    FREE(gallivm->module_name);
 226
 227    if (!use_mcjit) {
 228       /* Don't free the TargetData, it's owned by the exec engine */
 229    } else {
 230       if (gallivm->target) {
 231          LLVMDisposeTargetData(gallivm->target);
 232       }
 233    }
 234
 235    if (gallivm->builder)
 236       LLVMDisposeBuilder(gallivm->builder);
 237
 238    /* The LLVMContext should be owned by the parent of gallivm. */
 239
 240    gallivm->engine = NULL;
 241    gallivm->target = NULL;
 242    gallivm->module = NULL;
 243    gallivm->module_name = NULL;
 244    gallivm->cgpassmgr = NULL;
 245    gallivm->passmgr = NULL;
 246    gallivm->context = NULL;
 247    gallivm->builder = NULL;
 248 }
 249
 250
 251 /**
 252  * Free LLVM-generated code.  Should be done AFTER gallivm_free_ir().
 253  */
 254 static void
 255 gallivm_free_code(struct gallivm_state *gallivm)
 256 {
 257    assert(!gallivm->module);
 258    assert(!gallivm->engine);
 259    lp_free_generated_code(gallivm->code);
 260    gallivm->code = NULL;
 261    lp_free_memory_manager(gallivm->memorymgr);
 262    gallivm->memorymgr = NULL;
 263 }
 264
 265
 266 static boolean
 267 init_gallivm_engine(struct gallivm_state *gallivm)
 268 {
 269    if (1) {
 270       enum LLVM_CodeGenOpt_Level optlevel;
 271       char *error = NULL;
 272       int ret;
 273
 274       if (gallivm_perf & GALLIVM_PERF_NO_OPT) {
 275          optlevel = None;
 276       }
 277       else {
 278          optlevel = Default;
 279       }
 280
 281       ret = lp_build_create_jit_compiler_for_module(&gallivm->engine,
 282                                                     &gallivm->code,
 283                                                     gallivm->module,
 284                                                     gallivm->memorymgr,
 285                                                     (unsigned) optlevel,
 286                                                     use_mcjit,
 287                                                     &error);
 288       if (ret) {
 289          _debug_printf("%s\n", error);
 290          LLVMDisposeMessage(error);
 291          goto fail;
 292       }
 293    }
 294
 295    if (!use_mcjit) {
 296       gallivm->target = LLVMGetExecutionEngineTargetData(gallivm->engine);
 297       if (!gallivm->target)
 298          goto fail;
 299    } else {
 300       if (0) {
 301           /*
 302            * Dump the data layout strings.
 303            */
 304
 305           LLVMTargetDataRef target = LLVMGetExecutionEngineTargetData(gallivm->engine);
 306           char *data_layout;
 307           char *engine_data_layout;
 308
 309           data_layout = LLVMCopyStringRepOfTargetData(gallivm->target);
 310           engine_data_layout = LLVMCopyStringRepOfTargetData(target);
 311
 312           if (1) {
 313              debug_printf("module target data = %s\n", data_layout);
 314              debug_printf("engine target data = %s\n", engine_data_layout);
 315           }
 316
 317           free(data_layout);
 318           free(engine_data_layout);
 319       }
 320    }
 321
 322    return TRUE;
 323
 324 fail:
 325    return FALSE;
 326 }
 327
 328
 329 /**
 330  * Allocate gallivm LLVM objects.
 331  * \return  TRUE for success, FALSE for failure
 332  */
 333 static boolean
 334 init_gallivm_state(struct gallivm_state *gallivm, const char *name,
 335                    LLVMContextRef context)
 336 {
 337    assert(!gallivm->context);
 338    assert(!gallivm->module);
 339
 340    if (!lp_build_init())
 341       return FALSE;
 342
 343    gallivm->context = context;
 344
 345    if (!gallivm->context)
 346       goto fail;
 347
 348    gallivm->module_name = NULL;
 349    if (name) {
 350       size_t size = strlen(name) + 1;
 351       gallivm->module_name = MALLOC(size);
 352       if (gallivm->module_name) {
 353          memcpy(gallivm->module_name, name, size);
 354       }
 355    }
 356
 357    gallivm->module = LLVMModuleCreateWithNameInContext(name,
 358                                                        gallivm->context);
 359    if (!gallivm->module)
 360       goto fail;
 361
 362    gallivm->builder = LLVMCreateBuilderInContext(gallivm->context);
 363    if (!gallivm->builder)
 364       goto fail;
 365
 366    gallivm->memorymgr = lp_get_default_memory_manager();
 367    if (!gallivm->memorymgr)
 368       goto fail;
 369
 370    /* FIXME: MC-JIT only allows compiling one module at a time, and it must be
 371     * complete when MC-JIT is created. So defer the MC-JIT engine creation for
 372     * now.
 373     */
 374    if (!use_mcjit) {
 375       if (!init_gallivm_engine(gallivm)) {
 376          goto fail;
 377       }
 378    } else {
 379       /*
 380        * MC-JIT engine compiles the module immediately on creation, so we can't
 381        * obtain the target data from it.  Instead we create a target data layout
 382        * from a string.
 383        *
 384        * The produced layout strings are not precisely the same, but should make
 385        * no difference for the kind of optimization passes we run.
 386        *
 387        * For reference this is the layout string on x64:
 388        *
 389        *   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
 390        *
 391        * See also:
 392        * - http://llvm.org/docs/LangRef.html#datalayout
 393        */
 394
 395       {
 396          const unsigned pointer_size = 8 * sizeof(void *);
 397          char layout[512];
 398          snprintf(layout, sizeof layout, "%c-p:%u:%u:%u-i64:64:64-a0:0:%u-s0:%u:%u",
 399 #ifdef PIPE_ARCH_LITTLE_ENDIAN
 400                        'e', // little endian
 401 #else
 402                        'E', // big endian
 403 #endif
 404                        pointer_size, pointer_size, pointer_size, // pointer size, abi alignment, preferred alignment
 405                        pointer_size, // aggregate preferred alignment
 406                        pointer_size, pointer_size); // stack objects abi alignment, preferred alignment
 407
 408          gallivm->target = LLVMCreateTargetData(layout);
 409          if (!gallivm->target) {
 410             return FALSE;
 411          }
 412       }
 413    }
 414
 415    if (!create_pass_manager(gallivm))
 416       goto fail;
 417
 418    return TRUE;
 419
 420 fail:
 421    gallivm_free_ir(gallivm);
 422    gallivm_free_code(gallivm);
 423    return FALSE;
 424 }
 425
 426
 427 boolean
 428 lp_build_init(void)
 429 {
 430    if (gallivm_initialized)
 431       return TRUE;
 432
 433
 434    /* LLVMLinkIn* are no-ops at runtime.  They just ensure the respective
 435     * component is linked at buildtime, which is sufficient for its static
 436     * constructors to be called at load time.
 437     */
 438 #if defined(USE_MCJIT)
 439 #  if USE_MCJIT
 440       LLVMLinkInMCJIT();
 441 #  else
 442       LLVMLinkInJIT();
 443 #  endif
 444 #else
 445    use_mcjit = debug_get_bool_option("GALLIVM_MCJIT", FALSE);
 446    LLVMLinkInJIT();
 447    LLVMLinkInMCJIT();
 448 #endif
 449
 450 #ifdef DEBUG
 451    gallivm_debug = debug_get_option_gallivm_debug();
 452 #endif
 453
 454    gallivm_perf = debug_get_flags_option("GALLIVM_PERF", lp_bld_perf_flags, 0 );
 455
 456    lp_set_target_options();
 457
 458    util_cpu_detect();
 459
 460    /* For simulating less capable machines */
 461 #ifdef DEBUG
 462    if (debug_get_bool_option("LP_FORCE_SSE2", FALSE)) {
 463       assert(util_cpu_caps.has_sse2);
 464       util_cpu_caps.has_sse3 = 0;
 465       util_cpu_caps.has_ssse3 = 0;
 466       util_cpu_caps.has_sse4_1 = 0;
 467       util_cpu_caps.has_sse4_2 = 0;
 468       util_cpu_caps.has_avx = 0;
 469       util_cpu_caps.has_avx2 = 0;
 470       util_cpu_caps.has_f16c = 0;
 471       util_cpu_caps.has_fma = 0;
 472    }
 473 #endif
 474
 475    /* AMD Bulldozer AVX's throughput is the same as SSE2; and because using
 476     * 8-wide vector needs more floating ops than 4-wide (due to padding), it is
 477     * actually more efficient to use 4-wide vectors on this processor.
 478     *
 479     * See also:
 480     * - http://www.anandtech.com/show/4955/the-bulldozer-review-amd-fx8150-tested/2
 481     */
 482    if (util_cpu_caps.has_avx &&
 483        util_cpu_caps.has_intel) {
 484       lp_native_vector_width = 256;
 485    } else {
 486       /* Leave it at 128, even when no SIMD extensions are available.
 487        * Really needs to be a multiple of 128 so can fit 4 floats.
 488        */
 489       lp_native_vector_width = 128;
 490    }
 491
 492    lp_native_vector_width = debug_get_num_option("LP_NATIVE_VECTOR_WIDTH",
 493                                                  lp_native_vector_width);
 494
 495    if (lp_native_vector_width <= 128) {
 496       /* Hide AVX support, as often LLVM AVX intrinsics are only guarded by
 497        * "util_cpu_caps.has_avx" predicate, and lack the
 498        * "lp_native_vector_width > 128" predicate. And also to ensure a more
 499        * consistent behavior, allowing one to test SSE2 on AVX machines.
 500        * XXX: should not play games with util_cpu_caps directly as it might
 501        * get used for other things outside llvm too.
 502        */
 503       util_cpu_caps.has_avx = 0;
 504       util_cpu_caps.has_avx2 = 0;
 505       util_cpu_caps.has_f16c = 0;
 506       util_cpu_caps.has_fma = 0;
 507    }
 508    if ((LLVM_VERSION_MAJOR == 3 && LLVM_VERSION_MINOR < 4) || !use_mcjit) {
 509       /* AVX2 support has only been tested with LLVM 3.4, and it requires
 510        * MCJIT. */
 511       util_cpu_caps.has_avx2 = 0;
 512    }
 513
 514 #ifdef PIPE_ARCH_PPC_64
 515    /* Set the NJ bit in VSCR to 0 so denormalized values are handled as
 516     * specified by IEEE standard (PowerISA 2.06 - Section 6.3). This guarantees
 517     * that some rounding and half-float to float handling does not round
 518     * incorrectly to 0.
 519     * XXX: should eventually follow same logic on all platforms.
 520     * Right now denorms get explicitly disabled (but elsewhere) for x86,
 521     * whereas ppc64 explicitly enables them...
 522     */
 523    if (util_cpu_caps.has_altivec) {
 524       unsigned short mask[] = { 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF,
 525                                 0xFFFF, 0xFFFF, 0xFFFE, 0xFFFF };
 526       __asm (
 527         "mfvscr %%v1\n"
 528         "vand   %0,%%v1,%0\n"
 529         "mtvscr %0"
 530         :
 531         : "r" (*mask)
 532       );
 533    }
 534 #endif
 535
 536    gallivm_initialized = TRUE;
 537
 538    return TRUE;
 539 }
 540
 541
 542
 543 /**
 544  * Create a new gallivm_state object.
 545  */
 546 struct gallivm_state *
 547 gallivm_create(const char *name, LLVMContextRef context)
 548 {
 549    struct gallivm_state *gallivm;
 550
 551    gallivm = CALLOC_STRUCT(gallivm_state);
 552    if (gallivm) {
 553       if (!init_gallivm_state(gallivm, name, context)) {
 554          FREE(gallivm);
 555          gallivm = NULL;
 556       }
 557    }
 558
 559    return gallivm;
 560 }
 561
 562
 563 /**
 564  * Destroy a gallivm_state object.
 565  */
 566 void
 567 gallivm_destroy(struct gallivm_state *gallivm)
 568 {
 569    gallivm_free_ir(gallivm);
 570    gallivm_free_code(gallivm);
 571    FREE(gallivm);
 572 }
 573
 574
 575 /**
 576  * Validate a function.
 577  * Verification is only done with debug builds.
 578  */
 579 void
 580 gallivm_verify_function(struct gallivm_state *gallivm,
 581                         LLVMValueRef func)
 582 {
 583    /* Verify the LLVM IR.  If invalid, dump and abort */
 584 #ifdef DEBUG
 585    if (LLVMVerifyFunction(func, LLVMPrintMessageAction)) {
 586       lp_debug_dump_value(func);
 587       assert(0);
 588       return;
 589    }
 590 #endif
 591
 592    if (gallivm_debug & GALLIVM_DEBUG_IR) {
 593       /* Print the LLVM IR to stderr */
 594       lp_debug_dump_value(func);
 595       debug_printf("\n");
 596    }
 597 }
 598
 599
 600 /**
 601  * Compile a module.
 602  * This does IR optimization on all functions in the module.
 603  */
 604 void
 605 gallivm_compile_module(struct gallivm_state *gallivm)
 606 {
 607    LLVMValueRef func;
 608    int64_t time_begin = 0;
 609
 610    assert(!gallivm->compiled);
 611
 612    if (gallivm->builder) {
 613       LLVMDisposeBuilder(gallivm->builder);
 614       gallivm->builder = NULL;
 615    }
 616
 617    /* Dump bitcode to a file */
 618    if (gallivm_debug & GALLIVM_DEBUG_DUMP_BC) {
 619       char filename[256];
 620       assert(gallivm->module_name);
 621       snprintf(filename, sizeof(filename), "ir_%s.bc", gallivm->module_name);
 622       LLVMWriteBitcodeToFile(gallivm->module, filename);
 623       debug_printf("%s written\n", filename);
 624       debug_printf("Invoke as \"opt %s %s | llc -O%d %s%s\"\n",
 625                    gallivm_debug & GALLIVM_PERF_NO_OPT ? "-mem2reg" :
 626                    "-sroa -early-cse -simplifycfg -reassociate "
 627                    "-mem2reg -constprop -instcombine -gvn",
 628                    filename, gallivm_debug & GALLIVM_PERF_NO_OPT ? 0 : 2,
 629                    (LLVM_VERSION_MAJOR > 3 || (LLVM_VERSION_MAJOR == 3 && LLVM_VERSION_MINOR >= 5)) ? "[-mcpu=<-mcpu option>] " : "",
 630                    "[-mattr=<-mattr option(s)>]");
 631    }
 632
 633    if (gallivm_debug & GALLIVM_DEBUG_PERF)
 634       time_begin = os_time_get();
 635
 636 #if GALLIVM_HAVE_CORO
 637    LLVMRunPassManager(gallivm->cgpassmgr, gallivm->module);
 638 #endif
 639    /* Run optimization passes */
 640    LLVMInitializeFunctionPassManager(gallivm->passmgr);
 641    func = LLVMGetFirstFunction(gallivm->module);
 642    while (func) {
 643       if (0) {
 644          debug_printf("optimizing func %s...\n", LLVMGetValueName(func));
 645       }
 646
 647    /* Disable frame pointer omission on debug/profile builds */
 648    /* XXX: And workaround http://llvm.org/PR21435 */
 649 #if (LLVM_VERSION_MAJOR > 3 || (LLVM_VERSION_MAJOR == 3 && LLVM_VERSION_MINOR >= 7)) && \
 650     (defined(DEBUG) || defined(PROFILE) || \
 651      defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64))
 652       LLVMAddTargetDependentFunctionAttr(func, "no-frame-pointer-elim", "true");
 653       LLVMAddTargetDependentFunctionAttr(func, "no-frame-pointer-elim-non-leaf", "true");
 654 #endif
 655
 656       LLVMRunFunctionPassManager(gallivm->passmgr, func);
 657       func = LLVMGetNextFunction(func);
 658    }
 659    LLVMFinalizeFunctionPassManager(gallivm->passmgr);
 660
 661    if (gallivm_debug & GALLIVM_DEBUG_PERF) {
 662       int64_t time_end = os_time_get();
 663       int time_msec = (int)((time_end - time_begin) / 1000);
 664       assert(gallivm->module_name);
 665       debug_printf("optimizing module %s took %d msec\n",
 666                    gallivm->module_name, time_msec);
 667    }
 668
 669    if (use_mcjit) {
 670       /* Setting the module's DataLayout to an empty string will cause the
 671        * ExecutionEngine to copy to the DataLayout string from its target
 672        * machine to the module.  As of LLVM 3.8 the module and the execution
 673        * engine are required to have the same DataLayout.
 674        *
 675        * We must make sure we do this after running the optimization passes,
 676        * because those passes need a correct datalayout string.  For example,
 677        * if those optimization passes see an empty datalayout, they will assume
 678        * this is a little endian target and will do optimizations that break big
 679        * endian machines.
 680        *
 681        * TODO: This is just a temporary work-around.  The correct solution is
 682        * for gallivm_init_state() to create a TargetMachine and pull the
 683        * DataLayout from there.  Currently, the TargetMachine used by llvmpipe
 684        * is being implicitly created by the EngineBuilder in
 685        * lp_build_create_jit_compiler_for_module()
 686        */
 687       LLVMSetDataLayout(gallivm->module, "");
 688       assert(!gallivm->engine);
 689       if (!init_gallivm_engine(gallivm)) {
 690          assert(0);
 691       }
 692    }
 693    assert(gallivm->engine);
 694
 695    ++gallivm->compiled;
 696
 697    if (gallivm_debug & GALLIVM_DEBUG_ASM) {
 698       LLVMValueRef llvm_func = LLVMGetFirstFunction(gallivm->module);
 699
 700       while (llvm_func) {
 701          /*
 702           * Need to filter out functions which don't have an implementation,
 703           * such as the intrinsics. May not be sufficient in case of IPO?
 704           * LLVMGetPointerToGlobal() will abort otherwise.
 705           */
 706          if (!LLVMIsDeclaration(llvm_func)) {
 707             void *func_code = LLVMGetPointerToGlobal(gallivm->engine, llvm_func);
 708             lp_disassemble(llvm_func, func_code);
 709          }
 710          llvm_func = LLVMGetNextFunction(llvm_func);
 711       }
 712    }
 713
 714 #if defined(PROFILE)
 715    {
 716       LLVMValueRef llvm_func = LLVMGetFirstFunction(gallivm->module);
 717
 718       while (llvm_func) {
 719          if (!LLVMIsDeclaration(llvm_func)) {
 720             void *func_code = LLVMGetPointerToGlobal(gallivm->engine, llvm_func);
 721             lp_profile(llvm_func, func_code);
 722          }
 723          llvm_func = LLVMGetNextFunction(llvm_func);
 724       }
 725    }
 726 #endif
 727 }
 728
 729
 730
 731 func_pointer
 732 gallivm_jit_function(struct gallivm_state *gallivm,
 733                      LLVMValueRef func)
 734 {
 735    void *code;
 736    func_pointer jit_func;
 737    int64_t time_begin = 0;
 738
 739    assert(gallivm->compiled);
 740    assert(gallivm->engine);
 741
 742    if (gallivm_debug & GALLIVM_DEBUG_PERF)
 743       time_begin = os_time_get();
 744
 745    code = LLVMGetPointerToGlobal(gallivm->engine, func);
 746    assert(code);
 747    jit_func = pointer_to_func(code);
 748
 749    if (gallivm_debug & GALLIVM_DEBUG_PERF) {
 750       int64_t time_end = os_time_get();
 751       int time_msec = (int)(time_end - time_begin) / 1000;
 752       debug_printf("   jitting func %s took %d msec\n",
 753                    LLVMGetValueName(func), time_msec);
 754    }
 755
 756    return jit_func;
 757 }