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 "os/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-c/Analysis.h>
  42 #include <llvm-c/Transforms/Scalar.h>
  43 #include <llvm-c/BitWriter.h>
  44
  45
  46 /* Only MCJIT is available as of LLVM SVN r216982 */
  47 #if HAVE_LLVM >= 0x0306
  48 #  define USE_MCJIT 1
  49 #elif defined(PIPE_ARCH_PPC_64) || defined(PIPE_ARCH_S390) || defined(PIPE_ARCH_ARM) || defined(PIPE_ARCH_AARCH64)
  50 #  define USE_MCJIT 1
  51 #else
  52 #  define USE_MCJIT 0
  53 #endif
  54
  55 #if USE_MCJIT
  56 void LLVMLinkInMCJIT();
  57 #endif
  58
  59 #ifdef DEBUG
  60 unsigned gallivm_debug = 0;
  61
  62 static const struct debug_named_value lp_bld_debug_flags[] = {
  63    { "tgsi",   GALLIVM_DEBUG_TGSI, NULL },
  64    { "ir",     GALLIVM_DEBUG_IR, NULL },
  65    { "asm",    GALLIVM_DEBUG_ASM, NULL },
  66    { "nopt",   GALLIVM_DEBUG_NO_OPT, NULL },
  67    { "perf",   GALLIVM_DEBUG_PERF, NULL },
  68    { "no_brilinear", GALLIVM_DEBUG_NO_BRILINEAR, NULL },
  69    { "no_rho_approx", GALLIVM_DEBUG_NO_RHO_APPROX, NULL },
  70    { "no_quad_lod", GALLIVM_DEBUG_NO_QUAD_LOD, NULL },
  71    { "gc",     GALLIVM_DEBUG_GC, NULL },
  72    DEBUG_NAMED_VALUE_END
  73 };
  74
  75 DEBUG_GET_ONCE_FLAGS_OPTION(gallivm_debug, "GALLIVM_DEBUG", lp_bld_debug_flags, 0)
  76 #endif
  77
  78
  79 static boolean gallivm_initialized = FALSE;
  80
  81 unsigned lp_native_vector_width;
  82
  83
  84 /*
  85  * Optimization values are:
  86  * - 0: None (-O0)
  87  * - 1: Less (-O1)
  88  * - 2: Default (-O2, -Os)
  89  * - 3: Aggressive (-O3)
  90  *
  91  * See also CodeGenOpt::Level in llvm/Target/TargetMachine.h
  92  */
  93 enum LLVM_CodeGenOpt_Level {
  94    None,        // -O0
  95    Less,        // -O1
  96    Default,     // -O2, -Os
  97    Aggressive   // -O3
  98 };
  99
 100
 101 /**
 102  * Create the LLVM (optimization) pass manager and install
 103  * relevant optimization passes.
 104  * \return  TRUE for success, FALSE for failure
 105  */
 106 static boolean
 107 create_pass_manager(struct gallivm_state *gallivm)
 108 {
 109    assert(!gallivm->passmgr);
 110    assert(gallivm->target);
 111
 112    gallivm->passmgr = LLVMCreateFunctionPassManagerForModule(gallivm->module);
 113    if (!gallivm->passmgr)
 114       return FALSE;
 115    /*
 116     * TODO: some per module pass manager with IPO passes might be helpful -
 117     * the generated texture functions may benefit from inlining if they are
 118     * simple, or constant propagation into them, etc.
 119     */
 120
 121    // Old versions of LLVM get the DataLayout from the pass manager.
 122    LLVMAddTargetData(gallivm->target, gallivm->passmgr);
 123
 124    /* Setting the module's DataLayout to an empty string will cause the
 125     * ExecutionEngine to copy to the DataLayout string from its target
 126     * machine to the module.  As of LLVM 3.8 the module and the execution
 127     * engine are required to have the same DataLayout.
 128     *
 129     * TODO: This is just a temporary work-around.  The correct solution is
 130     * for gallivm_init_state() to create a TargetMachine and pull the
 131     * DataLayout from there.  Currently, the TargetMachine used by llvmpipe
 132     * is being implicitly created by the EngineBuilder in
 133     * lp_build_create_jit_compiler_for_module()
 134     */
 135
 136 #if HAVE_LLVM < 0x0308
 137    {
 138       char *td_str;
 139       // New ones from the Module.
 140       td_str = LLVMCopyStringRepOfTargetData(gallivm->target);
 141       LLVMSetDataLayout(gallivm->module, td_str);
 142       free(td_str);
 143    }
 144 #else
 145    LLVMSetDataLayout(gallivm->module, "");
 146 #endif
 147
 148    if ((gallivm_debug & GALLIVM_DEBUG_NO_OPT) == 0) {
 149       /* These are the passes currently listed in llvm-c/Transforms/Scalar.h,
 150        * but there are more on SVN.
 151        * TODO: Add more passes.
 152        */
 153       LLVMAddScalarReplAggregatesPass(gallivm->passmgr);
 154       LLVMAddLICMPass(gallivm->passmgr);
 155       LLVMAddCFGSimplificationPass(gallivm->passmgr);
 156       LLVMAddReassociatePass(gallivm->passmgr);
 157       LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
 158       LLVMAddConstantPropagationPass(gallivm->passmgr);
 159       LLVMAddInstructionCombiningPass(gallivm->passmgr);
 160       LLVMAddGVNPass(gallivm->passmgr);
 161    }
 162    else {
 163       /* We need at least this pass to prevent the backends to fail in
 164        * unexpected ways.
 165        */
 166       LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
 167    }
 168
 169    return TRUE;
 170 }
 171
 172
 173 /**
 174  * Free gallivm object's LLVM allocations, but not any generated code
 175  * nor the gallivm object itself.
 176  */
 177 void
 178 gallivm_free_ir(struct gallivm_state *gallivm)
 179 {
 180    if (gallivm->passmgr) {
 181       LLVMDisposePassManager(gallivm->passmgr);
 182    }
 183
 184    if (gallivm->engine) {
 185       /* This will already destroy any associated module */
 186       LLVMDisposeExecutionEngine(gallivm->engine);
 187    } else if (gallivm->module) {
 188       LLVMDisposeModule(gallivm->module);
 189    }
 190
 191 #if !USE_MCJIT
 192    /* Don't free the TargetData, it's owned by the exec engine */
 193 #else
 194    if (gallivm->target) {
 195       LLVMDisposeTargetData(gallivm->target);
 196    }
 197 #endif
 198
 199    if (gallivm->builder)
 200       LLVMDisposeBuilder(gallivm->builder);
 201
 202    /* The LLVMContext should be owned by the parent of gallivm. */
 203
 204    gallivm->engine = NULL;
 205    gallivm->target = NULL;
 206    gallivm->module = NULL;
 207    gallivm->passmgr = NULL;
 208    gallivm->context = NULL;
 209    gallivm->builder = NULL;
 210 }
 211
 212
 213 /**
 214  * Free LLVM-generated code.  Should be done AFTER gallivm_free_ir().
 215  */
 216 static void
 217 gallivm_free_code(struct gallivm_state *gallivm)
 218 {
 219    assert(!gallivm->module);
 220    assert(!gallivm->engine);
 221    lp_free_generated_code(gallivm->code);
 222    gallivm->code = NULL;
 223    lp_free_memory_manager(gallivm->memorymgr);
 224    gallivm->memorymgr = NULL;
 225 }
 226
 227
 228 static boolean
 229 init_gallivm_engine(struct gallivm_state *gallivm)
 230 {
 231    if (1) {
 232       enum LLVM_CodeGenOpt_Level optlevel;
 233       char *error = NULL;
 234       int ret;
 235
 236       if (gallivm_debug & GALLIVM_DEBUG_NO_OPT) {
 237          optlevel = None;
 238       }
 239       else {
 240          optlevel = Default;
 241       }
 242
 243       ret = lp_build_create_jit_compiler_for_module(&gallivm->engine,
 244                                                     &gallivm->code,
 245                                                     gallivm->module,
 246                                                     gallivm->memorymgr,
 247                                                     (unsigned) optlevel,
 248                                                     USE_MCJIT,
 249                                                     &error);
 250       if (ret) {
 251          _debug_printf("%s\n", error);
 252          LLVMDisposeMessage(error);
 253          goto fail;
 254       }
 255    }
 256
 257 #if !USE_MCJIT
 258    gallivm->target = LLVMGetExecutionEngineTargetData(gallivm->engine);
 259    if (!gallivm->target)
 260       goto fail;
 261 #else
 262    if (0) {
 263        /*
 264         * Dump the data layout strings.
 265         */
 266
 267        LLVMTargetDataRef target = LLVMGetExecutionEngineTargetData(gallivm->engine);
 268        char *data_layout;
 269        char *engine_data_layout;
 270
 271        data_layout = LLVMCopyStringRepOfTargetData(gallivm->target);
 272        engine_data_layout = LLVMCopyStringRepOfTargetData(target);
 273
 274        if (1) {
 275           debug_printf("module target data = %s\n", data_layout);
 276           debug_printf("engine target data = %s\n", engine_data_layout);
 277        }
 278
 279        free(data_layout);
 280        free(engine_data_layout);
 281    }
 282 #endif
 283
 284    return TRUE;
 285
 286 fail:
 287    return FALSE;
 288 }
 289
 290
 291 /**
 292  * Allocate gallivm LLVM objects.
 293  * \return  TRUE for success, FALSE for failure
 294  */
 295 static boolean
 296 init_gallivm_state(struct gallivm_state *gallivm, const char *name,
 297                    LLVMContextRef context)
 298 {
 299    assert(!gallivm->context);
 300    assert(!gallivm->module);
 301
 302    if (!lp_build_init())
 303       return FALSE;
 304
 305    gallivm->context = context;
 306
 307    if (!gallivm->context)
 308       goto fail;
 309
 310    gallivm->module = LLVMModuleCreateWithNameInContext(name,
 311                                                        gallivm->context);
 312    if (!gallivm->module)
 313       goto fail;
 314
 315    gallivm->builder = LLVMCreateBuilderInContext(gallivm->context);
 316    if (!gallivm->builder)
 317       goto fail;
 318
 319    gallivm->memorymgr = lp_get_default_memory_manager();
 320    if (!gallivm->memorymgr)
 321       goto fail;
 322
 323    /* FIXME: MC-JIT only allows compiling one module at a time, and it must be
 324     * complete when MC-JIT is created. So defer the MC-JIT engine creation for
 325     * now.
 326     */
 327 #if !USE_MCJIT
 328    if (!init_gallivm_engine(gallivm)) {
 329       goto fail;
 330    }
 331 #else
 332    /*
 333     * MC-JIT engine compiles the module immediately on creation, so we can't
 334     * obtain the target data from it.  Instead we create a target data layout
 335     * from a string.
 336     *
 337     * The produced layout strings are not precisely the same, but should make
 338     * no difference for the kind of optimization passes we run.
 339     *
 340     * For reference this is the layout string on x64:
 341     *
 342     *   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
 343     *
 344     * See also:
 345     * - http://llvm.org/docs/LangRef.html#datalayout
 346     */
 347
 348    {
 349       const unsigned pointer_size = 8 * sizeof(void *);
 350       char layout[512];
 351       util_snprintf(layout, sizeof layout, "%c-p:%u:%u:%u-i64:64:64-a0:0:%u-s0:%u:%u",
 352 #ifdef PIPE_ARCH_LITTLE_ENDIAN
 353                     'e', // little endian
 354 #else
 355                     'E', // big endian
 356 #endif
 357                     pointer_size, pointer_size, pointer_size, // pointer size, abi alignment, preferred alignment
 358                     pointer_size, // aggregate preferred alignment
 359                     pointer_size, pointer_size); // stack objects abi alignment, preferred alignment
 360
 361       gallivm->target = LLVMCreateTargetData(layout);
 362       if (!gallivm->target) {
 363          return FALSE;
 364       }
 365    }
 366 #endif
 367
 368    if (!create_pass_manager(gallivm))
 369       goto fail;
 370
 371    return TRUE;
 372
 373 fail:
 374    gallivm_free_ir(gallivm);
 375    gallivm_free_code(gallivm);
 376    return FALSE;
 377 }
 378
 379
 380 boolean
 381 lp_build_init(void)
 382 {
 383    if (gallivm_initialized)
 384       return TRUE;
 385
 386 #ifdef DEBUG
 387    gallivm_debug = debug_get_option_gallivm_debug();
 388 #endif
 389
 390    lp_set_target_options();
 391
 392 #if USE_MCJIT
 393    LLVMLinkInMCJIT();
 394 #else
 395    LLVMLinkInJIT();
 396 #endif
 397
 398    util_cpu_detect();
 399
 400    /* AMD Bulldozer AVX's throughput is the same as SSE2; and because using
 401     * 8-wide vector needs more floating ops than 4-wide (due to padding), it is
 402     * actually more efficient to use 4-wide vectors on this processor.
 403     *
 404     * See also:
 405     * - http://www.anandtech.com/show/4955/the-bulldozer-review-amd-fx8150-tested/2
 406     */
 407    if (util_cpu_caps.has_avx &&
 408        util_cpu_caps.has_intel) {
 409       lp_native_vector_width = 256;
 410    } else {
 411       /* Leave it at 128, even when no SIMD extensions are available.
 412        * Really needs to be a multiple of 128 so can fit 4 floats.
 413        */
 414       lp_native_vector_width = 128;
 415    }
 416
 417    lp_native_vector_width = debug_get_num_option("LP_NATIVE_VECTOR_WIDTH",
 418                                                  lp_native_vector_width);
 419
 420    if (lp_native_vector_width <= 128) {
 421       /* Hide AVX support, as often LLVM AVX intrinsics are only guarded by
 422        * "util_cpu_caps.has_avx" predicate, and lack the
 423        * "lp_native_vector_width > 128" predicate. And also to ensure a more
 424        * consistent behavior, allowing one to test SSE2 on AVX machines.
 425        * XXX: should not play games with util_cpu_caps directly as it might
 426        * get used for other things outside llvm too.
 427        */
 428       util_cpu_caps.has_avx = 0;
 429       util_cpu_caps.has_avx2 = 0;
 430    }
 431
 432 #ifdef PIPE_ARCH_PPC_64
 433    /* Set the NJ bit in VSCR to 0 so denormalized values are handled as
 434     * specified by IEEE standard (PowerISA 2.06 - Section 6.3). This guarantees
 435     * that some rounding and half-float to float handling does not round
 436     * incorrectly to 0.
 437     * XXX: should eventually follow same logic on all platforms.
 438     * Right now denorms get explicitly disabled (but elsewhere) for x86,
 439     * whereas ppc64 explicitly enables them...
 440     */
 441    if (util_cpu_caps.has_altivec) {
 442       unsigned short mask[] = { 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF,
 443                                 0xFFFF, 0xFFFF, 0xFFFE, 0xFFFF };
 444       __asm (
 445         "mfvscr %%v1\n"
 446         "vand   %0,%%v1,%0\n"
 447         "mtvscr %0"
 448         :
 449         : "r" (*mask)
 450       );
 451    }
 452 #endif
 453
 454    gallivm_initialized = TRUE;
 455
 456 #if 0
 457    /* For simulating less capable machines */
 458    util_cpu_caps.has_sse3 = 0;
 459    util_cpu_caps.has_ssse3 = 0;
 460    util_cpu_caps.has_sse4_1 = 0;
 461    util_cpu_caps.has_avx = 0;
 462    util_cpu_caps.has_f16c = 0;
 463 #endif
 464
 465    return TRUE;
 466 }
 467
 468
 469
 470 /**
 471  * Create a new gallivm_state object.
 472  */
 473 struct gallivm_state *
 474 gallivm_create(const char *name, LLVMContextRef context)
 475 {
 476    struct gallivm_state *gallivm;
 477
 478    gallivm = CALLOC_STRUCT(gallivm_state);
 479    if (gallivm) {
 480       if (!init_gallivm_state(gallivm, name, context)) {
 481          FREE(gallivm);
 482          gallivm = NULL;
 483       }
 484    }
 485
 486    return gallivm;
 487 }
 488
 489
 490 /**
 491  * Destroy a gallivm_state object.
 492  */
 493 void
 494 gallivm_destroy(struct gallivm_state *gallivm)
 495 {
 496    gallivm_free_ir(gallivm);
 497    gallivm_free_code(gallivm);
 498    FREE(gallivm);
 499 }
 500
 501
 502 /**
 503  * Validate a function.
 504  * Verification is only done with debug builds.
 505  */
 506 void
 507 gallivm_verify_function(struct gallivm_state *gallivm,
 508                         LLVMValueRef func)
 509 {
 510    /* Verify the LLVM IR.  If invalid, dump and abort */
 511 #ifdef DEBUG
 512    if (LLVMVerifyFunction(func, LLVMPrintMessageAction)) {
 513       lp_debug_dump_value(func);
 514       assert(0);
 515       return;
 516    }
 517 #endif
 518
 519    if (gallivm_debug & GALLIVM_DEBUG_IR) {
 520       /* Print the LLVM IR to stderr */
 521       lp_debug_dump_value(func);
 522       debug_printf("\n");
 523    }
 524 }
 525
 526
 527 /**
 528  * Compile a module.
 529  * This does IR optimization on all functions in the module.
 530  */
 531 void
 532 gallivm_compile_module(struct gallivm_state *gallivm)
 533 {
 534    LLVMValueRef func;
 535    int64_t time_begin = 0;
 536
 537    assert(!gallivm->compiled);
 538
 539    if (gallivm->builder) {
 540       LLVMDisposeBuilder(gallivm->builder);
 541       gallivm->builder = NULL;
 542    }
 543
 544    if (gallivm_debug & GALLIVM_DEBUG_PERF)
 545       time_begin = os_time_get();
 546
 547    /* Run optimization passes */
 548    LLVMInitializeFunctionPassManager(gallivm->passmgr);
 549    func = LLVMGetFirstFunction(gallivm->module);
 550    while (func) {
 551       if (0) {
 552          debug_printf("optimizing func %s...\n", LLVMGetValueName(func));
 553       }
 554
 555    /* Disable frame pointer omission on debug/profile builds */
 556    /* XXX: And workaround http://llvm.org/PR21435 */
 557 #if HAVE_LLVM >= 0x0307 && \
 558     (defined(DEBUG) || defined(PROFILE) || \
 559      defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64))
 560       LLVMAddTargetDependentFunctionAttr(func, "no-frame-pointer-elim", "true");
 561       LLVMAddTargetDependentFunctionAttr(func, "no-frame-pointer-elim-non-leaf", "true");
 562 #endif
 563
 564       LLVMRunFunctionPassManager(gallivm->passmgr, func);
 565       func = LLVMGetNextFunction(func);
 566    }
 567    LLVMFinalizeFunctionPassManager(gallivm->passmgr);
 568
 569    if (gallivm_debug & GALLIVM_DEBUG_PERF) {
 570       int64_t time_end = os_time_get();
 571       int time_msec = (int)(time_end - time_begin) / 1000;
 572       debug_printf("optimizing module %s took %d msec\n",
 573                    lp_get_module_id(gallivm->module), time_msec);
 574    }
 575
 576    /* Dump byte code to a file */
 577    if (0) {
 578       LLVMWriteBitcodeToFile(gallivm->module, "llvmpipe.bc");
 579       debug_printf("llvmpipe.bc written\n");
 580       debug_printf("Invoke as \"llc -o - llvmpipe.bc\"\n");
 581    }
 582
 583 #if USE_MCJIT
 584    assert(!gallivm->engine);
 585    if (!init_gallivm_engine(gallivm)) {
 586       assert(0);
 587    }
 588 #endif
 589    assert(gallivm->engine);
 590
 591    ++gallivm->compiled;
 592
 593    if (gallivm_debug & GALLIVM_DEBUG_ASM) {
 594       LLVMValueRef llvm_func = LLVMGetFirstFunction(gallivm->module);
 595
 596       while (llvm_func) {
 597          /*
 598           * Need to filter out functions which don't have an implementation,
 599           * such as the intrinsics. May not be sufficient in case of IPO?
 600           * LLVMGetPointerToGlobal() will abort otherwise.
 601           */
 602          if (!LLVMIsDeclaration(llvm_func)) {
 603             void *func_code = LLVMGetPointerToGlobal(gallivm->engine, llvm_func);
 604             lp_disassemble(llvm_func, func_code);
 605          }
 606          llvm_func = LLVMGetNextFunction(llvm_func);
 607       }
 608    }
 609
 610 #if defined(PROFILE)
 611    {
 612       LLVMValueRef llvm_func = LLVMGetFirstFunction(gallivm->module);
 613
 614       while (llvm_func) {
 615          if (!LLVMIsDeclaration(llvm_func)) {
 616             void *func_code = LLVMGetPointerToGlobal(gallivm->engine, llvm_func);
 617             lp_profile(llvm_func, func_code);
 618          }
 619          llvm_func = LLVMGetNextFunction(llvm_func);
 620       }
 621    }
 622 #endif
 623 }
 624
 625
 626
 627 func_pointer
 628 gallivm_jit_function(struct gallivm_state *gallivm,
 629                      LLVMValueRef func)
 630 {
 631    void *code;
 632    func_pointer jit_func;
 633
 634    assert(gallivm->compiled);
 635    assert(gallivm->engine);
 636
 637    code = LLVMGetPointerToGlobal(gallivm->engine, func);
 638    assert(code);
 639    jit_func = pointer_to_func(code);
 640
 641    return jit_func;
 642 }