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/u_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
  49 #define USE_MCJIT 1
  50 #define HAVE_AVX 1
  51
  52 #else
  53
  54 /**
  55  * AVX is supported in:
  56  * - standard JIT from LLVM 3.2 onwards
  57  * - MC-JIT from LLVM 3.1
  58  *   - MC-JIT supports limited OSes (MacOSX and Linux)
  59  * - standard JIT in LLVM 3.1, with backports
  60  */
  61 #if defined(PIPE_ARCH_PPC_64) || defined(PIPE_ARCH_S390) || defined(PIPE_ARCH_ARM) || defined(PIPE_ARCH_AARCH64)
  62 #  define USE_MCJIT 1
  63 #  define HAVE_AVX 0
  64 #elif HAVE_LLVM >= 0x0302 || (HAVE_LLVM == 0x0301 && defined(HAVE_JIT_AVX_SUPPORT))
  65 #  define USE_MCJIT 0
  66 #  define HAVE_AVX 1
  67 #elif HAVE_LLVM == 0x0301 && (defined(PIPE_OS_LINUX) || defined(PIPE_OS_APPLE))
  68 #  define USE_MCJIT 1
  69 #  define HAVE_AVX 1
  70 #else
  71 #  define USE_MCJIT 0
  72 #  define HAVE_AVX 0
  73 #endif
  74
  75 #endif /* HAVE_LLVM >= 0x0306 */
  76
  77 #if USE_MCJIT
  78 void LLVMLinkInMCJIT();
  79 #endif
  80
  81 #ifdef DEBUG
  82 unsigned gallivm_debug = 0;
  83
  84 static const struct debug_named_value lp_bld_debug_flags[] = {
  85    { "tgsi",   GALLIVM_DEBUG_TGSI, NULL },
  86    { "ir",     GALLIVM_DEBUG_IR, NULL },
  87    { "asm",    GALLIVM_DEBUG_ASM, NULL },
  88    { "nopt",   GALLIVM_DEBUG_NO_OPT, NULL },
  89    { "perf",   GALLIVM_DEBUG_PERF, NULL },
  90    { "no_brilinear", GALLIVM_DEBUG_NO_BRILINEAR, NULL },
  91    { "no_rho_approx", GALLIVM_DEBUG_NO_RHO_APPROX, NULL },
  92    { "no_quad_lod", GALLIVM_DEBUG_NO_QUAD_LOD, NULL },
  93    { "gc",     GALLIVM_DEBUG_GC, NULL },
  94    DEBUG_NAMED_VALUE_END
  95 };
  96
  97 DEBUG_GET_ONCE_FLAGS_OPTION(gallivm_debug, "GALLIVM_DEBUG", lp_bld_debug_flags, 0)
  98 #endif
  99
 100
 101 static boolean gallivm_initialized = FALSE;
 102
 103 unsigned lp_native_vector_width;
 104
 105
 106 /*
 107  * Optimization values are:
 108  * - 0: None (-O0)
 109  * - 1: Less (-O1)
 110  * - 2: Default (-O2, -Os)
 111  * - 3: Aggressive (-O3)
 112  *
 113  * See also CodeGenOpt::Level in llvm/Target/TargetMachine.h
 114  */
 115 enum LLVM_CodeGenOpt_Level {
 116    None,        // -O0
 117    Less,        // -O1
 118    Default,     // -O2, -Os
 119    Aggressive   // -O3
 120 };
 121
 122
 123 /**
 124  * Create the LLVM (optimization) pass manager and install
 125  * relevant optimization passes.
 126  * \return  TRUE for success, FALSE for failure
 127  */
 128 static boolean
 129 create_pass_manager(struct gallivm_state *gallivm)
 130 {
 131    char *td_str;
 132    assert(!gallivm->passmgr);
 133    assert(gallivm->target);
 134
 135    gallivm->passmgr = LLVMCreateFunctionPassManagerForModule(gallivm->module);
 136    if (!gallivm->passmgr)
 137       return FALSE;
 138
 139    // Old versions of LLVM get the DataLayout from the pass manager.
 140    LLVMAddTargetData(gallivm->target, gallivm->passmgr);
 141
 142    // New ones from the Module.
 143    td_str = LLVMCopyStringRepOfTargetData(gallivm->target);
 144    LLVMSetDataLayout(gallivm->module, td_str);
 145    free(td_str);
 146
 147    if ((gallivm_debug & GALLIVM_DEBUG_NO_OPT) == 0) {
 148       /* These are the passes currently listed in llvm-c/Transforms/Scalar.h,
 149        * but there are more on SVN.
 150        * TODO: Add more passes.
 151        */
 152       LLVMAddScalarReplAggregatesPass(gallivm->passmgr);
 153       LLVMAddLICMPass(gallivm->passmgr);
 154       LLVMAddCFGSimplificationPass(gallivm->passmgr);
 155       LLVMAddReassociatePass(gallivm->passmgr);
 156       LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
 157       LLVMAddConstantPropagationPass(gallivm->passmgr);
 158       LLVMAddInstructionCombiningPass(gallivm->passmgr);
 159       LLVMAddGVNPass(gallivm->passmgr);
 160    }
 161    else {
 162       /* We need at least this pass to prevent the backends to fail in
 163        * unexpected ways.
 164        */
 165       LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
 166    }
 167
 168    return TRUE;
 169 }
 170
 171
 172 /**
 173  * Free gallivm object's LLVM allocations, but not any generated code
 174  * nor the gallivm object itself.
 175  */
 176 void
 177 gallivm_free_ir(struct gallivm_state *gallivm)
 178 {
 179    if (gallivm->passmgr) {
 180       LLVMDisposePassManager(gallivm->passmgr);
 181    }
 182
 183    if (gallivm->engine) {
 184       /* This will already destroy any associated module */
 185       LLVMDisposeExecutionEngine(gallivm->engine);
 186    } else if (gallivm->module) {
 187       LLVMDisposeModule(gallivm->module);
 188    }
 189
 190 #if !USE_MCJIT
 191    /* Don't free the TargetData, it's owned by the exec engine */
 192 #else
 193    if (gallivm->target) {
 194       LLVMDisposeTargetData(gallivm->target);
 195    }
 196 #endif
 197
 198    if (gallivm->builder)
 199       LLVMDisposeBuilder(gallivm->builder);
 200
 201    /* The LLVMContext should be owned by the parent of gallivm. */
 202
 203    gallivm->engine = NULL;
 204    gallivm->target = NULL;
 205    gallivm->module = NULL;
 206    gallivm->passmgr = NULL;
 207    gallivm->context = NULL;
 208    gallivm->builder = NULL;
 209 }
 210
 211
 212 /**
 213  * Free LLVM-generated code.  Should be done AFTER gallivm_free_ir().
 214  */
 215 static void
 216 gallivm_free_code(struct gallivm_state *gallivm)
 217 {
 218    assert(!gallivm->module);
 219    assert(!gallivm->engine);
 220    lp_free_generated_code(gallivm->code);
 221    gallivm->code = NULL;
 222 }
 223
 224
 225 static boolean
 226 init_gallivm_engine(struct gallivm_state *gallivm)
 227 {
 228    if (1) {
 229       enum LLVM_CodeGenOpt_Level optlevel;
 230       char *error = NULL;
 231       int ret;
 232
 233       if (gallivm_debug & GALLIVM_DEBUG_NO_OPT) {
 234          optlevel = None;
 235       }
 236       else {
 237          optlevel = Default;
 238       }
 239
 240       ret = lp_build_create_jit_compiler_for_module(&gallivm->engine,
 241                                                     &gallivm->code,
 242                                                     gallivm->module,
 243                                                     (unsigned) optlevel,
 244                                                     USE_MCJIT,
 245                                                     &error);
 246       if (ret) {
 247          _debug_printf("%s\n", error);
 248          LLVMDisposeMessage(error);
 249          goto fail;
 250       }
 251    }
 252
 253 #if !USE_MCJIT
 254    gallivm->target = LLVMGetExecutionEngineTargetData(gallivm->engine);
 255    if (!gallivm->target)
 256       goto fail;
 257 #else
 258    if (0) {
 259        /*
 260         * Dump the data layout strings.
 261         */
 262
 263        LLVMTargetDataRef target = LLVMGetExecutionEngineTargetData(gallivm->engine);
 264        char *data_layout;
 265        char *engine_data_layout;
 266
 267        data_layout = LLVMCopyStringRepOfTargetData(gallivm->target);
 268        engine_data_layout = LLVMCopyStringRepOfTargetData(target);
 269
 270        if (1) {
 271           debug_printf("module target data = %s\n", data_layout);
 272           debug_printf("engine target data = %s\n", engine_data_layout);
 273        }
 274
 275        free(data_layout);
 276        free(engine_data_layout);
 277    }
 278 #endif
 279
 280    return TRUE;
 281
 282 fail:
 283    return FALSE;
 284 }
 285
 286
 287 /**
 288  * Allocate gallivm LLVM objects.
 289  * \return  TRUE for success, FALSE for failure
 290  */
 291 static boolean
 292 init_gallivm_state(struct gallivm_state *gallivm, const char *name,
 293                    LLVMContextRef context)
 294 {
 295    assert(!gallivm->context);
 296    assert(!gallivm->module);
 297
 298    if (!lp_build_init())
 299       return FALSE;
 300
 301    gallivm->context = context;
 302
 303    if (!gallivm->context)
 304       goto fail;
 305
 306    gallivm->module = LLVMModuleCreateWithNameInContext(name,
 307                                                        gallivm->context);
 308    if (!gallivm->module)
 309       goto fail;
 310
 311    gallivm->builder = LLVMCreateBuilderInContext(gallivm->context);
 312    if (!gallivm->builder)
 313       goto fail;
 314
 315    /* FIXME: MC-JIT only allows compiling one module at a time, and it must be
 316     * complete when MC-JIT is created. So defer the MC-JIT engine creation for
 317     * now.
 318     */
 319 #if !USE_MCJIT
 320    if (!init_gallivm_engine(gallivm)) {
 321       goto fail;
 322    }
 323 #else
 324    /*
 325     * MC-JIT engine compiles the module immediately on creation, so we can't
 326     * obtain the target data from it.  Instead we create a target data layout
 327     * from a string.
 328     *
 329     * The produced layout strings are not precisely the same, but should make
 330     * no difference for the kind of optimization passes we run.
 331     *
 332     * For reference this is the layout string on x64:
 333     *
 334     *   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
 335     *
 336     * See also:
 337     * - http://llvm.org/docs/LangRef.html#datalayout
 338     */
 339
 340    {
 341       const unsigned pointer_size = 8 * sizeof(void *);
 342       char layout[512];
 343       util_snprintf(layout, sizeof layout, "%c-p:%u:%u:%u-i64:64:64-a0:0:%u-s0:%u:%u",
 344 #ifdef PIPE_ARCH_LITTLE_ENDIAN
 345                     'e', // little endian
 346 #else
 347                     'E', // big endian
 348 #endif
 349                     pointer_size, pointer_size, pointer_size, // pointer size, abi alignment, preferred alignment
 350                     pointer_size, // aggregate preferred alignment
 351                     pointer_size, pointer_size); // stack objects abi alignment, preferred alignment
 352
 353       gallivm->target = LLVMCreateTargetData(layout);
 354       if (!gallivm->target) {
 355          return FALSE;
 356       }
 357    }
 358 #endif
 359
 360    if (!create_pass_manager(gallivm))
 361       goto fail;
 362
 363    return TRUE;
 364
 365 fail:
 366    gallivm_free_ir(gallivm);
 367    gallivm_free_code(gallivm);
 368    return FALSE;
 369 }
 370
 371
 372 boolean
 373 lp_build_init(void)
 374 {
 375    if (gallivm_initialized)
 376       return TRUE;
 377
 378    /* XXX: Remove this once lp_bld_misc.cpp has been adapted to the removal
 379     * of JITMemoryManager
 380     */
 381 #if HAVE_LLVM >= 0x0306
 382    return FALSE;
 383 #endif
 384
 385 #ifdef DEBUG
 386    gallivm_debug = debug_get_option_gallivm_debug();
 387 #endif
 388
 389    lp_set_target_options();
 390
 391 #if USE_MCJIT
 392    LLVMLinkInMCJIT();
 393 #else
 394    LLVMLinkInJIT();
 395 #endif
 396
 397    util_cpu_detect();
 398
 399    /* AMD Bulldozer AVX's throughput is the same as SSE2; and because using
 400     * 8-wide vector needs more floating ops than 4-wide (due to padding), it is
 401     * actually more efficient to use 4-wide vectors on this processor.
 402     *
 403     * See also:
 404     * - http://www.anandtech.com/show/4955/the-bulldozer-review-amd-fx8150-tested/2
 405     */
 406    if (HAVE_AVX &&
 407        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    if (!HAVE_AVX) {
 433       /*
 434        * note these instructions are VEX-only, so can only emit if we use
 435        * avx (don't want to base it on has_avx & has_f16c later as that would
 436        * omit it unnecessarily on amd cpus, see above).
 437        */
 438       util_cpu_caps.has_f16c = 0;
 439       util_cpu_caps.has_xop = 0;
 440    }
 441
 442 #ifdef PIPE_ARCH_PPC_64
 443    /* Set the NJ bit in VSCR to 0 so denormalized values are handled as
 444     * specified by IEEE standard (PowerISA 2.06 - Section 6.3). This guarantees
 445     * that some rounding and half-float to float handling does not round
 446     * incorrectly to 0.
 447     * XXX: should eventually follow same logic on all platforms.
 448     * Right now denorms get explicitly disabled (but elsewhere) for x86,
 449     * whereas ppc64 explicitly enables them...
 450     */
 451    if (util_cpu_caps.has_altivec) {
 452       unsigned short mask[] = { 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF,
 453                                 0xFFFF, 0xFFFF, 0xFFFE, 0xFFFF };
 454       __asm (
 455         "mfvscr %%v1\n"
 456         "vand   %0,%%v1,%0\n"
 457         "mtvscr %0"
 458         :
 459         : "r" (*mask)
 460       );
 461    }
 462 #endif
 463
 464    gallivm_initialized = TRUE;
 465
 466 #if 0
 467    /* For simulating less capable machines */
 468    util_cpu_caps.has_sse3 = 0;
 469    util_cpu_caps.has_ssse3 = 0;
 470    util_cpu_caps.has_sse4_1 = 0;
 471    util_cpu_caps.has_avx = 0;
 472    util_cpu_caps.has_f16c = 0;
 473 #endif
 474
 475    return TRUE;
 476 }
 477
 478
 479
 480 /**
 481  * Create a new gallivm_state object.
 482  */
 483 struct gallivm_state *
 484 gallivm_create(const char *name, LLVMContextRef context)
 485 {
 486    struct gallivm_state *gallivm;
 487
 488    gallivm = CALLOC_STRUCT(gallivm_state);
 489    if (gallivm) {
 490       if (!init_gallivm_state(gallivm, name, context)) {
 491          FREE(gallivm);
 492          gallivm = NULL;
 493       }
 494    }
 495
 496    return gallivm;
 497 }
 498
 499
 500 /**
 501  * Destroy a gallivm_state object.
 502  */
 503 void
 504 gallivm_destroy(struct gallivm_state *gallivm)
 505 {
 506    gallivm_free_ir(gallivm);
 507    gallivm_free_code(gallivm);
 508    FREE(gallivm);
 509 }
 510
 511
 512 /**
 513  * Validate a function.
 514  * Verification is only done with debug builds.
 515  */
 516 void
 517 gallivm_verify_function(struct gallivm_state *gallivm,
 518                         LLVMValueRef func)
 519 {
 520    /* Verify the LLVM IR.  If invalid, dump and abort */
 521 #ifdef DEBUG
 522    if (LLVMVerifyFunction(func, LLVMPrintMessageAction)) {
 523       lp_debug_dump_value(func);
 524       assert(0);
 525       return;
 526    }
 527 #endif
 528
 529    if (gallivm_debug & GALLIVM_DEBUG_IR) {
 530       /* Print the LLVM IR to stderr */
 531       lp_debug_dump_value(func);
 532       debug_printf("\n");
 533    }
 534 }
 535
 536
 537 /**
 538  * Compile a module.
 539  * This does IR optimization on all functions in the module.
 540  */
 541 void
 542 gallivm_compile_module(struct gallivm_state *gallivm)
 543 {
 544    LLVMValueRef func;
 545    int64_t time_begin;
 546
 547    assert(!gallivm->compiled);
 548
 549    if (gallivm->builder) {
 550       LLVMDisposeBuilder(gallivm->builder);
 551       gallivm->builder = NULL;
 552    }
 553
 554    if (gallivm_debug & GALLIVM_DEBUG_PERF)
 555       time_begin = os_time_get();
 556
 557    /* Run optimization passes */
 558    LLVMInitializeFunctionPassManager(gallivm->passmgr);
 559    func = LLVMGetFirstFunction(gallivm->module);
 560    while (func) {
 561       if (0) {
 562          debug_printf("optimizing func %s...\n", LLVMGetValueName(func));
 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
 594
 595
 596 func_pointer
 597 gallivm_jit_function(struct gallivm_state *gallivm,
 598                      LLVMValueRef func)
 599 {
 600    void *code;
 601    func_pointer jit_func;
 602
 603    assert(gallivm->compiled);
 604    assert(gallivm->engine);
 605
 606    code = LLVMGetPointerToGlobal(gallivm->engine, func);
 607    assert(code);
 608    jit_func = pointer_to_func(code);
 609
 610    if (gallivm_debug & GALLIVM_DEBUG_ASM) {
 611       lp_disassemble(func, code);
 612    }
 613
 614 #if defined(PROFILE)
 615    lp_profile(func, code);
 616 #endif
 617
 618    return jit_func;
 619 }