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    lp_free_memory_manager(gallivm->memorymgr);
 223    gallivm->memorymgr = NULL;
 224 }
 225
 226
 227 static boolean
 228 init_gallivm_engine(struct gallivm_state *gallivm)
 229 {
 230    if (1) {
 231       enum LLVM_CodeGenOpt_Level optlevel;
 232       char *error = NULL;
 233       int ret;
 234
 235       if (gallivm_debug & GALLIVM_DEBUG_NO_OPT) {
 236          optlevel = None;
 237       }
 238       else {
 239          optlevel = Default;
 240       }
 241
 242       ret = lp_build_create_jit_compiler_for_module(&gallivm->engine,
 243                                                     &gallivm->code,
 244                                                     gallivm->module,
 245                                                     gallivm->memorymgr,
 246                                                     (unsigned) optlevel,
 247                                                     USE_MCJIT,
 248                                                     &error);
 249       if (ret) {
 250          _debug_printf("%s\n", error);
 251          LLVMDisposeMessage(error);
 252          goto fail;
 253       }
 254    }
 255
 256 #if !USE_MCJIT
 257    gallivm->target = LLVMGetExecutionEngineTargetData(gallivm->engine);
 258    if (!gallivm->target)
 259       goto fail;
 260 #else
 261    if (0) {
 262        /*
 263         * Dump the data layout strings.
 264         */
 265
 266        LLVMTargetDataRef target = LLVMGetExecutionEngineTargetData(gallivm->engine);
 267        char *data_layout;
 268        char *engine_data_layout;
 269
 270        data_layout = LLVMCopyStringRepOfTargetData(gallivm->target);
 271        engine_data_layout = LLVMCopyStringRepOfTargetData(target);
 272
 273        if (1) {
 274           debug_printf("module target data = %s\n", data_layout);
 275           debug_printf("engine target data = %s\n", engine_data_layout);
 276        }
 277
 278        free(data_layout);
 279        free(engine_data_layout);
 280    }
 281 #endif
 282
 283    return TRUE;
 284
 285 fail:
 286    return FALSE;
 287 }
 288
 289
 290 /**
 291  * Allocate gallivm LLVM objects.
 292  * \return  TRUE for success, FALSE for failure
 293  */
 294 static boolean
 295 init_gallivm_state(struct gallivm_state *gallivm, const char *name,
 296                    LLVMContextRef context)
 297 {
 298    assert(!gallivm->context);
 299    assert(!gallivm->module);
 300
 301    if (!lp_build_init())
 302       return FALSE;
 303
 304    gallivm->context = context;
 305
 306    if (!gallivm->context)
 307       goto fail;
 308
 309    gallivm->module = LLVMModuleCreateWithNameInContext(name,
 310                                                        gallivm->context);
 311    if (!gallivm->module)
 312       goto fail;
 313
 314    gallivm->builder = LLVMCreateBuilderInContext(gallivm->context);
 315    if (!gallivm->builder)
 316       goto fail;
 317
 318 #if HAVE_LLVM < 0x0306
 319    gallivm->memorymgr = lp_get_default_memory_manager();
 320    if (!gallivm->memorymgr)
 321       goto fail;
 322 #else
 323    gallivm->memorymgr = 0;
 324 #endif
 325
 326    /* FIXME: MC-JIT only allows compiling one module at a time, and it must be
 327     * complete when MC-JIT is created. So defer the MC-JIT engine creation for
 328     * now.
 329     */
 330 #if !USE_MCJIT
 331    if (!init_gallivm_engine(gallivm)) {
 332       goto fail;
 333    }
 334 #else
 335    /*
 336     * MC-JIT engine compiles the module immediately on creation, so we can't
 337     * obtain the target data from it.  Instead we create a target data layout
 338     * from a string.
 339     *
 340     * The produced layout strings are not precisely the same, but should make
 341     * no difference for the kind of optimization passes we run.
 342     *
 343     * For reference this is the layout string on x64:
 344     *
 345     *   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
 346     *
 347     * See also:
 348     * - http://llvm.org/docs/LangRef.html#datalayout
 349     */
 350
 351    {
 352       const unsigned pointer_size = 8 * sizeof(void *);
 353       char layout[512];
 354       util_snprintf(layout, sizeof layout, "%c-p:%u:%u:%u-i64:64:64-a0:0:%u-s0:%u:%u",
 355 #ifdef PIPE_ARCH_LITTLE_ENDIAN
 356                     'e', // little endian
 357 #else
 358                     'E', // big endian
 359 #endif
 360                     pointer_size, pointer_size, pointer_size, // pointer size, abi alignment, preferred alignment
 361                     pointer_size, // aggregate preferred alignment
 362                     pointer_size, pointer_size); // stack objects abi alignment, preferred alignment
 363
 364       gallivm->target = LLVMCreateTargetData(layout);
 365       if (!gallivm->target) {
 366          return FALSE;
 367       }
 368    }
 369 #endif
 370
 371    if (!create_pass_manager(gallivm))
 372       goto fail;
 373
 374    return TRUE;
 375
 376 fail:
 377    gallivm_free_ir(gallivm);
 378    gallivm_free_code(gallivm);
 379    return FALSE;
 380 }
 381
 382
 383 boolean
 384 lp_build_init(void)
 385 {
 386    if (gallivm_initialized)
 387       return TRUE;
 388
 389    /* XXX: Remove this once lp_bld_misc.cpp has been adapted to the removal
 390     * of JITMemoryManager
 391     */
 392 #if HAVE_LLVM >= 0x0306
 393    return FALSE;
 394 #endif
 395
 396 #ifdef DEBUG
 397    gallivm_debug = debug_get_option_gallivm_debug();
 398 #endif
 399
 400    lp_set_target_options();
 401
 402 #if USE_MCJIT
 403    LLVMLinkInMCJIT();
 404 #else
 405    LLVMLinkInJIT();
 406 #endif
 407
 408    util_cpu_detect();
 409
 410    /* AMD Bulldozer AVX's throughput is the same as SSE2; and because using
 411     * 8-wide vector needs more floating ops than 4-wide (due to padding), it is
 412     * actually more efficient to use 4-wide vectors on this processor.
 413     *
 414     * See also:
 415     * - http://www.anandtech.com/show/4955/the-bulldozer-review-amd-fx8150-tested/2
 416     */
 417    if (HAVE_AVX &&
 418        util_cpu_caps.has_avx &&
 419        util_cpu_caps.has_intel) {
 420       lp_native_vector_width = 256;
 421    } else {
 422       /* Leave it at 128, even when no SIMD extensions are available.
 423        * Really needs to be a multiple of 128 so can fit 4 floats.
 424        */
 425       lp_native_vector_width = 128;
 426    }
 427
 428    lp_native_vector_width = debug_get_num_option("LP_NATIVE_VECTOR_WIDTH",
 429                                                  lp_native_vector_width);
 430
 431    if (lp_native_vector_width <= 128) {
 432       /* Hide AVX support, as often LLVM AVX intrinsics are only guarded by
 433        * "util_cpu_caps.has_avx" predicate, and lack the
 434        * "lp_native_vector_width > 128" predicate. And also to ensure a more
 435        * consistent behavior, allowing one to test SSE2 on AVX machines.
 436        * XXX: should not play games with util_cpu_caps directly as it might
 437        * get used for other things outside llvm too.
 438        */
 439       util_cpu_caps.has_avx = 0;
 440       util_cpu_caps.has_avx2 = 0;
 441    }
 442
 443    if (!HAVE_AVX) {
 444       /*
 445        * note these instructions are VEX-only, so can only emit if we use
 446        * avx (don't want to base it on has_avx & has_f16c later as that would
 447        * omit it unnecessarily on amd cpus, see above).
 448        */
 449       util_cpu_caps.has_f16c = 0;
 450       util_cpu_caps.has_xop = 0;
 451    }
 452
 453 #ifdef PIPE_ARCH_PPC_64
 454    /* Set the NJ bit in VSCR to 0 so denormalized values are handled as
 455     * specified by IEEE standard (PowerISA 2.06 - Section 6.3). This guarantees
 456     * that some rounding and half-float to float handling does not round
 457     * incorrectly to 0.
 458     * XXX: should eventually follow same logic on all platforms.
 459     * Right now denorms get explicitly disabled (but elsewhere) for x86,
 460     * whereas ppc64 explicitly enables them...
 461     */
 462    if (util_cpu_caps.has_altivec) {
 463       unsigned short mask[] = { 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF,
 464                                 0xFFFF, 0xFFFF, 0xFFFE, 0xFFFF };
 465       __asm (
 466         "mfvscr %%v1\n"
 467         "vand   %0,%%v1,%0\n"
 468         "mtvscr %0"
 469         :
 470         : "r" (*mask)
 471       );
 472    }
 473 #endif
 474
 475    gallivm_initialized = TRUE;
 476
 477 #if 0
 478    /* For simulating less capable machines */
 479    util_cpu_caps.has_sse3 = 0;
 480    util_cpu_caps.has_ssse3 = 0;
 481    util_cpu_caps.has_sse4_1 = 0;
 482    util_cpu_caps.has_avx = 0;
 483    util_cpu_caps.has_f16c = 0;
 484 #endif
 485
 486    return TRUE;
 487 }
 488
 489
 490
 491 /**
 492  * Create a new gallivm_state object.
 493  */
 494 struct gallivm_state *
 495 gallivm_create(const char *name, LLVMContextRef context)
 496 {
 497    struct gallivm_state *gallivm;
 498
 499    gallivm = CALLOC_STRUCT(gallivm_state);
 500    if (gallivm) {
 501       if (!init_gallivm_state(gallivm, name, context)) {
 502          FREE(gallivm);
 503          gallivm = NULL;
 504       }
 505    }
 506
 507    return gallivm;
 508 }
 509
 510
 511 /**
 512  * Destroy a gallivm_state object.
 513  */
 514 void
 515 gallivm_destroy(struct gallivm_state *gallivm)
 516 {
 517    gallivm_free_ir(gallivm);
 518    gallivm_free_code(gallivm);
 519    FREE(gallivm);
 520 }
 521
 522
 523 /**
 524  * Validate a function.
 525  * Verification is only done with debug builds.
 526  */
 527 void
 528 gallivm_verify_function(struct gallivm_state *gallivm,
 529                         LLVMValueRef func)
 530 {
 531    /* Verify the LLVM IR.  If invalid, dump and abort */
 532 #ifdef DEBUG
 533    if (LLVMVerifyFunction(func, LLVMPrintMessageAction)) {
 534       lp_debug_dump_value(func);
 535       assert(0);
 536       return;
 537    }
 538 #endif
 539
 540    if (gallivm_debug & GALLIVM_DEBUG_IR) {
 541       /* Print the LLVM IR to stderr */
 542       lp_debug_dump_value(func);
 543       debug_printf("\n");
 544    }
 545 }
 546
 547
 548 /**
 549  * Compile a module.
 550  * This does IR optimization on all functions in the module.
 551  */
 552 void
 553 gallivm_compile_module(struct gallivm_state *gallivm)
 554 {
 555    LLVMValueRef func;
 556    int64_t time_begin;
 557
 558    assert(!gallivm->compiled);
 559
 560    if (gallivm->builder) {
 561       LLVMDisposeBuilder(gallivm->builder);
 562       gallivm->builder = NULL;
 563    }
 564
 565    if (gallivm_debug & GALLIVM_DEBUG_PERF)
 566       time_begin = os_time_get();
 567
 568    /* Run optimization passes */
 569    LLVMInitializeFunctionPassManager(gallivm->passmgr);
 570    func = LLVMGetFirstFunction(gallivm->module);
 571    while (func) {
 572       if (0) {
 573          debug_printf("optimizing func %s...\n", LLVMGetValueName(func));
 574       }
 575       LLVMRunFunctionPassManager(gallivm->passmgr, func);
 576       func = LLVMGetNextFunction(func);
 577    }
 578    LLVMFinalizeFunctionPassManager(gallivm->passmgr);
 579
 580    if (gallivm_debug & GALLIVM_DEBUG_PERF) {
 581       int64_t time_end = os_time_get();
 582       int time_msec = (int)(time_end - time_begin) / 1000;
 583       debug_printf("optimizing module %s took %d msec\n",
 584                    lp_get_module_id(gallivm->module), time_msec);
 585    }
 586
 587    /* Dump byte code to a file */
 588    if (0) {
 589       LLVMWriteBitcodeToFile(gallivm->module, "llvmpipe.bc");
 590       debug_printf("llvmpipe.bc written\n");
 591       debug_printf("Invoke as \"llc -o - llvmpipe.bc\"\n");
 592    }
 593
 594 #if USE_MCJIT
 595    assert(!gallivm->engine);
 596    if (!init_gallivm_engine(gallivm)) {
 597       assert(0);
 598    }
 599 #endif
 600    assert(gallivm->engine);
 601
 602    ++gallivm->compiled;
 603 }
 604
 605
 606
 607 func_pointer
 608 gallivm_jit_function(struct gallivm_state *gallivm,
 609                      LLVMValueRef func)
 610 {
 611    void *code;
 612    func_pointer jit_func;
 613
 614    assert(gallivm->compiled);
 615    assert(gallivm->engine);
 616
 617    code = LLVMGetPointerToGlobal(gallivm->engine, func);
 618    assert(code);
 619    jit_func = pointer_to_func(code);
 620
 621    if (gallivm_debug & GALLIVM_DEBUG_ASM) {
 622       lp_disassemble(func, code);
 623    }
 624
 625 #if defined(PROFILE)
 626    lp_profile(func, code);
 627 #endif
 628
 629    return jit_func;
 630 }