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 "lp_bld.h"
  36 #include "lp_bld_debug.h"
  37 #include "lp_bld_misc.h"
  38 #include "lp_bld_init.h"
  39
  40 #include <llvm-c/Analysis.h>
  41 #include <llvm-c/Transforms/Scalar.h>
  42 #include <llvm-c/BitWriter.h>
  43
  44
  45 /**
  46  * AVX is supported in:
  47  * - standard JIT from LLVM 3.2 onwards
  48  * - MC-JIT from LLVM 3.1
  49  *   - MC-JIT supports limited OSes (MacOSX and Linux)
  50  * - standard JIT in LLVM 3.1, with backports
  51  */
  52 #if defined(PIPE_ARCH_PPC_64) || defined(PIPE_ARCH_S390)
  53 #  define USE_MCJIT 1
  54 #  define HAVE_AVX 0
  55 #elif HAVE_LLVM >= 0x0302 || (HAVE_LLVM == 0x0301 && defined(HAVE_JIT_AVX_SUPPORT))
  56 #  define USE_MCJIT 0
  57 #  define HAVE_AVX 1
  58 #elif HAVE_LLVM == 0x0301 && (defined(PIPE_OS_LINUX) || defined(PIPE_OS_APPLE))
  59 #  define USE_MCJIT 1
  60 #  define HAVE_AVX 1
  61 #else
  62 #  define USE_MCJIT 0
  63 #  define HAVE_AVX 0
  64 #endif
  65
  66
  67 #if USE_MCJIT
  68 void LLVMLinkInMCJIT();
  69 #endif
  70
  71
  72 #ifdef DEBUG
  73 unsigned gallivm_debug = 0;
  74
  75 static const struct debug_named_value lp_bld_debug_flags[] = {
  76    { "tgsi",   GALLIVM_DEBUG_TGSI, NULL },
  77    { "ir",     GALLIVM_DEBUG_IR, NULL },
  78    { "asm",    GALLIVM_DEBUG_ASM, NULL },
  79    { "nopt",   GALLIVM_DEBUG_NO_OPT, NULL },
  80    { "perf",   GALLIVM_DEBUG_PERF, NULL },
  81    { "no_brilinear", GALLIVM_DEBUG_NO_BRILINEAR, NULL },
  82    { "no_rho_approx", GALLIVM_DEBUG_NO_RHO_APPROX, NULL },
  83    { "gc",     GALLIVM_DEBUG_GC, NULL },
  84    DEBUG_NAMED_VALUE_END
  85 };
  86
  87 DEBUG_GET_ONCE_FLAGS_OPTION(gallivm_debug, "GALLIVM_DEBUG", lp_bld_debug_flags, 0)
  88 #endif
  89
  90
  91 static boolean gallivm_initialized = FALSE;
  92
  93 unsigned lp_native_vector_width;
  94
  95
  96 /*
  97  * Optimization values are:
  98  * - 0: None (-O0)
  99  * - 1: Less (-O1)
 100  * - 2: Default (-O2, -Os)
 101  * - 3: Aggressive (-O3)
 102  *
 103  * See also CodeGenOpt::Level in llvm/Target/TargetMachine.h
 104  */
 105 enum LLVM_CodeGenOpt_Level {
 106 #if HAVE_LLVM >= 0x207
 107    None,        // -O0
 108    Less,        // -O1
 109    Default,     // -O2, -Os
 110    Aggressive   // -O3
 111 #else
 112    Default,
 113    None,
 114    Aggressive
 115 #endif
 116 };
 117
 118
 119 #if HAVE_LLVM <= 0x0206
 120 /**
 121  * LLVM 2.6 permits only one ExecutionEngine to be created.  So use the
 122  * same gallivm state everywhere.
 123  */
 124 static struct gallivm_state *GlobalGallivm = NULL;
 125 #endif
 126
 127
 128 /**
 129  * Create the LLVM (optimization) pass manager and install
 130  * relevant optimization passes.
 131  * \return  TRUE for success, FALSE for failure
 132  */
 133 static boolean
 134 create_pass_manager(struct gallivm_state *gallivm)
 135 {
 136    assert(!gallivm->passmgr);
 137    assert(gallivm->target);
 138
 139    gallivm->passmgr = LLVMCreateFunctionPassManager(gallivm->provider);
 140    if (!gallivm->passmgr)
 141       return FALSE;
 142
 143    LLVMAddTargetData(gallivm->target, gallivm->passmgr);
 144
 145    if ((gallivm_debug & GALLIVM_DEBUG_NO_OPT) == 0) {
 146       /* These are the passes currently listed in llvm-c/Transforms/Scalar.h,
 147        * but there are more on SVN.
 148        * TODO: Add more passes.
 149        */
 150       LLVMAddScalarReplAggregatesPass(gallivm->passmgr);
 151       LLVMAddLICMPass(gallivm->passmgr);
 152       LLVMAddCFGSimplificationPass(gallivm->passmgr);
 153       LLVMAddReassociatePass(gallivm->passmgr);
 154
 155       if (HAVE_LLVM >= 0x207 && sizeof(void*) == 4) {
 156          /* For LLVM >= 2.7 and 32-bit build, use this order of passes to
 157           * avoid generating bad code.
 158           * Test with piglit glsl-vs-sqrt-zero test.
 159           */
 160          LLVMAddConstantPropagationPass(gallivm->passmgr);
 161          LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
 162       }
 163       else {
 164          LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
 165          LLVMAddConstantPropagationPass(gallivm->passmgr);
 166       }
 167
 168       if (util_cpu_caps.has_sse4_1) {
 169          /* FIXME: There is a bug in this pass, whereby the combination
 170           * of fptosi and sitofp (necessary for trunc/floor/ceil/round
 171           * implementation) somehow becomes invalid code.
 172           */
 173          LLVMAddInstructionCombiningPass(gallivm->passmgr);
 174       }
 175       LLVMAddGVNPass(gallivm->passmgr);
 176    }
 177    else {
 178       /* We need at least this pass to prevent the backends to fail in
 179        * unexpected ways.
 180        */
 181       LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
 182    }
 183
 184    return TRUE;
 185 }
 186
 187
 188 /**
 189  * Free gallivm object's LLVM allocations, but not the gallivm object itself.
 190  */
 191 static void
 192 free_gallivm_state(struct gallivm_state *gallivm)
 193 {
 194 #if HAVE_LLVM >= 0x207 /* XXX or 0x208? */
 195    /* This leads to crashes w/ some versions of LLVM */
 196    LLVMModuleRef mod;
 197    char *error;
 198
 199    if (gallivm->engine && gallivm->provider)
 200       LLVMRemoveModuleProvider(gallivm->engine, gallivm->provider,
 201                                &mod, &error);
 202 #endif
 203
 204    if (gallivm->passmgr) {
 205       LLVMDisposePassManager(gallivm->passmgr);
 206    }
 207
 208 #if 0
 209    /* XXX this seems to crash with all versions of LLVM */
 210    if (gallivm->provider)
 211       LLVMDisposeModuleProvider(gallivm->provider);
 212 #endif
 213
 214    if (HAVE_LLVM >= 0x207 && gallivm->engine) {
 215       /* This will already destroy any associated module */
 216       LLVMDisposeExecutionEngine(gallivm->engine);
 217    } else {
 218       LLVMDisposeModule(gallivm->module);
 219    }
 220
 221 #if !USE_MCJIT
 222    /* Don't free the TargetData, it's owned by the exec engine */
 223 #else
 224    if (gallivm->target) {
 225       LLVMDisposeTargetData(gallivm->target);
 226    }
 227 #endif
 228
 229    /* Never free the LLVM context.
 230     */
 231 #if 0
 232    if (gallivm->context)
 233       LLVMContextDispose(gallivm->context);
 234 #endif
 235
 236    if (gallivm->builder)
 237       LLVMDisposeBuilder(gallivm->builder);
 238
 239    gallivm->engine = NULL;
 240    gallivm->target = NULL;
 241    gallivm->module = NULL;
 242    gallivm->provider = NULL;
 243    gallivm->passmgr = NULL;
 244    gallivm->context = NULL;
 245    gallivm->builder = NULL;
 246 }
 247
 248
 249 static boolean
 250 init_gallivm_engine(struct gallivm_state *gallivm)
 251 {
 252    if (1) {
 253       /* We can only create one LLVMExecutionEngine (w/ LLVM 2.6 anyway) */
 254       enum LLVM_CodeGenOpt_Level optlevel;
 255       char *error = NULL;
 256       int ret;
 257
 258       if (gallivm_debug & GALLIVM_DEBUG_NO_OPT) {
 259          optlevel = None;
 260       }
 261       else {
 262          optlevel = Default;
 263       }
 264
 265 #if HAVE_LLVM >= 0x0301
 266       ret = lp_build_create_jit_compiler_for_module(&gallivm->engine,
 267                                                     gallivm->module,
 268                                                     (unsigned) optlevel,
 269                                                     USE_MCJIT,
 270                                                     &error);
 271 #else
 272       ret = LLVMCreateJITCompiler(&gallivm->engine, gallivm->provider,
 273                                   (unsigned) optlevel, &error);
 274 #endif
 275       if (ret) {
 276          _debug_printf("%s\n", error);
 277          LLVMDisposeMessage(error);
 278          goto fail;
 279       }
 280    }
 281
 282    LLVMAddModuleProvider(gallivm->engine, gallivm->provider);//new
 283
 284 #if !USE_MCJIT
 285    gallivm->target = LLVMGetExecutionEngineTargetData(gallivm->engine);
 286    if (!gallivm->target)
 287       goto fail;
 288 #else
 289    if (0) {
 290        /*
 291         * Dump the data layout strings.
 292         */
 293
 294        LLVMTargetDataRef target = LLVMGetExecutionEngineTargetData(gallivm->engine);
 295        char *data_layout;
 296        char *engine_data_layout;
 297
 298        data_layout = LLVMCopyStringRepOfTargetData(gallivm->target);
 299        engine_data_layout = LLVMCopyStringRepOfTargetData(target);
 300
 301        if (1) {
 302           debug_printf("module target data = %s\n", data_layout);
 303           debug_printf("engine target data = %s\n", engine_data_layout);
 304        }
 305
 306        free(data_layout);
 307        free(engine_data_layout);
 308    }
 309 #endif
 310
 311    return TRUE;
 312
 313 fail:
 314    return FALSE;
 315 }
 316
 317
 318 /**
 319  * Singleton
 320  *
 321  * We must never free LLVM contexts, because LLVM has several global caches
 322  * which pointing/derived from objects owned by the context, causing false
 323  * memory leaks and false cache hits when these objects are destroyed.
 324  *
 325  * TODO: For thread safety on multi-threaded OpenGL we should use one LLVM
 326  * context per thread, and put them in a pool when threads are destroyed.
 327  */
 328 static LLVMContextRef gallivm_context = NULL;
 329
 330
 331 /**
 332  * Allocate gallivm LLVM objects.
 333  * \return  TRUE for success, FALSE for failure
 334  */
 335 static boolean
 336 init_gallivm_state(struct gallivm_state *gallivm)
 337 {
 338    assert(!gallivm->context);
 339    assert(!gallivm->module);
 340    assert(!gallivm->provider);
 341
 342    lp_build_init();
 343
 344    if (!gallivm_context) {
 345       gallivm_context = LLVMContextCreate();
 346    }
 347    gallivm->context = gallivm_context;
 348    if (!gallivm->context)
 349       goto fail;
 350
 351    gallivm->module = LLVMModuleCreateWithNameInContext("gallivm",
 352                                                        gallivm->context);
 353    if (!gallivm->module)
 354       goto fail;
 355
 356    gallivm->provider =
 357       LLVMCreateModuleProviderForExistingModule(gallivm->module);
 358    if (!gallivm->provider)
 359       goto fail;
 360
 361    gallivm->builder = LLVMCreateBuilderInContext(gallivm->context);
 362    if (!gallivm->builder)
 363       goto fail;
 364
 365    /* FIXME: MC-JIT only allows compiling one module at a time, and it must be
 366     * complete when MC-JIT is created. So defer the MC-JIT engine creation for
 367     * now.
 368     */
 369 #if !USE_MCJIT
 370    if (!init_gallivm_engine(gallivm)) {
 371       goto fail;
 372    }
 373 #else
 374    /*
 375     * MC-JIT engine compiles the module immediately on creation, so we can't
 376     * obtain the target data from it.  Instead we create a target data layout
 377     * from a string.
 378     *
 379     * The produced layout strings are not precisely the same, but should make
 380     * no difference for the kind of optimization passes we run.
 381     *
 382     * For reference this is the layout string on x64:
 383     *
 384     *   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
 385     *
 386     * See also:
 387     * - http://llvm.org/docs/LangRef.html#datalayout
 388     */
 389
 390    {
 391       const unsigned pointer_size = 8 * sizeof(void *);
 392       char layout[512];
 393       util_snprintf(layout, sizeof layout, "%c-p:%u:%u:%u-i64:64:64-a0:0:%u-s0:%u:%u",
 394 #ifdef PIPE_ARCH_LITTLE_ENDIAN
 395                     'e', // little endian
 396 #else
 397                     'E', // big endian
 398 #endif
 399                     pointer_size, pointer_size, pointer_size, // pointer size, abi alignment, preferred alignment
 400                     pointer_size, // aggregate preferred alignment
 401                     pointer_size, pointer_size); // stack objects abi alignment, preferred alignment
 402
 403       gallivm->target = LLVMCreateTargetData(layout);
 404       if (!gallivm->target) {
 405          return FALSE;
 406       }
 407    }
 408 #endif
 409
 410    if (!create_pass_manager(gallivm))
 411       goto fail;
 412
 413    return TRUE;
 414
 415 fail:
 416    free_gallivm_state(gallivm);
 417    return FALSE;
 418 }
 419
 420
 421 void
 422 lp_build_init(void)
 423 {
 424    if (gallivm_initialized)
 425       return;
 426
 427 #ifdef DEBUG
 428    gallivm_debug = debug_get_option_gallivm_debug();
 429 #endif
 430
 431    lp_set_target_options();
 432
 433 #if USE_MCJIT
 434    LLVMLinkInMCJIT();
 435 #else
 436    LLVMLinkInJIT();
 437 #endif
 438
 439    util_cpu_detect();
 440
 441    /* AMD Bulldozer AVX's throughput is the same as SSE2; and because using
 442     * 8-wide vector needs more floating ops than 4-wide (due to padding), it is
 443     * actually more efficient to use 4-wide vectors on this processor.
 444     *
 445     * See also:
 446     * - http://www.anandtech.com/show/4955/the-bulldozer-review-amd-fx8150-tested/2
 447     */
 448    if (HAVE_AVX &&
 449        util_cpu_caps.has_avx &&
 450        util_cpu_caps.has_intel) {
 451       lp_native_vector_width = 256;
 452    } else {
 453       /* Leave it at 128, even when no SIMD extensions are available.
 454        * Really needs to be a multiple of 128 so can fit 4 floats.
 455        */
 456       lp_native_vector_width = 128;
 457    }
 458
 459    lp_native_vector_width = debug_get_num_option("LP_NATIVE_VECTOR_WIDTH",
 460                                                  lp_native_vector_width);
 461
 462    if (lp_native_vector_width <= 128) {
 463       /* Hide AVX support, as often LLVM AVX instrinsics are only guarded by
 464        * "util_cpu_caps.has_avx" predicate, and lack the
 465        * "lp_native_vector_width > 128" predicate. And also to ensure a more
 466        * consistent behavior, allowing one to test SSE2 on AVX machines.
 467        */
 468       util_cpu_caps.has_avx = 0;
 469    }
 470
 471    if (!HAVE_AVX) {
 472       /*
 473        * note these instructions are VEX-only, so can only emit if we use
 474        * avx (don't want to base it on has_avx & has_f16c later as that would
 475        * omit it unnecessarily on amd cpus, see above).
 476        */
 477       util_cpu_caps.has_f16c = 0;
 478    }
 479
 480 #ifdef PIPE_ARCH_PPC_64
 481    /* Set the NJ bit in VSCR to 0 so denormalized values are handled as
 482     * specified by IEEE standard (PowerISA 2.06 - Section 6.3). This garantees
 483     * that some rounding and half-float to float handling does not round
 484     * incorrectly to 0.
 485     */
 486    if (util_cpu_caps.has_altivec) {
 487       unsigned short mask[] = { 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF,
 488                                 0xFFFF, 0xFFFF, 0xFFFE, 0xFFFF };
 489       __asm (
 490         "mfvscr %%v1\n"
 491         "vand   %0,%%v1,%0\n"
 492         "mtvscr %0"
 493         :
 494         : "r" (*mask)
 495       );
 496    }
 497 #endif
 498
 499    gallivm_initialized = TRUE;
 500
 501 #if 0
 502    /* For simulating less capable machines */
 503    util_cpu_caps.has_sse3 = 0;
 504    util_cpu_caps.has_ssse3 = 0;
 505    util_cpu_caps.has_sse4_1 = 0;
 506    util_cpu_caps.has_avx = 0;
 507    util_cpu_caps.has_f16c = 0;
 508 #endif
 509 }
 510
 511
 512
 513 /**
 514  * Create a new gallivm_state object.
 515  * Note that we return a singleton.
 516  */
 517 struct gallivm_state *
 518 gallivm_create(void)
 519 {
 520    struct gallivm_state *gallivm;
 521
 522 #if HAVE_LLVM <= 0x206
 523    if (GlobalGallivm) {
 524       return GlobalGallivm;
 525    }
 526 #endif
 527
 528    gallivm = CALLOC_STRUCT(gallivm_state);
 529    if (gallivm) {
 530       if (!init_gallivm_state(gallivm)) {
 531          FREE(gallivm);
 532          gallivm = NULL;
 533       }
 534    }
 535
 536 #if HAVE_LLVM <= 0x206
 537    GlobalGallivm = gallivm;
 538 #endif
 539
 540    return gallivm;
 541 }
 542
 543
 544 /**
 545  * Destroy a gallivm_state object.
 546  */
 547 void
 548 gallivm_destroy(struct gallivm_state *gallivm)
 549 {
 550 #if HAVE_LLVM <= 0x0206
 551    /* No-op: don't destroy the singleton */
 552    (void) gallivm;
 553 #else
 554    free_gallivm_state(gallivm);
 555    FREE(gallivm);
 556 #endif
 557 }
 558
 559
 560 /**
 561  * Validate and optimze a function.
 562  */
 563 static void
 564 gallivm_optimize_function(struct gallivm_state *gallivm,
 565                           LLVMValueRef func)
 566 {
 567    if (0) {
 568       debug_printf("optimizing %s...\n", LLVMGetValueName(func));
 569    }
 570
 571    assert(gallivm->passmgr);
 572
 573    /* Apply optimizations to LLVM IR */
 574    LLVMRunFunctionPassManager(gallivm->passmgr, func);
 575
 576    if (0) {
 577       if (gallivm_debug & GALLIVM_DEBUG_IR) {
 578          /* Print the LLVM IR to stderr */
 579          lp_debug_dump_value(func);
 580          debug_printf("\n");
 581       }
 582    }
 583 }
 584
 585
 586 /**
 587  * Validate a function.
 588  */
 589 void
 590 gallivm_verify_function(struct gallivm_state *gallivm,
 591                         LLVMValueRef func)
 592 {
 593    /* Verify the LLVM IR.  If invalid, dump and abort */
 594 #ifdef DEBUG
 595    if (LLVMVerifyFunction(func, LLVMPrintMessageAction)) {
 596       lp_debug_dump_value(func);
 597       assert(0);
 598       return;
 599    }
 600 #endif
 601
 602    gallivm_optimize_function(gallivm, func);
 603
 604    if (gallivm_debug & GALLIVM_DEBUG_IR) {
 605       /* Print the LLVM IR to stderr */
 606       lp_debug_dump_value(func);
 607       debug_printf("\n");
 608    }
 609 }
 610
 611
 612 void
 613 gallivm_compile_module(struct gallivm_state *gallivm)
 614 {
 615 #if HAVE_LLVM > 0x206
 616    assert(!gallivm->compiled);
 617 #endif
 618
 619    /* Dump byte code to a file */
 620    if (0) {
 621       LLVMWriteBitcodeToFile(gallivm->module, "llvmpipe.bc");
 622       debug_printf("llvmpipe.bc written\n");
 623       debug_printf("Invoke as \"llc -o - llvmpipe.bc\"\n");
 624    }
 625
 626 #if USE_MCJIT
 627    assert(!gallivm->engine);
 628    if (!init_gallivm_engine(gallivm)) {
 629       assert(0);
 630    }
 631 #endif
 632    assert(gallivm->engine);
 633
 634    ++gallivm->compiled;
 635 }
 636
 637
 638
 639 func_pointer
 640 gallivm_jit_function(struct gallivm_state *gallivm,
 641                      LLVMValueRef func)
 642 {
 643    void *code;
 644    func_pointer jit_func;
 645
 646    assert(gallivm->compiled);
 647    assert(gallivm->engine);
 648
 649    code = LLVMGetPointerToGlobal(gallivm->engine, func);
 650    assert(code);
 651    jit_func = pointer_to_func(code);
 652
 653    if (gallivm_debug & GALLIVM_DEBUG_ASM) {
 654       lp_disassemble(func, code);
 655    }
 656
 657 #if defined(PROFILE)
 658    lp_profile(func, code);
 659 #endif
 660
 661    /* Free the function body to save memory */
 662    lp_func_delete_body(func);
 663
 664    return jit_func;
 665 }
 666
 667
 668 /**
 669  * Free the function (and its machine code).
 670  */
 671 void
 672 gallivm_free_function(struct gallivm_state *gallivm,
 673                       LLVMValueRef func,
 674                       const void *code)
 675 {
 676 #if !USE_MCJIT
 677    if (code) {
 678       LLVMFreeMachineCodeForFunction(gallivm->engine, func);
 679    }
 680
 681    LLVMDeleteFunction(func);
 682 #endif
 683 }