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