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 /**
  47  * AVX is supported in:
  48  * - standard JIT from LLVM 3.2 onwards
  49  * - MC-JIT from LLVM 3.1
  50  *   - MC-JIT supports limited OSes (MacOSX and Linux)
  51  * - standard JIT in LLVM 3.1, with backports
  52  */
  53 #if defined(PIPE_ARCH_PPC_64) || defined(PIPE_ARCH_S390) || defined(PIPE_ARCH_ARM) || defined(PIPE_ARCH_AARCH64)
  54 #  define USE_MCJIT 1
  55 #  define HAVE_AVX 0
  56 #elif HAVE_LLVM >= 0x0302 || (HAVE_LLVM == 0x0301 && defined(HAVE_JIT_AVX_SUPPORT))
  57 #  define USE_MCJIT 0
  58 #  define HAVE_AVX 1
  59 #elif HAVE_LLVM == 0x0301 && (defined(PIPE_OS_LINUX) || defined(PIPE_OS_APPLE))
  60 #  define USE_MCJIT 1
  61 #  define HAVE_AVX 1
  62 #else
  63 #  define USE_MCJIT 0
  64 #  define HAVE_AVX 0
  65 #endif
  66
  67
  68 #if USE_MCJIT
  69 void LLVMLinkInMCJIT();
  70 #endif
  71
  72 /*
  73  * LLVM has several global caches which pointing/derived from objects
  74  * owned by the context, so if we freeing contexts causes
  75  * memory leaks and false cache hits when these objects are destroyed.
  76  *
  77  * TODO: For thread safety on multi-threaded OpenGL we should use one LLVM
  78  * context per thread, and put them in a pool when threads are destroyed.
  79  */
  80 #define USE_GLOBAL_CONTEXT 1
  81
  82
  83 #ifdef DEBUG
  84 unsigned gallivm_debug = 0;
  85
  86 static const struct debug_named_value lp_bld_debug_flags[] = {
  87    { "tgsi",   GALLIVM_DEBUG_TGSI, NULL },
  88    { "ir",     GALLIVM_DEBUG_IR, NULL },
  89    { "asm",    GALLIVM_DEBUG_ASM, NULL },
  90    { "nopt",   GALLIVM_DEBUG_NO_OPT, NULL },
  91    { "perf",   GALLIVM_DEBUG_PERF, NULL },
  92    { "no_brilinear", GALLIVM_DEBUG_NO_BRILINEAR, NULL },
  93    { "no_rho_approx", GALLIVM_DEBUG_NO_RHO_APPROX, NULL },
  94    { "no_quad_lod", GALLIVM_DEBUG_NO_QUAD_LOD, NULL },
  95    { "gc",     GALLIVM_DEBUG_GC, NULL },
  96    DEBUG_NAMED_VALUE_END
  97 };
  98
  99 DEBUG_GET_ONCE_FLAGS_OPTION(gallivm_debug, "GALLIVM_DEBUG", lp_bld_debug_flags, 0)
 100 #endif
 101
 102
 103 static boolean gallivm_initialized = FALSE;
 104
 105 unsigned lp_native_vector_width;
 106
 107
 108 /*
 109  * Optimization values are:
 110  * - 0: None (-O0)
 111  * - 1: Less (-O1)
 112  * - 2: Default (-O2, -Os)
 113  * - 3: Aggressive (-O3)
 114  *
 115  * See also CodeGenOpt::Level in llvm/Target/TargetMachine.h
 116  */
 117 enum LLVM_CodeGenOpt_Level {
 118    None,        // -O0
 119    Less,        // -O1
 120    Default,     // -O2, -Os
 121    Aggressive   // -O3
 122 };
 123
 124
 125 /**
 126  * Create the LLVM (optimization) pass manager and install
 127  * relevant optimization passes.
 128  * \return  TRUE for success, FALSE for failure
 129  */
 130 static boolean
 131 create_pass_manager(struct gallivm_state *gallivm)
 132 {
 133    assert(!gallivm->passmgr);
 134    assert(gallivm->target);
 135
 136    gallivm->passmgr = LLVMCreateFunctionPassManagerForModule(gallivm->module);
 137    if (!gallivm->passmgr)
 138       return FALSE;
 139
 140    LLVMAddTargetData(gallivm->target, gallivm->passmgr);
 141
 142    if ((gallivm_debug & GALLIVM_DEBUG_NO_OPT) == 0) {
 143       /* These are the passes currently listed in llvm-c/Transforms/Scalar.h,
 144        * but there are more on SVN.
 145        * TODO: Add more passes.
 146        */
 147       LLVMAddScalarReplAggregatesPass(gallivm->passmgr);
 148       LLVMAddLICMPass(gallivm->passmgr);
 149       LLVMAddCFGSimplificationPass(gallivm->passmgr);
 150       LLVMAddReassociatePass(gallivm->passmgr);
 151       LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
 152       LLVMAddConstantPropagationPass(gallivm->passmgr);
 153       LLVMAddInstructionCombiningPass(gallivm->passmgr);
 154       LLVMAddGVNPass(gallivm->passmgr);
 155    }
 156    else {
 157       /* We need at least this pass to prevent the backends to fail in
 158        * unexpected ways.
 159        */
 160       LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
 161    }
 162
 163    return TRUE;
 164 }
 165
 166
 167 /**
 168  * Free gallivm object's LLVM allocations, but not any generated code
 169  * nor the gallivm object itself.
 170  */
 171 void
 172 gallivm_free_ir(struct gallivm_state *gallivm)
 173 {
 174    if (gallivm->passmgr) {
 175       LLVMDisposePassManager(gallivm->passmgr);
 176    }
 177
 178    if (gallivm->engine) {
 179       /* This will already destroy any associated module */
 180       LLVMDisposeExecutionEngine(gallivm->engine);
 181    } else if (gallivm->module) {
 182       LLVMDisposeModule(gallivm->module);
 183    }
 184
 185 #if !USE_MCJIT
 186    /* Don't free the TargetData, it's owned by the exec engine */
 187 #else
 188    if (gallivm->target) {
 189       LLVMDisposeTargetData(gallivm->target);
 190    }
 191 #endif
 192
 193    if (gallivm->builder)
 194       LLVMDisposeBuilder(gallivm->builder);
 195
 196    if (!USE_GLOBAL_CONTEXT && gallivm->context)
 197       LLVMContextDispose(gallivm->context);
 198
 199    gallivm->engine = NULL;
 200    gallivm->target = NULL;
 201    gallivm->module = NULL;
 202    gallivm->passmgr = NULL;
 203    gallivm->context = NULL;
 204    gallivm->builder = NULL;
 205 }
 206
 207
 208 /**
 209  * Free LLVM-generated code.  Should be done AFTER gallivm_free_ir().
 210  */
 211 static void
 212 gallivm_free_code(struct gallivm_state *gallivm)
 213 {
 214    assert(!gallivm->module);
 215    assert(!gallivm->engine);
 216    lp_free_generated_code(gallivm->code);
 217    gallivm->code = NULL;
 218 }
 219
 220
 221 static boolean
 222 init_gallivm_engine(struct gallivm_state *gallivm)
 223 {
 224    if (1) {
 225       enum LLVM_CodeGenOpt_Level optlevel;
 226       char *error = NULL;
 227       int ret;
 228
 229       if (gallivm_debug & GALLIVM_DEBUG_NO_OPT) {
 230          optlevel = None;
 231       }
 232       else {
 233          optlevel = Default;
 234       }
 235
 236       ret = lp_build_create_jit_compiler_for_module(&gallivm->engine,
 237                                                     &gallivm->code,
 238                                                     gallivm->module,
 239                                                     (unsigned) optlevel,
 240                                                     USE_MCJIT,
 241                                                     &error);
 242       if (ret) {
 243          _debug_printf("%s\n", error);
 244          LLVMDisposeMessage(error);
 245          goto fail;
 246       }
 247    }
 248
 249 #if !USE_MCJIT
 250    gallivm->target = LLVMGetExecutionEngineTargetData(gallivm->engine);
 251    if (!gallivm->target)
 252       goto fail;
 253 #else
 254    if (0) {
 255        /*
 256         * Dump the data layout strings.
 257         */
 258
 259        LLVMTargetDataRef target = LLVMGetExecutionEngineTargetData(gallivm->engine);
 260        char *data_layout;
 261        char *engine_data_layout;
 262
 263        data_layout = LLVMCopyStringRepOfTargetData(gallivm->target);
 264        engine_data_layout = LLVMCopyStringRepOfTargetData(target);
 265
 266        if (1) {
 267           debug_printf("module target data = %s\n", data_layout);
 268           debug_printf("engine target data = %s\n", engine_data_layout);
 269        }
 270
 271        free(data_layout);
 272        free(engine_data_layout);
 273    }
 274 #endif
 275
 276    return TRUE;
 277
 278 fail:
 279    return FALSE;
 280 }
 281
 282
 283 /**
 284  * Allocate gallivm LLVM objects.
 285  * \return  TRUE for success, FALSE for failure
 286  */
 287 static boolean
 288 init_gallivm_state(struct gallivm_state *gallivm, const char *name)
 289 {
 290    assert(!gallivm->context);
 291    assert(!gallivm->module);
 292
 293    lp_build_init();
 294
 295    if (USE_GLOBAL_CONTEXT) {
 296       gallivm->context = LLVMGetGlobalContext();
 297    } else {
 298       gallivm->context = LLVMContextCreate();
 299    }
 300    if (!gallivm->context)
 301       goto fail;
 302
 303    gallivm->module = LLVMModuleCreateWithNameInContext(name,
 304                                                        gallivm->context);
 305    if (!gallivm->module)
 306       goto fail;
 307
 308    gallivm->builder = LLVMCreateBuilderInContext(gallivm->context);
 309    if (!gallivm->builder)
 310       goto fail;
 311
 312    /* FIXME: MC-JIT only allows compiling one module at a time, and it must be
 313     * complete when MC-JIT is created. So defer the MC-JIT engine creation for
 314     * now.
 315     */
 316 #if !USE_MCJIT
 317    if (!init_gallivm_engine(gallivm)) {
 318       goto fail;
 319    }
 320 #else
 321    /*
 322     * MC-JIT engine compiles the module immediately on creation, so we can't
 323     * obtain the target data from it.  Instead we create a target data layout
 324     * from a string.
 325     *
 326     * The produced layout strings are not precisely the same, but should make
 327     * no difference for the kind of optimization passes we run.
 328     *
 329     * For reference this is the layout string on x64:
 330     *
 331     *   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
 332     *
 333     * See also:
 334     * - http://llvm.org/docs/LangRef.html#datalayout
 335     */
 336
 337    {
 338       const unsigned pointer_size = 8 * sizeof(void *);
 339       char layout[512];
 340       util_snprintf(layout, sizeof layout, "%c-p:%u:%u:%u-i64:64:64-a0:0:%u-s0:%u:%u",
 341 #ifdef PIPE_ARCH_LITTLE_ENDIAN
 342                     'e', // little endian
 343 #else
 344                     'E', // big endian
 345 #endif
 346                     pointer_size, pointer_size, pointer_size, // pointer size, abi alignment, preferred alignment
 347                     pointer_size, // aggregate preferred alignment
 348                     pointer_size, pointer_size); // stack objects abi alignment, preferred alignment
 349
 350       gallivm->target = LLVMCreateTargetData(layout);
 351       if (!gallivm->target) {
 352          return FALSE;
 353       }
 354    }
 355 #endif
 356
 357    if (!create_pass_manager(gallivm))
 358       goto fail;
 359
 360    return TRUE;
 361
 362 fail:
 363    gallivm_free_ir(gallivm);
 364    gallivm_free_code(gallivm);
 365    return FALSE;
 366 }
 367
 368
 369 void
 370 lp_build_init(void)
 371 {
 372    if (gallivm_initialized)
 373       return;
 374
 375 #ifdef DEBUG
 376    gallivm_debug = debug_get_option_gallivm_debug();
 377 #endif
 378
 379    lp_set_target_options();
 380
 381 #if USE_MCJIT
 382    LLVMLinkInMCJIT();
 383 #else
 384    LLVMLinkInJIT();
 385 #endif
 386
 387    util_cpu_detect();
 388
 389    /* AMD Bulldozer AVX's throughput is the same as SSE2; and because using
 390     * 8-wide vector needs more floating ops than 4-wide (due to padding), it is
 391     * actually more efficient to use 4-wide vectors on this processor.
 392     *
 393     * See also:
 394     * - http://www.anandtech.com/show/4955/the-bulldozer-review-amd-fx8150-tested/2
 395     */
 396    if (HAVE_AVX &&
 397        util_cpu_caps.has_avx &&
 398        util_cpu_caps.has_intel) {
 399       lp_native_vector_width = 256;
 400    } else {
 401       /* Leave it at 128, even when no SIMD extensions are available.
 402        * Really needs to be a multiple of 128 so can fit 4 floats.
 403        */
 404       lp_native_vector_width = 128;
 405    }
 406
 407    lp_native_vector_width = debug_get_num_option("LP_NATIVE_VECTOR_WIDTH",
 408                                                  lp_native_vector_width);
 409
 410    if (lp_native_vector_width <= 128) {
 411       /* Hide AVX support, as often LLVM AVX intrinsics are only guarded by
 412        * "util_cpu_caps.has_avx" predicate, and lack the
 413        * "lp_native_vector_width > 128" predicate. And also to ensure a more
 414        * consistent behavior, allowing one to test SSE2 on AVX machines.
 415        * XXX: should not play games with util_cpu_caps directly as it might
 416        * get used for other things outside llvm too.
 417        */
 418       util_cpu_caps.has_avx = 0;
 419       util_cpu_caps.has_avx2 = 0;
 420    }
 421
 422    if (!HAVE_AVX) {
 423       /*
 424        * note these instructions are VEX-only, so can only emit if we use
 425        * avx (don't want to base it on has_avx & has_f16c later as that would
 426        * omit it unnecessarily on amd cpus, see above).
 427        */
 428       util_cpu_caps.has_f16c = 0;
 429       util_cpu_caps.has_xop = 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
 466
 467
 468 /**
 469  * Create a new gallivm_state object.
 470  */
 471 struct gallivm_state *
 472 gallivm_create(const char *name)
 473 {
 474    struct gallivm_state *gallivm;
 475
 476    gallivm = CALLOC_STRUCT(gallivm_state);
 477    if (gallivm) {
 478       if (!init_gallivm_state(gallivm, name)) {
 479          FREE(gallivm);
 480          gallivm = NULL;
 481       }
 482    }
 483
 484    return gallivm;
 485 }
 486
 487
 488 /**
 489  * Destroy a gallivm_state object.
 490  */
 491 void
 492 gallivm_destroy(struct gallivm_state *gallivm)
 493 {
 494    gallivm_free_ir(gallivm);
 495    gallivm_free_code(gallivm);
 496    FREE(gallivm);
 497 }
 498
 499
 500 /**
 501  * Validate a function.
 502  * Verification is only done with debug builds.
 503  */
 504 void
 505 gallivm_verify_function(struct gallivm_state *gallivm,
 506                         LLVMValueRef func)
 507 {
 508    /* Verify the LLVM IR.  If invalid, dump and abort */
 509 #ifdef DEBUG
 510    if (LLVMVerifyFunction(func, LLVMPrintMessageAction)) {
 511       lp_debug_dump_value(func);
 512       assert(0);
 513       return;
 514    }
 515 #endif
 516
 517    if (gallivm_debug & GALLIVM_DEBUG_IR) {
 518       /* Print the LLVM IR to stderr */
 519       lp_debug_dump_value(func);
 520       debug_printf("\n");
 521    }
 522 }
 523
 524
 525 /**
 526  * Compile a module.
 527  * This does IR optimization on all functions in the module.
 528  */
 529 void
 530 gallivm_compile_module(struct gallivm_state *gallivm)
 531 {
 532    LLVMValueRef func;
 533    int64_t time_begin;
 534
 535    assert(!gallivm->compiled);
 536
 537    if (gallivm->builder) {
 538       LLVMDisposeBuilder(gallivm->builder);
 539       gallivm->builder = NULL;
 540    }
 541
 542    if (gallivm_debug & GALLIVM_DEBUG_PERF)
 543       time_begin = os_time_get();
 544
 545    /* Run optimization passes */
 546    LLVMInitializeFunctionPassManager(gallivm->passmgr);
 547    func = LLVMGetFirstFunction(gallivm->module);
 548    while (func) {
 549       if (0) {
 550          debug_printf("optimizing func %s...\n", LLVMGetValueName(func));
 551       }
 552       LLVMRunFunctionPassManager(gallivm->passmgr, func);
 553       func = LLVMGetNextFunction(func);
 554    }
 555    LLVMFinalizeFunctionPassManager(gallivm->passmgr);
 556
 557    if (gallivm_debug & GALLIVM_DEBUG_PERF) {
 558       int64_t time_end = os_time_get();
 559       int time_msec = (int)(time_end - time_begin) / 1000;
 560       debug_printf("optimizing module %s took %d msec\n",
 561                    lp_get_module_id(gallivm->module), time_msec);
 562    }
 563
 564    /* Dump byte code to a file */
 565    if (0) {
 566       LLVMWriteBitcodeToFile(gallivm->module, "llvmpipe.bc");
 567       debug_printf("llvmpipe.bc written\n");
 568       debug_printf("Invoke as \"llc -o - llvmpipe.bc\"\n");
 569    }
 570
 571 #if USE_MCJIT
 572    assert(!gallivm->engine);
 573    if (!init_gallivm_engine(gallivm)) {
 574       assert(0);
 575    }
 576 #endif
 577    assert(gallivm->engine);
 578
 579    ++gallivm->compiled;
 580 }
 581
 582
 583
 584 func_pointer
 585 gallivm_jit_function(struct gallivm_state *gallivm,
 586                      LLVMValueRef func)
 587 {
 588    void *code;
 589    func_pointer jit_func;
 590
 591    assert(gallivm->compiled);
 592    assert(gallivm->engine);
 593
 594    code = LLVMGetPointerToGlobal(gallivm->engine, func);
 595    assert(code);
 596    jit_func = pointer_to_func(code);
 597
 598    if (gallivm_debug & GALLIVM_DEBUG_ASM) {
 599       lp_disassemble(func, code);
 600    }
 601
 602 #if defined(PROFILE)
 603    lp_profile(func, code);
 604 #endif
 605
 606    return jit_func;
 607 }