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    assert(!gallivm->passmgr);
 110    assert(gallivm->target);
 111
 112    gallivm->passmgr = LLVMCreateFunctionPassManagerForModule(gallivm->module);
 113    if (!gallivm->passmgr)
 114       return FALSE;
 115    /*
 116     * TODO: some per module pass manager with IPO passes might be helpful -
 117     * the generated texture functions may benefit from inlining if they are
 118     * simple, or constant propagation into them, etc.
 119     */
 120
 121 #if HAVE_LLVM < 0x0309
 122    // Old versions of LLVM get the DataLayout from the pass manager.
 123    LLVMAddTargetData(gallivm->target, gallivm->passmgr);
 124 #endif
 125
 126    /* Setting the module's DataLayout to an empty string will cause the
 127     * ExecutionEngine to copy to the DataLayout string from its target
 128     * machine to the module.  As of LLVM 3.8 the module and the execution
 129     * engine are required to have the same DataLayout.
 130     *
 131     * TODO: This is just a temporary work-around.  The correct solution is
 132     * for gallivm_init_state() to create a TargetMachine and pull the
 133     * DataLayout from there.  Currently, the TargetMachine used by llvmpipe
 134     * is being implicitly created by the EngineBuilder in
 135     * lp_build_create_jit_compiler_for_module()
 136     */
 137
 138 #if HAVE_LLVM < 0x0308
 139    {
 140       char *td_str;
 141       // New ones from the Module.
 142       td_str = LLVMCopyStringRepOfTargetData(gallivm->target);
 143       LLVMSetDataLayout(gallivm->module, td_str);
 144       free(td_str);
 145    }
 146 #else
 147    LLVMSetDataLayout(gallivm->module, "");
 148 #endif
 149
 150    if ((gallivm_debug & GALLIVM_DEBUG_NO_OPT) == 0) {
 151       /* These are the passes currently listed in llvm-c/Transforms/Scalar.h,
 152        * but there are more on SVN.
 153        * TODO: Add more passes.
 154        */
 155       LLVMAddScalarReplAggregatesPass(gallivm->passmgr);
 156       LLVMAddLICMPass(gallivm->passmgr);
 157       LLVMAddCFGSimplificationPass(gallivm->passmgr);
 158       LLVMAddReassociatePass(gallivm->passmgr);
 159       LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
 160       LLVMAddConstantPropagationPass(gallivm->passmgr);
 161       LLVMAddInstructionCombiningPass(gallivm->passmgr);
 162       LLVMAddGVNPass(gallivm->passmgr);
 163    }
 164    else {
 165       /* We need at least this pass to prevent the backends to fail in
 166        * unexpected ways.
 167        */
 168       LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
 169    }
 170
 171    return TRUE;
 172 }
 173
 174
 175 /**
 176  * Free gallivm object's LLVM allocations, but not any generated code
 177  * nor the gallivm object itself.
 178  */
 179 void
 180 gallivm_free_ir(struct gallivm_state *gallivm)
 181 {
 182    if (gallivm->passmgr) {
 183       LLVMDisposePassManager(gallivm->passmgr);
 184    }
 185
 186    if (gallivm->engine) {
 187       /* This will already destroy any associated module */
 188       LLVMDisposeExecutionEngine(gallivm->engine);
 189    } else if (gallivm->module) {
 190       LLVMDisposeModule(gallivm->module);
 191    }
 192
 193 #if !USE_MCJIT
 194    /* Don't free the TargetData, it's owned by the exec engine */
 195 #else
 196    if (gallivm->target) {
 197       LLVMDisposeTargetData(gallivm->target);
 198    }
 199 #endif
 200
 201    if (gallivm->builder)
 202       LLVMDisposeBuilder(gallivm->builder);
 203
 204    /* The LLVMContext should be owned by the parent of gallivm. */
 205
 206    gallivm->engine = NULL;
 207    gallivm->target = NULL;
 208    gallivm->module = NULL;
 209    gallivm->passmgr = NULL;
 210    gallivm->context = NULL;
 211    gallivm->builder = NULL;
 212 }
 213
 214
 215 /**
 216  * Free LLVM-generated code.  Should be done AFTER gallivm_free_ir().
 217  */
 218 static void
 219 gallivm_free_code(struct gallivm_state *gallivm)
 220 {
 221    assert(!gallivm->module);
 222    assert(!gallivm->engine);
 223    lp_free_generated_code(gallivm->code);
 224    gallivm->code = NULL;
 225    lp_free_memory_manager(gallivm->memorymgr);
 226    gallivm->memorymgr = NULL;
 227 }
 228
 229
 230 static boolean
 231 init_gallivm_engine(struct gallivm_state *gallivm)
 232 {
 233    if (1) {
 234       enum LLVM_CodeGenOpt_Level optlevel;
 235       char *error = NULL;
 236       int ret;
 237
 238       if (gallivm_debug & GALLIVM_DEBUG_NO_OPT) {
 239          optlevel = None;
 240       }
 241       else {
 242          optlevel = Default;
 243       }
 244
 245       ret = lp_build_create_jit_compiler_for_module(&gallivm->engine,
 246                                                     &gallivm->code,
 247                                                     gallivm->module,
 248                                                     gallivm->memorymgr,
 249                                                     (unsigned) optlevel,
 250                                                     USE_MCJIT,
 251                                                     &error);
 252       if (ret) {
 253          _debug_printf("%s\n", error);
 254          LLVMDisposeMessage(error);
 255          goto fail;
 256       }
 257    }
 258
 259 #if !USE_MCJIT
 260    gallivm->target = LLVMGetExecutionEngineTargetData(gallivm->engine);
 261    if (!gallivm->target)
 262       goto fail;
 263 #else
 264    if (0) {
 265        /*
 266         * Dump the data layout strings.
 267         */
 268
 269        LLVMTargetDataRef target = LLVMGetExecutionEngineTargetData(gallivm->engine);
 270        char *data_layout;
 271        char *engine_data_layout;
 272
 273        data_layout = LLVMCopyStringRepOfTargetData(gallivm->target);
 274        engine_data_layout = LLVMCopyStringRepOfTargetData(target);
 275
 276        if (1) {
 277           debug_printf("module target data = %s\n", data_layout);
 278           debug_printf("engine target data = %s\n", engine_data_layout);
 279        }
 280
 281        free(data_layout);
 282        free(engine_data_layout);
 283    }
 284 #endif
 285
 286    return TRUE;
 287
 288 fail:
 289    return FALSE;
 290 }
 291
 292
 293 /**
 294  * Allocate gallivm LLVM objects.
 295  * \return  TRUE for success, FALSE for failure
 296  */
 297 static boolean
 298 init_gallivm_state(struct gallivm_state *gallivm, const char *name,
 299                    LLVMContextRef context)
 300 {
 301    assert(!gallivm->context);
 302    assert(!gallivm->module);
 303
 304    if (!lp_build_init())
 305       return FALSE;
 306
 307    gallivm->context = context;
 308
 309    if (!gallivm->context)
 310       goto fail;
 311
 312    gallivm->module = LLVMModuleCreateWithNameInContext(name,
 313                                                        gallivm->context);
 314    if (!gallivm->module)
 315       goto fail;
 316
 317    gallivm->builder = LLVMCreateBuilderInContext(gallivm->context);
 318    if (!gallivm->builder)
 319       goto fail;
 320
 321    gallivm->memorymgr = lp_get_default_memory_manager();
 322    if (!gallivm->memorymgr)
 323       goto fail;
 324
 325    /* FIXME: MC-JIT only allows compiling one module at a time, and it must be
 326     * complete when MC-JIT is created. So defer the MC-JIT engine creation for
 327     * now.
 328     */
 329 #if !USE_MCJIT
 330    if (!init_gallivm_engine(gallivm)) {
 331       goto fail;
 332    }
 333 #else
 334    /*
 335     * MC-JIT engine compiles the module immediately on creation, so we can't
 336     * obtain the target data from it.  Instead we create a target data layout
 337     * from a string.
 338     *
 339     * The produced layout strings are not precisely the same, but should make
 340     * no difference for the kind of optimization passes we run.
 341     *
 342     * For reference this is the layout string on x64:
 343     *
 344     *   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
 345     *
 346     * See also:
 347     * - http://llvm.org/docs/LangRef.html#datalayout
 348     */
 349
 350    {
 351       const unsigned pointer_size = 8 * sizeof(void *);
 352       char layout[512];
 353       util_snprintf(layout, sizeof layout, "%c-p:%u:%u:%u-i64:64:64-a0:0:%u-s0:%u:%u",
 354 #ifdef PIPE_ARCH_LITTLE_ENDIAN
 355                     'e', // little endian
 356 #else
 357                     'E', // big endian
 358 #endif
 359                     pointer_size, pointer_size, pointer_size, // pointer size, abi alignment, preferred alignment
 360                     pointer_size, // aggregate preferred alignment
 361                     pointer_size, pointer_size); // stack objects abi alignment, preferred alignment
 362
 363       gallivm->target = LLVMCreateTargetData(layout);
 364       if (!gallivm->target) {
 365          return FALSE;
 366       }
 367    }
 368 #endif
 369
 370    if (!create_pass_manager(gallivm))
 371       goto fail;
 372
 373    return TRUE;
 374
 375 fail:
 376    gallivm_free_ir(gallivm);
 377    gallivm_free_code(gallivm);
 378    return FALSE;
 379 }
 380
 381
 382 boolean
 383 lp_build_init(void)
 384 {
 385    if (gallivm_initialized)
 386       return TRUE;
 387
 388 #ifdef DEBUG
 389    gallivm_debug = debug_get_option_gallivm_debug();
 390 #endif
 391
 392    lp_set_target_options();
 393
 394 #if USE_MCJIT
 395    LLVMLinkInMCJIT();
 396 #else
 397    LLVMLinkInJIT();
 398 #endif
 399
 400    util_cpu_detect();
 401
 402    /* For simulating less capable machines */
 403 #ifdef DEBUG
 404    if (debug_get_bool_option("LP_FORCE_SSE2", FALSE)) {
 405       assert(util_cpu_caps.has_sse2);
 406       util_cpu_caps.has_sse3 = 0;
 407       util_cpu_caps.has_ssse3 = 0;
 408       util_cpu_caps.has_sse4_1 = 0;
 409       util_cpu_caps.has_sse4_2 = 0;
 410       util_cpu_caps.has_avx = 0;
 411       util_cpu_caps.has_avx2 = 0;
 412       util_cpu_caps.has_f16c = 0;
 413    }
 414 #endif
 415
 416    /* AMD Bulldozer AVX's throughput is the same as SSE2; and because using
 417     * 8-wide vector needs more floating ops than 4-wide (due to padding), it is
 418     * actually more efficient to use 4-wide vectors on this processor.
 419     *
 420     * See also:
 421     * - http://www.anandtech.com/show/4955/the-bulldozer-review-amd-fx8150-tested/2
 422     */
 423    if (util_cpu_caps.has_avx &&
 424        util_cpu_caps.has_intel) {
 425       lp_native_vector_width = 256;
 426    } else {
 427       /* Leave it at 128, even when no SIMD extensions are available.
 428        * Really needs to be a multiple of 128 so can fit 4 floats.
 429        */
 430       lp_native_vector_width = 128;
 431    }
 432
 433    lp_native_vector_width = debug_get_num_option("LP_NATIVE_VECTOR_WIDTH",
 434                                                  lp_native_vector_width);
 435
 436    if (lp_native_vector_width <= 128) {
 437       /* Hide AVX support, as often LLVM AVX intrinsics are only guarded by
 438        * "util_cpu_caps.has_avx" predicate, and lack the
 439        * "lp_native_vector_width > 128" predicate. And also to ensure a more
 440        * consistent behavior, allowing one to test SSE2 on AVX machines.
 441        * XXX: should not play games with util_cpu_caps directly as it might
 442        * get used for other things outside llvm too.
 443        */
 444       util_cpu_caps.has_avx = 0;
 445       util_cpu_caps.has_avx2 = 0;
 446       util_cpu_caps.has_f16c = 0;
 447    }
 448
 449 #ifdef PIPE_ARCH_PPC_64
 450    /* Set the NJ bit in VSCR to 0 so denormalized values are handled as
 451     * specified by IEEE standard (PowerISA 2.06 - Section 6.3). This guarantees
 452     * that some rounding and half-float to float handling does not round
 453     * incorrectly to 0.
 454     * XXX: should eventually follow same logic on all platforms.
 455     * Right now denorms get explicitly disabled (but elsewhere) for x86,
 456     * whereas ppc64 explicitly enables them...
 457     */
 458    if (util_cpu_caps.has_altivec) {
 459       unsigned short mask[] = { 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF,
 460                                 0xFFFF, 0xFFFF, 0xFFFE, 0xFFFF };
 461       __asm (
 462         "mfvscr %%v1\n"
 463         "vand   %0,%%v1,%0\n"
 464         "mtvscr %0"
 465         :
 466         : "r" (*mask)
 467       );
 468    }
 469 #endif
 470
 471    gallivm_initialized = TRUE;
 472
 473    return TRUE;
 474 }
 475
 476
 477
 478 /**
 479  * Create a new gallivm_state object.
 480  */
 481 struct gallivm_state *
 482 gallivm_create(const char *name, LLVMContextRef context)
 483 {
 484    struct gallivm_state *gallivm;
 485
 486    gallivm = CALLOC_STRUCT(gallivm_state);
 487    if (gallivm) {
 488       if (!init_gallivm_state(gallivm, name, context)) {
 489          FREE(gallivm);
 490          gallivm = NULL;
 491       }
 492    }
 493
 494    return gallivm;
 495 }
 496
 497
 498 /**
 499  * Destroy a gallivm_state object.
 500  */
 501 void
 502 gallivm_destroy(struct gallivm_state *gallivm)
 503 {
 504    gallivm_free_ir(gallivm);
 505    gallivm_free_code(gallivm);
 506    FREE(gallivm);
 507 }
 508
 509
 510 /**
 511  * Validate a function.
 512  * Verification is only done with debug builds.
 513  */
 514 void
 515 gallivm_verify_function(struct gallivm_state *gallivm,
 516                         LLVMValueRef func)
 517 {
 518    /* Verify the LLVM IR.  If invalid, dump and abort */
 519 #ifdef DEBUG
 520    if (LLVMVerifyFunction(func, LLVMPrintMessageAction)) {
 521       lp_debug_dump_value(func);
 522       assert(0);
 523       return;
 524    }
 525 #endif
 526
 527    if (gallivm_debug & GALLIVM_DEBUG_IR) {
 528       /* Print the LLVM IR to stderr */
 529       lp_debug_dump_value(func);
 530       debug_printf("\n");
 531    }
 532 }
 533
 534
 535 /**
 536  * Compile a module.
 537  * This does IR optimization on all functions in the module.
 538  */
 539 void
 540 gallivm_compile_module(struct gallivm_state *gallivm)
 541 {
 542    LLVMValueRef func;
 543    int64_t time_begin = 0;
 544
 545    assert(!gallivm->compiled);
 546
 547    if (gallivm->builder) {
 548       LLVMDisposeBuilder(gallivm->builder);
 549       gallivm->builder = NULL;
 550    }
 551
 552    if (gallivm_debug & GALLIVM_DEBUG_PERF)
 553       time_begin = os_time_get();
 554
 555    /* Run optimization passes */
 556    LLVMInitializeFunctionPassManager(gallivm->passmgr);
 557    func = LLVMGetFirstFunction(gallivm->module);
 558    while (func) {
 559       if (0) {
 560          debug_printf("optimizing func %s...\n", LLVMGetValueName(func));
 561       }
 562
 563    /* Disable frame pointer omission on debug/profile builds */
 564    /* XXX: And workaround http://llvm.org/PR21435 */
 565 #if HAVE_LLVM >= 0x0307 && \
 566     (defined(DEBUG) || defined(PROFILE) || \
 567      defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64))
 568       LLVMAddTargetDependentFunctionAttr(func, "no-frame-pointer-elim", "true");
 569       LLVMAddTargetDependentFunctionAttr(func, "no-frame-pointer-elim-non-leaf", "true");
 570 #endif
 571
 572       LLVMRunFunctionPassManager(gallivm->passmgr, func);
 573       func = LLVMGetNextFunction(func);
 574    }
 575    LLVMFinalizeFunctionPassManager(gallivm->passmgr);
 576
 577    if (gallivm_debug & GALLIVM_DEBUG_PERF) {
 578       int64_t time_end = os_time_get();
 579       int time_msec = (int)(time_end - time_begin) / 1000;
 580       debug_printf("optimizing module %s took %d msec\n",
 581                    lp_get_module_id(gallivm->module), time_msec);
 582    }
 583
 584    /* Dump byte code to a file */
 585    if (0) {
 586       LLVMWriteBitcodeToFile(gallivm->module, "llvmpipe.bc");
 587       debug_printf("llvmpipe.bc written\n");
 588       debug_printf("Invoke as \"llc -o - llvmpipe.bc\"\n");
 589    }
 590
 591 #if USE_MCJIT
 592    assert(!gallivm->engine);
 593    if (!init_gallivm_engine(gallivm)) {
 594       assert(0);
 595    }
 596 #endif
 597    assert(gallivm->engine);
 598
 599    ++gallivm->compiled;
 600
 601    if (gallivm_debug & GALLIVM_DEBUG_ASM) {
 602       LLVMValueRef llvm_func = LLVMGetFirstFunction(gallivm->module);
 603
 604       while (llvm_func) {
 605          /*
 606           * Need to filter out functions which don't have an implementation,
 607           * such as the intrinsics. May not be sufficient in case of IPO?
 608           * LLVMGetPointerToGlobal() will abort otherwise.
 609           */
 610          if (!LLVMIsDeclaration(llvm_func)) {
 611             void *func_code = LLVMGetPointerToGlobal(gallivm->engine, llvm_func);
 612             lp_disassemble(llvm_func, func_code);
 613          }
 614          llvm_func = LLVMGetNextFunction(llvm_func);
 615       }
 616    }
 617
 618 #if defined(PROFILE)
 619    {
 620       LLVMValueRef llvm_func = LLVMGetFirstFunction(gallivm->module);
 621
 622       while (llvm_func) {
 623          if (!LLVMIsDeclaration(llvm_func)) {
 624             void *func_code = LLVMGetPointerToGlobal(gallivm->engine, llvm_func);
 625             lp_profile(llvm_func, func_code);
 626          }
 627          llvm_func = LLVMGetNextFunction(llvm_func);
 628       }
 629    }
 630 #endif
 631 }
 632
 633
 634
 635 func_pointer
 636 gallivm_jit_function(struct gallivm_state *gallivm,
 637                      LLVMValueRef func)
 638 {
 639    void *code;
 640    func_pointer jit_func;
 641
 642    assert(gallivm->compiled);
 643    assert(gallivm->engine);
 644
 645    code = LLVMGetPointerToGlobal(gallivm->engine, func);
 646    assert(code);
 647    jit_func = pointer_to_func(code);
 648
 649    return jit_func;
 650 }