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) || defined(PIPE_ARCH_ARM) || defined(PIPE_ARCH_AARCH64)
  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    { "no_quad_lod", GALLIVM_DEBUG_NO_QUAD_LOD, NULL },
  84    { "gc",     GALLIVM_DEBUG_GC, NULL },
  85    DEBUG_NAMED_VALUE_END
  86 };
  87
  88 DEBUG_GET_ONCE_FLAGS_OPTION(gallivm_debug, "GALLIVM_DEBUG", lp_bld_debug_flags, 0)
  89 #endif
  90
  91
  92 static boolean gallivm_initialized = FALSE;
  93
  94 unsigned lp_native_vector_width;
  95
  96
  97 /*
  98  * Optimization values are:
  99  * - 0: None (-O0)
 100  * - 1: Less (-O1)
 101  * - 2: Default (-O2, -Os)
 102  * - 3: Aggressive (-O3)
 103  *
 104  * See also CodeGenOpt::Level in llvm/Target/TargetMachine.h
 105  */
 106 enum LLVM_CodeGenOpt_Level {
 107    None,        // -O0
 108    Less,        // -O1
 109    Default,     // -O2, -Os
 110    Aggressive   // -O3
 111 };
 112
 113
 114 /**
 115  * Create the LLVM (optimization) pass manager and install
 116  * relevant optimization passes.
 117  * \return  TRUE for success, FALSE for failure
 118  */
 119 static boolean
 120 create_pass_manager(struct gallivm_state *gallivm)
 121 {
 122    assert(!gallivm->passmgr);
 123    assert(gallivm->target);
 124
 125    gallivm->passmgr = LLVMCreateFunctionPassManagerForModule(gallivm->module);
 126    if (!gallivm->passmgr)
 127       return FALSE;
 128
 129    LLVMAddTargetData(gallivm->target, gallivm->passmgr);
 130
 131    if ((gallivm_debug & GALLIVM_DEBUG_NO_OPT) == 0) {
 132       /* These are the passes currently listed in llvm-c/Transforms/Scalar.h,
 133        * but there are more on SVN.
 134        * TODO: Add more passes.
 135        */
 136       LLVMAddScalarReplAggregatesPass(gallivm->passmgr);
 137       LLVMAddLICMPass(gallivm->passmgr);
 138       LLVMAddCFGSimplificationPass(gallivm->passmgr);
 139       LLVMAddReassociatePass(gallivm->passmgr);
 140
 141       if (sizeof(void*) == 4) {
 142          /* XXX: For LLVM >= 2.7 and 32-bit build, use this order of passes to
 143           * avoid generating bad code.
 144           * Test with piglit glsl-vs-sqrt-zero test.
 145           */
 146          LLVMAddConstantPropagationPass(gallivm->passmgr);
 147          LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
 148       }
 149       else {
 150          LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
 151          LLVMAddConstantPropagationPass(gallivm->passmgr);
 152       }
 153
 154       if (util_cpu_caps.has_sse4_1) {
 155          /* FIXME: There is a bug in this pass, whereby the combination
 156           * of fptosi and sitofp (necessary for trunc/floor/ceil/round
 157           * implementation) somehow becomes invalid code.
 158           */
 159          LLVMAddInstructionCombiningPass(gallivm->passmgr);
 160       }
 161       LLVMAddGVNPass(gallivm->passmgr);
 162    }
 163    else {
 164       /* We need at least this pass to prevent the backends to fail in
 165        * unexpected ways.
 166        */
 167       LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
 168    }
 169
 170    return TRUE;
 171 }
 172
 173
 174 /**
 175  * Free gallivm object's LLVM allocations, but not the gallivm object itself.
 176  */
 177 static void
 178 free_gallivm_state(struct gallivm_state *gallivm)
 179 {
 180    if (gallivm->passmgr) {
 181       LLVMDisposePassManager(gallivm->passmgr);
 182    }
 183
 184    if (gallivm->engine) {
 185       /* This will already destroy any associated module */
 186       LLVMDisposeExecutionEngine(gallivm->engine);
 187    } else if (gallivm->module) {
 188       LLVMDisposeModule(gallivm->module);
 189    }
 190
 191 #if !USE_MCJIT
 192    /* Don't free the TargetData, it's owned by the exec engine */
 193 #else
 194    if (gallivm->target) {
 195       LLVMDisposeTargetData(gallivm->target);
 196    }
 197 #endif
 198
 199    /* Never free the LLVM context.
 200     */
 201 #if 0
 202    if (gallivm->context)
 203       LLVMContextDispose(gallivm->context);
 204 #endif
 205
 206    if (gallivm->builder)
 207       LLVMDisposeBuilder(gallivm->builder);
 208
 209    gallivm->engine = NULL;
 210    gallivm->target = NULL;
 211    gallivm->module = NULL;
 212    gallivm->passmgr = NULL;
 213    gallivm->context = NULL;
 214    gallivm->builder = NULL;
 215 }
 216
 217
 218 static boolean
 219 init_gallivm_engine(struct gallivm_state *gallivm)
 220 {
 221    if (1) {
 222       enum LLVM_CodeGenOpt_Level optlevel;
 223       char *error = NULL;
 224       int ret;
 225
 226       if (gallivm_debug & GALLIVM_DEBUG_NO_OPT) {
 227          optlevel = None;
 228       }
 229       else {
 230          optlevel = Default;
 231       }
 232
 233       ret = lp_build_create_jit_compiler_for_module(&gallivm->engine,
 234                                                     gallivm->module,
 235                                                     (unsigned) optlevel,
 236                                                     USE_MCJIT,
 237                                                     &error);
 238       if (ret) {
 239          _debug_printf("%s\n", error);
 240          LLVMDisposeMessage(error);
 241          goto fail;
 242       }
 243    }
 244
 245 #if !USE_MCJIT
 246    gallivm->target = LLVMGetExecutionEngineTargetData(gallivm->engine);
 247    if (!gallivm->target)
 248       goto fail;
 249 #else
 250    if (0) {
 251        /*
 252         * Dump the data layout strings.
 253         */
 254
 255        LLVMTargetDataRef target = LLVMGetExecutionEngineTargetData(gallivm->engine);
 256        char *data_layout;
 257        char *engine_data_layout;
 258
 259        data_layout = LLVMCopyStringRepOfTargetData(gallivm->target);
 260        engine_data_layout = LLVMCopyStringRepOfTargetData(target);
 261
 262        if (1) {
 263           debug_printf("module target data = %s\n", data_layout);
 264           debug_printf("engine target data = %s\n", engine_data_layout);
 265        }
 266
 267        free(data_layout);
 268        free(engine_data_layout);
 269    }
 270 #endif
 271
 272    return TRUE;
 273
 274 fail:
 275    return FALSE;
 276 }
 277
 278
 279 /**
 280  * Singleton
 281  *
 282  * We must never free LLVM contexts, because LLVM has several global caches
 283  * which pointing/derived from objects owned by the context, causing false
 284  * memory leaks and false cache hits when these objects are destroyed.
 285  *
 286  * TODO: For thread safety on multi-threaded OpenGL we should use one LLVM
 287  * context per thread, and put them in a pool when threads are destroyed.
 288  */
 289 static LLVMContextRef gallivm_context = NULL;
 290
 291
 292 /**
 293  * Allocate gallivm LLVM objects.
 294  * \return  TRUE for success, FALSE for failure
 295  */
 296 static boolean
 297 init_gallivm_state(struct gallivm_state *gallivm)
 298 {
 299    assert(!gallivm->context);
 300    assert(!gallivm->module);
 301
 302    lp_build_init();
 303
 304    if (!gallivm_context) {
 305       gallivm_context = LLVMContextCreate();
 306    }
 307    gallivm->context = gallivm_context;
 308    if (!gallivm->context)
 309       goto fail;
 310
 311    gallivm->module = LLVMModuleCreateWithNameInContext("gallivm",
 312                                                        gallivm->context);
 313    if (!gallivm->module)
 314       goto fail;
 315
 316    gallivm->builder = LLVMCreateBuilderInContext(gallivm->context);
 317    if (!gallivm->builder)
 318       goto fail;
 319
 320    /* FIXME: MC-JIT only allows compiling one module at a time, and it must be
 321     * complete when MC-JIT is created. So defer the MC-JIT engine creation for
 322     * now.
 323     */
 324 #if !USE_MCJIT
 325    if (!init_gallivm_engine(gallivm)) {
 326       goto fail;
 327    }
 328 #else
 329    /*
 330     * MC-JIT engine compiles the module immediately on creation, so we can't
 331     * obtain the target data from it.  Instead we create a target data layout
 332     * from a string.
 333     *
 334     * The produced layout strings are not precisely the same, but should make
 335     * no difference for the kind of optimization passes we run.
 336     *
 337     * For reference this is the layout string on x64:
 338     *
 339     *   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
 340     *
 341     * See also:
 342     * - http://llvm.org/docs/LangRef.html#datalayout
 343     */
 344
 345    {
 346       const unsigned pointer_size = 8 * sizeof(void *);
 347       char layout[512];
 348       util_snprintf(layout, sizeof layout, "%c-p:%u:%u:%u-i64:64:64-a0:0:%u-s0:%u:%u",
 349 #ifdef PIPE_ARCH_LITTLE_ENDIAN
 350                     'e', // little endian
 351 #else
 352                     'E', // big endian
 353 #endif
 354                     pointer_size, pointer_size, pointer_size, // pointer size, abi alignment, preferred alignment
 355                     pointer_size, // aggregate preferred alignment
 356                     pointer_size, pointer_size); // stack objects abi alignment, preferred alignment
 357
 358       gallivm->target = LLVMCreateTargetData(layout);
 359       if (!gallivm->target) {
 360          return FALSE;
 361       }
 362    }
 363 #endif
 364
 365    if (!create_pass_manager(gallivm))
 366       goto fail;
 367
 368    return TRUE;
 369
 370 fail:
 371    free_gallivm_state(gallivm);
 372    return FALSE;
 373 }
 374
 375
 376 void
 377 lp_build_init(void)
 378 {
 379    if (gallivm_initialized)
 380       return;
 381
 382 #ifdef DEBUG
 383    gallivm_debug = debug_get_option_gallivm_debug();
 384 #endif
 385
 386    lp_set_target_options();
 387
 388 #if USE_MCJIT
 389    LLVMLinkInMCJIT();
 390 #else
 391    LLVMLinkInJIT();
 392 #endif
 393
 394    util_cpu_detect();
 395
 396    /* AMD Bulldozer AVX's throughput is the same as SSE2; and because using
 397     * 8-wide vector needs more floating ops than 4-wide (due to padding), it is
 398     * actually more efficient to use 4-wide vectors on this processor.
 399     *
 400     * See also:
 401     * - http://www.anandtech.com/show/4955/the-bulldozer-review-amd-fx8150-tested/2
 402     */
 403    if (HAVE_AVX &&
 404        util_cpu_caps.has_avx &&
 405        util_cpu_caps.has_intel) {
 406       lp_native_vector_width = 256;
 407    } else {
 408       /* Leave it at 128, even when no SIMD extensions are available.
 409        * Really needs to be a multiple of 128 so can fit 4 floats.
 410        */
 411       lp_native_vector_width = 128;
 412    }
 413
 414    lp_native_vector_width = debug_get_num_option("LP_NATIVE_VECTOR_WIDTH",
 415                                                  lp_native_vector_width);
 416
 417    if (lp_native_vector_width <= 128) {
 418       /* Hide AVX support, as often LLVM AVX intrinsics are only guarded by
 419        * "util_cpu_caps.has_avx" predicate, and lack the
 420        * "lp_native_vector_width > 128" predicate. And also to ensure a more
 421        * consistent behavior, allowing one to test SSE2 on AVX machines.
 422        * XXX: should not play games with util_cpu_caps directly as it might
 423        * get used for other things outside llvm too.
 424        */
 425       util_cpu_caps.has_avx = 0;
 426       util_cpu_caps.has_avx2 = 0;
 427    }
 428
 429    if (!HAVE_AVX) {
 430       /*
 431        * note these instructions are VEX-only, so can only emit if we use
 432        * avx (don't want to base it on has_avx & has_f16c later as that would
 433        * omit it unnecessarily on amd cpus, see above).
 434        */
 435       util_cpu_caps.has_f16c = 0;
 436       util_cpu_caps.has_xop = 0;
 437    }
 438
 439 #ifdef PIPE_ARCH_PPC_64
 440    /* Set the NJ bit in VSCR to 0 so denormalized values are handled as
 441     * specified by IEEE standard (PowerISA 2.06 - Section 6.3). This guarantees
 442     * that some rounding and half-float to float handling does not round
 443     * incorrectly to 0.
 444     * XXX: should eventually follow same logic on all platforms.
 445     * Right now denorms get explicitly disabled (but elsewhere) for x86,
 446     * whereas ppc64 explicitly enables them...
 447     */
 448    if (util_cpu_caps.has_altivec) {
 449       unsigned short mask[] = { 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF,
 450                                 0xFFFF, 0xFFFF, 0xFFFE, 0xFFFF };
 451       __asm (
 452         "mfvscr %%v1\n"
 453         "vand   %0,%%v1,%0\n"
 454         "mtvscr %0"
 455         :
 456         : "r" (*mask)
 457       );
 458    }
 459 #endif
 460
 461    gallivm_initialized = TRUE;
 462
 463 #if 0
 464    /* For simulating less capable machines */
 465    util_cpu_caps.has_sse3 = 0;
 466    util_cpu_caps.has_ssse3 = 0;
 467    util_cpu_caps.has_sse4_1 = 0;
 468    util_cpu_caps.has_avx = 0;
 469    util_cpu_caps.has_f16c = 0;
 470 #endif
 471 }
 472
 473
 474
 475 /**
 476  * Create a new gallivm_state object.
 477  * Note that we return a singleton.
 478  */
 479 struct gallivm_state *
 480 gallivm_create(void)
 481 {
 482    struct gallivm_state *gallivm;
 483
 484    gallivm = CALLOC_STRUCT(gallivm_state);
 485    if (gallivm) {
 486       if (!init_gallivm_state(gallivm)) {
 487          FREE(gallivm);
 488          gallivm = NULL;
 489       }
 490    }
 491
 492    return gallivm;
 493 }
 494
 495
 496 /**
 497  * Destroy a gallivm_state object.
 498  */
 499 void
 500 gallivm_destroy(struct gallivm_state *gallivm)
 501 {
 502    free_gallivm_state(gallivm);
 503    FREE(gallivm);
 504 }
 505
 506
 507 /**
 508  * Validate and optimze a function.
 509  */
 510 static void
 511 gallivm_optimize_function(struct gallivm_state *gallivm,
 512                           LLVMValueRef func)
 513 {
 514    if (0) {
 515       debug_printf("optimizing %s...\n", LLVMGetValueName(func));
 516    }
 517
 518    assert(gallivm->passmgr);
 519
 520    /* Apply optimizations to LLVM IR */
 521    LLVMRunFunctionPassManager(gallivm->passmgr, func);
 522
 523    if (0) {
 524       if (gallivm_debug & GALLIVM_DEBUG_IR) {
 525          /* Print the LLVM IR to stderr */
 526          lp_debug_dump_value(func);
 527          debug_printf("\n");
 528       }
 529    }
 530 }
 531
 532
 533 /**
 534  * Validate a function.
 535  */
 536 void
 537 gallivm_verify_function(struct gallivm_state *gallivm,
 538                         LLVMValueRef func)
 539 {
 540    /* Verify the LLVM IR.  If invalid, dump and abort */
 541 #ifdef DEBUG
 542    if (LLVMVerifyFunction(func, LLVMPrintMessageAction)) {
 543       lp_debug_dump_value(func);
 544       assert(0);
 545       return;
 546    }
 547 #endif
 548
 549    gallivm_optimize_function(gallivm, func);
 550
 551    if (gallivm_debug & GALLIVM_DEBUG_IR) {
 552       /* Print the LLVM IR to stderr */
 553       lp_debug_dump_value(func);
 554       debug_printf("\n");
 555    }
 556 }
 557
 558
 559 void
 560 gallivm_compile_module(struct gallivm_state *gallivm)
 561 {
 562    assert(!gallivm->compiled);
 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    /* Free the function body to save memory */
 607    lp_func_delete_body(func);
 608
 609    return jit_func;
 610 }
 611
 612
 613 /**
 614  * Free the function (and its machine code).
 615  */
 616 void
 617 gallivm_free_function(struct gallivm_state *gallivm,
 618                       LLVMValueRef func,
 619                       const void *code)
 620 {
 621 #if !USE_MCJIT
 622    if (code) {
 623       LLVMFreeMachineCodeForFunction(gallivm->engine, func);
 624    }
 625
 626    LLVMDeleteFunction(func);
 627 #endif
 628 }