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  * LLVM has several global caches which pointing/derived from objects
  73  * owned by the context, so if we freeing contexts causes
  74  * memory leaks and false cache hits when these objects are destroyed.
  75  *
  76  * TODO: For thread safety on multi-threaded OpenGL we should use one LLVM
  77  * context per thread, and put them in a pool when threads are destroyed.
  78  */
  79 #define USE_GLOBAL_CONTEXT 1
  80
  81
  82 #ifdef DEBUG
  83 unsigned gallivm_debug = 0;
  84
  85 static const struct debug_named_value lp_bld_debug_flags[] = {
  86    { "tgsi",   GALLIVM_DEBUG_TGSI, NULL },
  87    { "ir",     GALLIVM_DEBUG_IR, NULL },
  88    { "asm",    GALLIVM_DEBUG_ASM, NULL },
  89    { "nopt",   GALLIVM_DEBUG_NO_OPT, NULL },
  90    { "perf",   GALLIVM_DEBUG_PERF, NULL },
  91    { "no_brilinear", GALLIVM_DEBUG_NO_BRILINEAR, NULL },
  92    { "no_rho_approx", GALLIVM_DEBUG_NO_RHO_APPROX, NULL },
  93    { "no_quad_lod", GALLIVM_DEBUG_NO_QUAD_LOD, NULL },
  94    { "gc",     GALLIVM_DEBUG_GC, NULL },
  95    DEBUG_NAMED_VALUE_END
  96 };
  97
  98 DEBUG_GET_ONCE_FLAGS_OPTION(gallivm_debug, "GALLIVM_DEBUG", lp_bld_debug_flags, 0)
  99 #endif
 100
 101
 102 static boolean gallivm_initialized = FALSE;
 103
 104 unsigned lp_native_vector_width;
 105
 106
 107 /*
 108  * Optimization values are:
 109  * - 0: None (-O0)
 110  * - 1: Less (-O1)
 111  * - 2: Default (-O2, -Os)
 112  * - 3: Aggressive (-O3)
 113  *
 114  * See also CodeGenOpt::Level in llvm/Target/TargetMachine.h
 115  */
 116 enum LLVM_CodeGenOpt_Level {
 117    None,        // -O0
 118    Less,        // -O1
 119    Default,     // -O2, -Os
 120    Aggressive   // -O3
 121 };
 122
 123
 124 /**
 125  * Create the LLVM (optimization) pass manager and install
 126  * relevant optimization passes.
 127  * \return  TRUE for success, FALSE for failure
 128  */
 129 static boolean
 130 create_pass_manager(struct gallivm_state *gallivm)
 131 {
 132    assert(!gallivm->passmgr);
 133    assert(gallivm->target);
 134
 135    gallivm->passmgr = LLVMCreateFunctionPassManagerForModule(gallivm->module);
 136    if (!gallivm->passmgr)
 137       return FALSE;
 138
 139    LLVMAddTargetData(gallivm->target, gallivm->passmgr);
 140
 141    if ((gallivm_debug & GALLIVM_DEBUG_NO_OPT) == 0) {
 142       /* These are the passes currently listed in llvm-c/Transforms/Scalar.h,
 143        * but there are more on SVN.
 144        * TODO: Add more passes.
 145        */
 146       LLVMAddScalarReplAggregatesPass(gallivm->passmgr);
 147       LLVMAddLICMPass(gallivm->passmgr);
 148       LLVMAddCFGSimplificationPass(gallivm->passmgr);
 149       LLVMAddReassociatePass(gallivm->passmgr);
 150
 151       if (sizeof(void*) == 4) {
 152          /* XXX: For LLVM >= 2.7 and 32-bit build, use this order of passes to
 153           * avoid generating bad code.
 154           * Test with piglit glsl-vs-sqrt-zero test.
 155           */
 156          LLVMAddConstantPropagationPass(gallivm->passmgr);
 157          LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
 158       }
 159       else {
 160          LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
 161          LLVMAddConstantPropagationPass(gallivm->passmgr);
 162       }
 163
 164       if (util_cpu_caps.has_sse4_1) {
 165          /* FIXME: There is a bug in this pass, whereby the combination
 166           * of fptosi and sitofp (necessary for trunc/floor/ceil/round
 167           * implementation) somehow becomes invalid code.
 168           */
 169          LLVMAddInstructionCombiningPass(gallivm->passmgr);
 170       }
 171       LLVMAddGVNPass(gallivm->passmgr);
 172    }
 173    else {
 174       /* We need at least this pass to prevent the backends to fail in
 175        * unexpected ways.
 176        */
 177       LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
 178    }
 179
 180    return TRUE;
 181 }
 182
 183
 184 /**
 185  * Free gallivm object's LLVM allocations, but not the gallivm object itself.
 186  */
 187 static void
 188 free_gallivm_state(struct gallivm_state *gallivm)
 189 {
 190    if (gallivm->passmgr) {
 191       LLVMDisposePassManager(gallivm->passmgr);
 192    }
 193
 194    if (gallivm->engine) {
 195       /* This will already destroy any associated module */
 196       LLVMDisposeExecutionEngine(gallivm->engine);
 197    } else if (gallivm->module) {
 198       LLVMDisposeModule(gallivm->module);
 199    }
 200
 201 #if !USE_MCJIT
 202    /* Don't free the TargetData, it's owned by the exec engine */
 203 #else
 204    if (gallivm->target) {
 205       LLVMDisposeTargetData(gallivm->target);
 206    }
 207 #endif
 208
 209    if (gallivm->builder)
 210       LLVMDisposeBuilder(gallivm->builder);
 211
 212    if (!USE_GLOBAL_CONTEXT && gallivm->context)
 213       LLVMContextDispose(gallivm->context);
 214
 215    lp_free_generated_code(gallivm->code);
 216
 217    gallivm->engine = NULL;
 218    gallivm->target = NULL;
 219    gallivm->module = NULL;
 220    gallivm->passmgr = NULL;
 221    gallivm->context = NULL;
 222    gallivm->builder = NULL;
 223    gallivm->code = NULL;
 224 }
 225
 226
 227 static boolean
 228 init_gallivm_engine(struct gallivm_state *gallivm)
 229 {
 230    if (1) {
 231       enum LLVM_CodeGenOpt_Level optlevel;
 232       char *error = NULL;
 233       int ret;
 234
 235       if (gallivm_debug & GALLIVM_DEBUG_NO_OPT) {
 236          optlevel = None;
 237       }
 238       else {
 239          optlevel = Default;
 240       }
 241
 242       ret = lp_build_create_jit_compiler_for_module(&gallivm->engine,
 243                                                     &gallivm->code,
 244                                                     gallivm->module,
 245                                                     (unsigned) optlevel,
 246                                                     USE_MCJIT,
 247                                                     &error);
 248       if (ret) {
 249          _debug_printf("%s\n", error);
 250          LLVMDisposeMessage(error);
 251          goto fail;
 252       }
 253    }
 254
 255 #if !USE_MCJIT
 256    gallivm->target = LLVMGetExecutionEngineTargetData(gallivm->engine);
 257    if (!gallivm->target)
 258       goto fail;
 259 #else
 260    if (0) {
 261        /*
 262         * Dump the data layout strings.
 263         */
 264
 265        LLVMTargetDataRef target = LLVMGetExecutionEngineTargetData(gallivm->engine);
 266        char *data_layout;
 267        char *engine_data_layout;
 268
 269        data_layout = LLVMCopyStringRepOfTargetData(gallivm->target);
 270        engine_data_layout = LLVMCopyStringRepOfTargetData(target);
 271
 272        if (1) {
 273           debug_printf("module target data = %s\n", data_layout);
 274           debug_printf("engine target data = %s\n", engine_data_layout);
 275        }
 276
 277        free(data_layout);
 278        free(engine_data_layout);
 279    }
 280 #endif
 281
 282    return TRUE;
 283
 284 fail:
 285    return FALSE;
 286 }
 287
 288
 289 /**
 290  * Allocate gallivm LLVM objects.
 291  * \return  TRUE for success, FALSE for failure
 292  */
 293 static boolean
 294 init_gallivm_state(struct gallivm_state *gallivm)
 295 {
 296    assert(!gallivm->context);
 297    assert(!gallivm->module);
 298
 299    lp_build_init();
 300
 301    if (USE_GLOBAL_CONTEXT) {
 302       gallivm->context = LLVMGetGlobalContext();
 303    } else {
 304       gallivm->context = LLVMContextCreate();
 305    }
 306    if (!gallivm->context)
 307       goto fail;
 308
 309    gallivm->module = LLVMModuleCreateWithNameInContext("gallivm",
 310                                                        gallivm->context);
 311    if (!gallivm->module)
 312       goto fail;
 313
 314    gallivm->builder = LLVMCreateBuilderInContext(gallivm->context);
 315    if (!gallivm->builder)
 316       goto fail;
 317
 318    /* FIXME: MC-JIT only allows compiling one module at a time, and it must be
 319     * complete when MC-JIT is created. So defer the MC-JIT engine creation for
 320     * now.
 321     */
 322 #if !USE_MCJIT
 323    if (!init_gallivm_engine(gallivm)) {
 324       goto fail;
 325    }
 326 #else
 327    /*
 328     * MC-JIT engine compiles the module immediately on creation, so we can't
 329     * obtain the target data from it.  Instead we create a target data layout
 330     * from a string.
 331     *
 332     * The produced layout strings are not precisely the same, but should make
 333     * no difference for the kind of optimization passes we run.
 334     *
 335     * For reference this is the layout string on x64:
 336     *
 337     *   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
 338     *
 339     * See also:
 340     * - http://llvm.org/docs/LangRef.html#datalayout
 341     */
 342
 343    {
 344       const unsigned pointer_size = 8 * sizeof(void *);
 345       char layout[512];
 346       util_snprintf(layout, sizeof layout, "%c-p:%u:%u:%u-i64:64:64-a0:0:%u-s0:%u:%u",
 347 #ifdef PIPE_ARCH_LITTLE_ENDIAN
 348                     'e', // little endian
 349 #else
 350                     'E', // big endian
 351 #endif
 352                     pointer_size, pointer_size, pointer_size, // pointer size, abi alignment, preferred alignment
 353                     pointer_size, // aggregate preferred alignment
 354                     pointer_size, pointer_size); // stack objects abi alignment, preferred alignment
 355
 356       gallivm->target = LLVMCreateTargetData(layout);
 357       if (!gallivm->target) {
 358          return FALSE;
 359       }
 360    }
 361 #endif
 362
 363    if (!create_pass_manager(gallivm))
 364       goto fail;
 365
 366    return TRUE;
 367
 368 fail:
 369    free_gallivm_state(gallivm);
 370    return FALSE;
 371 }
 372
 373
 374 void
 375 lp_build_init(void)
 376 {
 377    if (gallivm_initialized)
 378       return;
 379
 380 #ifdef DEBUG
 381    gallivm_debug = debug_get_option_gallivm_debug();
 382 #endif
 383
 384    lp_set_target_options();
 385
 386 #if USE_MCJIT
 387    LLVMLinkInMCJIT();
 388 #else
 389    LLVMLinkInJIT();
 390 #endif
 391
 392    util_cpu_detect();
 393
 394    /* AMD Bulldozer AVX's throughput is the same as SSE2; and because using
 395     * 8-wide vector needs more floating ops than 4-wide (due to padding), it is
 396     * actually more efficient to use 4-wide vectors on this processor.
 397     *
 398     * See also:
 399     * - http://www.anandtech.com/show/4955/the-bulldozer-review-amd-fx8150-tested/2
 400     */
 401    if (HAVE_AVX &&
 402        util_cpu_caps.has_avx &&
 403        util_cpu_caps.has_intel) {
 404       lp_native_vector_width = 256;
 405    } else {
 406       /* Leave it at 128, even when no SIMD extensions are available.
 407        * Really needs to be a multiple of 128 so can fit 4 floats.
 408        */
 409       lp_native_vector_width = 128;
 410    }
 411
 412    lp_native_vector_width = debug_get_num_option("LP_NATIVE_VECTOR_WIDTH",
 413                                                  lp_native_vector_width);
 414
 415    if (lp_native_vector_width <= 128) {
 416       /* Hide AVX support, as often LLVM AVX intrinsics are only guarded by
 417        * "util_cpu_caps.has_avx" predicate, and lack the
 418        * "lp_native_vector_width > 128" predicate. And also to ensure a more
 419        * consistent behavior, allowing one to test SSE2 on AVX machines.
 420        * XXX: should not play games with util_cpu_caps directly as it might
 421        * get used for other things outside llvm too.
 422        */
 423       util_cpu_caps.has_avx = 0;
 424       util_cpu_caps.has_avx2 = 0;
 425    }
 426
 427    if (!HAVE_AVX) {
 428       /*
 429        * note these instructions are VEX-only, so can only emit if we use
 430        * avx (don't want to base it on has_avx & has_f16c later as that would
 431        * omit it unnecessarily on amd cpus, see above).
 432        */
 433       util_cpu_caps.has_f16c = 0;
 434       util_cpu_caps.has_xop = 0;
 435    }
 436
 437 #ifdef PIPE_ARCH_PPC_64
 438    /* Set the NJ bit in VSCR to 0 so denormalized values are handled as
 439     * specified by IEEE standard (PowerISA 2.06 - Section 6.3). This guarantees
 440     * that some rounding and half-float to float handling does not round
 441     * incorrectly to 0.
 442     * XXX: should eventually follow same logic on all platforms.
 443     * Right now denorms get explicitly disabled (but elsewhere) for x86,
 444     * whereas ppc64 explicitly enables them...
 445     */
 446    if (util_cpu_caps.has_altivec) {
 447       unsigned short mask[] = { 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF,
 448                                 0xFFFF, 0xFFFF, 0xFFFE, 0xFFFF };
 449       __asm (
 450         "mfvscr %%v1\n"
 451         "vand   %0,%%v1,%0\n"
 452         "mtvscr %0"
 453         :
 454         : "r" (*mask)
 455       );
 456    }
 457 #endif
 458
 459    gallivm_initialized = TRUE;
 460
 461 #if 0
 462    /* For simulating less capable machines */
 463    util_cpu_caps.has_sse3 = 0;
 464    util_cpu_caps.has_ssse3 = 0;
 465    util_cpu_caps.has_sse4_1 = 0;
 466    util_cpu_caps.has_avx = 0;
 467    util_cpu_caps.has_f16c = 0;
 468 #endif
 469 }
 470
 471
 472
 473 /**
 474  * Create a new gallivm_state object.
 475  * Note that we return a singleton.
 476  */
 477 struct gallivm_state *
 478 gallivm_create(void)
 479 {
 480    struct gallivm_state *gallivm;
 481
 482    gallivm = CALLOC_STRUCT(gallivm_state);
 483    if (gallivm) {
 484       if (!init_gallivm_state(gallivm)) {
 485          FREE(gallivm);
 486          gallivm = NULL;
 487       }
 488    }
 489
 490    return gallivm;
 491 }
 492
 493
 494 /**
 495  * Destroy a gallivm_state object.
 496  */
 497 void
 498 gallivm_destroy(struct gallivm_state *gallivm)
 499 {
 500    free_gallivm_state(gallivm);
 501    FREE(gallivm);
 502 }
 503
 504
 505 /**
 506  * Validate a function.
 507  */
 508 void
 509 gallivm_verify_function(struct gallivm_state *gallivm,
 510                         LLVMValueRef func)
 511 {
 512    /* Verify the LLVM IR.  If invalid, dump and abort */
 513 #ifdef DEBUG
 514    if (LLVMVerifyFunction(func, LLVMPrintMessageAction)) {
 515       lp_debug_dump_value(func);
 516       assert(0);
 517       return;
 518    }
 519 #endif
 520
 521    if (gallivm_debug & GALLIVM_DEBUG_IR) {
 522       /* Print the LLVM IR to stderr */
 523       lp_debug_dump_value(func);
 524       debug_printf("\n");
 525    }
 526 }
 527
 528
 529 void
 530 gallivm_compile_module(struct gallivm_state *gallivm)
 531 {
 532    LLVMValueRef func;
 533
 534    assert(!gallivm->compiled);
 535
 536    if (gallivm->builder) {
 537       LLVMDisposeBuilder(gallivm->builder);
 538       gallivm->builder = NULL;
 539    }
 540
 541    /* Run optimization passes */
 542    LLVMInitializeFunctionPassManager(gallivm->passmgr);
 543    func = LLVMGetFirstFunction(gallivm->module);
 544    while (func) {
 545       if (0) {
 546          debug_printf("optimizing %s...\n", LLVMGetValueName(func));
 547       }
 548       LLVMRunFunctionPassManager(gallivm->passmgr, func);
 549       func = LLVMGetNextFunction(func);
 550    }
 551    LLVMFinalizeFunctionPassManager(gallivm->passmgr);
 552
 553    /* Dump byte code to a file */
 554    if (0) {
 555       LLVMWriteBitcodeToFile(gallivm->module, "llvmpipe.bc");
 556       debug_printf("llvmpipe.bc written\n");
 557       debug_printf("Invoke as \"llc -o - llvmpipe.bc\"\n");
 558    }
 559
 560 #if USE_MCJIT
 561    assert(!gallivm->engine);
 562    if (!init_gallivm_engine(gallivm)) {
 563       assert(0);
 564    }
 565 #endif
 566    assert(gallivm->engine);
 567
 568    ++gallivm->compiled;
 569 }
 570
 571
 572
 573 func_pointer
 574 gallivm_jit_function(struct gallivm_state *gallivm,
 575                      LLVMValueRef func)
 576 {
 577    void *code;
 578    func_pointer jit_func;
 579
 580    assert(gallivm->compiled);
 581    assert(gallivm->engine);
 582
 583    code = LLVMGetPointerToGlobal(gallivm->engine, func);
 584    assert(code);
 585    jit_func = pointer_to_func(code);
 586
 587    if (gallivm_debug & GALLIVM_DEBUG_ASM) {
 588       lp_disassemble(func, code);
 589    }
 590
 591 #if defined(PROFILE)
 592    lp_profile(func, code);
 593 #endif
 594
 595    /* Free the function body to save memory */
 596    lp_func_delete_body(func);
 597
 598    return jit_func;
 599 }
 600
 601
 602 /**
 603  * Free the function (and its machine code).
 604  */
 605 void
 606 gallivm_free_function(struct gallivm_state *gallivm,
 607                       LLVMValueRef func,
 608                       const void *code)
 609 {
 610 #if !USE_MCJIT
 611    if (code) {
 612       LLVMFreeMachineCodeForFunction(gallivm->engine, func);
 613    }
 614
 615    LLVMDeleteFunction(func);
 616 #endif
 617 }