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 any generated code
 186  * nor the gallivm object itself.
 187  */
 188 void
 189 gallivm_free_ir(struct gallivm_state *gallivm)
 190 {
 191    if (gallivm->passmgr) {
 192       LLVMDisposePassManager(gallivm->passmgr);
 193    }
 194
 195    if (gallivm->engine) {
 196       /* This will already destroy any associated module */
 197       LLVMDisposeExecutionEngine(gallivm->engine);
 198    } else if (gallivm->module) {
 199       LLVMDisposeModule(gallivm->module);
 200    }
 201
 202 #if !USE_MCJIT
 203    /* Don't free the TargetData, it's owned by the exec engine */
 204 #else
 205    if (gallivm->target) {
 206       LLVMDisposeTargetData(gallivm->target);
 207    }
 208 #endif
 209
 210    if (gallivm->builder)
 211       LLVMDisposeBuilder(gallivm->builder);
 212
 213    if (!USE_GLOBAL_CONTEXT && gallivm->context)
 214       LLVMContextDispose(gallivm->context);
 215
 216    gallivm->engine = NULL;
 217    gallivm->target = NULL;
 218    gallivm->module = NULL;
 219    gallivm->passmgr = NULL;
 220    gallivm->context = NULL;
 221    gallivm->builder = NULL;
 222 }
 223
 224
 225 /**
 226  * Free LLVM-generated code.  Should be done AFTER gallivm_free_ir().
 227  */
 228 static void
 229 gallivm_free_code(struct gallivm_state *gallivm)
 230 {
 231    assert(!gallivm->module);
 232    assert(!gallivm->engine);
 233    lp_free_generated_code(gallivm->code);
 234    gallivm->code = NULL;
 235 }
 236
 237
 238 static boolean
 239 init_gallivm_engine(struct gallivm_state *gallivm)
 240 {
 241    if (1) {
 242       enum LLVM_CodeGenOpt_Level optlevel;
 243       char *error = NULL;
 244       int ret;
 245
 246       if (gallivm_debug & GALLIVM_DEBUG_NO_OPT) {
 247          optlevel = None;
 248       }
 249       else {
 250          optlevel = Default;
 251       }
 252
 253       ret = lp_build_create_jit_compiler_for_module(&gallivm->engine,
 254                                                     &gallivm->code,
 255                                                     gallivm->module,
 256                                                     (unsigned) optlevel,
 257                                                     USE_MCJIT,
 258                                                     &error);
 259       if (ret) {
 260          _debug_printf("%s\n", error);
 261          LLVMDisposeMessage(error);
 262          goto fail;
 263       }
 264    }
 265
 266 #if !USE_MCJIT
 267    gallivm->target = LLVMGetExecutionEngineTargetData(gallivm->engine);
 268    if (!gallivm->target)
 269       goto fail;
 270 #else
 271    if (0) {
 272        /*
 273         * Dump the data layout strings.
 274         */
 275
 276        LLVMTargetDataRef target = LLVMGetExecutionEngineTargetData(gallivm->engine);
 277        char *data_layout;
 278        char *engine_data_layout;
 279
 280        data_layout = LLVMCopyStringRepOfTargetData(gallivm->target);
 281        engine_data_layout = LLVMCopyStringRepOfTargetData(target);
 282
 283        if (1) {
 284           debug_printf("module target data = %s\n", data_layout);
 285           debug_printf("engine target data = %s\n", engine_data_layout);
 286        }
 287
 288        free(data_layout);
 289        free(engine_data_layout);
 290    }
 291 #endif
 292
 293    return TRUE;
 294
 295 fail:
 296    return FALSE;
 297 }
 298
 299
 300 /**
 301  * Allocate gallivm LLVM objects.
 302  * \return  TRUE for success, FALSE for failure
 303  */
 304 static boolean
 305 init_gallivm_state(struct gallivm_state *gallivm)
 306 {
 307    assert(!gallivm->context);
 308    assert(!gallivm->module);
 309
 310    lp_build_init();
 311
 312    if (USE_GLOBAL_CONTEXT) {
 313       gallivm->context = LLVMGetGlobalContext();
 314    } else {
 315       gallivm->context = LLVMContextCreate();
 316    }
 317    if (!gallivm->context)
 318       goto fail;
 319
 320    gallivm->module = LLVMModuleCreateWithNameInContext("gallivm",
 321                                                        gallivm->context);
 322    if (!gallivm->module)
 323       goto fail;
 324
 325    gallivm->builder = LLVMCreateBuilderInContext(gallivm->context);
 326    if (!gallivm->builder)
 327       goto fail;
 328
 329    /* FIXME: MC-JIT only allows compiling one module at a time, and it must be
 330     * complete when MC-JIT is created. So defer the MC-JIT engine creation for
 331     * now.
 332     */
 333 #if !USE_MCJIT
 334    if (!init_gallivm_engine(gallivm)) {
 335       goto fail;
 336    }
 337 #else
 338    /*
 339     * MC-JIT engine compiles the module immediately on creation, so we can't
 340     * obtain the target data from it.  Instead we create a target data layout
 341     * from a string.
 342     *
 343     * The produced layout strings are not precisely the same, but should make
 344     * no difference for the kind of optimization passes we run.
 345     *
 346     * For reference this is the layout string on x64:
 347     *
 348     *   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
 349     *
 350     * See also:
 351     * - http://llvm.org/docs/LangRef.html#datalayout
 352     */
 353
 354    {
 355       const unsigned pointer_size = 8 * sizeof(void *);
 356       char layout[512];
 357       util_snprintf(layout, sizeof layout, "%c-p:%u:%u:%u-i64:64:64-a0:0:%u-s0:%u:%u",
 358 #ifdef PIPE_ARCH_LITTLE_ENDIAN
 359                     'e', // little endian
 360 #else
 361                     'E', // big endian
 362 #endif
 363                     pointer_size, pointer_size, pointer_size, // pointer size, abi alignment, preferred alignment
 364                     pointer_size, // aggregate preferred alignment
 365                     pointer_size, pointer_size); // stack objects abi alignment, preferred alignment
 366
 367       gallivm->target = LLVMCreateTargetData(layout);
 368       if (!gallivm->target) {
 369          return FALSE;
 370       }
 371    }
 372 #endif
 373
 374    if (!create_pass_manager(gallivm))
 375       goto fail;
 376
 377    return TRUE;
 378
 379 fail:
 380    gallivm_free_ir(gallivm);
 381    gallivm_free_code(gallivm);
 382    return FALSE;
 383 }
 384
 385
 386 void
 387 lp_build_init(void)
 388 {
 389    if (gallivm_initialized)
 390       return;
 391
 392 #ifdef DEBUG
 393    gallivm_debug = debug_get_option_gallivm_debug();
 394 #endif
 395
 396    lp_set_target_options();
 397
 398 #if USE_MCJIT
 399    LLVMLinkInMCJIT();
 400 #else
 401    LLVMLinkInJIT();
 402 #endif
 403
 404    util_cpu_detect();
 405
 406    /* AMD Bulldozer AVX's throughput is the same as SSE2; and because using
 407     * 8-wide vector needs more floating ops than 4-wide (due to padding), it is
 408     * actually more efficient to use 4-wide vectors on this processor.
 409     *
 410     * See also:
 411     * - http://www.anandtech.com/show/4955/the-bulldozer-review-amd-fx8150-tested/2
 412     */
 413    if (HAVE_AVX &&
 414        util_cpu_caps.has_avx &&
 415        util_cpu_caps.has_intel) {
 416       lp_native_vector_width = 256;
 417    } else {
 418       /* Leave it at 128, even when no SIMD extensions are available.
 419        * Really needs to be a multiple of 128 so can fit 4 floats.
 420        */
 421       lp_native_vector_width = 128;
 422    }
 423
 424    lp_native_vector_width = debug_get_num_option("LP_NATIVE_VECTOR_WIDTH",
 425                                                  lp_native_vector_width);
 426
 427    if (lp_native_vector_width <= 128) {
 428       /* Hide AVX support, as often LLVM AVX intrinsics are only guarded by
 429        * "util_cpu_caps.has_avx" predicate, and lack the
 430        * "lp_native_vector_width > 128" predicate. And also to ensure a more
 431        * consistent behavior, allowing one to test SSE2 on AVX machines.
 432        * XXX: should not play games with util_cpu_caps directly as it might
 433        * get used for other things outside llvm too.
 434        */
 435       util_cpu_caps.has_avx = 0;
 436       util_cpu_caps.has_avx2 = 0;
 437    }
 438
 439    if (!HAVE_AVX) {
 440       /*
 441        * note these instructions are VEX-only, so can only emit if we use
 442        * avx (don't want to base it on has_avx & has_f16c later as that would
 443        * omit it unnecessarily on amd cpus, see above).
 444        */
 445       util_cpu_caps.has_f16c = 0;
 446       util_cpu_caps.has_xop = 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 #if 0
 474    /* For simulating less capable machines */
 475    util_cpu_caps.has_sse3 = 0;
 476    util_cpu_caps.has_ssse3 = 0;
 477    util_cpu_caps.has_sse4_1 = 0;
 478    util_cpu_caps.has_avx = 0;
 479    util_cpu_caps.has_f16c = 0;
 480 #endif
 481 }
 482
 483
 484
 485 /**
 486  * Create a new gallivm_state object.
 487  * Note that we return a singleton.
 488  */
 489 struct gallivm_state *
 490 gallivm_create(void)
 491 {
 492    struct gallivm_state *gallivm;
 493
 494    gallivm = CALLOC_STRUCT(gallivm_state);
 495    if (gallivm) {
 496       if (!init_gallivm_state(gallivm)) {
 497          FREE(gallivm);
 498          gallivm = NULL;
 499       }
 500    }
 501
 502    return gallivm;
 503 }
 504
 505
 506 /**
 507  * Destroy a gallivm_state object.
 508  */
 509 void
 510 gallivm_destroy(struct gallivm_state *gallivm)
 511 {
 512    gallivm_free_ir(gallivm);
 513    gallivm_free_code(gallivm);
 514    FREE(gallivm);
 515 }
 516
 517
 518 /**
 519  * Validate a function.
 520  */
 521 void
 522 gallivm_verify_function(struct gallivm_state *gallivm,
 523                         LLVMValueRef func)
 524 {
 525    /* Verify the LLVM IR.  If invalid, dump and abort */
 526 #ifdef DEBUG
 527    if (LLVMVerifyFunction(func, LLVMPrintMessageAction)) {
 528       lp_debug_dump_value(func);
 529       assert(0);
 530       return;
 531    }
 532 #endif
 533
 534    if (gallivm_debug & GALLIVM_DEBUG_IR) {
 535       /* Print the LLVM IR to stderr */
 536       lp_debug_dump_value(func);
 537       debug_printf("\n");
 538    }
 539 }
 540
 541
 542 void
 543 gallivm_compile_module(struct gallivm_state *gallivm)
 544 {
 545    LLVMValueRef func;
 546
 547    assert(!gallivm->compiled);
 548
 549    if (gallivm->builder) {
 550       LLVMDisposeBuilder(gallivm->builder);
 551       gallivm->builder = NULL;
 552    }
 553
 554    /* Run optimization passes */
 555    LLVMInitializeFunctionPassManager(gallivm->passmgr);
 556    func = LLVMGetFirstFunction(gallivm->module);
 557    while (func) {
 558       if (0) {
 559          debug_printf("optimizing %s...\n", LLVMGetValueName(func));
 560       }
 561       LLVMRunFunctionPassManager(gallivm->passmgr, func);
 562       func = LLVMGetNextFunction(func);
 563    }
 564    LLVMFinalizeFunctionPassManager(gallivm->passmgr);
 565
 566    /* Dump byte code to a file */
 567    if (0) {
 568       LLVMWriteBitcodeToFile(gallivm->module, "llvmpipe.bc");
 569       debug_printf("llvmpipe.bc written\n");
 570       debug_printf("Invoke as \"llc -o - llvmpipe.bc\"\n");
 571    }
 572
 573 #if USE_MCJIT
 574    assert(!gallivm->engine);
 575    if (!init_gallivm_engine(gallivm)) {
 576       assert(0);
 577    }
 578 #endif
 579    assert(gallivm->engine);
 580
 581    ++gallivm->compiled;
 582 }
 583
 584
 585
 586 func_pointer
 587 gallivm_jit_function(struct gallivm_state *gallivm,
 588                      LLVMValueRef func)
 589 {
 590    void *code;
 591    func_pointer jit_func;
 592
 593    assert(gallivm->compiled);
 594    assert(gallivm->engine);
 595
 596    code = LLVMGetPointerToGlobal(gallivm->engine, func);
 597    assert(code);
 598    jit_func = pointer_to_func(code);
 599
 600    if (gallivm_debug & GALLIVM_DEBUG_ASM) {
 601       lp_disassemble(func, code);
 602    }
 603
 604 #if defined(PROFILE)
 605    lp_profile(func, code);
 606 #endif
 607
 608    return jit_func;
 609 }