/************************************************************************** * * Copyright 2009 VMware, Inc. * All Rights Reserved. * * Permission is hereby granted, free of charge, to any person obtaining a * copy of this software and associated documentation files (the * "Software"), to deal in the Software without restriction, including * without limitation the rights to use, copy, modify, merge, publish, * distribute, sub license, and/or sell copies of the Software, and to * permit persons to whom the Software is furnished to do so, subject to * the following conditions: * * The above copyright notice and this permission notice (including the * next paragraph) shall be included in all copies or substantial portions * of the Software. * * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. * **************************************************************************/ #include "pipe/p_config.h" #include "pipe/p_compiler.h" #include "util/u_cpu_detect.h" #include "util/u_debug.h" #include "util/u_memory.h" #include "util/simple_list.h" #include "os/os_time.h" #include "lp_bld.h" #include "lp_bld_debug.h" #include "lp_bld_misc.h" #include "lp_bld_init.h" #include #include #include /* Only MCJIT is available as of LLVM SVN r216982 */ #if HAVE_LLVM >= 0x0306 # define USE_MCJIT 1 #elif defined(PIPE_ARCH_PPC_64) || defined(PIPE_ARCH_S390) || defined(PIPE_ARCH_ARM) || defined(PIPE_ARCH_AARCH64) # define USE_MCJIT 1 #endif #if defined(USE_MCJIT) static const bool use_mcjit = USE_MCJIT; #else static bool use_mcjit = FALSE; #endif #ifdef DEBUG unsigned gallivm_debug = 0; static const struct debug_named_value lp_bld_debug_flags[] = { { "tgsi", GALLIVM_DEBUG_TGSI, NULL }, { "ir", GALLIVM_DEBUG_IR, NULL }, { "asm", GALLIVM_DEBUG_ASM, NULL }, { "nopt", GALLIVM_DEBUG_NO_OPT, NULL }, { "perf", GALLIVM_DEBUG_PERF, NULL }, { "no_brilinear", GALLIVM_DEBUG_NO_BRILINEAR, NULL }, { "no_rho_approx", GALLIVM_DEBUG_NO_RHO_APPROX, NULL }, { "no_quad_lod", GALLIVM_DEBUG_NO_QUAD_LOD, NULL }, { "gc", GALLIVM_DEBUG_GC, NULL }, { "dumpbc", GALLIVM_DEBUG_DUMP_BC, NULL }, DEBUG_NAMED_VALUE_END }; DEBUG_GET_ONCE_FLAGS_OPTION(gallivm_debug, "GALLIVM_DEBUG", lp_bld_debug_flags, 0) #endif static boolean gallivm_initialized = FALSE; unsigned lp_native_vector_width; /* * Optimization values are: * - 0: None (-O0) * - 1: Less (-O1) * - 2: Default (-O2, -Os) * - 3: Aggressive (-O3) * * See also CodeGenOpt::Level in llvm/Target/TargetMachine.h */ enum LLVM_CodeGenOpt_Level { None, // -O0 Less, // -O1 Default, // -O2, -Os Aggressive // -O3 }; /** * Create the LLVM (optimization) pass manager and install * relevant optimization passes. * \return TRUE for success, FALSE for failure */ static boolean create_pass_manager(struct gallivm_state *gallivm) { assert(!gallivm->passmgr); assert(gallivm->target); gallivm->passmgr = LLVMCreateFunctionPassManagerForModule(gallivm->module); if (!gallivm->passmgr) return FALSE; /* * TODO: some per module pass manager with IPO passes might be helpful - * the generated texture functions may benefit from inlining if they are * simple, or constant propagation into them, etc. */ #if HAVE_LLVM < 0x0309 // Old versions of LLVM get the DataLayout from the pass manager. LLVMAddTargetData(gallivm->target, gallivm->passmgr); #endif { char *td_str; // New ones from the Module. td_str = LLVMCopyStringRepOfTargetData(gallivm->target); LLVMSetDataLayout(gallivm->module, td_str); free(td_str); } if ((gallivm_debug & GALLIVM_DEBUG_NO_OPT) == 0) { /* These are the passes currently listed in llvm-c/Transforms/Scalar.h, * but there are more on SVN. * TODO: Add more passes. */ LLVMAddScalarReplAggregatesPass(gallivm->passmgr); LLVMAddLICMPass(gallivm->passmgr); LLVMAddCFGSimplificationPass(gallivm->passmgr); LLVMAddReassociatePass(gallivm->passmgr); LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr); LLVMAddConstantPropagationPass(gallivm->passmgr); LLVMAddInstructionCombiningPass(gallivm->passmgr); LLVMAddGVNPass(gallivm->passmgr); } else { /* We need at least this pass to prevent the backends to fail in * unexpected ways. */ LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr); } return TRUE; } /** * Free gallivm object's LLVM allocations, but not any generated code * nor the gallivm object itself. */ void gallivm_free_ir(struct gallivm_state *gallivm) { if (gallivm->passmgr) { LLVMDisposePassManager(gallivm->passmgr); } if (gallivm->engine) { /* This will already destroy any associated module */ LLVMDisposeExecutionEngine(gallivm->engine); } else if (gallivm->module) { LLVMDisposeModule(gallivm->module); } FREE(gallivm->module_name); if (!use_mcjit) { /* Don't free the TargetData, it's owned by the exec engine */ } else { if (gallivm->target) { LLVMDisposeTargetData(gallivm->target); } } if (gallivm->builder) LLVMDisposeBuilder(gallivm->builder); /* The LLVMContext should be owned by the parent of gallivm. */ gallivm->engine = NULL; gallivm->target = NULL; gallivm->module = NULL; gallivm->module_name = NULL; gallivm->passmgr = NULL; gallivm->context = NULL; gallivm->builder = NULL; } /** * Free LLVM-generated code. Should be done AFTER gallivm_free_ir(). */ static void gallivm_free_code(struct gallivm_state *gallivm) { assert(!gallivm->module); assert(!gallivm->engine); lp_free_generated_code(gallivm->code); gallivm->code = NULL; lp_free_memory_manager(gallivm->memorymgr); gallivm->memorymgr = NULL; } static boolean init_gallivm_engine(struct gallivm_state *gallivm) { if (1) { enum LLVM_CodeGenOpt_Level optlevel; char *error = NULL; int ret; if (gallivm_debug & GALLIVM_DEBUG_NO_OPT) { optlevel = None; } else { optlevel = Default; } ret = lp_build_create_jit_compiler_for_module(&gallivm->engine, &gallivm->code, gallivm->module, gallivm->memorymgr, (unsigned) optlevel, use_mcjit, &error); if (ret) { _debug_printf("%s\n", error); LLVMDisposeMessage(error); goto fail; } } if (!use_mcjit) { gallivm->target = LLVMGetExecutionEngineTargetData(gallivm->engine); if (!gallivm->target) goto fail; } else { if (0) { /* * Dump the data layout strings. */ LLVMTargetDataRef target = LLVMGetExecutionEngineTargetData(gallivm->engine); char *data_layout; char *engine_data_layout; data_layout = LLVMCopyStringRepOfTargetData(gallivm->target); engine_data_layout = LLVMCopyStringRepOfTargetData(target); if (1) { debug_printf("module target data = %s\n", data_layout); debug_printf("engine target data = %s\n", engine_data_layout); } free(data_layout); free(engine_data_layout); } } return TRUE; fail: return FALSE; } /** * Allocate gallivm LLVM objects. * \return TRUE for success, FALSE for failure */ static boolean init_gallivm_state(struct gallivm_state *gallivm, const char *name, LLVMContextRef context) { assert(!gallivm->context); assert(!gallivm->module); if (!lp_build_init()) return FALSE; gallivm->context = context; if (!gallivm->context) goto fail; gallivm->module_name = NULL; if (name) { size_t size = strlen(name) + 1; gallivm->module_name = MALLOC(size); if (gallivm->module_name) { memcpy(gallivm->module_name, name, size); } } gallivm->module = LLVMModuleCreateWithNameInContext(name, gallivm->context); if (!gallivm->module) goto fail; gallivm->builder = LLVMCreateBuilderInContext(gallivm->context); if (!gallivm->builder) goto fail; gallivm->memorymgr = lp_get_default_memory_manager(); if (!gallivm->memorymgr) goto fail; /* FIXME: MC-JIT only allows compiling one module at a time, and it must be * complete when MC-JIT is created. So defer the MC-JIT engine creation for * now. */ if (!use_mcjit) { if (!init_gallivm_engine(gallivm)) { goto fail; } } else { /* * MC-JIT engine compiles the module immediately on creation, so we can't * obtain the target data from it. Instead we create a target data layout * from a string. * * The produced layout strings are not precisely the same, but should make * no difference for the kind of optimization passes we run. * * For reference this is the layout string on x64: * * 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 * * See also: * - http://llvm.org/docs/LangRef.html#datalayout */ { const unsigned pointer_size = 8 * sizeof(void *); char layout[512]; util_snprintf(layout, sizeof layout, "%c-p:%u:%u:%u-i64:64:64-a0:0:%u-s0:%u:%u", #ifdef PIPE_ARCH_LITTLE_ENDIAN 'e', // little endian #else 'E', // big endian #endif pointer_size, pointer_size, pointer_size, // pointer size, abi alignment, preferred alignment pointer_size, // aggregate preferred alignment pointer_size, pointer_size); // stack objects abi alignment, preferred alignment gallivm->target = LLVMCreateTargetData(layout); if (!gallivm->target) { return FALSE; } } } if (!create_pass_manager(gallivm)) goto fail; return TRUE; fail: gallivm_free_ir(gallivm); gallivm_free_code(gallivm); return FALSE; } boolean lp_build_init(void) { if (gallivm_initialized) return TRUE; /* LLVMLinkIn* are no-ops at runtime. They just ensure the respective * component is linked at buildtime, which is sufficient for its static * constructors to be called at load time. */ #if defined(USE_MCJIT) # if USE_MCJIT LLVMLinkInMCJIT(); # else LLVMLinkInJIT(); # endif #else use_mcjit = debug_get_bool_option("GALLIVM_MCJIT", FALSE); LLVMLinkInJIT(); LLVMLinkInMCJIT(); #endif #ifdef DEBUG gallivm_debug = debug_get_option_gallivm_debug(); #endif lp_set_target_options(); util_cpu_detect(); /* For simulating less capable machines */ #ifdef DEBUG if (debug_get_bool_option("LP_FORCE_SSE2", FALSE)) { assert(util_cpu_caps.has_sse2); util_cpu_caps.has_sse3 = 0; util_cpu_caps.has_ssse3 = 0; util_cpu_caps.has_sse4_1 = 0; util_cpu_caps.has_sse4_2 = 0; util_cpu_caps.has_avx = 0; util_cpu_caps.has_avx2 = 0; util_cpu_caps.has_f16c = 0; util_cpu_caps.has_fma = 0; } #endif /* AMD Bulldozer AVX's throughput is the same as SSE2; and because using * 8-wide vector needs more floating ops than 4-wide (due to padding), it is * actually more efficient to use 4-wide vectors on this processor. * * See also: * - http://www.anandtech.com/show/4955/the-bulldozer-review-amd-fx8150-tested/2 */ if (util_cpu_caps.has_avx && util_cpu_caps.has_intel) { lp_native_vector_width = 256; } else { /* Leave it at 128, even when no SIMD extensions are available. * Really needs to be a multiple of 128 so can fit 4 floats. */ lp_native_vector_width = 128; } lp_native_vector_width = debug_get_num_option("LP_NATIVE_VECTOR_WIDTH", lp_native_vector_width); if (lp_native_vector_width <= 128) { /* Hide AVX support, as often LLVM AVX intrinsics are only guarded by * "util_cpu_caps.has_avx" predicate, and lack the * "lp_native_vector_width > 128" predicate. And also to ensure a more * consistent behavior, allowing one to test SSE2 on AVX machines. * XXX: should not play games with util_cpu_caps directly as it might * get used for other things outside llvm too. */ util_cpu_caps.has_avx = 0; util_cpu_caps.has_avx2 = 0; util_cpu_caps.has_f16c = 0; util_cpu_caps.has_fma = 0; } if (HAVE_LLVM < 0x0304 || !use_mcjit) { /* AVX2 support has only been tested with LLVM 3.4, and it requires * MCJIT. */ util_cpu_caps.has_avx2 = 0; } #ifdef PIPE_ARCH_PPC_64 /* Set the NJ bit in VSCR to 0 so denormalized values are handled as * specified by IEEE standard (PowerISA 2.06 - Section 6.3). This guarantees * that some rounding and half-float to float handling does not round * incorrectly to 0. * XXX: should eventually follow same logic on all platforms. * Right now denorms get explicitly disabled (but elsewhere) for x86, * whereas ppc64 explicitly enables them... */ if (util_cpu_caps.has_altivec) { unsigned short mask[] = { 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFE, 0xFFFF }; __asm ( "mfvscr %%v1\n" "vand %0,%%v1,%0\n" "mtvscr %0" : : "r" (*mask) ); } #endif gallivm_initialized = TRUE; return TRUE; } /** * Create a new gallivm_state object. */ struct gallivm_state * gallivm_create(const char *name, LLVMContextRef context) { struct gallivm_state *gallivm; gallivm = CALLOC_STRUCT(gallivm_state); if (gallivm) { if (!init_gallivm_state(gallivm, name, context)) { FREE(gallivm); gallivm = NULL; } } return gallivm; } /** * Destroy a gallivm_state object. */ void gallivm_destroy(struct gallivm_state *gallivm) { gallivm_free_ir(gallivm); gallivm_free_code(gallivm); FREE(gallivm); } /** * Validate a function. * Verification is only done with debug builds. */ void gallivm_verify_function(struct gallivm_state *gallivm, LLVMValueRef func) { /* Verify the LLVM IR. If invalid, dump and abort */ #ifdef DEBUG if (LLVMVerifyFunction(func, LLVMPrintMessageAction)) { lp_debug_dump_value(func); assert(0); return; } #endif if (gallivm_debug & GALLIVM_DEBUG_IR) { /* Print the LLVM IR to stderr */ lp_debug_dump_value(func); debug_printf("\n"); } } /** * Compile a module. * This does IR optimization on all functions in the module. */ void gallivm_compile_module(struct gallivm_state *gallivm) { LLVMValueRef func; int64_t time_begin = 0; assert(!gallivm->compiled); if (gallivm->builder) { LLVMDisposeBuilder(gallivm->builder); gallivm->builder = NULL; } if (gallivm_debug & GALLIVM_DEBUG_PERF) time_begin = os_time_get(); /* Run optimization passes */ LLVMInitializeFunctionPassManager(gallivm->passmgr); func = LLVMGetFirstFunction(gallivm->module); while (func) { if (0) { debug_printf("optimizing func %s...\n", LLVMGetValueName(func)); } /* Disable frame pointer omission on debug/profile builds */ /* XXX: And workaround http://llvm.org/PR21435 */ #if HAVE_LLVM >= 0x0307 && \ (defined(DEBUG) || defined(PROFILE) || \ defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)) LLVMAddTargetDependentFunctionAttr(func, "no-frame-pointer-elim", "true"); LLVMAddTargetDependentFunctionAttr(func, "no-frame-pointer-elim-non-leaf", "true"); #endif LLVMRunFunctionPassManager(gallivm->passmgr, func); func = LLVMGetNextFunction(func); } LLVMFinalizeFunctionPassManager(gallivm->passmgr); if (gallivm_debug & GALLIVM_DEBUG_PERF) { int64_t time_end = os_time_get(); int time_msec = (int)(time_end - time_begin) / 1000; assert(gallivm->module_name); debug_printf("optimizing module %s took %d msec\n", gallivm->module_name, time_msec); } /* Dump byte code to a file */ if (gallivm_debug & GALLIVM_DEBUG_DUMP_BC) { char filename[256]; assert(gallivm->module_name); util_snprintf(filename, sizeof(filename), "ir_%s.bc", gallivm->module_name); LLVMWriteBitcodeToFile(gallivm->module, filename); debug_printf("%s written\n", filename); debug_printf("Invoke as \"llc %s%s -o - %s\"\n", (HAVE_LLVM >= 0x0305) ? "[-mcpu=<-mcpu option>] " : "", "[-mattr=<-mattr option(s)>]", filename); } if (use_mcjit) { /* Setting the module's DataLayout to an empty string will cause the * ExecutionEngine to copy to the DataLayout string from its target * machine to the module. As of LLVM 3.8 the module and the execution * engine are required to have the same DataLayout. * * We must make sure we do this after running the optimization passes, * because those passes need a correct datalayout string. For example, * if those optimization passes see an empty datalayout, they will assume * this is a little endian target and will do optimizations that break big * endian machines. * * TODO: This is just a temporary work-around. The correct solution is * for gallivm_init_state() to create a TargetMachine and pull the * DataLayout from there. Currently, the TargetMachine used by llvmpipe * is being implicitly created by the EngineBuilder in * lp_build_create_jit_compiler_for_module() */ LLVMSetDataLayout(gallivm->module, ""); assert(!gallivm->engine); if (!init_gallivm_engine(gallivm)) { assert(0); } } assert(gallivm->engine); ++gallivm->compiled; if (gallivm_debug & GALLIVM_DEBUG_ASM) { LLVMValueRef llvm_func = LLVMGetFirstFunction(gallivm->module); while (llvm_func) { /* * Need to filter out functions which don't have an implementation, * such as the intrinsics. May not be sufficient in case of IPO? * LLVMGetPointerToGlobal() will abort otherwise. */ if (!LLVMIsDeclaration(llvm_func)) { void *func_code = LLVMGetPointerToGlobal(gallivm->engine, llvm_func); lp_disassemble(llvm_func, func_code); } llvm_func = LLVMGetNextFunction(llvm_func); } } #if defined(PROFILE) { LLVMValueRef llvm_func = LLVMGetFirstFunction(gallivm->module); while (llvm_func) { if (!LLVMIsDeclaration(llvm_func)) { void *func_code = LLVMGetPointerToGlobal(gallivm->engine, llvm_func); lp_profile(llvm_func, func_code); } llvm_func = LLVMGetNextFunction(llvm_func); } } #endif } func_pointer gallivm_jit_function(struct gallivm_state *gallivm, LLVMValueRef func) { void *code; func_pointer jit_func; int64_t time_begin = 0; assert(gallivm->compiled); assert(gallivm->engine); if (gallivm_debug & GALLIVM_DEBUG_PERF) time_begin = os_time_get(); code = LLVMGetPointerToGlobal(gallivm->engine, func); assert(code); jit_func = pointer_to_func(code); if (gallivm_debug & GALLIVM_DEBUG_PERF) { int64_t time_end = os_time_get(); int time_msec = (int)(time_end - time_begin) / 1000; debug_printf(" jitting func %s took %d msec\n", LLVMGetValueName(func), time_msec); } return jit_func; }