#include "util/u_cpu_detect.h"
#include "util/u_debug.h"
#include "util/u_memory.h"
-#include "util/u_simple_list.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 <llvm-c/BitWriter.h>
-/**
- * AVX is supported in:
- * - standard JIT from LLVM 3.2 onwards
- * - MC-JIT from LLVM 3.1
- * - MC-JIT supports limited OSes (MacOSX and Linux)
- * - standard JIT in LLVM 3.1, with backports
- */
-#if defined(PIPE_ARCH_PPC_64)
+/* Only MCJIT is available as of LLVM SVN r216982 */
+#if HAVE_LLVM >= 0x0306
# define USE_MCJIT 1
-# define HAVE_AVX 0
-#elif HAVE_LLVM >= 0x0302 || (HAVE_LLVM == 0x0301 && defined(HAVE_JIT_AVX_SUPPORT))
-# define USE_MCJIT 0
-# define HAVE_AVX 1
-#elif HAVE_LLVM == 0x0301 && (defined(PIPE_OS_LINUX) || defined(PIPE_OS_APPLE))
+#elif defined(PIPE_ARCH_PPC_64) || defined(PIPE_ARCH_S390) || defined(PIPE_ARCH_ARM) || defined(PIPE_ARCH_AARCH64)
# define USE_MCJIT 1
-# define HAVE_AVX 1
#else
# define USE_MCJIT 0
-# define HAVE_AVX 0
#endif
-
#if USE_MCJIT
void LLVMLinkInMCJIT();
#endif
-
#ifdef DEBUG
unsigned gallivm_debug = 0;
{ "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 },
DEBUG_NAMED_VALUE_END
};
* See also CodeGenOpt::Level in llvm/Target/TargetMachine.h
*/
enum LLVM_CodeGenOpt_Level {
-#if HAVE_LLVM >= 0x207
None, // -O0
Less, // -O1
Default, // -O2, -Os
Aggressive // -O3
-#else
- Default,
- None,
- Aggressive
-#endif
};
-#if HAVE_LLVM <= 0x0206
-/**
- * LLVM 2.6 permits only one ExecutionEngine to be created. So use the
- * same gallivm state everywhere.
- */
-static struct gallivm_state *GlobalGallivm = NULL;
-#endif
-
-
/**
* Create the LLVM (optimization) pass manager and install
* relevant optimization passes.
static boolean
create_pass_manager(struct gallivm_state *gallivm)
{
+ char *td_str;
assert(!gallivm->passmgr);
assert(gallivm->target);
- gallivm->passmgr = LLVMCreateFunctionPassManager(gallivm->provider);
+ gallivm->passmgr = LLVMCreateFunctionPassManagerForModule(gallivm->module);
if (!gallivm->passmgr)
return FALSE;
+ // Old versions of LLVM get the DataLayout from the pass manager.
LLVMAddTargetData(gallivm->target, gallivm->passmgr);
+ // 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);
-
- if (HAVE_LLVM >= 0x207 && sizeof(void*) == 4) {
- /* For LLVM >= 2.7 and 32-bit build, use this order of passes to
- * avoid generating bad code.
- * Test with piglit glsl-vs-sqrt-zero test.
- */
- LLVMAddConstantPropagationPass(gallivm->passmgr);
- LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
- }
- else {
- LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
- LLVMAddConstantPropagationPass(gallivm->passmgr);
- }
-
- if (util_cpu_caps.has_sse4_1) {
- /* FIXME: There is a bug in this pass, whereby the combination
- * of fptosi and sitofp (necessary for trunc/floor/ceil/round
- * implementation) somehow becomes invalid code.
- */
- LLVMAddInstructionCombiningPass(gallivm->passmgr);
- }
+ LLVMAddReassociatePass(gallivm->passmgr);
+ LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
+ LLVMAddConstantPropagationPass(gallivm->passmgr);
+ LLVMAddInstructionCombiningPass(gallivm->passmgr);
LLVMAddGVNPass(gallivm->passmgr);
}
else {
/**
- * Free gallivm object's LLVM allocations, but not the gallivm object itself.
+ * Free gallivm object's LLVM allocations, but not any generated code
+ * nor the gallivm object itself.
*/
-static void
-free_gallivm_state(struct gallivm_state *gallivm)
+void
+gallivm_free_ir(struct gallivm_state *gallivm)
{
-#if HAVE_LLVM >= 0x207 /* XXX or 0x208? */
- /* This leads to crashes w/ some versions of LLVM */
- LLVMModuleRef mod;
- char *error;
-
- if (gallivm->engine && gallivm->provider)
- LLVMRemoveModuleProvider(gallivm->engine, gallivm->provider,
- &mod, &error);
-#endif
-
if (gallivm->passmgr) {
LLVMDisposePassManager(gallivm->passmgr);
}
-#if 0
- /* XXX this seems to crash with all versions of LLVM */
- if (gallivm->provider)
- LLVMDisposeModuleProvider(gallivm->provider);
-#endif
-
- if (HAVE_LLVM >= 0x207 && gallivm->engine) {
+ if (gallivm->engine) {
/* This will already destroy any associated module */
LLVMDisposeExecutionEngine(gallivm->engine);
- } else {
+ } else if (gallivm->module) {
LLVMDisposeModule(gallivm->module);
}
}
#endif
- /* Never free the LLVM context.
- */
-#if 0
- if (gallivm->context)
- LLVMContextDispose(gallivm->context);
-#endif
-
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->provider = 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) {
- /* We can only create one LLVMExecutionEngine (w/ LLVM 2.6 anyway) */
enum LLVM_CodeGenOpt_Level optlevel;
char *error = NULL;
int ret;
optlevel = Default;
}
-#if USE_MCJIT
- ret = lp_build_create_mcjit_compiler_for_module(&gallivm->engine,
- gallivm->module,
- (unsigned) optlevel,
- &error);
-#else
- ret = LLVMCreateJITCompiler(&gallivm->engine, gallivm->provider,
- (unsigned) optlevel, &error);
-#endif
+ 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 defined(DEBUG) || defined(PROFILE)
- lp_register_oprofile_jit_event_listener(gallivm->engine);
-#endif
}
- LLVMAddModuleProvider(gallivm->engine, gallivm->provider);//new
-
#if !USE_MCJIT
gallivm->target = LLVMGetExecutionEngineTargetData(gallivm->engine);
if (!gallivm->target)
}
-/**
- * Singleton
- *
- * We must never free LLVM contexts, because LLVM has several global caches
- * which pointing/derived from objects owned by the context, causing false
- * memory leaks and false cache hits when these objects are destroyed.
- *
- * TODO: For thread safety on multi-threaded OpenGL we should use one LLVM
- * context per thread, and put them in a pool when threads are destroyed.
- */
-static LLVMContextRef gallivm_context = NULL;
-
-
/**
* Allocate gallivm LLVM objects.
* \return TRUE for success, FALSE for failure
*/
static boolean
-init_gallivm_state(struct gallivm_state *gallivm)
+init_gallivm_state(struct gallivm_state *gallivm, const char *name,
+ LLVMContextRef context)
{
assert(!gallivm->context);
assert(!gallivm->module);
- assert(!gallivm->provider);
- lp_build_init();
+ if (!lp_build_init())
+ return FALSE;
+
+ gallivm->context = context;
- if (!gallivm_context) {
- gallivm_context = LLVMContextCreate();
- }
- gallivm->context = gallivm_context;
if (!gallivm->context)
goto fail;
- gallivm->module = LLVMModuleCreateWithNameInContext("gallivm",
+ gallivm->module = LLVMModuleCreateWithNameInContext(name,
gallivm->context);
if (!gallivm->module)
goto fail;
- gallivm->provider =
- LLVMCreateModuleProviderForExistingModule(gallivm->module);
- if (!gallivm->provider)
- 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.
return TRUE;
fail:
- free_gallivm_state(gallivm);
+ gallivm_free_ir(gallivm);
+ gallivm_free_code(gallivm);
return FALSE;
}
-void
+boolean
lp_build_init(void)
{
if (gallivm_initialized)
- return;
+ return TRUE;
#ifdef DEBUG
gallivm_debug = debug_get_option_gallivm_debug();
* See also:
* - http://www.anandtech.com/show/4955/the-bulldozer-review-amd-fx8150-tested/2
*/
- if (HAVE_AVX &&
- util_cpu_caps.has_avx &&
+ if (util_cpu_caps.has_avx &&
util_cpu_caps.has_intel) {
lp_native_vector_width = 256;
} else {
lp_native_vector_width);
if (lp_native_vector_width <= 128) {
- /* Hide AVX support, as often LLVM AVX instrinsics are only guarded by
+ /* 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;
}
+#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;
#if 0
util_cpu_caps.has_ssse3 = 0;
util_cpu_caps.has_sse4_1 = 0;
util_cpu_caps.has_avx = 0;
+ util_cpu_caps.has_f16c = 0;
#endif
+
+ return TRUE;
}
/**
* Create a new gallivm_state object.
- * Note that we return a singleton.
*/
struct gallivm_state *
-gallivm_create(void)
+gallivm_create(const char *name, LLVMContextRef context)
{
struct gallivm_state *gallivm;
-#if HAVE_LLVM <= 0x206
- if (GlobalGallivm) {
- return GlobalGallivm;
- }
-#endif
-
gallivm = CALLOC_STRUCT(gallivm_state);
if (gallivm) {
- if (!init_gallivm_state(gallivm)) {
+ if (!init_gallivm_state(gallivm, name, context)) {
FREE(gallivm);
gallivm = NULL;
}
}
-#if HAVE_LLVM <= 0x206
- GlobalGallivm = gallivm;
-#endif
-
return gallivm;
}
void
gallivm_destroy(struct gallivm_state *gallivm)
{
-#if HAVE_LLVM <= 0x0206
- /* No-op: don't destroy the singleton */
- (void) gallivm;
-#else
- free_gallivm_state(gallivm);
+ gallivm_free_ir(gallivm);
+ gallivm_free_code(gallivm);
FREE(gallivm);
-#endif
-}
-
-
-/**
- * Validate and optimze a function.
- */
-static void
-gallivm_optimize_function(struct gallivm_state *gallivm,
- LLVMValueRef func)
-{
- if (0) {
- debug_printf("optimizing %s...\n", LLVMGetValueName(func));
- }
-
- assert(gallivm->passmgr);
-
- /* Apply optimizations to LLVM IR */
- LLVMRunFunctionPassManager(gallivm->passmgr, func);
-
- if (0) {
- if (gallivm_debug & GALLIVM_DEBUG_IR) {
- /* Print the LLVM IR to stderr */
- lp_debug_dump_value(func);
- debug_printf("\n");
- }
- }
}
/**
* Validate a function.
+ * Verification is only done with debug builds.
*/
void
gallivm_verify_function(struct gallivm_state *gallivm,
}
#endif
- gallivm_optimize_function(gallivm, func);
-
if (gallivm_debug & GALLIVM_DEBUG_IR) {
/* Print the LLVM IR to stderr */
lp_debug_dump_value(func);
}
+/**
+ * Compile a module.
+ * This does IR optimization on all functions in the module.
+ */
void
gallivm_compile_module(struct gallivm_state *gallivm)
{
-#if HAVE_LLVM > 0x206
+ LLVMValueRef func;
+ int64_t time_begin = 0;
+
assert(!gallivm->compiled);
-#endif
+
+ 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));
+ }
+ 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;
+ debug_printf("optimizing module %s took %d msec\n",
+ lp_get_module_id(gallivm->module), time_msec);
+ }
/* Dump byte code to a file */
if (0) {
}
+
func_pointer
gallivm_jit_function(struct gallivm_state *gallivm,
LLVMValueRef func)
jit_func = pointer_to_func(code);
if (gallivm_debug & GALLIVM_DEBUG_ASM) {
- lp_disassemble(code);
+ lp_disassemble(func, code);
}
- /* Free the function body to save memory */
- lp_func_delete_body(func);
+#if defined(PROFILE)
+ lp_profile(func, code);
+#endif
return jit_func;
}
-
-
-/**
- * Free the function (and its machine code).
- */
-void
-gallivm_free_function(struct gallivm_state *gallivm,
- LLVMValueRef func,
- const void *code)
-{
-#if !USE_MCJIT
- if (code) {
- LLVMFreeMachineCodeForFunction(gallivm->engine, func);
- }
-
- LLVMDeleteFunction(func);
-#endif
-}