#include "util/u_debug.h"
#include "util/u_memory.h"
#include "util/simple_list.h"
-#include "os/os_time.h"
+#include "util/os_time.h"
#include "lp_bld.h"
#include "lp_bld_debug.h"
#include "lp_bld_misc.h"
#include "lp_bld_init.h"
+#include <llvm/Config/llvm-config.h>
#include <llvm-c/Analysis.h>
#include <llvm-c/Transforms/Scalar.h>
+#if LLVM_VERSION_MAJOR >= 7
+#include <llvm-c/Transforms/Utils.h>
+#endif
#include <llvm-c/BitWriter.h>
-
-
-/* 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
-#else
-static bool USE_MCJIT = 0;
+#if GALLIVM_HAVE_CORO
+#if LLVM_VERSION_MAJOR <= 8 && (defined(PIPE_ARCH_AARCH64) || defined (PIPE_ARCH_ARM) || defined(PIPE_ARCH_S390))
+#include <llvm-c/Transforms/IPO.h>
+#endif
+#include <llvm-c/Transforms/Coroutines.h>
#endif
+unsigned gallivm_perf = 0;
+
+static const struct debug_named_value lp_bld_perf_flags[] = {
+ { "no_brilinear", GALLIVM_PERF_NO_BRILINEAR, "disable brilinear optimization" },
+ { "no_rho_approx", GALLIVM_PERF_NO_RHO_APPROX, "disable rho_approx optimization" },
+ { "no_quad_lod", GALLIVM_PERF_NO_QUAD_LOD, "disable quad_lod optimization" },
+ { "no_aos_sampling", GALLIVM_PERF_NO_AOS_SAMPLING, "disable aos sampling optimization" },
+ { "nopt", GALLIVM_PERF_NO_OPT, "disable optimization passes to speed up shader compilation" },
+ { "no_filter_hacks", GALLIVM_PERF_NO_BRILINEAR | GALLIVM_PERF_NO_RHO_APPROX |
+ GALLIVM_PERF_NO_QUAD_LOD, "disable filter optimization hacks" },
+ DEBUG_NAMED_VALUE_END
+};
#ifdef DEBUG
unsigned gallivm_debug = 0;
{ "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
};
gallivm->passmgr = LLVMCreateFunctionPassManagerForModule(gallivm->module);
if (!gallivm->passmgr)
return FALSE;
+
+#if GALLIVM_HAVE_CORO
+ gallivm->cgpassmgr = LLVMCreatePassManager();
+#endif
/*
* 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
-
- /* 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.
- *
- * 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()
- */
-
-#if HAVE_LLVM < 0x0308
{
char *td_str;
// New ones from the Module.
LLVMSetDataLayout(gallivm->module, td_str);
free(td_str);
}
-#else
- LLVMSetDataLayout(gallivm->module, "");
+
+#if GALLIVM_HAVE_CORO
+#if LLVM_VERSION_MAJOR <= 8 && (defined(PIPE_ARCH_AARCH64) || defined (PIPE_ARCH_ARM) || defined(PIPE_ARCH_S390))
+ LLVMAddArgumentPromotionPass(gallivm->cgpassmgr);
+ LLVMAddFunctionAttrsPass(gallivm->cgpassmgr);
+#endif
+ LLVMAddCoroEarlyPass(gallivm->cgpassmgr);
+ LLVMAddCoroSplitPass(gallivm->cgpassmgr);
+ LLVMAddCoroElidePass(gallivm->cgpassmgr);
#endif
- 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.
+ if ((gallivm_perf & GALLIVM_PERF_NO_OPT) == 0) {
+ /*
+ * TODO: Evaluate passes some more - keeping in mind
+ * both quality of generated code and compile times.
+ */
+ /*
+ * NOTE: if you change this, don't forget to change the output
+ * with GALLIVM_DEBUG_DUMP_BC in gallivm_compile_module.
*/
LLVMAddScalarReplAggregatesPass(gallivm->passmgr);
- LLVMAddLICMPass(gallivm->passmgr);
+ LLVMAddEarlyCSEPass(gallivm->passmgr);
LLVMAddCFGSimplificationPass(gallivm->passmgr);
+ /*
+ * FIXME: LICM is potentially quite useful. However, for some
+ * rather crazy shaders the compile time can reach _hours_ per shader,
+ * due to licm implying lcssa (since llvm 3.5), which can take forever.
+ * Even for sane shaders, the cost of licm is rather high (and not just
+ * due to lcssa, licm itself too), though mostly only in cases when it
+ * can actually move things, so having to disable it is a pity.
+ * LLVMAddLICMPass(gallivm->passmgr);
+ */
LLVMAddReassociatePass(gallivm->passmgr);
LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
+#if LLVM_VERSION_MAJOR <= 11
LLVMAddConstantPropagationPass(gallivm->passmgr);
+#endif
LLVMAddInstructionCombiningPass(gallivm->passmgr);
LLVMAddGVNPass(gallivm->passmgr);
}
*/
LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
}
+#if GALLIVM_HAVE_CORO
+ LLVMAddCoroCleanupPass(gallivm->passmgr);
+#endif
return TRUE;
}
LLVMDisposePassManager(gallivm->passmgr);
}
+#if GALLIVM_HAVE_CORO
+ if (gallivm->cgpassmgr) {
+ LLVMDisposePassManager(gallivm->cgpassmgr);
+ }
+#endif
+
if (gallivm->engine) {
/* This will already destroy any associated module */
LLVMDisposeExecutionEngine(gallivm->engine);
LLVMDisposeModule(gallivm->module);
}
+ if (gallivm->cache) {
+ lp_free_objcache(gallivm->cache->jit_obj_cache);
+ free(gallivm->cache->data);
+ }
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->target) {
+ LLVMDisposeTargetData(gallivm->target);
}
if (gallivm->builder)
gallivm->target = NULL;
gallivm->module = NULL;
gallivm->module_name = NULL;
+ gallivm->cgpassmgr = NULL;
gallivm->passmgr = NULL;
gallivm->context = NULL;
gallivm->builder = NULL;
+ gallivm->cache = NULL;
}
char *error = NULL;
int ret;
- if (gallivm_debug & GALLIVM_DEBUG_NO_OPT) {
+ if (gallivm_perf & GALLIVM_PERF_NO_OPT) {
optlevel = None;
}
else {
ret = lp_build_create_jit_compiler_for_module(&gallivm->engine,
&gallivm->code,
+ gallivm->cache,
gallivm->module,
gallivm->memorymgr,
(unsigned) optlevel,
- USE_MCJIT,
&error);
if (ret) {
_debug_printf("%s\n", error);
}
}
- if (!USE_MCJIT) {
- gallivm->target = LLVMGetExecutionEngineTargetData(gallivm->engine);
- if (!gallivm->target)
- goto fail;
- } else {
- if (0) {
- /*
- * Dump the data layout strings.
- */
+ if (0) {
+ /*
+ * Dump the data layout strings.
+ */
- LLVMTargetDataRef target = LLVMGetExecutionEngineTargetData(gallivm->engine);
- char *data_layout;
- char *engine_data_layout;
+ LLVMTargetDataRef target = LLVMGetExecutionEngineTargetData(gallivm->engine);
+ char *data_layout;
+ char *engine_data_layout;
- data_layout = LLVMCopyStringRepOfTargetData(gallivm->target);
- engine_data_layout = LLVMCopyStringRepOfTargetData(target);
+ 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);
- }
+ 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);
- }
+ free(data_layout);
+ free(engine_data_layout);
}
return TRUE;
*/
static boolean
init_gallivm_state(struct gallivm_state *gallivm, const char *name,
- LLVMContextRef context)
+ LLVMContextRef context, struct lp_cached_code *cache)
{
assert(!gallivm->context);
assert(!gallivm->module);
return FALSE;
gallivm->context = context;
-
+ gallivm->cache = cache;
if (!gallivm->context)
goto fail;
* 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
+ /*
+ * 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];
+ snprintf(layout, sizeof layout, "%c-p:%u:%u:%u-i64:64:64-a0:0:%u-s0:%u:%u",
+#if UTIL_ARCH_LITTLE_ENDIAN
+ 'e', // little endian
#else
- 'E', // big endian
+ '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
+ 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;
- }
+ gallivm->target = LLVMCreateTargetData(layout);
+ if (!gallivm->target) {
+ return FALSE;
}
}
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.
+ */
LLVMLinkInMCJIT();
-#if !defined(USE_MCJIT)
- USE_MCJIT = debug_get_bool_option("GALLIVM_MCJIT", 0);
- LLVMLinkInJIT();
-#endif
#ifdef DEBUG
gallivm_debug = debug_get_option_gallivm_debug();
#endif
+ gallivm_perf = debug_get_flags_option("GALLIVM_PERF", lp_bld_perf_flags, 0 );
+
lp_set_target_options();
util_cpu_detect();
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) {
+ if (util_cpu_caps.has_avx2 || util_cpu_caps.has_avx) {
lp_native_vector_width = 256;
} else {
/* Leave it at 128, even when no SIMD extensions are available.
*/
lp_native_vector_width = 128;
}
-
+
lp_native_vector_width = debug_get_num_option("LP_NATIVE_VECTOR_WIDTH",
lp_native_vector_width);
+#if LLVM_VERSION_MAJOR < 4
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
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
#ifdef PIPE_ARCH_PPC_64
/* Set the NJ bit in VSCR to 0 so denormalized values are handled as
* Create a new gallivm_state object.
*/
struct gallivm_state *
-gallivm_create(const char *name, LLVMContextRef context)
+gallivm_create(const char *name, LLVMContextRef context,
+ struct lp_cached_code *cache)
{
struct gallivm_state *gallivm;
gallivm = CALLOC_STRUCT(gallivm_state);
if (gallivm) {
- if (!init_gallivm_state(gallivm, name, context)) {
+ if (!init_gallivm_state(gallivm, name, context, cache)) {
FREE(gallivm);
gallivm = NULL;
}
}
+ assert(gallivm != NULL);
return gallivm;
}
gallivm->builder = NULL;
}
+ if (gallivm->cache && gallivm->cache->data_size) {
+ goto skip_cached;
+ }
+
+ /* Dump bitcode to a file */
+ if (gallivm_debug & GALLIVM_DEBUG_DUMP_BC) {
+ char filename[256];
+ assert(gallivm->module_name);
+ snprintf(filename, sizeof(filename), "ir_%s.bc", gallivm->module_name);
+ LLVMWriteBitcodeToFile(gallivm->module, filename);
+ debug_printf("%s written\n", filename);
+ debug_printf("Invoke as \"opt %s %s | llc -O%d %s%s\"\n",
+ gallivm_debug & GALLIVM_PERF_NO_OPT ? "-mem2reg" :
+ "-sroa -early-cse -simplifycfg -reassociate "
+ "-mem2reg -constprop -instcombine -gvn",
+ filename, gallivm_debug & GALLIVM_PERF_NO_OPT ? 0 : 2,
+ "[-mcpu=<-mcpu option>] ",
+ "[-mattr=<-mattr option(s)>]");
+ }
+
if (gallivm_debug & GALLIVM_DEBUG_PERF)
time_begin = os_time_get();
+#if GALLIVM_HAVE_CORO
+ LLVMRunPassManager(gallivm->cgpassmgr, gallivm->module);
+#endif
/* Run optimization passes */
LLVMInitializeFunctionPassManager(gallivm->passmgr);
func = LLVMGetFirstFunction(gallivm->module);
/* 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))
+#if 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
if (gallivm_debug & GALLIVM_DEBUG_PERF) {
int64_t time_end = os_time_get();
- int time_msec = (int)(time_end - time_begin) / 1000;
+ 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 (0) {
- LLVMWriteBitcodeToFile(gallivm->module, "llvmpipe.bc");
- debug_printf("llvmpipe.bc written\n");
- debug_printf("Invoke as \"llc -o - llvmpipe.bc\"\n");
- }
-
- if (USE_MCJIT) {
- assert(!gallivm->engine);
- if (!init_gallivm_engine(gallivm)) {
- assert(0);
- }
+ /* 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()
+ */
+ skip_cached:
+ LLVMSetDataLayout(gallivm->module, "");
+ assert(!gallivm->engine);
+ if (!init_gallivm_engine(gallivm)) {
+ assert(0);
}
assert(gallivm->engine);
++gallivm->compiled;
+ if (gallivm->debug_printf_hook)
+ LLVMAddGlobalMapping(gallivm->engine, gallivm->debug_printf_hook, debug_printf);
+
if (gallivm_debug & GALLIVM_DEBUG_ASM) {
LLVMValueRef llvm_func = LLVMGetFirstFunction(gallivm->module);
{
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;
}
projects / mesa.git / blobdiff