#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 <llvm-c/Analysis.h>
#include <llvm-c/Transforms/Scalar.h>
+#if HAVE_LLVM >= 0x0700
+#include <llvm-c/Transforms/Utils.h>
+#endif
#include <llvm-c/BitWriter.h>
# 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
-# define USE_MCJIT 0
#endif
-#if USE_MCJIT
-void LLVMLinkInMCJIT();
+#if defined(USE_MCJIT)
+static const bool use_mcjit = USE_MCJIT;
+#else
+static bool use_mcjit = FALSE;
#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
};
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, "");
-#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);
LLVMAddConstantPropagationPass(gallivm->passmgr);
LLVMDisposeModule(gallivm->module);
}
-#if !USE_MCJIT
- /* Don't free the TargetData, it's owned by the exec engine */
-#else
- if (gallivm->target) {
- LLVMDisposeTargetData(gallivm->target);
+ 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);
+ }
}
-#endif
if (gallivm->builder)
LLVMDisposeBuilder(gallivm->builder);
gallivm->engine = NULL;
gallivm->target = NULL;
gallivm->module = NULL;
+ gallivm->module_name = NULL;
gallivm->passmgr = NULL;
gallivm->context = NULL;
gallivm->builder = NULL;
char *error = NULL;
int ret;
- if (gallivm_debug & GALLIVM_DEBUG_NO_OPT) {
+ if (gallivm_perf & GALLIVM_PERF_NO_OPT) {
optlevel = None;
}
else {
gallivm->module,
gallivm->memorymgr,
(unsigned) optlevel,
- USE_MCJIT,
+ 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 (!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;
+ 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);
+ }
}
-#endif
return TRUE;
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)
* 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
- */
+ 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",
+ {
+ 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",
#ifdef PIPE_ARCH_LITTLE_ENDIAN
- 'e', // 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;
+ }
}
}
-#endif
if (!create_pass_manager(gallivm))
goto fail;
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();
+ gallivm_perf = debug_get_flags_option("GALLIVM_PERF", lp_bld_perf_flags, 0 );
-#if USE_MCJIT
- LLVMLinkInMCJIT();
-#else
- LLVMLinkInJIT();
-#endif
+ 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
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
gallivm->builder = NULL;
}
+ /* 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,
+ (HAVE_LLVM >= 0x0305) ? "[-mcpu=<-mcpu option>] " : "",
+ "[-mattr=<-mattr option(s)>]");
+ }
+
if (gallivm_debug & GALLIVM_DEBUG_PERF)
time_begin = os_time_get();
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",
- lp_get_module_id(gallivm->module), 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");
+ gallivm->module_name, time_msec);
}
-#if USE_MCJIT
- assert(!gallivm->engine);
- if (!init_gallivm_engine(gallivm)) {
- assert(0);
+ 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);
+ }
}
-#endif
assert(gallivm->engine);
++gallivm->compiled;
{
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;
}