Mercurial > hg > truffle
changeset 13683:de839ec35cc7
schedule lambda method compilation and execution on GPU (PTX) when possible; fix a couple of bugs.
line wrap: on
line diff
--- a/graal/com.oracle.graal.asm.ptx/src/com/oracle/graal/asm/ptx/PTXAssembler.java Fri Jan 17 17:26:47 2014 +0100 +++ b/graal/com.oracle.graal.asm.ptx/src/com/oracle/graal/asm/ptx/PTXAssembler.java Fri Jan 17 16:03:13 2014 -0500 @@ -700,4 +700,7 @@ bra(str); } + public void nullCheck(Register r) { + // setp(....); + } }
--- a/graal/com.oracle.graal.compiler.ptx.test/src/com/oracle/graal/compiler/ptx/test/BasicPTXTest.java Fri Jan 17 17:26:47 2014 +0100 +++ b/graal/com.oracle.graal.compiler.ptx.test/src/com/oracle/graal/compiler/ptx/test/BasicPTXTest.java Fri Jan 17 16:03:13 2014 -0500 @@ -43,15 +43,15 @@ test("staticIntKernel", 'a', 42); } + public static int staticIntKernel(char p0, int p1) { + return p1 + p0; + } + @Test public void testVirtualIntKernel() { test("virtualIntKernel", 'a', 42); } - public static int staticIntKernel(char p0, int p1) { - return p1 + p0; - } - public int virtualIntKernel(char p0, int p1) { return p1 + p0; }
--- a/graal/com.oracle.graal.compiler.ptx/src/com/oracle/graal/compiler/ptx/PTXLIRGenerator.java Fri Jan 17 17:26:47 2014 +0100 +++ b/graal/com.oracle.graal.compiler.ptx/src/com/oracle/graal/compiler/ptx/PTXLIRGenerator.java Fri Jan 17 16:03:13 2014 -0500 @@ -827,7 +827,8 @@ @Override public void emitNullCheck(ValueNode v, DeoptimizingNode deopting) { - throw GraalInternalError.unimplemented("PTXLIRGenerator.emitNullCheck()"); + assert v.kind() == Kind.Object; + append(new PTXMove.NullCheckOp(load(operand(v)), state(deopting))); } @Override
--- a/graal/com.oracle.graal.hotspot.ptx/src/com/oracle/graal/hotspot/ptx/PTXHotSpotBackend.java Fri Jan 17 17:26:47 2014 +0100 +++ b/graal/com.oracle.graal.hotspot.ptx/src/com/oracle/graal/hotspot/ptx/PTXHotSpotBackend.java Fri Jan 17 16:03:13 2014 -0500 @@ -94,6 +94,8 @@ public PTXHotSpotBackend(HotSpotGraalRuntime runtime, HotSpotProviders providers) { super(runtime, providers); + CompilerToGPU compilerToGPU = getRuntime().getCompilerToGPU(); + deviceInitialized = OmitDeviceInit || compilerToGPU.deviceInit(); } @Override @@ -110,7 +112,6 @@ public void completeInitialization() { HotSpotHostForeignCallsProvider hostForeignCalls = (HotSpotHostForeignCallsProvider) getRuntime().getHostProviders().getForeignCalls(); CompilerToGPU compilerToGPU = getRuntime().getCompilerToGPU(); - deviceInitialized = OmitDeviceInit || compilerToGPU.deviceInit(); if (deviceInitialized) { long launchKernel = compilerToGPU.getLaunchKernelAddress(); hostForeignCalls.registerForeignCall(LAUNCH_KERNEL, launchKernel, NativeCall, DESTROYS_REGISTERS, NOT_LEAF, NOT_REEXECUTABLE, ANY_LOCATION);
--- a/graal/com.oracle.graal.hotspot.ptx/src/com/oracle/graal/hotspot/ptx/PTXHotSpotCodeCacheProvider.java Fri Jan 17 17:26:47 2014 +0100 +++ b/graal/com.oracle.graal.hotspot.ptx/src/com/oracle/graal/hotspot/ptx/PTXHotSpotCodeCacheProvider.java Fri Jan 17 16:03:13 2014 -0500 @@ -25,12 +25,18 @@ import com.oracle.graal.api.code.*; import com.oracle.graal.hotspot.*; import com.oracle.graal.hotspot.meta.*; +import java.util.Arrays; public class PTXHotSpotCodeCacheProvider extends HotSpotCodeCacheProvider { public PTXHotSpotCodeCacheProvider(HotSpotGraalRuntime runtime, TargetDescription target) { super(runtime, target); + } + @Override + public String disassemble(CompilationResult compResult, InstalledCode installedCode) { + byte[] code = installedCode == null ? Arrays.copyOf(compResult.getTargetCode(), compResult.getTargetCodeSize()) : installedCode.getCode(); + return new String(code); } @Override
--- a/graal/com.oracle.graal.hotspot.ptx/src/com/oracle/graal/hotspot/ptx/PTXHotSpotForeignCallsProvider.java Fri Jan 17 17:26:47 2014 +0100 +++ b/graal/com.oracle.graal.hotspot.ptx/src/com/oracle/graal/hotspot/ptx/PTXHotSpotForeignCallsProvider.java Fri Jan 17 16:03:13 2014 -0500 @@ -33,10 +33,12 @@ throw GraalInternalError.unimplemented(); } + @Override public LocationIdentity[] getKilledLocations(ForeignCallDescriptor descriptor) { throw GraalInternalError.unimplemented(); } + @Override public boolean canDeoptimize(ForeignCallDescriptor descriptor) { throw GraalInternalError.unimplemented(); }
--- a/graal/com.oracle.graal.hotspot/src/com/oracle/graal/hotspot/bridge/VMToCompilerImpl.java Fri Jan 17 17:26:47 2014 +0100 +++ b/graal/com.oracle.graal.hotspot/src/com/oracle/graal/hotspot/bridge/VMToCompilerImpl.java Fri Jan 17 16:03:13 2014 -0500 @@ -556,8 +556,9 @@ int id = allocateCompileTaskId(method, entryBCI); HotSpotBackend backend = runtime.getHostBackend(); CompilationTask task = new CompilationTask(backend, method, entryBCI, id); + boolean isLambdaMethod = (method.getName().contains("lambda$main$") & method.isSynthetic()); - if (blocking) { + if (blocking || isLambdaMethod) { task.runCompilation(true); } else { try {
--- a/graal/com.oracle.graal.lir.ptx/src/com/oracle/graal/lir/ptx/PTXMove.java Fri Jan 17 17:26:47 2014 +0100 +++ b/graal/com.oracle.graal.lir.ptx/src/com/oracle/graal/lir/ptx/PTXMove.java Fri Jan 17 16:03:13 2014 -0500 @@ -32,7 +32,7 @@ import com.oracle.graal.asm.ptx.*; import com.oracle.graal.graph.*; import com.oracle.graal.lir.*; -import com.oracle.graal.lir.StandardOp.MoveOp; +import com.oracle.graal.lir.StandardOp.*; import com.oracle.graal.lir.asm.*; public class PTXMove { @@ -239,4 +239,29 @@ protected static void compareAndSwap(CompilationResultBuilder crb, PTXAssembler masm, AllocatableValue result, PTXAddressValue address, AllocatableValue cmpValue, AllocatableValue newValue) { throw new InternalError("NYI"); } + + public static class NullCheckOp extends PTXLIRInstruction implements NullCheck { + + @Use({REG}) protected AllocatableValue input; + @State protected LIRFrameState state; + + public NullCheckOp(Variable input, LIRFrameState state) { + this.input = input; + this.state = state; + } + + @Override + public void emitCode(CompilationResultBuilder crb, PTXMacroAssembler masm) { + crb.recordImplicitException(masm.codeBuffer.position(), state); + masm.nullCheck(asRegister(input)); + } + + public Value getCheckedValue() { + return input; + } + + public LIRFrameState getState() { + return state; + } + } }
--- a/src/gpu/ptx/vm/gpu_ptx.cpp Fri Jan 17 17:26:47 2014 +0100 +++ b/src/gpu/ptx/vm/gpu_ptx.cpp Fri Jan 17 16:03:13 2014 -0500 @@ -331,7 +331,7 @@ } if (TraceGPUInteraction) { - tty->print_cr("[CUDA] Got function handle for %s", name); + tty->print_cr("[CUDA] Got function handle for %s kernel address %p", name, cu_function); } return cu_function; @@ -375,7 +375,9 @@ return 0L; } // Push device_return_value to kernelParams - gpu::Ptx::CUdeviceptr* returnValuePtr = (gpu::Ptx::CUdeviceptr*) (address) parametersAndReturnValueBuffer + parametersAndReturnValueBufferSize - sizeof(device_return_value); + gpu::Ptx::CUdeviceptr* returnValuePtr = (gpu::Ptx::CUdeviceptr*) + ((address) parametersAndReturnValueBuffer + + parametersAndReturnValueBufferSize - sizeof(device_return_value)); *returnValuePtr = device_return_value; } @@ -417,7 +419,7 @@ } thread->set_vm_result(return_val); } else if (returnTypeSize > 0) { - status = gpu::Ptx::_cuda_cu_memcpy_dtoh(&primitiveReturnValue, device_return_value, T_LONG_BYTE_SIZE); + status = gpu::Ptx::_cuda_cu_memcpy_dtoh(&primitiveReturnValue, device_return_value, returnTypeSize); if (status != GRAAL_CUDA_SUCCESS) { tty->print_cr("[CUDA] *** Error (%d) Failed to copy value from device argument", status); SharedRuntime::throw_and_post_jvmti_exception(thread, vmSymbols::java_lang_Exception(), "[CUDA] Failed to copy value from device argument");
--- a/src/share/vm/oops/method.cpp Fri Jan 17 17:26:47 2014 +0100 +++ b/src/share/vm/oops/method.cpp Fri Jan 17 16:03:13 2014 -0500 @@ -2025,3 +2025,32 @@ guarantee(md == NULL || md->is_methodData(), "should be method data"); } + +#ifdef GRAAL + +// Return true if the name of the method indicates that this is a +// lambda method other than <init>. Lambda method is one with a name +// that starts with lambda$ and is synthetic. + +bool Method::is_lambda() const { + Symbol * klass_name = method_holder()->name(); + Symbol * method_name = name(); + ResourceMark rm; + if (klass_name != NULL) { + if (klass_name != NULL && method_name != NULL) { + const char* lambdaPrefix = "lambda$main$"; + char* methodPrefix = strstr(method_name->as_C_string(), lambdaPrefix); + if (methodPrefix != 0) { + if ((strncmp(lambdaPrefix, methodPrefix, strlen(lambdaPrefix)) == 0) && + is_synthetic()) { + //tty->print_cr("[Check] %s::%s", klass_name->as_C_string(), method_name->as_C_string()); + return true; + } else { + return false; + } + } + } + //} + } +} +#endif
--- a/src/share/vm/oops/method.hpp Fri Jan 17 17:26:47 2014 +0100 +++ b/src/share/vm/oops/method.hpp Fri Jan 17 16:03:13 2014 -0500 @@ -617,6 +617,11 @@ // simultaneously. Use with caution. bool has_compiled_code() const { return code() != NULL; } +#ifdef GRAAL + // Return true if the name of the method indicates that this is a + // lambda method other than <init>. + bool is_lambda() const; +#endif // sizing static int header_size() { return sizeof(Method)/HeapWordSize; } static int size(bool is_native);
--- a/src/share/vm/runtime/compilationPolicy.cpp Fri Jan 17 17:26:47 2014 +0100 +++ b/src/share/vm/runtime/compilationPolicy.cpp Fri Jan 17 16:03:13 2014 -0500 @@ -45,6 +45,9 @@ #include "runtime/vm_operations.hpp" #include "utilities/events.hpp" #include "utilities/globalDefinitions.hpp" +#ifdef GRAAL +#include "runtime/gpu.hpp" +#endif CompilationPolicy* CompilationPolicy::_policy; elapsedTimer CompilationPolicy::_accumulated_time; @@ -101,6 +104,32 @@ if (ReplayCompiles) return false; if (m->has_compiled_code()) return false; // already compiled + +#ifdef GRAAL + // Check if this is a Lambda method that can be compiled to a GPU. + if (m->is_lambda()) { + // If GPU is available and the necessary linkage is available + // rerurn true indicatin that this method must be compiled. + if (gpu::is_available() && gpu::has_gpu_linkage()) { + if (TraceGPUInteraction) { + tty->print("Compiling Lambda method"); + m->print_short_name(); + switch (gpu::get_target_il_type()) { + case gpu::PTX : + tty->print_cr(" to PTX"); + break; + case gpu::HSAIL : + tty->print_cr(" to HSAIL"); + break; + default : + tty->print_cr(" to Unknown GPU!!!"); + } + } + return true; + } + } +#endif + if (!can_be_compiled(m, comp_level)) return false; return !UseInterpreter || // must compile all methods