truffle: src/gpu/hsail/vm/gpu_hsail.cpp comparison

comparison src/gpu/hsail/vm/gpu_hsail.cpp @ 14969:a6c144380ce7

HSAIL: added UseHSAILDeoptimization VM option for disabling HSAIL deopt support Contributed-by: Eric Caspole <eric.caspole@amd.com>

author	Doug Simon <doug.simon@oracle.com>
date	Fri, 04 Apr 2014 12:22:49 +0200
parents	3e9a960f0da1
children	2cae21d9f122

comparison

equal deleted inserted replaced

-:169caf662ac7
+:a6c144380ce7
 if (lookFor == false) {
 tty->print_cr("-%d", len-1);
 }
 }
-// for experimentation
-static bool useDeoptInfo = true;
 jboolean Hsail::execute_kernel_void_1d_internal(address kernel, int dimX, jobject args_handle, methodHandle& mh, nmethod *nm, jobject oops_save_handle, TRAPS) {
 ResourceMark rm(THREAD);
 objArrayOop argsArray = (objArrayOop) JNIHandles::resolve(args_handle);
 // Reset the kernel arguments
 _okra_clearargs(kernel);
 HSAILDeoptimizationInfo* e;
-if (useDeoptInfo) {
+if (UseHSAILDeoptimization) {
 e = new (ResourceObj::C_HEAP, mtInternal) HSAILDeoptimizationInfo();
 e->set_never_ran_array(NEW_C_HEAP_ARRAY(jboolean, dimX, mtInternal));
 memset(e->never_ran_array(), 0, dimX * sizeof(jboolean));
 }
 thread->set_gpu_exception_method(mh());
 THROW_MSG_0(vmSymbols::java_lang_NullPointerException(), buf);
 }
 // Run the kernel
-bool success = _okra_execute_with_range(kernel, dimX);
+bool success = false;
-// check if any workitem requested a deopt
+{
-// currently we only support at most one such workitem
+TraceTime t1("execute kernel", TraceGPUInteraction);
+success = _okra_execute_with_range(kernel, dimX);
+}
-int deoptcode = e->deopt_occurred();
-if (useDeoptInfo &&  deoptcode != 0) {
+if (UseHSAILDeoptimization) {
-if (deoptcode != 1) {
+// check if any workitem requested a deopt
-// error condition detected in deopt code
+// currently we only support at most one such workitem
-char msg[200];
+int deoptcode = e->deopt_occurred();
-sprintf(msg, "deopt error detected, slot for workitem %d was not empty", -1*(deoptcode + 1));
+if (deoptcode != 0) {
-guarantee(deoptcode == 1, msg);
+if (deoptcode != 1) {
-}
+// error condition detected in deopt code
-if (TraceGPUInteraction) {
+char msg[200];
-tty->print_cr("deopt happened.");
+sprintf(msg, "deopt error detected, slot for workitem %d was not empty", -1 * (deoptcode + 1));
-HSAILKernelDeoptimization * pdeopt = &e->_deopt_save_states[0];
+guarantee(deoptcode == 1, msg);
-tty->print_cr("first deopter was workitem %d", pdeopt->workitem());
+}
-}
+{
-// Before handling any deopting workitems, save the pointers from
+TraceTime t3("handle deoptimizing workitems", TraceGPUInteraction);
-// the hsail frames in oops_save so they get adjusted by any
-// GC. Need to do this before leaving thread_in_vm mode.
+if (TraceGPUInteraction) {
-// resolve handle only needed once here (not exiting vm mode)
+tty->print_cr("deopt happened.");
-objArrayOop oopsSaveArray = (objArrayOop) JNIHandles::resolve(oops_save_handle);
+HSAILKernelDeoptimization * pdeopt = &e->_deopt_save_states[0];
+tty->print_cr("first deopter was workitem %d", pdeopt->workitem());
-// since slots are allocated from the beginning, we know how far to look
-assert(e->num_deopts() < MAX_DEOPT_SAVE_STATES_SIZE, "deopt save state overflow");
-for (int k = 0; k < e->num_deopts(); k++) {
-HSAILKernelDeoptimization * pdeopt = &e->_deopt_save_states[k];
-jint workitem = pdeopt->workitem();
-if (workitem != -1) {
-// this is a workitem that deopted
-HSAILFrame *hsailFrame = pdeopt->first_frame();
-int dregOopMap = hsailFrame->dreg_oops_map();
-for (int bit = 0; bit < 16; bit++) {
-if ((dregOopMap & (1 << bit)) != 0) {
-// the dregister at this bit is an oop, save it in the array
-int index = k * 16 + bit;
-void* saved_oop = (void*) hsailFrame->get_d_reg(bit);
-oopsSaveArray->obj_at_put(index, (oop) saved_oop);
-}
 }
-}
-}
+// Before handling any deopting workitems, save the pointers from
+// the hsail frames in oops_save so they get adjusted by any
-// Handle any deopting workitems.
+// GC. Need to do this before leaving thread_in_vm mode.
-int count_deoptimized = 0;
+// resolve handle only needed once here (not exiting vm mode)
-for (int k = 0; k < e->num_deopts(); k++) {
+objArrayOop oopsSaveArray = (objArrayOop) JNIHandles::resolve(oops_save_handle);
-HSAILKernelDeoptimization * pdeopt = &e->_deopt_save_states[k];
+// since slots are allocated from the beginning, we know how far to look
-jint workitem = pdeopt->workitem();
+assert(e->num_deopts() < MAX_DEOPT_SAVE_STATES_SIZE, "deopt save state overflow");
-if (workitem != -1) {
+for (int k = 0; k < e->num_deopts(); k++) {
-int deoptId = pdeopt->pc_offset();
+HSAILKernelDeoptimization * pdeopt = &e->_deopt_save_states[k];
-HSAILFrame *hsailFrame = pdeopt->first_frame();
+jint workitem = pdeopt->workitem();
+if (workitem != -1) {
-// update the hsailFrame from the oopsSaveArray
+// this is a workitem that deopted
-// re-resolve the handle
+HSAILFrame *hsailFrame = pdeopt->first_frame();
-oopsSaveArray = (objArrayOop) JNIHandles::resolve(oops_save_handle);
+int dregOopMap = hsailFrame->dreg_oops_map();
+for (int bit = 0; bit < 16; bit++) {
-int dregOopMap = hsailFrame->dreg_oops_map();
+if ((dregOopMap & (1 << bit)) != 0) {
-for (int bit = 0; bit < 16; bit++) {
+// the dregister at this bit is an oop, save it in the array
-if ((dregOopMap & (1 << bit)) != 0) {
+int index = k * 16 + bit;
-// the dregister at this bit is an oop, retrieve it from array and put back in frame
+void* saved_oop = (void*) hsailFrame->get_d_reg(bit);
-int index = k * 16 + bit;
+oopsSaveArray->obj_at_put(index, (oop) saved_oop);
-void * dregValue = (void *) oopsSaveArray->obj_at(index);
-void * oldDregValue = (void *) hsailFrame->get_d_reg(bit);
-assert((oldDregValue != 0 ? dregValue != 0 : dregValue == 0) , "bad dregValue retrieved");
-if (TraceGPUInteraction) {
-if (dregValue != oldDregValue) {
-tty->print_cr("oop moved for $d%d, workitem %d, slot %d, old=%p, new=%p", bit, workitem, k, oldDregValue, dregValue);
 }
-}
-hsailFrame->put_d_reg(bit, (jlong) dregValue);
-}
-}
-JavaValue result(T_VOID);
-JavaCallArguments javaArgs;
-javaArgs.set_alternative_target(nm);
-javaArgs.push_int(deoptId);
-javaArgs.push_long((jlong) hsailFrame);
-// override the deoptimization action with Action_none until we decide
-// how to handle the other actions.
-int myActionReason = Deoptimization::make_trap_request(Deoptimization::trap_request_reason(pdeopt->reason()), Deoptimization::Action_none);
-javaArgs.push_int(myActionReason);
-javaArgs.push_oop((oop)NULL);
-if (TraceGPUInteraction) {
-int dregOopMap = hsailFrame->dreg_oops_map();
-tty->print_cr("[HSAIL] Deoptimizing to host for workitem=%d (slot=%d) with deoptId=%d, frame=" INTPTR_FORMAT ", actionAndReason=%d, dregOopMap=%04x", workitem, k, deoptId, hsailFrame, myActionReason, dregOopMap);
-// show the registers containing references
-for (int bit = 0; bit < 16; bit++) {
-if ((dregOopMap & (1 << bit)) != 0) {
-tty->print_cr("  oop $d%d = %p", bit, hsailFrame->get_d_reg(bit));
 }
 }
 }
-JavaCalls::call(&result, mh, &javaArgs, THREAD);
-count_deoptimized++;
+// Handle any deopting workitems.
-}
+int count_deoptimized = 0;
-}
+for (int k = 0; k < e->num_deopts(); k++) {
-if (TraceGPUInteraction) {
+HSAILKernelDeoptimization * pdeopt = &e->_deopt_save_states[k];
-tty->print_cr("[HSAIL] Deoptimizing to host completed for %d workitems", count_deoptimized);
-}
+jint workitem = pdeopt->workitem();
+if (workitem != -1) {
-// Handle any never_ran workitems if there were any
+int deoptId = pdeopt->pc_offset();
-int count_never_ran = 0;
+HSAILFrame *hsailFrame = pdeopt->first_frame();
-bool handleNeverRansHere = true;
-// turn off verbose trace stuff for javacall arg setup
+// update the hsailFrame from the oopsSaveArray
-bool savedTraceGPUInteraction = TraceGPUInteraction;
+// re-resolve the handle
-TraceGPUInteraction = false;
+oopsSaveArray = (objArrayOop) JNIHandles::resolve(oops_save_handle);
-jboolean *never_ran_array = e->never_ran_array();
-if (handleNeverRansHere) {
+int dregOopMap = hsailFrame->dreg_oops_map();
-for (int k = 0; k < dimX; k++) {
+for (int bit = 0; bit < 16; bit++) {
-if (never_ran_array[k]) {
+if ((dregOopMap & (1 << bit)) != 0) {
-// run it as a javaCall
+// the dregister at this bit is an oop, retrieve it from array and put back in frame
-KlassHandle methKlass = mh->method_holder();
+int index = k * 16 + bit;
-Thread* THREAD = Thread::current();
+void * dregValue = (void *) oopsSaveArray->obj_at(index);
-JavaValue result(T_VOID);
+void * oldDregValue = (void *) hsailFrame->get_d_reg(bit);
-JavaCallArguments javaArgs;
+assert((oldDregValue != 0 ? dregValue != 0 : dregValue == 0), "bad dregValue retrieved");
-// re-resolve the args_handle here
+if (TraceGPUInteraction) {
-objArrayOop resolvedArgsArray = (objArrayOop) JNIHandles::resolve(args_handle);
+if (dregValue != oldDregValue) {
-// This object sets up the javaCall arguments
+tty->print_cr("oop moved for $d%d, workitem %d, slot %d, old=%p, new=%p", bit, workitem, k, oldDregValue, dregValue);
-// the way argsArray is set up, this should work for instance methods as well
+}
-// (the receiver will be the first oop pushed)
+}
-HSAILJavaCallArguments hjca(&javaArgs, k, mh->signature(), resolvedArgsArray, mh->is_static());
+hsailFrame->put_d_reg(bit, (jlong) dregValue);
-if (mh->is_static()) {
+}
-JavaCalls::call_static(&result, methKlass, mh->name(), mh->signature(), &javaArgs, THREAD);
+}
-} else {
-JavaCalls::call_virtual(&result, methKlass, mh->name(), mh->signature(), &javaArgs, THREAD);
+JavaValue result(T_VOID);
+JavaCallArguments javaArgs;
+javaArgs.set_alternative_target(nm);
+javaArgs.push_int(deoptId);
+javaArgs.push_long((jlong) hsailFrame);
+// override the deoptimization action with Action_none until we decide
+// how to handle the other actions.
+int myActionReason = Deoptimization::make_trap_request(Deoptimization::trap_request_reason(pdeopt->reason()), Deoptimization::Action_none);
+javaArgs.push_int(myActionReason);
+javaArgs.push_oop((oop) NULL);
+if (TraceGPUInteraction) {
+int dregOopMap = hsailFrame->dreg_oops_map();
+tty->print_cr("[HSAIL] Deoptimizing to host for workitem=%d (slot=%d) with deoptId=%d, frame=" INTPTR_FORMAT ", actionAndReason=%d, dregOopMap=%04x", workitem, k, deoptId, hsailFrame, myActionReason, dregOopMap);
+// show the registers containing references
+for (int bit = 0; bit < 16; bit++) {
+if ((dregOopMap & (1 << bit)) != 0) {
+tty->print_cr("  oop $d%d = %p", bit, hsailFrame->get_d_reg(bit));
+}
+}
+}
+JavaCalls::call(&result, mh, &javaArgs, THREAD);
+count_deoptimized++;
 }
-count_never_ran++;
+}
+if (TraceGPUInteraction) {
+tty->print_cr("[HSAIL] Deoptimizing to host completed for %d workitems", count_deoptimized);
 }
 }
-TraceGPUInteraction = savedTraceGPUInteraction;
-if (TraceGPUInteraction) {
+{
-tty->print_cr("%d workitems never ran, have been run via JavaCall", count_never_ran);
+TraceTime t3("handle never-rans", TraceGPUInteraction);
-showRanges(never_ran_array, dimX);
+// Handle any never_ran workitems if there were any
+int count_never_ran = 0;
+bool handleNeverRansHere = true;
+// turn off verbose trace stuff for javacall arg setup
+bool savedTraceGPUInteraction = TraceGPUInteraction;
+TraceGPUInteraction = false;
+jboolean *never_ran_array = e->never_ran_array();
+if (handleNeverRansHere) {
+for (int k = 0; k < dimX; k++) {
+if (never_ran_array[k]) {
+// run it as a javaCall
+KlassHandle methKlass = mh->method_holder();
+Thread* THREAD = Thread::current();
+JavaValue result(T_VOID);
+JavaCallArguments javaArgs;
+// re-resolve the args_handle here
+objArrayOop resolvedArgsArray = (objArrayOop) JNIHandles::resolve(args_handle);
+// This object sets up the javaCall arguments
+// the way argsArray is set up, this should work for instance methods as well
+// (the receiver will be the first oop pushed)
+HSAILJavaCallArguments hjca(&javaArgs, k, mh->signature(), resolvedArgsArray, mh->is_static());
+if (mh->is_static()) {
+JavaCalls::call_static(&result, methKlass, mh->name(), mh->signature(), &javaArgs, THREAD);
+} else {
+JavaCalls::call_virtual(&result, methKlass, mh->name(), mh->signature(), &javaArgs, THREAD);
+}
+count_never_ran++;
+}
+}
+TraceGPUInteraction = savedTraceGPUInteraction;
+if (TraceGPUInteraction) {
+tty->print_cr("%d workitems never ran, have been run via JavaCall", count_never_ran);
+showRanges(never_ran_array, dimX);
+}
+} // end of never-ran handling
 }
-} // end of never-ran handling
+}
-}
-if (useDeoptInfo) {
 FREE_C_HEAP_ARRAY(jboolean, e->never_ran_array(), mtInternal);
 delete e;
 }
 return success;
 }

Mercurial > hg > truffle

comparison src/gpu/hsail/vm/gpu_hsail.cpp @ 14969:a6c144380ce7