[OpenMP][Archer][TSan] Use TSan fibers to implement loop-level analysi for OpenMP ws-loops (4a11bb40) · Commits · llvm-doe / llvm-project

compiler-rt/lib/tsan/rtl/tsan_rtl.h

+1 −0

Original line number	Diff line number	Diff line
		@@ -801,6 +801,7 @@ void FuncExit(ThreadState *thr) {
		#if !SANITIZER_GO
		DCHECK_LT(thr->shadow_stack_pos, thr->shadow_stack_end);
		#endif
		CHECK_GT(thr->shadow_stack_pos, thr->shadow_stack);
		thr->shadow_stack_pos--;
		}

openmp/runtime/src/ompt-specific.cpp

+0 −17

Original line number	Diff line number	Diff line
		@@ -483,23 +483,6 @@ int __ompt_get_task_memory_internal(void *addr, size_t size, int blocknum) {
		*addr = taskdata;
		*size = taskdata->td_size_alloc;
		return 0;

		/* void *ret_addr;
		int64_t ret_size = taskdata->td_size_alloc - sizeof(kmp_taskdata_t);

		// kmp_task_t->data1 is an optional member
		if (taskdata->td_flags.destructors_thunk)
		ret_addr = &task->data1 + 1;
		else
		ret_addr = &task->part_id + 1;

		ret_size -= (char )(ret_addr) - (char )(task);
		if (ret_size < 0)
		return 0;

		*addr = ret_addr;
		*size = (size_t)ret_size;
		return 1;*/
		}

		//----------------------------------------------------------

openmp/tools/archer/ompt-tsan.cpp

+70 −1

Original line number	Diff line number	Diff line
		@@ -65,6 +65,7 @@ public:
		int ignore_serial{0};
		std::atomic<int> all_memory{0};
		int tasking{0};
		int dispatch_fibers{0};
		int stack_size{1024};
		std::atomic<int> untieds{0};

		@@ -103,6 +104,8 @@ public:
		continue;
		if (sscanf(token.c_str(), "tasking=%d", &tasking))
		continue;
		if (sscanf(token.c_str(), "dispatch_fibers=%d", &dispatch_fibers))
		continue;
		if (sscanf(token.c_str(), "stack_size=%d", &stack_size))
		continue;
		if (sscanf(token.c_str(), "ignore_serial=%d", &ignore_serial))
		@@ -505,6 +508,31 @@ struct TaskData final : DataPoolEntry<TaskData> {
		size_t PrivateDataSize{0};
		void *PrivateDataAddr{nullptr};

		std::vector<void *> DispatchFibers;
		std::size_t DispatchFiberIdx{0};
		void *DispatchOriginalFiber{nullptr};

		void activateDispatchFibers(const std::size_t size, void *fiber = nullptr) {
		DispatchFibers.resize(size);
		DispatchOriginalFiber = fiber;
		for (auto &RFiber : DispatchFibers)
		RFiber = TsanCreateFiber(1);
		}

		void *switchDispatchFiber() {
		void *Res = DispatchFibers.at(DispatchFiberIdx);
		DispatchFiberIdx = (DispatchFiberIdx + 1) % DispatchFibers.size();
		TsanSwitchToFiber(Res, 1);
		return Res;
		}

		void deactivateDispatchFibers() {
		if (DispatchOriginalFiber)
		TsanSwitchToFiber(DispatchOriginalFiber, 1);
		for (auto &Fiber : DispatchFibers)
		TsanDestroyFiber(Fiber);
		}

		const void *CodePtr{nullptr};
		/// Count how often this structure has been put into child tasks + 1.
		std::atomic_int RefCount{1};
		@@ -662,7 +690,7 @@ static void ompt_tsan_thread_begin(ompt_thread_t thread_type,
		DependencyDataPool::ThreadDataPool = new DependencyDataPool;
		TsanNewMemory(DependencyDataPool::ThreadDataPool,
		sizeof(DependencyDataPool::ThreadDataPool));
		if (archer_flags->tasking) {
		if (archer_flags->tasking \|\| archer_flags->dispatch_fibers) {
		TsanGetCurrentFiber();
		}

		@@ -712,6 +740,43 @@ static void ompt_tsan_parallel_end(ompt_data_t *parallel_data,
		#endif
		}

		static void ompt_tsan_dispatch(ompt_data_t *parallel_data,
		ompt_data_t *task_data, ompt_dispatch_t kind,
		ompt_data_t instance) {
		auto *Data = ToTaskData(task_data);
		switch (kind) {
		case ompt_dispatch_section:
		case ompt_dispatch_ws_loop_chunk:
		case ompt_dispatch_distribute_chunk:
		Data->switchDispatchFiber();
		break;
		case ompt_dispatch_taskloop_chunk:
		case ompt_dispatch_iteration:
		break;
		}
		}

		static void ompt_tsan_work(ompt_work_t work_type,
		ompt_scope_endpoint_t endpoint,
		ompt_data_t parallel_data, ompt_data_t task_data,
		uint64_t count, const void *codeptr_ra) {
		auto *Data = ToTaskData(task_data);
		switch (endpoint) {
		case ompt_scope_begin:
		Data->activateDispatchFibers(std::min(archer_flags->dispatch_fibers == -1
		? static_cast<uint64_t>(13)
		: count,
		count),
		TsanGetCurrentFiber());
		break;
		case ompt_scope_end:
		Data->deactivateDispatchFibers();
		break;
		case ompt_scope_beginend:
		break;
		}
		}

		static void ompt_tsan_implicit_task(ompt_scope_endpoint_t endpoint,
		ompt_data_t *parallel_data,
		ompt_data_t *task_data,
		@@ -1291,6 +1356,10 @@ static int ompt_tsan_initialize(ompt_function_lookup_t lookup, int device_num,
		SET_CALLBACK(implicit_task);
		SET_CALLBACK(sync_region);
		SET_CALLBACK(parallel_end);
		if (__archer::archer_flags->dispatch_fibers) {
		SET_CALLBACK(dispatch);
		SET_CALLBACK(work);
		}

		SET_CALLBACK(task_create);
		SET_CALLBACK(task_schedule);

openmp/tools/archer/tests/parallel/parallel-loop-level-fp.c

0 → 100644

+40 −0

Original line number	Diff line number	Diff line
		/*
		* loop-level.c -- Archer testcase
		*/
		//===----------------------------------------------------------------------===//
		//
		// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
		//
		// See tools/archer/LICENSE.txt for details.
		// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
		//
		//===----------------------------------------------------------------------===//

		// RUN: %libarcher-compile && env ARCHER_OPTIONS=dispatch_fibers=1 %libarcher-run \| FileCheck %s
		// RUN: %libarcher-compile && env ARCHER_OPTIONS=dispatch_fibers=1:ignore_serial=1 %libarcher-run \| FileCheck %s
		// REQUIRES: tsan
		// XFAIL: *
		#include <omp.h>
		#include <stdio.h>
		#include <unistd.h>

		#define NUM_THREADS 2

		int main(int argc, char *argv[]) {
		int vars[NUM_THREADS] = {0};
		int i;
		const int len = 10;

		#pragma omp parallel for num_threads(NUM_THREADS) shared(vars) \
		schedule(dynamic, 1)
		for (i = 0; i < len; i++) {
		vars[omp_get_thread_num()]++;
		}

		fprintf(stderr, "DONE\n");
		return 0;
		}

		// CHECK-NOT: ThreadSanitizer: data race
		// CHECK-NOT: ThreadSanitizer: reported
		// CHECK: DONE

openmp/tools/archer/tests/races/parallel-for-antidep-dynamic.c

0 → 100644

+48 −0

Original line number	Diff line number	Diff line
		/*
		* parallel-for-antidep-dynamic.c -- Archer testcase
		*/
		//===----------------------------------------------------------------------===//
		//
		// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
		//
		// See tools/archer/LICENSE.txt for details.
		// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
		//
		//===----------------------------------------------------------------------===//
		// RUN: %libarcher-compile && env ARCHER_OPTIONS=dispatch_fibers=1 %libarcher-run-race \| FileCheck %s
		// RUN: %libarcher-compile && env ARCHER_OPTIONS=dispatch_fibers=1:ignore_serial=1 %libarcher-run-race \| FileCheck %s
		// REQUIRES: tsan
		#include <assert.h>
		#include <stdio.h>
		#include <stdlib.h>

		#define LEN 6

		int main(int argc, char *argv[]) {
		int indeces[LEN] = {1, 3, 5, 7, 9, 13};
		double base[26];
		double *p = base;
		double *q = p + 12;
		int i;

		for (i = 1; i < 26; ++i)
		base[i] = 0.5 * i;

		#pragma omp parallel for num_threads(2) schedule(dynamic, 1)
		for (i = 0; i < LEN; ++i) {
		int idx = indeces[i];
		p[idx] += 1.0 + i;
		q[idx] += 3.0 + i;
		}

		fprintf(stderr, "DONE.\n");
		return 0;
		}

		// CHECK: WARNING: ThreadSanitizer: data race
		// CHECK: Write of size 8
		// CHECK-NEXT: #0 {{.*}}parallel-for-antidep-dynamic.c:3{{4\|5}}
		// CHECK: Previous write of size 8
		// CHECK-NEXT: #0 {{.*}}parallel-for-antidep-dynamic.c:3{{4\|5}}
		// CHECK: DONE
		// CHECK: ThreadSanitizer: reported {{[0-9]+}} warnings