[llvm-branch-commits] [compiler-rt] [compiler-rt] Add 'csan' library for the concurrency sanitizer (PR #225782)

[Yaxun Liu via llvm-branch-commits]

================
@@ -0,0 +1,456 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+///
+/// \file
+/// Watchpoint-based data race detector for GPU targets.
+///
+//===----------------------------------------------------------------------===//
+
+#include <gpuintrin.h>
+
+#include "csan_offload_packet.h"
+#include "csan_watch.h"
+#include "sanitizer_common/sanitizer_internal_defs.h"
+#include "shared/rpc.h"
+
+using namespace __sanitizer;
+
+#define INTERFACE extern "C" SANITIZER_INTERFACE_ATTRIBUTE
+
+extern "C" {
+// Externally initialized by the sanitizer, keeps one table per active device.
+[[gnu::visibility("protected")]] u64 *__csan_watchpoint_table = nullptr;
+}
+
+static constexpr u64 SAMPLE_DELAY_MIN_NS = 1000;
+static constexpr u64 SAMPLE_DELAY_MAX_NS = 10000;
+static constexpr u32 WP_CHANCE = 8;
+static_assert((CSAN_WATCHPOINT_TABLE_ENTRIES &
+               (CSAN_WATCHPOINT_TABLE_ENTRIES - 1)) == 0,
+              "watchpoint table size must be a power of two");
+static_assert(WP_CHANCE >= 2 && (WP_CHANCE & (WP_CHANCE - 1)) == 0,
+              "WP_CHANCE must be a power of two");
+
+// The GPU case does
+static constexpr u32 GPU_MAX_ACCESS_SIZE = 16;
+static constexpr u32 GPU_WATCHPOINT_ENTRIES = CSAN_WATCHPOINT_TABLE_ENTRIES;
+static constexpr u32 GPU_CHECK_ADJACENT_SLOTS = 0;
+static_assert(GPU_WATCHPOINT_ENTRIES * sizeof(u64) == 2 * 1024 * 1024,
+              "GPU watchpoint table must be 2 MiB");
+using GpuWatchpointTable =
+    __csan::WatchpointTable<GPU_MAX_ACCESS_SIZE, GPU_CHECK_ADJACENT_SLOTS>;
+
+static GpuWatchpointTable get_watchpoints() {
+  return GpuWatchpointTable(__csan_watchpoint_table);
+}
+
+// LDS addresses share the global watchpoint table. The original address is
+// combined with the block's linear ID to create a unique global address.
+static constexpr u32 LDS_OFFSET_BITS = 20;
+static constexpr u64 LDS_OFFSET_MASK = (1ull << LDS_OFFSET_BITS) - 1;
+static constexpr u64 LDS_FLAG = 1ull << (GpuWatchpointTable::AddressBits - 1);
+static constexpr u64 GLOBAL_ADDRESS_MASK = LDS_FLAG - 1;
+static constexpr u64 LDS_MAX_BLOCKS =
+    1ull << (GpuWatchpointTable::AddressBits - 1 - LDS_OFFSET_BITS);
+static_assert((LDS_FLAG | ((LDS_MAX_BLOCKS - 1) << LDS_OFFSET_BITS) |
+               LDS_OFFSET_MASK) == GpuWatchpointTable::AddressMask,
+              "LDS key fields must fill the address bits");
+
+[[gnu::visibility("protected"),
+  gnu::weak]] rpc::Client client asm("__llvm_rpc_client");
+
+static u64 __csan_num_data_races = 0;
+
+INTERFACE u64 __csan_get_num_data_races() {
+  return __atomic_load_n(&__csan_num_data_races, __ATOMIC_RELAXED);
+}
+
+// Shallow deduplication check to save the host thread work. Keyed on both the
+// PC and the race kind so each distinct kind of race at a PC is reported once.
+static bool should_report(void *pc, unsigned kind) {
+  static u64 seen[64] = {};
+  const u64 token = (reinterpret_cast<uptr>(pc) >> 4) ^
----------------
yxsamliu wrote:

Should the deduplication key also include `peer_pc`? With only the current PC and race kind, different races from the same instruction may suppress each other before the host can deduplicate the PC pair. If this is intentional, could we document the tradeoff?

https://github.com/llvm/llvm-project/pull/225782