[clang] [compiler-rt] [llvm] [PGO][AMDGPU] Add offload profiling with uniformity-aware optimization (PR #177665)

[Yaxun Liu via cfe-commits]

================
@@ -0,0 +1,463 @@
+//===- InstrProfilingPlatformROCm.c - Profile data ROCm platform ---------===//
+//
+// 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
+//
+//===----------------------------------------------------------------------===//
+
+#include "InstrProfiling.h"
+#include "InstrProfilingInternal.h"
+#include "InstrProfilingPort.h"
+#include <dlfcn.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+static int ProcessDeviceOffloadPrf(void *DeviceOffloadPrf, int TUIndex);
+
+static int IsVerboseMode() {
+  static int IsVerbose = -1;
+  if (IsVerbose == -1)
+    IsVerbose = getenv("LLVM_PROFILE_VERBOSE") != NULL;
+  return IsVerbose;
+}
+
+/* -------------------------------------------------------------------------- */
+/*  Dynamic loading of HIP runtime symbols                                   */
+/* -------------------------------------------------------------------------- */
+
+typedef int (*hipMemcpyFromSymbolTy)(void *, const void *, size_t, size_t, int);
+typedef int (*hipGetSymbolAddressTy)(void **, const void *);
+typedef int (*hipMemcpyTy)(void *, void *, size_t, int);
+typedef int (*hipModuleGetGlobalTy)(void **, size_t *, void *, const char *);
+
+static hipMemcpyFromSymbolTy pHipMemcpyFromSymbol = NULL;
+static hipGetSymbolAddressTy pHipGetSymbolAddress = NULL;
+static hipMemcpyTy pHipMemcpy = NULL;
+static hipModuleGetGlobalTy pHipModuleGetGlobal = NULL;
+
+/* -------------------------------------------------------------------------- */
+/*  Device-to-host copies                                                     */
+/*  Keep HIP-only to avoid an HSA dependency.                                 */
+/* -------------------------------------------------------------------------- */
+
+static void EnsureHipLoaded(void) {
+  static int Initialized = 0;
+  if (Initialized)
+    return;
+  Initialized = 1;
+
+  void *Handle = dlopen("libamdhip64.so", RTLD_LAZY | RTLD_LOCAL);
----------------
yxsamliu wrote:

Fixed by using weak references for dlopen/dlsym/dlerror. The profile runtime is a static library, so adding -ldl globally would affect all profiling users. With weak references, programs without -ldl link successfully with dl* resolving to NULL. This only affects HIP-specific profile collection in InstrProfilingPlatformROCm.c. HIP offload pgo should still work since it includes -ldl by default. OpenMP is not affected since it does not use InstrProfilingPlatformROCm.c for data collection. It uses its own data collection which does not need -ldl.

Could you verify this fix resolves the link error on your end? Thanks.


https://github.com/llvm/llvm-project/pull/177665