Add malloc/new profiling on linux

src/LinuxMallocProfiling.cpp

Index: src/LinuxMallocProfiling.cpp
diff --git a/src/LinuxMallocProfiling.cpp b/src/LinuxMallocProfiling.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..ce02644b541b5aa521aa41defe7cd09d7983f496
--- /dev/null
+++ b/src/LinuxMallocProfiling.cpp
@@ -0,0 +1,83 @@
+//===- subzero/src/LinuxMallocProfiling.cpp - malloc/new tracing  ---------===//
+//
+//                        The Subzero Code Generator
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+///
+/// \file
+/// \brief malloc/new/...caller tracing.
+///
+//===----------------------------------------------------------------------===//
+
+#include "LinuxMallocProfiling.h"
+
+#include <malloc.h>

[Jim Stichnoth, 2016/03/12 00:12:58]

alphabetize these.

[sehr, 2016/03/15 00:36:45]

Done.

+#include <dlfcn.h>
+#include <iostream>
+#include <map>
+
+extern "C" void *__libc_malloc(size_t size);
+
+namespace {
+std::map<void*, uint64_t> *Callers;

[Jim Stichnoth, 2016/03/12 00:12:58]

Use std::unordered_map<>, no?

Also, please use a typedef instead.

using MapType = std::unordered_map<void *, uint64_t>;

[Jim Stichnoth, 2016/03/12 00:36:15]

Probably should add a note somewhere that this is not thread-safe?  Could even
check the flags for threads=0.

[sehr, 2016/03/15 00:36:45]

Added here and a warning comes out if a build including the code is run with
threads!=0.

[sehr, 2016/03/15 00:36:45]

Done.

+// The Callers structure allocates memory, which would perturb the tracing.
+// InAllocatorFunction is true when we are tracing a new/malloc/...
+bool InAllocatorFunction = false;
+
+void *InternalAllocator(size_t size, void *caller) {

[Jim Stichnoth, 2016/03/12 00:12:58]

LLVM coding style would be:

void *internalAllocator(size_t Size, void *Caller) {

[sehr, 2016/03/15 00:36:45]

Done.

+  if (Callers != nullptr && !InAllocatorFunction) {
+    InAllocatorFunction = true;
+    auto SetI = Callers->find(caller);

[Jim Stichnoth, 2016/03/12 00:12:58]

Can't you just simply:

  ++(*Callers)[caller];

?

[sehr, 2016/03/15 00:36:45]

Done.

+    if (SetI == Callers->end()) {
+      (*Callers)[caller] = 1;
+    } else {
+      (*Callers)[caller]++;
+    }
+    InAllocatorFunction = false;
+  }
+  return __libc_malloc(size);
+}
+} // end anonymous namespace

[Jim Stichnoth, 2016/03/12 00:12:58]

end of anonymous namespace

[sehr, 2016/03/15 00:36:45]

Done.

+
+void *operator new (size_t size) {
+  void *caller = __builtin_return_address(0);
+  return InternalAllocator(size, caller);
+}
+
+void *operator new[] (size_t size) {
+  void *caller = __builtin_return_address(0);
+  return InternalAllocator(size, caller);
+}
+
+extern "C" void *malloc(size_t size) {
+  void *caller = __builtin_return_address(0);
+  return InternalAllocator(size, caller);
+}
+
+namespace Ice {
+
+void LinuxMallocProfiling::Start() {
+  Callers = new std::map<void*, uint64_t>();
+}
+
+void LinuxMallocProfiling::End() {
+  for (auto C : *Callers) {

[Karl, 2016/03/11 23:25:11]

Should you use?
    for (auto &C : *Callers) {

[sehr, 2016/03/15 00:36:45]

Done.

+    Dl_info dli;
+    dladdr(C.first, &dli);
+
+    std::cout << C.second << " ";
+    if (dli.dli_sname == NULL) {
+      std::cout << C.first;
+    } else {
+      std::cout << dli.dli_sname;
+    }
+    std::cout << std::endl;

[Jim Stichnoth, 2016/03/12 00:12:58]

I believe LLVM coding style says use "\n" instead of std::endl.

[sehr, 2016/03/15 00:36:45]

Done.

+  }
+  delete Callers;

[Jim Stichnoth, 2016/03/12 00:12:58]

Our code base favors unique_ptr<> over explicit delete.

[sehr, 2016/03/15 00:36:45]

The Callers map needs to be accessible to non-member functions, and unique_ptr
would need to be a constructed object in static data, which is also
unacceptable.  Suggestions?

+  Callers = nullptr;
+}
+
+} // end namespace Ice

[Jim Stichnoth, 2016/03/12 00:12:58]

end of namespace Ice

[sehr, 2016/03/15 00:36:45]

Done.