(Do not land) Very hacky patch for using -fprofile-functions in the Win-Clang build

tools/cygprofile/cygprofile2.cc

+#include <stdio.h>
+#include <algorithm>
+#include <unordered_set>
+#include <unordered_map>
+
+// Compiling printf with -finstrument-functions seems to break it?
+
+#ifdef _MSC_VER
+#include <windows.h>
+#include <process.h>
+#include <dbghelp.h>
+
+#if 0
+extern "C" unsigned long __stdcall GetCurrentProcessId(void);
+
+extern "C" unsigned long _beginthread( void( __cdecl *start_address )( void * ), unsigned stack_size, void *arglist );
+#endif
+
+namespace {
+
+// We can easily run out of memory if this is too high.
+#define N (50*1024*1024)
+//#define N (1*1024*1024)
+void *tracebuf[N];
+int n = 0;
+
+char symbuf[1024*1024];
+
+void dump(void*) {
+  HMODULE dbghelp = LoadLibraryA("dbghelp.dll");
+  decltype(::SymFromAddr) *symfromaddr =
+    reinterpret_cast<decltype(::SymFromAddr)*>(GetProcAddress(dbghelp, "SymFromAddr")); 
+  decltype(::SymInitialize) *syminitialize =
+    reinterpret_cast<decltype(::SymInitialize)*>(GetProcAddress(dbghelp, "SymInitialize")); 
+  decltype(::SymSetOptions) *symsetoptions =
+    reinterpret_cast<decltype(::SymSetOptions)*>(GetProcAddress(dbghelp, "SymSetOptions")); 
+
+
+  char fname[128];
+  sprintf(fname, "/src/tmp/cygprofile%lu.txt", GetCurrentProcessId());
+  FILE *f = fopen(fname, "a");
+
+  syminitialize(GetCurrentProcess(), NULL, TRUE);
+  symsetoptions(SYMOPT_DEFERRED_LOADS | SYMOPT_PUBLICS_ONLY);
+
+  std::unordered_set<void*> seen;
+
+#if 0
+  std::unordered_map<void*, unsigned> freq;
+  for (int i = 0; i < N; i++)
+    ++freq[tracebuf[i]];
+  std::vector<void*> funcs;
+  for (auto i : freq)
+    funcs.push_back(i.first);
+  std::sort(funcs.begin(), funcs.end(), [&](void* a, void* b) {
+      return freq[b] < freq[a];
+      });
+  std::unordered_set<std::string> seen_names;
+  for (void *func : funcs) {
+    PSYMBOL_INFO symbol = (PSYMBOL_INFO)symbuf;
+    symbol->SizeOfStruct = sizeof(SYMBOL_INFO);
+    symbol->MaxNameLen = MAX_SYM_NAME;
+    DWORD64 offset = 0;
+    if (symfromaddr(GetCurrentProcess(), (DWORD64)func, &offset, symbol)) {
+      char *name = symbol->Name;
+      if (name[0] == '_')
+        name++;
+      if (seen_names.count(name))
+        continue;
+      seen_names.insert(name);
+      fprintf(f, "%s (%u)\n", name, freq[func]);
+    }
+  }
+#endif
+
+#if 1
+  // Due to inlining, multiple functions can end up with the same symbol.
+  std::unordered_set<std::string> seen_names;
+
+  for (int i = 0; i < N; i++) {
+    if (seen.count(tracebuf[i]))
+      continue;
+    seen.insert(tracebuf[i]);
+
+    //fprintf(f, "0x%p\n", tracebuf[i]);
+
+    PSYMBOL_INFO symbol = (PSYMBOL_INFO)symbuf;
+    symbol->SizeOfStruct = sizeof(SYMBOL_INFO);
+    symbol->MaxNameLen = MAX_SYM_NAME;
+    DWORD64 offset = 0;
+    if (symfromaddr(GetCurrentProcess(), (DWORD64)tracebuf[i], &offset, symbol)) {
+      char *name = symbol->Name;
+      if (name[0] == '_')
+        name++;
+      if (seen_names.count(name))
+        continue;
+      seen_names.insert(name);
+      fprintf(f, "%s\n", name);
+    }
+  }
+#endif
+
+  fclose(f);
+}
+
+}
+#endif
+
+extern "C" {
+
+// Set attributes to be sure we don't instrument ourselves.
+void __cyg_profile_func_enter(void* this_fn, void* call_site)
+    __attribute__((no_instrument_function));
+void __cyg_profile_func_exit(void* this_fn, void* call_site)
+    __attribute__((no_instrument_function));
+
+void __cyg_profile_func_enter(void* this_fn, void* callee_unused) {
+  // This gets pulled into NaCL binaries and all kinds of places.
+#ifdef _MSC_VER
+  static int done = 0;
+
+  if (done)
+    return;
+
+  tracebuf[n++] = this_fn;
+
+  if (n == N) {
+    done = 1;
+    _beginthread(dump, 0, nullptr);
+    return;
+  }
+#endif
+}
+
+void __cyg_profile_func_exit(void* this_fn, void* call_site) {}
+
+}  // extern "C"
+