[sanitizer][win] Fix `Atexit()` on MinGW asan_dynamic runtime

[alvinhochun]

Summary

Some functions of asan depends on Atexit() handlers. On Windows, this
is implemented in ad3ec82bb1c7217b0187a1e16bb22642e194ce94 to queue the
handlers in a vector then register them with atexit() only after the
CRT is fully initialized. However, this is broken on MinGW with
asan_dynamic runtime due to different initialization order. This change
fixes the issue by making sure that Atexit() can call atexit()
directly past the pre-initialization phase.

This fixes two asan test cases on MinGW.

[mstorsjo]

compiler-rt/lib/sanitizer_common/sanitizer_win.cpp

731

So... is the root issue that C++ global constructors are executed after the .CRT$XID initializer in MSVC environments, but before in mingw environments? I believe they're supposed to be compatible, so if there's such a discrepancy, maybe we should also consider trying to fix it in mingw-w64 crt?

[alvinhochun]

compiler-rt/lib/sanitizer_common/sanitizer_win.cpp

731

No, C++ global constructors are called after .CRT$XID initializer on MinGW too. This is what I got from stepping through it:

1. RunAtexit is called quite early.
2. The AsanInitializer constructor is called quite late by __main -> __do_global_ctors (gccmain.c). This constructor calls AsanInitFromRtl -> AsanInitInternal -> Atexit(asan_atexit).

I do suspect on MSVC target there is something else other than the C++ global constructor that calls __asan_init early. Perhaps it may have something to do with linking with static vs dynamic CRT? (Static asan has other initializers.)

[alvinhochun]

Reformat

compiler-rt/lib/sanitizer_common/sanitizer_win.cpp

 
 void ListOfModules::fallbackInit() { clear(); }
 
 // We can't use atexit() directly at __asan_init time as the CRT is not fully
 // initialized at this point.  Place the functions into a vector and use
 // atexit() as soon as it is ready for use (i.e. after .CRT$XIC initializers).
 InternalMmapVectorNoCtor<void (*)(void)> atexit_functions;
 
-int Atexit(void (*function)(void)) {
+static int queueAtexit(void (*function)(void)) {
   atexit_functions.push_back(function);
   return 0;
 }
 
+// If Atexit() is being called after RunAtexit() has already been run, it needs
+// to be able to call atexit() directly. Here we use a function ponter to
+// switch out its behaviour.
+// An example of where this is needed is the asan_dynamic runtime on MinGW-w64.
+// On this environment, __asan_init is called during global constructor phase,
+// way after calling the .CRT$XID initializer.
+static int (*volatile queueOrCallAtExit)(void (*)(void)) = &queueAtexit;
+
+int Atexit(void (*function)(void)) { return queueOrCallAtExit(function); }
+
 static int RunAtexit() {
   TraceLoggingUnregister(g_asan_provider);
+  queueOrCallAtExit = &atexit;
   int ret = 0;
   for (uptr i = 0; i < atexit_functions.size(); ++i) {
     ret |= atexit(atexit_functions[i]);
   }
   return ret;
 }
 
 #pragma section(".CRT$XID", long, read)

compiler-rt/test/asan/TestCases/Windows/coverage-basic.cpp

 // RUN: rm -rf %t-dir
 // RUN: mkdir %t-dir && cd %t-dir
 // RUN: %clangxx_asan -fsanitize-coverage=func,trace-pc-guard %s -o test.exe
 // RUN: %env_asan_opts=coverage=1 %run ./test.exe
 //
 // RUN: %sancov print *.sancov | FileCheck %s
 
-// FIXME: Investigate failure on MinGW.
-// XFAIL: target={{.*-windows-gnu}}
-
 #include <stdio.h>
 
 void foo() { fputs("FOO", stderr); }
 void bar() { fputs("BAR", stderr); }
 
 int main(int argc, char **argv) {
   if (argc == 2) {
     foo();

compiler-rt/test/asan/TestCases/atexit_stats.cpp

 // Make sure we report atexit stats.
 // RUN: %clangxx_asan -O3 %s -o %t
 // RUN: %env_asan_opts=atexit=1:print_stats=1 %run %t 2>&1 | FileCheck %s
 //
 // No atexit output in older versions of Android due to
 // https://code.google.com/p/address-sanitizer/issues/detail?id=263
 // UNSUPPORTED: android
 
-// FIXME: Investigate failure on MinGW.
-// XFAIL: target={{.*-windows-gnu}}
-
 #include <stdlib.h>
 #if !defined(__APPLE__) && !defined(__FreeBSD__) && !defined(__NetBSD__)
 #include <malloc.h>
 #endif
 int *p1 = (int*)malloc(900);
 int *p2 = (int*)malloc(90000);
 int *p3 = (int*)malloc(9000000);
 int main() { }
 
 // CHECK: AddressSanitizer exit stats: