[PATCH] D48891: [libFuzzer] Make -fsanitize=memory,fuzzer work.

[Kostya Serebryany via Phabricator via llvm-commits]

kcc added inline comments.


================
Comment at: compiler-rt/lib/fuzzer/FuzzerDriver.cpp:541
     EF->LLVMFuzzerInitialize(argc, argv);
+  DisableInterceptors = new ScopedDisableMsanInterceptorChecks();
   const Vector<std::string> Args(*argv, *argv + *argc);
----------------
errr. this sounds like an overkill. 
If you never destruct this, then just call __msan_scoped_disable_interceptor_checks


================
Comment at: compiler-rt/lib/fuzzer/FuzzerInternal.h:155
 
+struct ScopedEnableMsanInterceptorChecks {
+  ScopedEnableMsanInterceptorChecks() {
----------------
you only ever need ScopedEnable, right? 
never ScopedDisable


================
Comment at: compiler-rt/lib/fuzzer/FuzzerLoop.cpp:182
     return; // Happens when running individual inputs.
-  MD.PrintMutationSequence();
-  Printf("; base unit: %s\n", Sha1ToString(BaseSha1).c_str());
-  size_t UnitSize = CurrentUnitSize;
-  if (UnitSize <= kMaxUnitSizeToPrint) {
-    PrintHexArray(CurrentUnitData, UnitSize, "\n");
-    PrintASCII(CurrentUnitData, UnitSize, "\n");
+  {
+    ScopedDisableMsanInterceptorChecks S;
----------------
Do you need this extra scope here? 


================
Comment at: compiler-rt/lib/fuzzer/FuzzerLoop.cpp:519
+  if (EF->__msan_unpoison)
+    EF->__msan_unpoison(Data, Size);
   // We copy the contents of Unit into a separate heap buffer
----------------
why Data, not DataCopy? 


================
Comment at: compiler-rt/test/fuzzer/msan.test:18
+RUN: %msan_compiler %S/UseAfterDtor.cpp -o %t
+RUN: MSAN_OPTIONS=poison_in_dtor=1 not %run %t -seed=1 -runs=10000000 2>&1 | FileCheck %s --check-prefix=REPORT
+
----------------
also add one test that would break if you accidentally disable and never enable back the interceptor checks  


https://reviews.llvm.org/D48891