[compiler-rt][asan|win] Fix flaky unittest due to large allocations

[etienneb]

Summary

Coverage is using large arrays which requires large allocations.
These allocations are flaky and often failing on win64.

We are using the 32-bits size until this gets a better fix.

[rnk]

lib/sanitizer_common/sanitizer_coverage_libcdep.cc

174–178

I don't think we need the #if:

static const uptr kTrEventArrayMaxSize =
    (SANITIZER_WORDSIZE == 32 || SANITIZER_WINDOWS) ? 1 << 22 : 1 << 30;

175–177

This comment should document what makes Windows special, and not just "the unittests fail unless we do this". Windows doesn't do overcommit (committed virtual memory costs swap), so programs can't reliably map such large amounts of virtual memory.

[etienneb]

lib/sanitizer_common/sanitizer_coverage_libcdep.cc

174–178

It won't affect linux 64-bit ?

174–178

ok, I saw your snipper.
fine.

[etienneb]

rnk comments.

[rnk]

lgtm

lib/sanitizer_common/sanitizer_coverage_libcdep.cc

private:
 
   // Tracing event array, size and current pointer.
   // We record all events (basic block entries) in a global buffer of u32
   // values. Each such value is the index in pc_array.
   // So far the tracing is highly experimental:
   //   - not thread-safe;
   //   - does not support long traces;
   //   - not tuned for performance.
-  static const uptr kTrEventArrayMaxSize = FIRST_32_SECOND_64(1 << 22, 1 << 30);
+  // Windows doesn't do overcommit (committed virtual memory costs swap), so
+  // programs can't reliably map such large amounts of virtual memory.
+  // TODO(etienneb): Find a way to support coverage of larger executable
+static const uptr kTrEventArrayMaxSize =
+    (SANITIZER_WORDSIZE == 32 || SANITIZER_WINDOWS) ? 1 << 22 : 1 << 30;
   u32 *tr_event_array;
   uptr tr_event_array_size;
   u32 *tr_event_pointer;
   static const uptr kTrPcArrayMaxSize    = FIRST_32_SECOND_64(1 << 22, 1 << 27);
 
   StaticSpinMutex mu;
 };