diff --git a/source/lib/bits.h b/source/lib/bits.h
index 094fa34496..5ef37b3519 100644
--- a/source/lib/bits.h
+++ b/source/lib/bits.h
@@ -122,14 +122,12 @@ inline T SetBitsTo(T num, size_t lo_idx, size_t hi_idx, size_t value)
 
 
 /**
- * @return number of 1-bits in mask
+ * @return number of 1-bits in mask.
+ * execution time is proportional to number of 1-bits in mask.
  **/
 template<typename T>
-inline size_t PopulationCount(T mask)
+inline size_t SparsePopulationCount(T mask)
 {
-	// note: a more complex but probably faster method is given at
-	// http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
-
 	size_t num1Bits = 0;
 	while(mask)
 	{
@@ -140,6 +138,24 @@ inline size_t PopulationCount(T mask)
 	return num1Bits;
 }
 
+/**
+ * @return number of 1-bits in mask.
+ * execution time is logarithmic in the total number of bits.
+ * supports up to 128-bit integers (if their arithmetic operators are defined).
+ * [http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel]
+ **/
+template<typename T>
+static inline size_t PopulationCount(T x)
+{
+	const T mask = T(~T(0));
+	x -= (x >> 1) & (mask/3);	// count 2 bits
+	x = (x & (mask/15*3)) + ((x >> 2) & (mask/15*3));	// count 4 bits
+	x = (x + (x >> 4)) & (mask/255*15);	// count 8 bits
+	return (x * (mask/255)) >> ((sizeof(T)-1)*CHAR_BIT);
+}
+
+
+
 /**
  * @return whether the given number is a power of two.
  **/
diff --git a/source/lib/debug.cpp b/source/lib/debug.cpp
index 46d1986bdb..8391c053e4 100644
--- a/source/lib/debug.cpp
+++ b/source/lib/debug.cpp
@@ -225,7 +225,7 @@ public:
 	{
 		va_list ap;
 		va_start(ap, fmt);
-		const int len = vswprintf_s(m_pos, m_charsLeft, fmt, ap);
+		const int len = vswprintf(m_pos, m_charsLeft, fmt, ap);
 		va_end(ap);
 		if(len < 0)
 			return false;
diff --git a/source/lib/sysdep/os/win/wposix/waio.cpp b/source/lib/sysdep/os/win/wposix/waio.cpp
index ae93ecde45..d7dc27e97f 100644
--- a/source/lib/sysdep/os/win/wposix/waio.cpp
+++ b/source/lib/sysdep/os/win/wposix/waio.cpp
@@ -455,7 +455,11 @@ Status waio_Preallocate(int fd, off_t size)
 	// allocate all space up front to reduce fragmentation
 	LARGE_INTEGER size64; size64.QuadPart = alignedSize;
 	WARN_IF_FALSE(SetFilePointerEx(hFile, size64, 0, FILE_BEGIN));
-	WARN_IF_FALSE(SetEndOfFile(hFile));
+	if(!SetEndOfFile(hFile))
+	{
+		debug_printf(L"Preallocate(%lld) failed: %d\n", size, GetLastError());
+		return ERR::FAIL;	// NOWARN (probably not enough disk space)
+	}
 
 	// avoid synchronous zero-fill (see discussion in header)
 	if(pSetFileValidData)
diff --git a/source/lib/sysdep/os/win/wvm.cpp b/source/lib/sysdep/os/win/wvm.cpp
index ab97bc902c..c3d0471293 100644
--- a/source/lib/sysdep/os/win/wvm.cpp
+++ b/source/lib/sysdep/os/win/wvm.cpp
@@ -54,7 +54,7 @@ static WUTIL_FUNC(pVirtualAllocExNuma, LPVOID, (HANDLE, LPVOID, SIZE_T, DWORD, D
 static DWORD WINAPI EmulateGetCurrentProcessorNumber(VOID)
 {
 	const u8 apicId = x86_x64_ApicId();
-	const DWORD processor = ProcessorFromApicId(apicId);
+	const DWORD processor = (DWORD)ProcessorFromApicId(apicId);
 	ASSERT(processor < os_cpu_MaxProcessors);
 	return processor;
 }