According to the spec it is 32-bit unsigned value for a PE32+ image just like a PE32 image.
However, in my testing with both 32-bit and 64-bit applications (PE32/PE32+ files) on Windows 7 SP1 Home Premium x64, the maximum file size for either is between 1.8-1.85GB.
I tested by creating a very basic C executable with Visual Studio (~8K for 32-bit and 9K for 64-bit) and the added an empty code section to the PE header until Windows would no longer load it, and then binary searched for the limit. Looking at the process with vmmap showed that almost all of the entire first 2GB of address space were the image (including any subsequently loaded DLLs such as kernel32.dll). The limit was the same for me with both 32 and 64-bit processes. The 64-bit process did have the flag set in it's NT Header's File Header's section stating that it could handle addresses >2GB. It also could allocate memory for non-image sections above the 2GB limit.
It seems like the image is required to fit in it's entirety in the lower 2GB of VA space for the process, which means the SizeOfImage is being treated a signed 32-bit integer by the loader effectively.
与恶龙缠斗过久,自身亦成为恶龙;凝视深渊过久,深渊将回以凝视…