[MLton] Problem with MinGW

Matthew Fluet fluet@cs.cornell.edu
Thu, 8 Jun 2006 18:22:35 -0400 (EDT)

>> The codegens are very different outside the Basis Library.
> I know this, of course. I just thought that as a windows system call was 
> returning a different error code, that the methods called before it must be 
> different. I presumed that the codegen doesn't particularly influence the 
> runtime methods invoked.

Yes, good point.  And you are correct that the dynamic behavior of which 
runtime methods are invoked should be pretty much the same in the 

>>> The parameters passed to the open3 method are identical in both cases. The 
>>> only difference is the stack. The first argument is 00 aligned for the C 
>>> codegen, and 1a aligned for the native codgen. However, I think this 
>>> should be ok... right?
>> Maybe yes, maybe no.  Could MinGW have a different calling convention than 
>> x86-linux and x86-cygwin?  In particular, does MinGW demand different stack 
>> alignment.  I could see that causing a problem very much like that which we 
>> are seeing.
> This is indeed the problem!
> MinGW functions seem perfectly happy with a stack aligned on 2 bytes. 
> However, some windows system calls (not all) fail when called this way. The 
> MinGW code also appears to preserve 4 byte alignment, so if you call a MinGW 
> compiled method (Posix_FileSys_open3) with offset 2, 6, a, or c, then the 
> windows system call gets it too.

Good.  It also explains why MinGW was working before with the native 
codegen but breaks on the x86_64 branch.  On the x86_64 branch, we are 
handling many more C-side types (like mode_t) in their native C 
representation.  On most x86 platforms, mode_t is 32bits, but on MinGW, it 
is 16bits.

On the other hand, I was a bit curious how we got off of 4 byte alignment. 
It turns out it is a bug in the native codegen's implementation of the C 
calling convention.  It promotes an 8bit integral value to 32bits before 
pushing on the stack, but it pushes a 16bit value without widening it. 
It should be widening all arguments to multiples of 32bits.  Previous to 
the revisions on the x86_64 branch, the C-side implementation of the Basis 
Library was predominantly using 32bit arguments.

I'll check in a fix.