mingw32 misaligns SSE code (leading to crash of compiled programs)

[stevenj]

Binary package hint: mingw32

Package: mingw32 version 4.2.1.dfsg-1ubuntu1
Debian Release: lenny/sid
  APT prefers hardy-updates
  APT policy: (500, 'hardy-updates'), (500, 'hardy-security'), (500, 'hardy')
Architecture: amd64 (x86_64)

Hi, when using this version of mingw32 to cross-compile FFTW for Win32,
we noticed that it miscompiles SSE code by generating misaligned data;
this causes the compiled code to crash when run. mingw32 in Debian
stable works fine. We've boiled it down to a simple test C file:

#include <xmmintrin.h>
union uvec {
  unsigned int x[4];
  __m128 v;
};
union uvec foo = { 1, 2, 3, 4};

Compile with:
 i586-mingw32msvc-gcc -msse -c --save-temps bug.c
and look at the resulting bug.s assembler output. It is:

 .file "bug.c"
.globl _foo
 .data
 .align 8
_foo:
 .long 1
 .long 2
 .long 3
 .long 4

Note the ".align 8" -- SSE code like this __m128 variable must be
16-byte aligned, so it should be ".align 16".

In contrast, i586-mingw32msvc-gcc 3.4.5 in the Debian stable version of
mingw correctly produces ".align 16". Similarly, Ubuntu's native gcc
4.2.4 correctly produces ".align 16". Hopefully, there is a newer
upstream version of mingw32 that fixes this.