Created attachment 24026 [details] A code snippet that reproduces the error g++ tries to save an xmm register using MOVAPD (SSE instruction) to the stack when switching between working chunks during openMP multithreading. This obviously causes a random segmentation fault when the stack pointer happens not be aligned to a 16-byte boundary. Please see the attached code. I compile it with g++ 4.5.2 (I'm using MinGW) and flags -O3 -msse3 -fopenmp. It is important that the optimization is on and the compiler tries to save the xmm register containing the constant zero between working chunks. This is the instruction that causes the segmentation fault if ebp-0x48 is not divisible by 0x10: 0040143a: movapd %xmm1,-0x48(%ebp) // right here 0040143f: call 0x4014bc <GOMP_loop_dynamic_next> PS. I have to admit that I'm completely new to GCC, openMP or SSE (I just learned about openMP today and I've playing with them for a couple hours only), so I might be just doing something really stupid.
On linux we assume a 16 byte aligned stack for threads, so this might hint at some MinGW specific issue either in our OMP runtime or with windows thread handling.
Kai, is this reproducable with current releases, and can you confirm Richard's suggestion about threads not having 16-byte aligned stacks?
I can confirm this bug for gcc 4.7.2 mingw64. The -mstackrealign command-line flag can be used as workaround as described on http://stackoverflow.com/questions/6716654/segmentation-fault-using-openmp-and-sse until this bug is properly patched.