According to 64-bit Windows calling convention, the return value is in %rax instead of %xmm0 in cdot kernel.
In zdot, the caller allocates a memory space for return value and sets this memory address to the first hidden parameter. Thus, the callee (zdot) should assign the result to this memory space and return the memory address in %rax.

It makes sure the return value as same as blas_cpu_numbers which is an internal global variable to store the number of threads in OpenBLAS.

When OpenBLAS is compiled with USE_OPENMP=1, it ignores OPENBLAS_NUM_THREADS and GOTO_NUM_THREADS flags.Therefore, you should use OMP_NUM_THREADS.

Without setting OMP_NUM_THREADS, a process will use maximal number of threads on a computing node. Thus, if there are 2 processes on the computing node, the thread will contend against other threads on CPU cores. As a result, the application will hang.

Used $(CURDIR) instead of pwd in generating shared library on Mac OSX.
Add more tips about this issue. Thank Dr Kane O'Donnell.

OSX cann't use relative path in shared library. Thank Mr.Kane for this patch. The detail is in this link (https://github.com/xianyi/OpenBLAS/issues/57).

In i386 calling convention, the caller put the address of return value of zdot into the first hidden parameter.
Thus, the callee should delete this address before return.
Actually, I have fixed the same bug on x86/zdot_sse2.S (issue #32). However, that is not a good implementation which uses 3 instructions. Mr. John told me used "ret $0x4" to skip the first hidden address (4 bytes).