In gcc 4.1.2, vec_ld() does not work correctly on board of CPU MPC74XX.
float temp[4];
__vector float Src;
Src = (__vector float)vec_ld(0, temp);
However, if float variable is aligned to 16 bytes, it works correctly:
float temp[4] __attribute__((aligned(16)));
Is this by design?
Yes, AltiVec loads and stores require 16 byte alignment. This is very well documented in the AltiVec manuals.
Unlike other SIMD architectures such as SSE however, note that AltiVec silently truncates unaligned addresses to the next lowest 16 byte boundary, rather than generating an exception, so your code will not crash, but it will not behave correctly if you attempt to load or store at an unaligned address.
In cases where you can not avoid unaligned loads you can load two adjacent aligned vectors and then use vec_lvsl + vec_perm to create the required vector:
float temp[4];
__vector float sr1, src2, src;
src1 = vec_ld(0, temp);
src2 = vec_ld(16, temp);
src = vec_perm(src1, src2, vec_lvsl(0, temp));
By the way, in Power8 they finally added support for unaligned load/store vector access. For details, see information on lxvd2x / lxvw4x and stxvd2x / stxvw4x instructions in section "7.6 VSX Instruction Set" of Power ISA 2.07 document.
Those who have access to IBM XL C/C++ Compiler, could use vec_xld2() / vec_xlw4() and vec_xstd2() / vec_xstw4() intrinsics.
As of version "g++ (GCC) 4.10.0 20140419 (experimental)", I am not aware of GCC equivalents, but I believe, users of GCC could access unaligned memory by pointer dereferencing:
signed int *data;
// ...
vector signed int r = *(vector signed int *)&(data[i]);
