As a workaround I tried to use inline assembly for loads and stores: #define asm_lvx( vec, idx, ptr ) \ asm ( "lvx %0,%1,%2" \ : "=v" (vec) \ : "r" (idx), "r" (ptr) ) But there's a problem, in lvx/stvx '0' in second term acts as literal 0, not as register number. Shouldn't gcc be able to understand those instructions and not use r0 ? Constraint letter "r" means any GPR. If you are going to create the address manually, you explicitly need to tell GCC only to use a base register by using constraint letter "b". The extra stack manipulation is a known problem. For the memcpy look, open a Bugzilla. David
