I wee need to improve spilling code for vectors. Look on the code generated for TestByteDoubleVect.test() form regression test (from 7178280 bug report). Initializing vectors (Replicate nodes) were pushed on stack because there are calls between those values usages (xmm registers are SOC). The problem is that values are loaded from stack inside a loop which causes performance problem. Find a way to load values back from stack after call or before loop where they are used.