The resolve stubs save off the FPU state before transitioning out the runtime and there's a line of code which uses finit to reset the FPU before transitioning out. This is ok if the VM can never execute an Java code inside the runtime call. For C2 this is true but C1 can use execute this code at an unresolved call site which can cause initialization. When this happens the Java code executed with the wrong FPU control word and bad things can happen. This is reproducible by simply using java -client -Xcomp -XX:+VerifyFPU foo on a fastdebug build.