On Thu, May 26, 2016 at 12:43:44PM -0700, David Miller wrote: > From: Catalin Marinas <catalin.marinas@xxxxxxx> > Date: Thu, 26 May 2016 15:20:58 +0100 > > > We can solve (a) by adding more __SC_WRAP annotations in the generic > > unistd.h. > ... > > I really think it's much more robust to clear the tops of the registers > by default. Then you won't be auditing constantly and adding more and > more wrappers. I think we could avoid adding a new __SC_WRAP by redefining __SYSCALL for ILP32 to always invoke a wrapper. But given the wrapper overhead, cache locality, I don't think we would notice any performance difference in either case. > You can't even quantify the performance gains for me in any precise > way. Whatever you gain by avoiding the 64-bit > decompostion/reconstitution for those few system calls with 64-bit > registers, you are losing by calling the wrappers for more common > system calls, more often. I hope Yury can provide some numbers. All being equal, I would go for the lowest code maintenance cost (which is probably less annotations and wrappers). > "it's more natural to pass 64-bit values in a register" is not a clear > justification for this change. It's more related to how we went about the ILP32 ABI. We initially asked for a 64-bit native ABI similar to x32 until the libc-alpha community raised some POSIX compliance issues on time structures. So we decided to go for a 32-bit-like ABI while keeping the syscall interface close to the AArch64/ILP32 procedure calling standard (64-bit values passed in a single register). And now we have this discussion, revisiting this decision again (which is perfectly fine, we better get it right before any merging plans; thanks for your input). -- Catalin -- To unsubscribe from this list: send the line "unsubscribe linux-arch" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
