On 9/16/14 4:22 PM, Ingo Molnar wrote: > > * Nadav Amit <namit@xxxxxxxxxxxxxxxxx> wrote: > >> The code that deals with x86 cpuid fields is hard to follow since it performs >> many bit operations and does not refer to cpuid field explicitly. To >> eliminate the need of openning a spec whenever dealing with cpuid fields, this >> patch-set introduces structs that reflect the various cpuid functions. >> >> Thanks for reviewing the patch-set. >> >> Nadav Amit (3): >> x86: Adding structs to reflect cpuid fields >> x86: Use new cpuid structs in cpuid functions >> KVM: x86: Using cpuid structs in KVM >> >> arch/x86/include/asm/cpuid_def.h | 163 +++++++++++++++++++++++++++++++++++++++ >> arch/x86/kernel/cpu/common.c | 56 ++++++++------ >> arch/x86/kvm/cpuid.c | 36 +++++---- >> 3 files changed, 219 insertions(+), 36 deletions(-) >> create mode 100644 arch/x86/include/asm/cpuid_def.h > > I personally like bitfields in theory (they provide type clarity > and abstract robustness, compared to open-coded bitmask numeric > literals that are often used in cpuid using code, obfuscating > cpuid usage), with the big caveat that for many years I didn't > like bitfields in practice: older versions of GCC did a really > poor job of optimizing them. > > So such a series would only be acceptable if it's demonstrated > that both 'latest' and 'reasonably old' GCC versions do a good > job in that department, compared to the old open-coded bitmask > ops ... > > Comparing the 'size vmlinux' output of before/after kernels would > probably be a good start in seeing the impact of such a change. > > If those results are positive then this technique could be > propagated to all cpuid using code in arch/x86/, of which > there's plenty. Thanks for the quick response. I was not aware GCC behaves this way. I made some small experiments with GCC-4.8 and GCC-4.4 and in brief my conclusions are: 1. The assembled code of bitmask and bitfields is indeed different. 2. GCC-4.8 and GCC-4.4 behave pretty much the same, yet GCC-4.8 appears to make better instructions reordering. 3. Loading/storing a single bitfield seems to be pretty much optimized (marginal advantage from code size point-of-view for bitmask, same number of instructions). 4. Loading/storing multiple bitfields seems to be somewhat under-optimized - multiple accesses to the original value result in ~30% more instructions and code-size. So you are correct - bitfields are less optimized. Nonetheless, since cpuid data is mostly used during startup, and otherwise a single bitfield is usually accessed in each function - I wonder whether it worth keeping the optimized but "obfuscate" code. Obviously, I can guess your answer to this question... Nadav -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
