> On Sep 3, 2020, at 12:48 PM, Ramana Radhakrishnan <ramana....@googlemail.com>
> wrote:
>
> On Thu, Sep 3, 2020 at 6:13 PM Kees Cook via Gcc-patches
> <gcc-patches@gcc.gnu.org> wrote:
>>
>> On Thu, Sep 03, 2020 at 09:29:54AM -0500, Qing Zhao wrote:
>>> On average, all the options starting with “used_…” (i.e, only the
>>> registers that are used in the routine will be zeroed) have very low
>>> runtime overheads, at most 1.72% for integer benchmarks, and 1.17% for FP
>>> benchmarks.
>>> If all the registers will be zeroed, the runtime overhead is bigger,
>>> all_arg is 5.7%, all_gpr is 3.5%, and all is 17.56% for integer benchmarks
>>> on average.
>>> Looks like the overhead of zeroing vector registers is much bigger.
>>>
>>> For ROP mitigation, -fzero-call-used-regs=used-gpr-arg should be enough,
>>> the runtime overhead with this is very small.
>>
>> That looks great; thanks for doing those tests!
>>
>> (And it seems like these benchmarks are kind of a "worst case" scenario
>> with regard to performance, yes? As in it's mostly tight call loops?)
>
>
> That's true of some of them but definitely not all - the GCC benchmark
> springs to mind in SPEC as having quite a flat profile, so I'd take a
> look there and probe a bit more in that one to see what happens. Don't
> ask me what else , that's all I have in my cache this evening :)
>
> I'd also query the "average" slowdown metric in those numbers as
> something that's being measured in a different way here. IIRC the SPEC
> scores for int and FP are computed with a geometric mean of the
> individual ratios of each of the benchmark. Thus I don't think the
> average of the slowdowns is enough to talk about slowdowns for the
> benchmark suite. A quick calculation of the arithmetic mean of column
> B in my head suggests that it's the arithmetic mean of all the
> slowdowns ?
>
> i.e. Slowdown (Geometric Mean (x, y, z, ....)) != Arithmetic mean (
> Slowdown (x), Slowdown (y) .....)
>
> So another metric to look at would be to look at the Slowdown of your
> estimated (probably non-reportable) SPEC scores as well to get a more
> "spec like" metric.
Please take a look at the new csv file at:
https://gitlab.com/x86-gcc/gcc/-/wikis/Zero-call-used-registers-data
<https://gitlab.com/x86-gcc/gcc/-/wikis/Zero-call-used-registers-data>
I just uploaded the slowdown data computed based on
Est.SPECrate(R)2017_int_base and Est.SPECrate(R)2017_fp_base. All data are
computed against “no”.
Compare this slowdown data to the one I computed previously as (Arithmetic mean
(Slowdown(x), slowdown(y)…), the numbers do change a little bit, however, the
basic information provided from the data keeps the same as before.
Let me know if you have further comments.
thanks.
Qing
>
> regards
> Ramana
>>
>> --
>> Kees Cook
