Sorry, my first look at this was incorrect. Changing the pointer to a
scalar does re-enable the optimization. But so does making it less than 4
words in size.
The optimization fails because the loop body gets translated to { tmp :=
Foo{}; testFoo[i] = tmp } by the order pass
(cmd/compile/internal/gc/order.go:417-423), and that confuses the later
optimization. It only does this transformation for large (>4 word) and
write-barrier-requiring types.
In any case, this is fixed at tip. It was fixed as part of redoing the
write barrier logic.
On Sunday, April 16, 2017 at 4:13:23 AM UTC-7, T L wrote:
>
>
>
> On Saturday, April 15, 2017 at 7:18:51 PM UTC+8, Konstantin Khomoutov
> wrote:
>>
>> On Sat, 15 Apr 2017 01:31:27 -0700 (PDT)
>> "'Keith Randall' via golang-nuts" <golan...@googlegroups.com> wrote:
>>
>> > > I've been looking into why some of our code wasn't resulting in the
>> > > memclr optimization originally introduced in response to
>> > > https://github.com/golang/go/issues/5373.
>> [...]
>> > > type Foo struct {
>> > > H string
>> > > A int64
>> > > B int64
>> > > C int64
>> > > }
>> [...]
>> > > For me (go 1.8) the above snippet results in asm that doesn't make
>> > > use of memclr. On the other hand, when I comment out field C in the
>> > > struct above then the code does make use of the optimization.
>> > > Similarly, when I change the first field from string to int64, it
>> > > also works.
>> [...]
>> > The value being zeroed must not have any pointers. This is a change
>> > from 1.7 and is part of the requirement for the hybrid write barrier
>> > <https://gist.github.com/aclements/4b5e2758310032dbdb030d7648b5ab32>
>> > (part of making garbage collection pauses smaller).
>>
>> Sorry, I might be missing the point completely but while this perfectly
>> explains using of the memclr optimization after removal of H but how
>> can it explain also enabling of memclr after removal of C which is not
>> a pointer?
>>
>
> Looks if H is present, it will optimized as memclrHasPointers, otherwise,
> it will optimized as memclrNoHeapPointers.
> But it is really weird that not optimization made if C is present.
>
>
>
--
You received this message because you are subscribed to the Google Groups
"golang-nuts" group.
To unsubscribe from this group and stop receiving emails from it, send an email
to golang-nuts+unsubscr...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.
