> On Mar 12, 2020, at 12:36 PM, Laszlo Ersek <ler...@redhat.com> wrote:
>
> On 03/12/20 15:03, Leif Lindholm wrote:
>> +Ard, Laszlo.
>>
>> I think it would make sense to move it to MdeModulePkg (or MdePkg) and
>> rename it BaseCompilerIntrinsicsLib (it *is* a BASE library).
>>
>> As I alluded in my reply to Ray - x86 also have this problem, but to a
>> lesser extent, and ended up creating library functions to call instead
>> of using plain C for certain operations.
>> Which was probably the right decision when it was restricted to a
>> very few corner cases.
>
> I think people that are interested in IA32/X64 are happier with explicit
> CopyMem() calls that are optimized (via one of the several BaseMemoryLib
> instances, such as SSE2, REP + string instructions, MMX, "smart"
> (chunked) C code, etc), than with a naive loop, such as the one in
> "ArmPkg/Library/CompilerIntrinsicsLib/memcpy.c", that gets silently
> plugged into an intrinsic (such as a structure assignment).
>
> I mean, compiler intrinsics exist in the first place because they
> implement language features in a fast / performant manner, behind the
> scenes. If we replace the internals of an intrinsic with a slow / naive
> implementation, then the intrinsic has no more right to exist, the
> compiler could / should just generate the normal naive code.
>
> I don't mind the code movement, but I'd like to avoid using
> BaseCompilerIntrinsicsLib on IA32/X64. On those arches, I think it would
> only be an improvement if it had a configurable backend, similarly how
> CopyMem() is currently configurable.
>
> ... I guess I've gotten very used to calling CopyMem(), in place of
> structure assignment.
>
Laszlo,
My brain has flipped too.
For x86 I find smaller structure assignments can cause the optimizer to
optimize away the memcpy/memset and you only see issue issues on a NOOPT build
since DEBUG builds tend to be size optimized. I tend to hit this issue when I
try to turn off optimizations to try and walk vendor code in the debugger. So
code review is the 1st line of defense.
Thanks,
Andrew Fish
-=-=-=-=-=-=-=-=-=-=-=-
Groups.io Links: You receive all messages sent to this group.
View/Reply Online (#55830): https://edk2.groups.io/g/devel/message/55830
Mute This Topic: https://groups.io/mt/71671270/21656
Group Owner: devel+ow...@edk2.groups.io
Unsubscribe: https://edk2.groups.io/g/devel/unsub [arch...@mail-archive.com]
-=-=-=-=-=-=-=-=-=-=-=-
