You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Investigate emitting movq for the CreateScalarUnsafe helper intrinsics that take a long/ulong on x86 #11626
For the CreateScalarUnsafe helper intrinsics that take a long/ulong, it may be beneficial to emit the movq xmm, [mem] instruction on x86, rather than falling back to emulating the behavior in software.
Given that the input is a regular long/ulong, but on x86 the instruction requires the second operand come from memory, this may require some additional work to make happen or we may find it not worthwhile to do at all.
Activity
tannergooding commentedon Dec 6, 2018
CC. @CarolEidt, @fiigii
As per https://github.com/dotnet/coreclr/pull/21351/files/ff31bc6a186a15cd9395928c546e82de077f7246#r239645439
CarolEidt commentedon Dec 6, 2018
@tannergooding - thanks for capturing this.
anthonycanino commentedon Sep 12, 2022
@tannergooding I think this is no longer applicable.
On SharpLab
gives
saucecontrol commentedon Sep 12, 2022
that's x64 code.
here's x86
anthonycanino commentedon Sep 12, 2022
Got it --- thanks for that. My mistake.