[flang][runtime] Speed up initialization & destruction #148087
Conversation
klausler
force-pushed
the
wq-speedup
branch
2 times, most recently
from
7597a02 to
b9be690
Compare
There was a problem hiding this comment.
| // address all elements. It genernalizes contiguity by also allowing | |
| // address all elements. It generalizes contiguity by also allowing |
There was a problem hiding this comment.
Thanks. I no longer always see this sort of thing.
There was a problem hiding this comment.
Should we mark such methods with force-inline attribute. I am not suggesting doing it in this PR, but I wonder if you tried it.
There was a problem hiding this comment.
I didn't try it, and I've run into other trouble because something somewhere has a link-time unresolved external reference now to Descriptor::Elements(). Updating soon to leave Elements out-of-line with a new inline InlineElements() that it can call.
There was a problem hiding this comment.
I have no clue how this change could have caused a linking issue..
There was a problem hiding this comment.
I couldn't track it down either. The unresolved reference is not within flang_rt.runtime or the generated code for the failing test.
flang-rt/lib/runtime/derived.cpp
Outdated
There was a problem hiding this comment.
As the chunk size grows, I guess, we can hit cache conflicts depending on the aliasing implementation. Do you know if there is any non-temporal memcpy implementation that we can try?
There was a problem hiding this comment.
Maybe I should use
char *to{rawInstance + *stride};
char *from{rawInstance};
for (std::size_t bytes{...}; bytes--; ) {
*to++ = *from++;
}
There was a problem hiding this comment.
Not sure about that. It will probably be the same as memcpy if the compiler vectorizes it. I think memcpy may do better with rep movs on x86. Let's keep it simple with memcpy. If the cache issue pops up anywhere I will experiment with https://clang.llvm.org/docs/LanguageExtensions.html#non-temporal-load-store-builtins
There was a problem hiding this comment.
memcpy is not defined for overlapping regions, unfortunately.
Rework derived type initialization in the runtime to just initialize the first element of any array, and then memcpy it to the others, rather than exercising the per-component paths for each element. Reword derived type destruction in the runtime to detect and exploit a fast path for allocatable components whose types themselves don't need nested destruction. Small tweaks were made in hot paths exposed by profiling in descriptor operations and derived type assignment.
2 participants
Rework derived type initialization in the runtime to just initialize the first element of any array, and then memcpy it to the others, rather than exercising the per-component paths for each element.
Reword derived type destruction in the runtime to detect and exploit a fast path for allocatable components whose types themselves don't need nested destruction.
Small tweaks were made in hot paths exposed by profiling in descriptor operations and derived type assignment.