BukeBeyond wrote:
We have fast AVX256 scanners that detect data change and generate a hash in a
united pass. When this code is inlined, it further optimizes to just a few
instructions with constant size structures.
There is a lot of optimization potential for these flat data structures. We
can we use vector instructions to process them in large chunks. However,
padding bytes with random data is a problem. Detecting padding locations,
especially with multiple inheritance is not trivial. __builtin_clear_padding()
solves this problem.
Now we can simply mark a flat structure with;
```
struct F1 : Flat<F1>
{
int i;
bool b;
};
```
where Flat<> simply calls __builtin_clear_padding((F1*) this); at construction.
LLVM unites any redundant writes from further inheritance. LLVM also combines
clear padding and initializated field bytes into larger writes.
These flat objects can be detected with a concept, and now have special powers,
like fast comparison and/or hashing, implemented generically outside of these
objects.
This is made possible by this work here.
Thank you!
https://github.com/llvm/llvm-project/pull/75371
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
