Dmitry Babokin
Dmitry Babokin
I'm looking at 32 bit support in other Linux distributions and see that the last `Ubuntu` to support 32 bit is `18.04`, the last `CentOS` to support 32 bit is...
Using generic LLVM for ISPC build "mostly works", meaning: - we don't have control over patches, meaning we don't know if all the issues we care about are fixed. For...
Using "stock" LLVM build is BAD idea. It misses a number of important patches. The recommended way to build ISPC is to build LLVM as part of ISPC build process,...
It's a matter of implicit or explicit decision making. IMHO it's better to specify explicitly that something is off, rather than rely on implicit config. For targets like WASM or...
This needs to be reported to SDE team. It looks like a bug in SDE.
Looks like we need to disable these targets. We should either disable them in CI, or do the check in `alloy.py`, which disables the run if it's 1) AMD paltform,...
So, basically it need to be a `memcpy` and we need to recognize it when for the types data to avoid unpacking/packing.
Could you provide more complete example? I can't reproduce it - that's how I'm trying it: [here](https://godbolt.org/z/8nx6s3dza) I suspect that compiler is not wrong, i.e. with 4 wide target all...
Thanks for the reproducer! > I agree that the statement is technically correct but is it really undefined behaviour to have a single program instance writing to a single location?...
It's expected, at least with `--math-lib=fast`. Fast implementation of libs do differ on different ISAs. It's very difficult / impossible to avoid not sacrificing performance.