karolherbst wrote:
> > But I also wasn't aware how short lived releases branches are as I just
> > realized that normally no more releases are planned, is that right?
>
> I'm not familiar with llvm release cadence. 22.x branch is probably not
> frozen yet because 23 is not released yet. So critical bug fix should be ok
> to land in 22.x branch.
Yeah, not sure if this is critical enough to make yet another release, but not
my call to make. But I've seen other PRs against 22.x that were closed
already, because of "we don't do another 22.x release".
But it does break certain builtins like `cos`, because there is no definition
for `__clc_flush_denormal_if_not_supported` in the final SPIR-V.
It also feels like the bug is actually something else, because it also looks
like a codegen or tooling issue, though this PR is rather trivial and does fix
the issue.
Like in the SPIR-V (without this PR) I'm seeing this:
```
%_Z37__clc_flush_denormal_if_not_supportedf = OpFunction %float None %4639
%4641 = OpFunctionParameter %float
OpFunctionEnd
```
And then multiple calls to it across the entire libclc:
```
%113249 = OpFunctionCall %float
%_Z37__clc_flush_denormal_if_not_supportedf %113248
%113251 = OpFunctionCall %float
%_Z37__clc_flush_denormal_if_not_supportedf %113250
%113253 = OpFunctionCall %float
%_Z37__clc_flush_denormal_if_not_supportedf %113252
%198384 = OpFunctionCall %float
%_Z37__clc_flush_denormal_if_not_supportedf %198383
%198386 = OpFunctionCall %float
%_Z37__clc_flush_denormal_if_not_supportedf %198385
%199563 = OpFunctionCall %float
%_Z37__clc_flush_denormal_if_not_supportedf %199562
%199565 = OpFunctionCall %float
%_Z37__clc_flush_denormal_if_not_supportedf %199564
%200742 = OpFunctionCall %float
%_Z37__clc_flush_denormal_if_not_supportedf %200741
%200744 = OpFunctionCall %float
%_Z37__clc_flush_denormal_if_not_supportedf %200743
%201921 = OpFunctionCall %float
%_Z37__clc_flush_denormal_if_not_supportedf %201920
%201923 = OpFunctionCall %float
%_Z37__clc_flush_denormal_if_not_supportedf %201922
```
And I'd have expected that this would already break at linking time or
something.
Or I guess those are just normal C rules, because no compilation unit actually
instantiated the definition of `__clc_flush_denormal_if_not_supported`. But
then again, I think it would be better if at build time it would already made
sure that all symbols are resolved so we won't run into this issue in the
future?
https://github.com/llvm/llvm-project/pull/207264
_______________________________________________
llvm-branch-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-branch-commits
