partial specialization doesn't preserve attributes on call

[nlewycky]

Bugzilla Link	3757
Resolution	WONTFIX
Resolved on	Jan 15, 2011 18:27
Version	trunk
OS	Linux
CC	@lattner

Extended Description

The partial specialization pass doesn't preserve attributes on the call or invoke instruction. That means things like tail call, but also all the readnone/readonly noalias, sret, etc. This could cause a miscompilation for anyone using partial specialization.

There's logic to copy attributes properly inside of DeadArgumentElimination.cpp RemoveDeadStuffFromFunction(F) but it looks non-trivial to refactor.

[lattner]

Someone should either fix this or remove partial specialization.

[llvmbot]

Not sure why I'm CC"d on this, but I took a look anyways. Here's a patch that should solve the problems described in this bug, but it doesn't seem to improve PartialSpecialization's overall failure at MultiSource.

[llvmbot]

Proposed Fix

[lattner]

Any update on this? Should the partial specialization pass just be removed?

[llvmbot]

I think so. Its value is dubious at best. To be valuable it would need to either tie into alias analysis and only specialize functions where the argument specialized improved alias results significantly or where the specialized argument allows removal of conditionals.

[llvmbot]

I think so. Its value is dubious at best. To be valuable it would need to
either tie into alias analysis and only specialize functions where the argument
specialized improved alias results significantly or where the specialized
argument allows removal of conditionals.

Or where it allows a call through a function pointer to be turned into a direct call. I think that all by itself justifies the pass's existence.

[lattner]

Over a year later, we should still either fix this or remove partial specialization.

[lattner]

Ah, I see that Andrew is the author and that his comment #​5 indicates that he's not attached to the pass. Andrew, is it ok if I zap it from mainline?

[llvmbot]

Ah, I see that Andrew is the author and that his comment #​5 indicates that he's
not attached to the pass. Andrew, is it ok if I zap it from mainline?

Zap it. There was a discussion on the llvm-dev a few months ago which seemed to indicate some people were using/improving it, but if not it should go.

[lattner]

Thanks Andrew, removed in r123554.