[CODEGEN] Add support for extern prologue/epilogue functions

[KireinaHoro]

This commmit adds support for adding prologue and/or epilogue function calls for a code block via the pragma schedule. A dummy use can be found here: https://gist.github.com/KireinaHoro/c959170557d75d81c52c116c3750d78a

The addition is proposed to support better control in the generated code when using the tensorize and inline asm flows to do codegen for something like Gemmini, which requires explicit fence instructions to block the execution flow until data has been fully flushed back to DRAM. With this change an external function that does the fence can be placed in the epilogue of a computation. The prologue pragma may be useful in similar situations when a device needs some sort of initialization, and is added for symmetry.

[KireinaHoro]

Newcomer here, not sure who to @ for review...

[tqchen]

Thanks for the contribution. It would be great if we can add an testcase to cover this usecase. In terms of implementations, such intrinsic is better to be lowered before the codegen phase, since they can be lowered to Call blocks, so that we can make use of most codegens(e.g. CodeGenC will work as well).

Given that this is a new feature, it would be useful to open a RFC thread on the https://discuss.tvm.ai/ to let the folks know about it is coming and discuss the API choices. For example I can see quite a few alternatives:

• Since this is more like a call_at_epilogue and call_at_prologue.
• It can be viewed as a context enter exit calls for a context

[tqchen]

ping @KireinaHoro :)

[KireinaHoro]

@tqchen I'd love to further work on this, but I'm currently blocked by [this discussion thread](https://discuss.tvm.ai/t/tensorize-with-stride-for-input-tensors/6018) and cannot perform further testing for the use case where all this idea started. Would you mind taking a look at that please? Thanks a lot.

[KireinaHoro]

I guess I'd create an RFC on the forum first and see what the community says. Thanks for the attention!

[KireinaHoro]

An RFC thread has been created here.

[tqchen]

Closed due to inactive status (to make sure every PR get visited instead of being forgotten). The RFC and proposal still makes sense. Thank you @KireinaHoro for bringing this up intially, and a separate PR is more than welcomed.