[Gemmini] Refactor lower-gemmini pass

[qzylalala]

1. We spilt lower-gemmini pass to two individual passes lower-gemmini and lower-gemmini-to-llvm.
2. The new lower-gemmini pass will lower gemmini operations to 3rd library(libgemmini) function call.
3. We also remain the code for old lower-gemmini pass, you can choose them using populate function in midend/include/Dialect/Gemmini/Transform.h.

• populateGemminiLegalizeForLLVMExportPatterns() for the old one.
• populateGemminiLegalizeForFuncExportPatterns() for the new one.

[linuxlonelyeagle]

@qzylalala Why does libgemmini exist? What is the benefit of this? What is the benefit of using assembly? If assembly is generated like this, then why write a compiler?The gemmini intrinsic op is at the same level as the op in the llvm dialect.So splitting into two passes I think is very strange.

[qzylalala]

@qzylalala Why does libgemmini exist? What is the benefit of this? What is the benefit of using assembly? If assembly is generated like this, then why write a compiler?The gemmini intrinsic op is at the same level as the op in the llvm dialect.So splitting into two passes I think is very strange.

I agree with your opinion. However, I need to clarify the current new requirement. Hongbin hopes that we can try to use the method of calling third-party libraries to complete the compilation（Another team need this feature）. Therefore, before transforming the gemmini operator into intrinsic op, I replace this process with a pass that calls third-party functions.

[linuxlonelyeagle]

@qzylalala I recommend you rewrite a pass.

[linuxlonelyeagle]

@qzylalala I recommend you rewrite a pass.

Rewrite a pass for your situation.

[linuxlonelyeagle]

The most important point is that no changes should be made to the existing lower-gemmini.My suggestion, either remove the backend and delete the existing lower-gemmini, or don't change the current lower-gemmini and rewrite the pass as needed.

[linuxlonelyeagle]

@zhanghb97 @qzylalala I think we need to have a discussion here. To give you my thoughts, if we do this as we do now, I think we no longer need to use the existing backend, and of course in doing so I don't think there is any need for the lower-gemmini pass to exist any more, which I think is actually acceptable because the backend does cause a lot of problems. If the backend still exists, then no changes should be made to the existing lower-gemmini, and in this PR, the matching pattern of lower-gemmini is changed.Outside of the first time, there are a lot of problems with this PR, compilation failure which is a problem, no endline in the code, and a lot of @qzylalala local computer with some settings.I think this PR needs to be stopped first, and there should be a discussion here before we can move on to the next step.

[qzylalala]

The most important point is that no changes should be made to the existing lower-gemmini.My suggestion, either remove the backend and delete the existing lower-gemmini, or don't change the current lower-gemmini and rewrite the pass as needed.

I also tend to not modify the original code, but to rewrite a new pass. In this way, the makefile does not need to be modified, and testing can also pass. After completing the functional testing, I will rearrange the project structure.