[NewPM][AMDGPU] Port amdgpu-simplifylib/amdgpu-usenative
authorArthur Eubanks <aeubanks@google.com>
Mon, 28 Dec 2020 17:27:28 +0000 (09:27 -0800)
committerArthur Eubanks <aeubanks@google.com>
Mon, 28 Dec 2020 18:38:51 +0000 (10:38 -0800)
commit9abc457724bd54014328a6f0b7ed230bacd9f610
tree56eaecc68e0602eb35db39079874c49858a0ee7b
parent7b00e9fae3853d4693e608cc52f6d6da5059f5ff
[NewPM][AMDGPU] Port amdgpu-simplifylib/amdgpu-usenative

And add them to the pipeline via
AMDGPUTargetMachine::registerPassBuilderCallbacks(), which mirrors
AMDGPUTargetMachine::adjustPassManager().

These passes can't be unconditionally added to PassRegistry.def since
they are only present when the AMDGPU backend is enabled. And there are
no target-specific headers in llvm/include, so parsing these pass names
must occur somewhere in the AMDGPU directory. I decided the best place
was inside the TargetMachine, since the PassBuilder invokes
TargetMachine::registerPassBuilderCallbacks() anyway. If we come up with
a cleaner solution for target-specific passes in the future that's fine,
but there aren't too many target-specific IR passes living in
target-specific directories so it shouldn't be too bad to change in the
future.

Reviewed By: ychen, arsenm

Differential Revision: https://reviews.llvm.org/D93863
llvm/lib/Target/AMDGPU/AMDGPU.h
llvm/lib/Target/AMDGPU/AMDGPULibCalls.cpp
llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.h
llvm/lib/Target/AMDGPU/CMakeLists.txt
llvm/test/CodeGen/AMDGPU/simplify-libcalls.ll
llvm/test/CodeGen/AMDGPU/simplify-libcalls2.ll
llvm/utils/gn/secondary/llvm/lib/Target/AMDGPU/BUILD.gn