[MLIR][GPU] Fix build files for mlir-opt.

[herhut]

Summary

The recent refactoring of build files broke building with the MIR CUDA
integration enabled. This fixes it by adding some additional
dependencies to mlir-opt.

[herhut]

There is probably a prettier way to do this but this unblocks building.

[merge_guards_bot]

Unit tests: unknown.

clang-tidy: unknown.

clang-format: unknown.

Build artifacts: console-log.txt

Pre-merge checks is in beta. Report issue. Please join beta or enable it for your project.

[stephenneuendorffer]

Thanks for this. I didn't realize I hadn't tested the GPU config.

mlir/tools/mlir-opt/CMakeLists.txt

21

Maybe WHOLE_ARCHIVE_LINK_LIBS?

[herhut]

I am setting up a build bot that will build the tree with the cuda parts enabled so we can see this easier.

mlir/tools/mlir-opt/CMakeLists.txt

21

Seems this landed already. We use FULL_LINK_LIBS is other places, as well, but we can do a global rename in a follow up.

[mehdi_amini]

In D74041#1859889, @herhut wrote:

I am setting up a build bot that will build the tree with the cuda parts enabled so we can see this easier.

I updated our build kite bot accordingly already: https://buildkite.com/mlir/mlir-core

[herhut]

In D74041#1859977, @mehdi_amini wrote:

In D74041#1859889, @herhut wrote:

I am setting up a build bot that will build the tree with the cuda parts enabled so we can see this easier.

I updated our build kite bot accordingly already: https://buildkite.com/mlir/mlir-core

Does it also run the CUDA tests that require a GPU (mlir-cuda-runner based ones)?

[mehdi_amini]

Does it also run the CUDA tests that require a GPU (mlir-cuda-runner based ones)?

No it isn't just a cloud VM, building and testing on a GPU machine would be quite a waste

mlir/lib/Conversion/GPUToCUDA/CMakeLists.txt

 if(MLIR_CUDA_CONVERSIONS_ENABLED)
-  llvm_map_components_to_libnames(nvptx "NVPTX")
-
   add_llvm_library(MLIRGPUtoCUDATransforms
     ConvertKernelFuncToCubin.cpp
     ConvertLaunchFuncToCudaCalls.cpp
   )
   target_link_libraries(MLIRGPUtoCUDATransforms
+    LLVMNVPTXCodeGen
+    LLVMNVPTXDesc
+    LLVMNVPTXInfo
     MLIRGPU
     MLIRLLVMIR
     MLIRNVVMIR
     MLIRPass
     MLIRTargetNVVMIR
-    ${nvptx}
   )
 endif()

mlir/tools/mlir-opt/CMakeLists.txt

 )
 add_llvm_library(MLIRMlirOptLib
   mlir-opt.cpp
 )
 target_link_libraries(MLIRMlirOptLib
   ${LIB_LIBS}
 )
 
-set(LIBS
+set(FULL_LINK_LIBS
   MLIRAnalysis
   MLIRAffineOps
   MLIRAffineToStandard
   MLIRDialect
   MLIRLoopsToGPU
   MLIRLinalgToLLVM
 
   MLIRLoopToStandard
set(FULL_LINK_LIBS
   MLIRTestIR
   MLIRTestPass
   MLIRTestTransforms
   MLIRSupport
   MLIRVectorOps
   MLIRVectorToLLVM
   MLIRVectorToLoops
 )
+set(LIBS
+  MLIRIR
+  MLIROptMain
+  LLVMSupport
+  LLVMCore
+  LLVMAsmParser
+)
+
 if(MLIR_CUDA_CONVERSIONS_ENABLED)
   list(APPEND LIBS
+    LLVMNVPTXCodeGen
+    LLVMNVPTXDesc
+    LLVMNVPTXInfo
+    MLIRTargetNVVMIR
+  )
+  # Order matters here, so insert at front.
+  list(INSERT FULL_LINK_LIBS 0
     MLIRGPUtoCUDATransforms
   )
 endif()
 add_llvm_tool(mlir-opt
  mlir-opt.cpp
 )
 llvm_update_compile_flags(mlir-opt)
 # It is necessary to use whole_archive_link to ensure that all static
-# initializers are called.  However, whole_archive_link libraries cannot
-# also be target_link_libraries
-whole_archive_link(mlir-opt ${LIBS})
-target_link_libraries(mlir-opt PRIVATE MLIRIR MLIROptMain LLVMSupport LLVMCore LLVMAsmParser)
+# initializers are called. However, whole_archive_link libraries cannot
+# also be target_link_libraries.
+whole_archive_link(mlir-opt ${FULL_LINK_LIBS})
+target_link_libraries(mlir-opt PRIVATE ${LIBS})