[Mlir-commits] [mlir] [MLIR][NVVM] Reduce the scope of the LLVM_HAS_NVPTX_TARGET guard (PR #97349)

[Fabian Mora]

================
@@ -48,9 +48,7 @@ void GpuModuleToBinaryPass::getDependentDialects(
   // Register all GPU related translations.
   registry.insert<gpu::GPUDialect>();
   registry.insert<LLVM::LLVMDialect>();
-#if LLVM_HAS_NVPTX_TARGET
   registry.insert<NVVM::NVVMDialect>();
-#endif
 #if MLIR_ENABLE_ROCM_CONVERSIONS
   registry.insert<ROCDL::ROCDLDialect>();
 #endif
----------------
fabianmcg wrote:

The only op generated by the pass is a `gpu.binary`, and `gpu` should be in the context because the pass only performs serialization if there's a `gpu.module` in the input.

All other dialects, `llvm`, `nvvm`, `rocdl`, `spirv`, have to be in the input or loaded by a `register*Translation*` call. And all the `register*Translation*` have to be called before invoking the pass.


https://github.com/llvm/llvm-project/pull/97349