[mlir] Add InitializeNativeTargetAsmParser to ExecutionEngine.

[nicolasvasilache]

Summary

This is required to allow python to work with lowerings that use inline_asm.

[nicolasvasilache]

mlir/lib/CAPI/ExecutionEngine/CMakeLists.txt

6 ↗

(On Diff #388761)

Hmm not too sure how to set this up properly in OSS, @ftynse do you have a suggestion?

[mehdi_amini]

utils/bazel/llvm-project-overlay/mlir/BUILD.bazel

4994

You didn't add an include, is this needed? Is this failing with a linker error without this?

[nicolasvasilache]

utils/bazel/llvm-project-overlay/mlir/BUILD.bazel

4994

Yes, this would fail with a linker error without this change.
Do you have a suggestion for the CMakeLists update?

[ftynse]

mlir/lib/CAPI/ExecutionEngine/CMakeLists.txt

6 ↗

(On Diff #388761)

Something like ${LLVM_NATIVE_ARCH}AsmParser would be more restricted.

Also, I'm pretty sure you want to add the dependency in MLIRExecutionEngine, not in its C API.

[nicolasvasilache]

Fix CMake

[mehdi_amini]

mlir/lib/CAPI/ExecutionEngine/CMakeLists.txt

6 ↗

(On Diff #388761)

Also, I'm pretty sure you want to add the dependency in MLIRExecutionEngine, not in its C API.

Why?

The call is added to the C API only, not to the C++ library.

[mehdi_amini]

mlir/lib/CAPI/ExecutionEngine/CMakeLists.txt

6 ↗

(On Diff #388761)

Ping?

[hpmorgan]

My team is working on MLIR-based projects which don't target the native (host) architecture, and there's some consternation about the increased build times arising from the added dependency on ${LLVM_NATIVE_ARCH}AsmParser in MLIRExecutionEngine (note that we don't use the execution engine either). Is this dependency on the native target unavoidable, or something that could be selectively disabled? Alternatively, could the MLIRExecutionEngine project be made optional in some way? Many thanks in advance!

[nicolasvasilache]

see https://reviews.llvm.org/D115666

[ftynse]

mlir/lib/CAPI/ExecutionEngine/CMakeLists.txt

6 ↗

(On Diff #388761)

Hm, I missed the fact that it was added to the C API. This only makes it stranger: I would expect the functionality that requires a library dependency to be implemented in C++, not in the C API glue.

mlir/lib/CAPI/ExecutionEngine/ExecutionEngine.cpp

 
 using namespace mlir;
 
 extern "C" MlirExecutionEngine
 mlirExecutionEngineCreate(MlirModule op, int optLevel, int numPaths,
                           const MlirStringRef *sharedLibPaths) {
   static bool initOnce = [] {
     llvm::InitializeNativeTarget();
+    llvm::InitializeNativeTargetAsmParser(); // needed for inline_asm
     llvm::InitializeNativeTargetAsmPrinter();
     return true;
   }();
   (void)initOnce;
 
   mlir::registerLLVMDialectTranslation(*unwrap(op)->getContext());
 
   auto tmBuilderOrError = llvm::orc::JITTargetMachineBuilder::detectHost();

mlir/lib/ExecutionEngine/CMakeLists.txt

add_mlir_library(MLIRExecutionEngine
 
   ADDITIONAL_HEADER_DIRS
   ${MLIR_MAIN_INCLUDE_DIR}/mlir/ExecutionEngine
 
   DEPENDS
   intrinsics_gen
 
   LINK_COMPONENTS
+  ${LLVM_NATIVE_ARCH}AsmParser
   Core
   Coroutines
   ExecutionEngine
   Object
   OrcJIT
   JITLink
   Analysis
   AggressiveInstCombine

utils/bazel/llvm-project-overlay/mlir/BUILD.bazel

cc_library(
     ]),
     hdrs = ["include/mlir/Translation.h"],
     includes = ["include"],
     deps = [
         ":Analysis",
         ":IR",
         ":Parser",
         ":Support",
+        "//llvm:AllTargetsAsmParsers",
         "//llvm:Support",
     ],
 )
 
 cc_library(
     name = "ToLLVMIRTranslation",
     srcs = [
         "lib/Target/LLVMIR/DebugTranslation.cpp",