[mlir] Expand Tanh operations in StdToLLVM conversion

[ezhulenev]

Remove test-expand-tanh pass

[ezhulenev]

Fix typo

[mehdi_amini]

Why does it make sense as a standalone pass rather than as part of a pipeline? This seems a bit "targeted" to go beyond a testing pass.

[ezhulenev]

Populate expand-tanh patterns in std-to-llvm lowering

[ezhulenev]

Update CMakeLists

[ftynse]

Could you rather explain what are you trying to achieve? This doesn't necessary make sense in the standard-to-llvm pass either (and the dependency from a conversion on an in-dialect transform looks weird to me). When this conversion was introduced, there were concerns about doing it inside the conversion because not all targets actually want it. In particular, NVVM and ROCm have intrinsics that tanh should be lowered to instead. The solution was to have it in a separate populate* function that the client can call when necessary.

[ftynse]

Slightly longer-term, I think tanh should move to the math that is being split out of the standard, at which point we can have this as either a pass on math dialect (when we need to expand, potentially along with other ops that can be expanded within the dialect) or as a conversion from the math dialect to target-specific intrinsics.

[ezhulenev]

In D95758#2533599, @ftynse wrote:

The solution was to have it in a separate populate* function that the client can call when necessary.

Yes, I also felt weird about adding Transforms->Conversion dependency. I think I'll add a target specific (TFRT cpu jit compilation) pass then (FYI cl/354793665)

mlir/lib/Conversion/StandardToLLVM/CMakeLists.txt

 add_mlir_conversion_library(MLIRStandardToLLVM
   StandardToLLVM.cpp
 
   ADDITIONAL_HEADER_DIRS
   ${MLIR_MAIN_INCLUDE_DIR}/mlir/Conversion/StandardToLLVM
 
   DEPENDS
   MLIRConversionPassIncGen
   intrinsics_gen
 
   LINK_COMPONENTS
   Core
 
   LINK_LIBS PUBLIC
   MLIRLLVMIR
+  MLIRStandardOpsTransforms
   MLIRTransforms
   )

mlir/lib/Conversion/StandardToLLVM/StandardToLLVM.cpp

 //
 //===----------------------------------------------------------------------===//
 
 #include "../PassDetail.h"
 #include "mlir/Conversion/StandardToLLVM/ConvertStandardToLLVM.h"
 #include "mlir/Conversion/StandardToLLVM/ConvertStandardToLLVMPass.h"
 #include "mlir/Dialect/LLVMIR/LLVMDialect.h"
 #include "mlir/Dialect/StandardOps/IR/Ops.h"
+#include "mlir/Dialect/StandardOps/Transforms/Passes.h"
 #include "mlir/IR/Attributes.h"
 #include "mlir/IR/BlockAndValueMapping.h"
 #include "mlir/IR/Builders.h"
 #include "mlir/IR/BuiltinOps.h"
 #include "mlir/IR/MLIRContext.h"
 #include "mlir/IR/PatternMatch.h"
 #include "mlir/IR/TypeUtilities.h"
 #include "mlir/Support/LogicalResult.h"
void runOnOperation() override {
     ModuleOp m = getOperation();
 
     LowerToLLVMOptions options = {useBarePtrCallConv, emitCWrappers,
                                   indexBitwidth, useAlignedAlloc,
                                   llvm::DataLayout(this->dataLayout)};
     LLVMTypeConverter typeConverter(&getContext(), options);
 
     OwningRewritePatternList patterns;
+    populateExpandTanhPattern(patterns, &getContext());
     populateStdToLLVMConversionPatterns(typeConverter, patterns);
 
     LLVMConversionTarget target(getContext());
     if (failed(applyPartialConversion(m, target, std::move(patterns))))
       signalPassFailure();
     m->setAttr(LLVM::LLVMDialect::getDataLayoutAttrName(),
                StringAttr::get(this->dataLayout, m.getContext()));
   }