[VMVX] Add support for arith.maxnumf and arith.minnumf lowering.

Similar to what is done in CPU backend, it adds a ArithExpandOps pass to the late stage of the pipeline. This drops the local llvm revert because the VMVX backend can emulate the ops. Note: the arith.maximum and arith.minimum ops are also expanded with the change. They become a seq of cmp + select ops. Signed-off-by: hanhanW <[email protected]>
iree-org · Jul 29, 2024 · 3a0dff0 · 3a0dff0
1 parent 45323df
commit 3a0dff0
Show file tree

Hide file tree

Showing 2 changed files with 2 additions and 1 deletion.
diff --git a/compiler/src/iree/compiler/Dialect/VMVX/Transforms/Passes.cpp b/compiler/src/iree/compiler/Dialect/VMVX/Transforms/Passes.cpp
@@ -87,6 +87,7 @@ buildVectorVMVXTransformPassPipeline(OpPassManager &variantPassManager) {
       .addPass(createCSEPass)
       .addPass([]() { return createConvertVectorToSCFPass(); })
       .addPass(createCanonicalizerPass)
+      .addPass(arith::createArithExpandOpsPass)
       .addPass(memref::createExpandOpsPass);
 
   // Handle tensor-type constants.

diff --git a/third_party/llvm-project b/third_party/llvm-project
+2 −2		mlir/lib/Dialect/Linalg/IR/LinalgOps.cpp
+1 −1		mlir/test/Dialect/Linalg/transform-op-decompose.mlir