[mlir][Linalg] Employ finer-grained control of C interface emission

Summary:
Linalg makes it possible to interface codegen with externally precompiled HPC libraries. The mechanism to allow such interop uses a normalized ABI and the emission of C interface wrappers.

The mechanism controlling these C interface emission is too aggressive and makes it very easy to obtained undefined symbols for external function (e.g. the ones coming from libm).

This revision uses the newly introduced llvm.emit_c_interface function attribute which allows controlling this behavior at a function granularity. As a consequence LinalgToLLVM does not need to activate the C wrapper emission when adding the StdToLLVM patterns.

Differential Revision: https://reviews.llvm.org/D77364

3 files changed, 12 insertions, 8 deletions

diff --git a/mlir/lib/Conversion/LinalgToLLVM/LinalgToLLVM.cpp b/mlir/lib/Conversion/LinalgToLLVM/LinalgToLLVM.cpp
index febf4eb..cb66ae9 100644
--- a/mlir/lib/Conversion/LinalgToLLVM/LinalgToLLVM.cpp
+++ b/mlir/lib/Conversion/LinalgToLLVM/LinalgToLLVM.cpp
@@ -400,8 +400,13 @@ static FlatSymbolRefAttr getLibraryCallSymbolRef(Operation *op,
   // Insert before module terminator.
   rewriter.setInsertionPoint(module.getBody(),
                              std::prev(module.getBody()->end()));
-  rewriter.create<FuncOp>(op->getLoc(), fnNameAttr.getValue(), libFnType,
-                          ArrayRef<NamedAttribute>{});
+  FuncOp funcOp =
+      rewriter.create<FuncOp>(op->getLoc(), fnNameAttr.getValue(), libFnType,
+                              ArrayRef<NamedAttribute>{});
+  // Insert a function attribute that will trigger the emission of the
+  // corresponding `_mlir_ciface_xxx` interface so that external libraries see
+  // a normalized ABI. This interface is added during std to llvm conversion.
+  funcOp.setAttr("llvm.emit_c_interface", UnitAttr::get(op->getContext()));
   return fnNameAttr;
 }
 
@@ -573,8 +578,7 @@ void ConvertLinalgToLLVMPass::runOnModule() {
   LLVMTypeConverter converter(&getContext());
   populateAffineToStdConversionPatterns(patterns, &getContext());
   populateLoopToStdConversionPatterns(patterns, &getContext());
-  populateStdToLLVMConversionPatterns(converter, patterns, /*useAlloca=*/false,
-                                      /*emitCWrappers=*/true);
+  populateStdToLLVMConversionPatterns(converter, patterns);
   populateVectorToLLVMMatrixConversionPatterns(converter, patterns);
   populateVectorToLLVMConversionPatterns(converter, patterns);
   populateLinalgToStandardConversionPatterns(patterns, &getContext());
diff --git a/mlir/test/mlir-cpu-runner/unranked_memref.mlir b/mlir/test/mlir-cpu-runner/unranked_memref.mlir
index 7447e9d..13d4db9 100644
--- a/mlir/test/mlir-cpu-runner/unranked_memref.mlir
+++ b/mlir/test/mlir-cpu-runner/unranked_memref.mlir
@@ -55,5 +55,5 @@ func @main() -> () {
     return
 }
 
-func @print_memref_i8(memref<*xi8>)
-func @print_memref_f32(memref<*xf32>)
+func @print_memref_i8(memref<*xi8>) attributes { llvm.emit_c_interface }
+func @print_memref_f32(memref<*xf32>) attributes { llvm.emit_c_interface }
diff --git a/mlir/test/mlir-cpu-runner/utils.mlir b/mlir/test/mlir-cpu-runner/utils.mlir
index 2a56920..d3ab617 100644
--- a/mlir/test/mlir-cpu-runner/utils.mlir
+++ b/mlir/test/mlir-cpu-runner/utils.mlir
@@ -50,7 +50,7 @@ func @print_3d() {
 //    PRINT-3D-NEXT: 2,    2,    4,    2,    2
 //    PRINT-3D-NEXT: 2,    2,    2,    2,    2
 
-func @print_memref_f32(memref<*xf32>)
+func @print_memref_f32(memref<*xf32>) attributes { llvm.emit_c_interface }
 
 !vector_type_C = type vector<4x4xf32>
 !matrix_type_CC = type memref<1x1x!vector_type_C>
@@ -71,4 +71,4 @@ func @vector_splat_2d() {
 // PRINT-VECTOR-SPLAT-2D: Memref base@ = {{.*}} rank = 2 offset = 0 sizes = [1, 1] strides = [1, 1] data =
 // PRINT-VECTOR-SPLAT-2D-NEXT: [((10, 10, 10, 10),   (10, 10, 10, 10),   (10, 10, 10, 10),   (10, 10, 10, 10))]
 
-func @print_memref_vector_4x4xf32(memref<?x?x!vector_type_C>)
+func @print_memref_vector_4x4xf32(memref<?x?x!vector_type_C>) attributes { llvm.emit_c_interface }