38 lines · plain
1// RUN: mlir-opt %s -split-input-file -test-single-fold | FileCheck %s2 3// The tests in this file verify that fold() methods can handle complex4// optimization scenarios without requiring multiple folding iterations.5// This is important because:6//7// 1. OpBuilder::createOrFold() only calls fold() once, so operations must8// be fully optimized in that single call9// 2. Multiple rounds of folding would incur higher performance costs,10// so it's more efficient to complete all optimizations in one pass11//12// These tests ensure that folding implementations are robust and complete,13// avoiding situations where operations are left in intermediate states14// that could be further optimized.15 16// CHECK-LABEL: fold_extract_in_single_pass17// CHECK-SAME: (%{{.*}}: vector<4xf16>, %[[ARG1:.+]]: f16)18func.func @fold_extract_in_single_pass(%arg0: vector<4xf16>, %arg1: f16) -> f16 {19 %0 = vector.insert %arg1, %arg0 [1] : f16 into vector<4xf16>20 %c1 = arith.constant 1 : index21 // Verify that the fold is finished in a single pass even if the index is dynamic.22 %1 = vector.extract %0[%c1] : f16 from vector<4xf16>23 // CHECK: return %[[ARG1]] : f1624 return %1 : f1625}26 27// -----28 29// CHECK-LABEL: fold_insert_in_single_pass30func.func @fold_insert_in_single_pass() -> vector<2xf16> {31 %cst = arith.constant dense<0.000000e+00> : vector<2xf16>32 %c1 = arith.constant 1 : index33 %c2 = arith.constant 2.5 : f1634 // Verify that the fold is finished in a single pass even if the index is dynamic.35 // CHECK: arith.constant dense<[0.000000e+00, 2.500000e+00]> : vector<2xf16>36 %0 = vector.insert %c2, %cst [%c1] : f16 into vector<2xf16>37 return %0 : vector<2xf16>38}