Kotlin KMP引入CocoaPod依赖导致arm64架构符号未定义问题

问题背景 #

我们开发了一款支持Android、iOS和JS平台的Kotlin KMP库，近期为其添加AI功能并在全平台使用TensorFlow Lite。针对iOS平台，通过CocoaPod引入TensorFlowLiteObjC依赖，配置如下：

cocoapods {
    summary = "..."
    homepage = "..."

    framework {
        baseName = "...."

        export(project(":module_a"))
        export(project(":module_b"))

        freeCompilerArgs += "-Xallocator=std"
        freeCompilerArgs += "-linker-options"
        freeCompilerArgs += "-application_extension"

        isStatic = false


        binaryOption("bundleShortVersionString", rootProject.extra["versionName"] as String)
        binaryOption("bundleVersion", (rootProject.extra["versionCode"] as Int).toString())
    }

    ios.deploymentTarget = "14"

    pod(
        name = "TensorFlowLiteObjC",
        version = "2.5.0",
        moduleName = "TFLTensorFlowLite"
    )
}

编译报错信息 #

执行./gradlew podPublishXCFramework命令编译库时，出现如下错误：

> Task :linkPodDebugFrameworkIosArm64 FAILED
w: Interop library /.../module_b-cinterop-TFLTensorFlowLite can't be exported with -Xexport-library
e: /Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/bin/ld invocation reported errors
Please try to disable compiler caches and rerun the build. To disable compiler caches, add the following line to the gradle.properties file in the project's root directory:
    
    kotlin.native.cacheKind.iosArm64=none
    
The /Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/bin/ld command returned non-zero exit code: 1.
output:
ld: warning: ignoring duplicate libraries: '-ldl'
Undefined symbols for architecture arm64:
  "_TfLiteInterpreterAllocateTensors", referenced from:
      -[TFLInterpreter allocateTensorsWithError:] in TFLTensorFlowLite[6](TFLInterpreter.o)
  "_TfLiteInterpreterCreate", referenced from:
      -[TFLInterpreter initWithModelPath:options:delegates:error:] in TFLTensorFlowLite[6](TFLInterpreter.o)
  "_TfLiteInterpreterDelete", referenced from:
      -[TFLInterpreter dealloc] in TFLTensorFlowLite[6](TFLInterpreter.o)
  "_TfLiteInterpreterGetInputTensor", referenced from:
      -[TFLInterpreter cTensorOfType:atIndex:error:] in TFLTensorFlowLite[6](TFLInterpreter.o)
  "_TfLiteInterpreterGetInputTensorCount", referenced from:
      -[TFLInterpreter initWithModelPath:options:delegates:error:] in TFLTensorFlowLite[6](TFLInterpreter.o)
  "_TfLiteInterpreterGetOutputTensor", referenced from:
      -[TFLInterpreter cTensorOfType:atIndex:error:] in TFLTensorFlowLite[6](TFLInterpreter.o)
  "_TfLiteInterpreterGetOutputTensorCount", referenced from:
      -[TFLInterpreter initWithModelPath:options:delegates:error:] in TFLTensorFlowLite[6](TFLInterpreter.o)
  "_TfLiteInterpreterInvoke", referenced from:
      -[TFLInterpreter invokeWithError:] in TFLTensorFlowLite[6](TFLInterpreter.o)
  "_TfLiteInterpreterOptionsAddDelegate", referenced from:
      -[TFLInterpreter initWithModelPath:options:delegates:error:] in TFLTensorFlowLite[6](TFLInterpreter.o)
      -[TFLInterpreter initWithModelPath:options:delegates:error:] in TFLTensorFlowLite[6](TFLInterpreter.o)
  "_TfLiteInterpreterOptionsCreate", referenced from:
      -[TFLInterpreter initWithModelPath:options:delegates:error:] in TFLTensorFlowLite[6](TFLInterpreter.o)
  "_TfLiteInterpreterOptionsDelete", referenced from:
      -[TFLInterpreter initWithModelPath:options:delegates:error:] in TFLTensorFlowLite[6](TFLInterpreter.o)
  "_TfLiteInterpreterOptionsSetErrorReporter", referenced from:
      -[TFLInterpreter initWithModelPath:options:delegates:error:] in TFLTensorFlowLite[6](TFLInterpreter.o)
  "_TfLiteInterpreterOptionsSetNumThreads", referenced from:
      -[TFLInterpreter initWithModelPath:options:delegates:error:] in TFLTensorFlowLite[6](TFLInterpreter.o)
  "_TfLiteInterpreterResizeInputTensor", referenced from:
      -[TFLInterpreter resizeInputTensorAtIndex:toShape:error:] in TFLTensorFlowLite[6](TFLInterpreter.o)
  "_TfLiteModelCreateFromFile", referenced from:
      -[TFLInterpreter initWithModelPath:options:delegates:error:] in TFLTensorFlowLite[6](TFLInterpreter.o)
  "_TfLiteModelDelete", referenced from:
      -[TFLInterpreter initWithModelPath:options:delegates:error:] in TFLTensorFlowLite[6](TFLInterpreter.o)
  "_TfLiteTensorByteSize", referenced from:
      -[TFLInterpreter copyData:toInputTensorAtIndex:error:] in TFLTensorFlowLite[6](TFLInterpreter.o)
      -[TFLInterpreter dataFromTensor:error:] in TFLTensorFlowLite[6](TFLInterpreter.o)
  "_TfLiteTensorCopyFromBuffer", referenced from:
      -[TFLInterpreter copyData:toInputTensorAtIndex:error:] in TFLTensorFlowLite[6](TFLInterpreter.o)
  "_TfLiteTensorData", referenced from:
      -[TFLInterpreter dataFromTensor:error:] in TFLTensorFlowLite[6](TFLInterpreter.o)
  "_TfLiteTensorDim", referenced from:
      -[TFLInterpreter shapeOfTensor:error:] in TFLTensorFlowLite[6](TFLInterpreter.o)
  "_TfLiteTensorName", referenced from:
      -[TFLInterpreter tensorOfType:atIndex:error:] in TFLTensorFlowLite[6](TFLInterpreter.o)
  "_TfLiteTensorNumDims", referenced from:
      -[TFLInterpreter shapeOfTensor:error:] in TFLTensorFlowLite[6](TFLInterpreter.o)
  "_TfLiteTensorQuantizationParams", referenced from:
      -[TFLInterpreter tensorOfType:atIndex:error:] in TFLTensorFlowLite[6](TFLInterpreter.o)
  "_TfLiteTensorType", referenced from:
      -[TFLInterpreter tensorOfType:atIndex:error:] in TFLTensorFlowLite[6](TFLInterpreter.o)
  "_TfLiteVersion", referenced from:
      __GLOBAL__sub_I_TFLInterpreter.mm in TFLTensorFlowLite[6](TFLInterpreter.o)
      __GLOBAL__sub_I_TFLInterpreter.mm in TFLTensorFlowLite[6](TFLInterpreter.o)
  "_TfLiteXNNPackDelegateCreate", referenced from:
      -[TFLInterpreter initWithModelPath:options:delegates:error:] in TFLTensorFlowLite[6](TFLInterpreter.o)
  "_TfLiteXNNPackDelegateDelete", referenced from:
      -[TFLInterpreter dealloc] in TFLTensorFlowLite[6](TFLInterpreter.o)
  "_TfLiteXNNPackDelegateOptionsDefault", referenced from:
      -[TFLInterpreter initWithModelPath:options:delegates:error:] in TFLTensorFlowLite[6](TFLInterpreter.o)
ld: symbol(s) not found for architecture arm64

FAILURE: Build failed with an exception.

已尝试的无效方案 #

• 在build.gradle.kts的cocoapods.framework{}区块中添加freeCompilerArgs += listOf("-linker-option", "-framework", "-linker-option", "TensorFlowLiteObjC")
• 添加linkerOpts("-framework", "TensorFlowLiteObjC")参数
• 升级Kotlin版本至2.0.0
• 禁用编译器缓存（添加kotlin.native.cacheKind.iosArm64=none到gradle.properties）

解决方案 #

方案1：切换为静态库编译 #

将framework配置中的isStatic = false改为isStatic = true。动态库在链接第三方依赖时易出现符号缺失问题，静态库会直接将依赖符号打包进最终产物，避免链接阶段找不到符号。

修改后的配置片段：

framework {
    baseName = "...."

    export(project(":module_a"))
    export(project(":module_b"))

    freeCompilerArgs += "-Xallocator=std"
    freeCompilerArgs += "-linker-options"
    freeCompilerArgs += "-application_extension"

    isStatic = true // 修改此处

    binaryOption("bundleShortVersionString", rootProject.extra["versionName"] as String)
    binaryOption("bundleVersion", (rootProject.extra["versionCode"] as Int).toString())
}

方案2：为cinterop库添加显式链接配置 #

若必须使用动态库，需为生成的cinterop库添加链接参数：

1. 在项目的src/iosMain/cinterop目录下创建TFLTensorFlowLite.def文件（名称需与pod的moduleName对应），添加内容：

language = Objective-C
modules = TFLTensorFlowLite
linkerOpts = -framework TensorFlowLiteObjC

2. 在cocoapods配置中指定该def文件：

pod(
    name = "TensorFlowLiteObjC",
    version = "2.5.0",
    moduleName = "TFLTensorFlowLite",
    defFile = file("src/iosMain/cinterop/TFLTensorFlowLite.def")
)

方案3：调整模块导出方式 #

错误提示提到Interop library /.../module_b-cinterop-TFLTensorFlowLite can't be exported with -Xexport-library，说明module_b间接依赖了TensorFlowLite的cinterop库且被export。将export改为api，避免将cinterop库导出到主框架：

// 替换原有的export为api
api(project(":module_a"))
api(project(":module_b"))

方案4：升级TensorFlowLiteObjC版本 #

当前使用的2.5.0版本较旧，尝试升级到最新稳定版（如2.16.1），新版本Pod包可能修复了链接相关问题：

pod(
    name = "TensorFlowLiteObjC",
    version = "2.16.1",
    moduleName = "TFLTensorFlowLite"
)

内容的提问来源于stack exchange，提问作者Milan