[KT-42234] Moved optimization flags to konan.properties

Combined with KT-40670, it simplifies tuning of LLVM optimization and compilation pipeline.

kotlin-native/backend.native/compiler/ir/backend.native/src/org/jetbrains/kotlin/backend/konan/OptimizationPipeline.kt

@@ -9,7 +9,6 @@ import org.jetbrains.kotlin.backend.konan.llvm.makeVisibilityHiddenLikeLlvmInter
 import org.jetbrains.kotlin.konan.target.CompilerOutputKind
 import org.jetbrains.kotlin.konan.target.Configurables
 import org.jetbrains.kotlin.konan.target.KonanTarget
-import org.jetbrains.kotlin.konan.target.ZephyrConfigurables
 
 private fun initializeLlvmGlobalPassRegistry() {
     val passRegistry = LLVMGetGlobalPassRegistry()
@@ -46,53 +45,26 @@ private class LlvmPipelineConfiguration(context: Context) {
 
     val targetTriple: String = context.llvm.targetTriple
 
-    // Some of these values are copied from corresponding runtime.bc
-    // which is using "generic" target CPU for many case.
-    // This approach is suboptimal because target-cpu="generic" limits
-    // the set of used cpu features.
-    // TODO: refactor KonanTarget so that we can explicitly specify
-    //  target cpu or arch+features combination in a single place.
-    val cpuModel: String = when (target) {
-        KonanTarget.IOS_ARM32 -> "generic"
-        KonanTarget.IOS_ARM64 -> "cyclone"
-        KonanTarget.IOS_X64 -> "core2"
-        KonanTarget.TVOS_ARM64 -> "cyclone"
-        KonanTarget.TVOS_X64 -> "core2"
-        KonanTarget.WATCHOS_X86 -> "i386"
-        KonanTarget.WATCHOS_X64 -> "core2"
-        KonanTarget.WATCHOS_ARM64,
-        KonanTarget.WATCHOS_ARM32 -> "cortex-a7"
-        KonanTarget.LINUX_X64 -> "x86-64"
-        KonanTarget.MINGW_X86 -> "pentium4"
-        KonanTarget.MINGW_X64 -> "x86-64"
-        KonanTarget.MACOS_X64 -> "core2"
-        KonanTarget.LINUX_ARM32_HFP -> "arm1136jf-s"
-        KonanTarget.LINUX_ARM64 -> "generic"
-        KonanTarget.ANDROID_ARM32 -> "arm7tdmi"
-        KonanTarget.ANDROID_ARM64 -> "cortex-a57"
-        KonanTarget.ANDROID_X64 -> "x86-64"
-        KonanTarget.ANDROID_X86 -> "i686"
-        KonanTarget.LINUX_MIPS32 -> "mips32r2"
-        KonanTarget.LINUX_MIPSEL32 -> "mips32r2"
-        KonanTarget.WASM32 -> "generic"
-        is KonanTarget.ZEPHYR -> (configurables as ZephyrConfigurables).targetCpu ?: run {
-            context.reportCompilationWarning("targetCpu for target $target was not set. Targeting `generic` cpu.")
-            "generic"
-        }
+    val cpuModel: String = configurables.targetCpu ?: run {
+        context.reportCompilationWarning("targetCpu for target $target was not set. Targeting `generic` cpu.")
+        "generic"
     }
 
-    val cpuFeatures: String = when (target) {
-        KonanTarget.LINUX_ARM32_HFP -> "+dsp,+strict-align,+vfp2,-crypto,-d16,-fp-armv8,-fp-only-sp,-fp16,-neon,-thumb-mode,-vfp3,-vfp4"
-        KonanTarget.ANDROID_ARM32 -> "+soft-float,+strict-align,-crypto,-neon,-thumb-mode"
-        else -> ""
-    }
+    val cpuFeatures: String = configurables.targetCpuFeatures ?: ""
 
     /**
      * Null value means that LLVM should use default inliner params
      * for the provided optimization and size level.
      */
     val customInlineThreshold: Int? = when {
-        context.shouldOptimize() -> INLINE_THRESHOLD_OPT
+        context.shouldOptimize() -> configurables.llvmInlineThreshold?.let {
+            it.toIntOrNull() ?: run {
+                context.reportCompilationWarning(
+                        "`llvmInlineThreshold` should be an integer. Got `$it` instead. Using default value."
+                )
+                null
+            }
+        }
         context.shouldContainDebugInfo() -> null
         else -> null
     }
@@ -122,15 +94,6 @@ private class LlvmPipelineConfiguration(context: Context) {
 
     val codeModel: LLVMCodeModel = LLVMCodeModel.LLVMCodeModelDefault
 
-    companion object {
-        // By default LLVM uses 250 for -03 builds.
-        // We use a smaller value since default value leads to
-        // unreasonably bloated runtime code without any measurable
-        // performance benefits.
-        // This value still has to be tuned for different targets, though.
-        private const val INLINE_THRESHOLD_OPT = 100
-    }
-
     enum class LlvmOptimizationLevel(val value: Int) {
         NONE(0),
         DEFAULT(1),

kotlin-native/konan/konan.properties

@@ -49,7 +49,14 @@ llvm.mingw_x64.user = msys2-mingw-w64-x86_64-clang-llvm-lld-compiler_rt-8.0.1
 llvm.macos_x64.dev = clang-llvm-apple-8.0.0-darwin-macos
 llvm.macos_x64.user = clang-llvm-apple-8.0.0-darwin-macos
 
+# By default LLVM uses 250 for -03 builds.
+# We use a smaller value since default value leads to
+# unreasonably bloated runtime code without any measurable
+# performance benefits.
+# This value still has to be tuned for different targets, though.
+llvmInlineThreshold = 100
 clangDebugFlags = -O0
+
 llvmVersion.linux_x64 = 8.0.0
 llvmVersion.mingw_x64 = 8.0.1
 llvmVersion.macos_x64 = 8.0.0
@@ -63,7 +70,7 @@ additionalToolsDir.macos_x64 = xcode-addon-xcode_12_0-macos_x64
 
 arch.macos_x64 = x86_64
 targetSysRoot.macos_x64 = target-sysroot-xcode_12_0-macos_x64
-
+targetCpu.macos_x64 = core2
 clangFlags.macos_x64 = -cc1 -emit-obj -disable-llvm-passes -x ir
 clangNooptFlags.macos_x64 = -O1
 clangOptFlags.macos_x64 = -O3
@@ -105,7 +112,7 @@ target-sysroot-xcode_12_0-ios_arm32.default = \
 arch.ios_arm32 = armv7
 # Shared with 64-bit version.
 targetSysRoot.ios_arm32 = target-sysroot-xcode_12_0-ios_arm64
-
+targetCpu.ios_arm32 = generic
 clangFlags.ios_arm32 = -cc1 -emit-obj -disable-llvm-optzns -x ir
 clangNooptFlags.ios_arm32 = -O1
 clangOptFlags.ios_arm32 = -O3
@@ -137,7 +144,7 @@ target-sysroot-xcode_12_0-ios_arm64.default = \
 
 arch.ios_arm64 = arm64
 targetSysRoot.ios_arm64 = target-sysroot-xcode_12_0-ios_arm64
-
+targetCpu.ios_arm64 = cyclone
 clangFlags.ios_arm64 = -cc1 -emit-obj -disable-llvm-passes -x ir
 clangNooptFlags.ios_arm64 = -O1
 clangOptFlags.ios_arm64 = -O3
@@ -167,7 +174,7 @@ target-sysroot-xcode_12_0-ios_x64.default = \
 
 arch.ios_x64 = x86_64
 targetSysRoot.ios_x64 = target-sysroot-xcode_12_0-ios_x64
-
+targetCpu.ios_x64 = core2
 clangFlags.ios_x64 = -cc1 -emit-obj -disable-llvm-passes -x ir
 clangNooptFlags.ios_x64 = -O1
 clangOptFlags.ios_x64 = -O3
@@ -194,7 +201,7 @@ target-sysroot-xcode_12_0-tvos_x64.default = \
 
 arch.tvos_x64 = x86_64
 targetSysRoot.tvos_x64 = target-sysroot-xcode_12_0-tvos_x64
-
+targetCpu.tvos_x64 = core2
 clangFlags.tvos_x64 = -cc1 -emit-obj -disable-llvm-passes -x ir
 clangNooptFlags.tvos_x64 = -O1
 clangOptFlags.tvos_x64 = -O3
@@ -221,7 +228,7 @@ target-sysroot-xcode_12_0-tvos_arm64.default = \
 
 arch.tvos_arm64 = arm64
 targetSysRoot.tvos_arm64 = target-sysroot-xcode_12_0-tvos_arm64
-
+targetCpu.tvos_arm64 = cyclone
 clangFlags.tvos_arm64 = -cc1 -emit-obj -disable-llvm-passes -x ir
 clangNooptFlags.tvos_arm64 = -O1
 clangOptFlags.tvos_arm64 = -O3
@@ -248,7 +255,7 @@ target-sysroot-xcode_12_0-watchos_arm32.default = \
 
 arch.watchos_arm32 = armv7k
 targetSysRoot.watchos_arm32 = target-sysroot-xcode_12_0-watchos_arm32
-
+targetCpu.watchos_arm32 = cortex-a7
 clangFlags.watchos_arm32 = -cc1 -emit-obj -disable-llvm-passes -x ir
 clangNooptFlags.watchos_arm32 = -O1
 clangOptFlags.watchos_arm32 = -O3
@@ -277,7 +284,7 @@ target-sysroot-xcode_12_0-watchos_arm64.default = \
 
 arch.watchos_arm64 = arm64_32
 targetSysRoot.watchos_arm64 = target-sysroot-xcode_12_0-watchos_arm32
-
+targetCpu.watchos_arm64 = cortex-a7
 clangFlags.watchos_arm64 = -cc1 -emit-obj -disable-llvm-passes -x ir -mllvm -aarch64-watch-bitcode-compatibility \
   -mllvm -arm-bitcode-compatibility -mllvm -fast-isel=false -mllvm -global-isel=false
 clangNooptFlags.watchos_arm64 = -O1
@@ -307,12 +314,12 @@ target-sysroot-xcode_12_0-watchos_x86.default = \
 
 arch.watchos_x86 = i386
 targetSysRoot.watchos_x86 = target-sysroot-xcode_12_0-watchos_x86
-
-# -target-cpu pentium4 makes sure that cogenerator knows which CPU flavour to emit code for.
+# -target-cpu pentium4 makes sure that code-generator knows which CPU flavour to emit code for.
 # Can be seen on optimized build and FP tests, where value is passed on stack
 # by Kotlin and taken from SSE registers by C code.
 # TODO: once this information is available in function attributes, we can remove this flag.
-clangFlags.watchos_x86 = -cc1 -emit-obj -disable-llvm-passes -x ir -target-cpu pentium4
+targetCpu.watchos_x86 = pentium4
+clangFlags.watchos_x86 = -cc1 -emit-obj -disable-llvm-passes -x ir -target-cpu $targetCpu.watchos_x86
 clangNooptFlags.watchos_x86 = -O1
 clangOptFlags.watchos_x86 = -O3
 clangDebugFlags.watchos_x86 = -O0
@@ -355,7 +362,8 @@ quadruple.linux_x64 = x86_64-unknown-linux-gnu
 targetSysRoot.linux_x64 = target-gcc-toolchain-3-linux-x86-64/x86_64-unknown-linux-gnu/sysroot
 # targetSysroot-relative.
 libGcc.linux_x64 = ../../lib/gcc/x86_64-unknown-linux-gnu/4.8.5
-clangFlags.linux_x64 = -cc1 -target-cpu x86-64 -emit-obj -disable-llvm-optzns -x ir
+targetCpu.linux_x64 = x86-64
+clangFlags.linux_x64 = -cc1 -target-cpu $targetCpu.linux_x64 -emit-obj -disable-llvm-optzns -x ir
 clangNooptFlags.linux_x64 = -O1
 clangOptFlags.linux_x64 = -O3 -ffunction-sections
 clangDebugFlags.linux_x64 = -O0
@@ -398,7 +406,9 @@ linkerKonanFlags.linux_arm32_hfp = -Bstatic -lstdc++ -Bdynamic -ldl -lm -lpthrea
   --defsym __cxa_demangle=Konan_cxa_demangle --no-threads
 # targetSysroot-relative.
 libGcc.linux_arm32_hfp = lib/gcc/arm-linux-gnueabihf/4.8.3
-clangFlags.linux_arm32_hfp = -cc1 -target-cpu arm1136jf-s -mfloat-abi hard -emit-obj -disable-llvm-optzns -x ir
+targetCpu.linux_arm32_hfp = arm1136jf-s
+targetCpuFeatures.linux_arm32_hfp = +dsp,+strict-align,+vfp2,-crypto,-d16,-fp-armv8,-fp-only-sp,-fp16,-neon,-thumb-mode,-vfp3,-vfp4
+clangFlags.linux_arm32_hfp = -cc1 -target-cpu $targetCpu.linux_arm32_hfp -mfloat-abi hard -emit-obj -disable-llvm-optzns -x ir
 clangNooptFlags.linux_arm32_hfp = -O1
 clangOptFlags.linux_arm32_hfp = -O3 -ffunction-sections
 clangDebugFlags.linux_arm32_hfp = -O0
@@ -440,6 +450,7 @@ linkerKonanFlags.linux_arm64 = -Bstatic -lstdc++ -Bdynamic -ldl -lm -lpthread \
   --defsym __cxa_demangle=Konan_cxa_demangle --no-threads
 # targetSysroot-relative.
 libGcc.linux_arm64 = usr/lib
+targetCpu.linux_arm64 = generic
 clangFlags.linux_arm64 = -cc1 -target-cpu cortex-a57 -emit-obj -disable-llvm-optzns -x ir
 clangNooptFlags.linux_arm64 = -O1
 clangOptFlags.linux_arm64 = -O3 -ffunction-sections
@@ -470,7 +481,8 @@ linkerKonanFlags.linux_mips32 = -Bstatic -lstdc++ -Bdynamic -ldl -lm -lpthread \
   --defsym __cxa_demangle=Konan_cxa_demangle  -z notext
 # targetSysroot-relative.
 libGcc.linux_mips32 = lib/gcc/mips-unknown-linux-gnu/4.9.4
-clangFlags.linux_mips32 = -cc1 -emit-obj -disable-llvm-optzns -x ir -target-cpu mips32r2
+targetCpu.linux_mips32 = mips32r2
+clangFlags.linux_mips32 = -cc1 -emit-obj -disable-llvm-optzns -x ir -target-cpu $targetCpu.linux_mips32
 clangNooptFlags.linux_mips32 = -O1
 clangOptFlags.linux_mips32 = -O3 -ffunction-sections
 clangDebugFlags.linux_mips32 = -O0
@@ -498,7 +510,8 @@ linkerKonanFlags.linux_mipsel32 = -Bstatic -lstdc++ -Bdynamic -ldl -lm -lpthread
   --defsym __cxa_demangle=Konan_cxa_demangle -z notext
 # targetSysroot-relative.
 libGcc.linux_mipsel32 = lib/gcc/mipsel-unknown-linux-gnu/4.9.4
-clangFlags.linux_mipsel32 = -cc1 -emit-obj -disable-llvm-optzns -x ir -target-cpu mips32r2
+targetCpu.linux_mipsel32 = mips32r2
+clangFlags.linux_mipsel32 = -cc1 -emit-obj -disable-llvm-optzns -x ir -target-cpu targetCpu.linux_mipsel32
 clangNooptFlags.linux_mipsel32 = -O1
 clangOptFlags.linux_mipsel32 = -O3 -ffunction-sections
 clangDebugFlags.linux_mipsel32 = -O0
@@ -528,7 +541,9 @@ dependencies.mingw_x64-android_arm32 = \
     libffi-3.2.1-mingw-w64-x86-64
 
 quadruple.android_arm32 = arm-linux-androideabi
-clangFlags.android_arm32 = -cc1 -target-cpu arm7tdmi -emit-obj -disable-llvm-optzns -x ir -femulated-tls -mrelocation-model pic
+targetCpu.android_arm32 = arm7tdmi
+targetCpuFeatures.android_arm32 = +soft-float,+strict-align,-crypto,-neon,-thumb-mode
+clangFlags.android_arm32 = -cc1 -target-cpu $targetCpu.android_arm32 -emit-obj -disable-llvm-optzns -x ir -femulated-tls -mrelocation-model pic
 clangOptFlags.android_arm32 = -O3 -ffunction-sections
 linkerNoDebugFlags.android_arm32 = -Wl,-S
 clangNooptFlags.android_arm32 = -O1
@@ -555,7 +570,8 @@ dependencies.mingw_x64-android_arm64 = \
     libffi-3.2.1-mingw-w64-x86-64
 
 quadruple.android_arm64 = aarch64-linux-android
-clangFlags.android_arm64 = -cc1 -target-cpu cortex-a57 -emit-obj -disable-llvm-passes -x ir -femulated-tls -mrelocation-model pic
+targetCpu.android_arm64 = cortex-a57
+clangFlags.android_arm64 = -cc1 -target-cpu $targetCpu.android_arm64 -emit-obj -disable-llvm-passes -x ir -femulated-tls -mrelocation-model pic
 clangOptFlags.android_arm64 = -O3 -ffunction-sections
 clangNooptFlags.android_arm64 = -O1
 targetSysRoot.android_arm64 = target-sysroot-1-android_ndk
@@ -582,12 +598,12 @@ dependencies.mingw_x64-android_x86 = \
     libffi-3.2.1-mingw-w64-x86-64
 
 quadruple.android_x86 = i686-linux-android
-
 # -target-cpu pentium4 makes sure that cogenerator knows which CPU flavour to emit code for.
 # Can be seen on optimized build and FP tests, where value is passed on stack
 # by Kotlin and taken from SSE registers by C code.
 # TODO: once this information is available in function attributes, we can set target CPU to i686.
-clangFlags.android_x86 = -cc1 -target-cpu pentium4 -emit-obj -disable-llvm-passes -x ir -femulated-tls -mrelocation-model pic
+targetCpu.android_x86 = pentium4
+clangFlags.android_x86 = -cc1 -target-cpu targetCpu.android_x86 -emit-obj -disable-llvm-passes -x ir -femulated-tls -mrelocation-model pic
 clangOptFlags.android_x86 = -O3 -ffunction-sections
 clangNooptFlags.android_x86 = -O1
 targetSysRoot.android_x86 = target-sysroot-1-android_ndk
@@ -614,7 +630,8 @@ dependencies.mingw_x64-android_x64 = \
     libffi-3.2.1-mingw-w64-x86-64
 
 quadruple.android_x64 = x86_64-linux-android
-clangFlags.android_x64 = -cc1 -target-cpu x86-64 -emit-obj -disable-llvm-passes -x ir -femulated-tls -mrelocation-model pic
+targetCpu.android_x64 = x86-64
+clangFlags.android_x64 = -cc1 -target-cpu $targetCpu.android_x64 -emit-obj -disable-llvm-passes -x ir -femulated-tls -mrelocation-model pic
 clangOptFlags.android_x64 = -O3 -ffunction-sections
 clangNooptFlags.android_x64 = -O1
 targetSysRoot.android_x64 = target-sysroot-1-android_ndk
@@ -641,10 +658,11 @@ dependencies.macos_x64-mingw_x64 = \
 
 quadruple.mingw_x64 = x86_64-w64-mingw32
 targetSysRoot.mingw_x64 = msys2-mingw-w64-x86_64-clang-llvm-lld-compiler_rt-8.0.1
+targetCpu.mingw_x64 = x86-64
 # For using with Universal Windows Platform (UWP) we need to use this slower option.
 # See https://youtrack.jetbrains.com/issue/KT-27654.
 # TODO: remove, once fixed in mingw, check with the bug testcase.
-clangFlags.mingw_x64 = -cc1 -emit-obj -disable-llvm-passes -x ir -femulated-tls -target-cpu x86-64
+clangFlags.mingw_x64 = -cc1 -emit-obj -disable-llvm-passes -x ir -femulated-tls -target-cpu $targetCpu.mingw_x64
 clangOptFlags.mingw_x64 = -O3 -ffunction-sections
 clangNooptFlags.mingw_x64 = -O1
 linkerNoDebugFlags.mingw_x64 = -Wl,-S
@@ -677,11 +695,12 @@ targetSysRoot.mingw_x86 = msys2-mingw-w64-i686-clang-llvm-lld-compiler_rt-8.0.1
 # For using with Universal Windows Platform (UWP) we need to use slower -emulated-tls option.
 # See https://youtrack.jetbrains.com/issue/KT-27654.
 # TODO: remove, once fixed in mingw, check with the bug testcase.
-# -target-cpu pentium4 makes sure that cogenerator knows which CPU flavour to emit code for.
+# -target-cpu pentium4 makes sure that code-generator knows which CPU flavour to emit code for.
 # Can be seen on optimized build and FP tests, where value is passed in coprocessor registers
 # by Kotlin and taken from SSE registers by C code.
 # TODO: once this information is available in function attributes, we can remove this flag.
-clangFlags.mingw_x86 = -cc1 -emit-obj -disable-llvm-passes -x ir -femulated-tls -target-cpu pentium4
+targetCpu.mingw_x86 = pentium4
+clangFlags.mingw_x86 = -cc1 -emit-obj -disable-llvm-passes -x ir -femulated-tls -target-cpu $targetCpu.mingw_x86
 clangOptFlags.mingw_x86 = -O3 -ffunction-sections
 clangNooptFlags.mingw_x86 = -O1
 linkerNoDebugFlags.mingw_x86 = -Wl,-S
@@ -716,6 +735,7 @@ dependencies.mingw_x64-wasm32 = \
 
 quadruple.wasm32 = wasm32-unknown-unknown
 targetSysRoot.wasm32 = target-sysroot-4-embedded
+targetCpu.wasm32 = generic
 clangFlags.wasm32 = -cc1 -emit-obj -disable-llvm-optzns -x ir
 clangNooptFlags.wasm32 = -O1
 clangOptFlags.wasm32 = -O3

kotlin-native/shared/src/main/kotlin/org/jetbrains/kotlin/konan/target/Configurables.kt

@@ -53,6 +53,10 @@ interface Configurables : TargetableExternalStorage {
     val absoluteTargetToolchain get() = absolute(targetToolchain)
     val absoluteLlvmHome get() = absolute(llvmHome)
 
+    val targetCpu get() = targetString("targetCpu")
+    val targetCpuFeatures get() = targetString("targetCpuFeatures")
+    val llvmInlineThreshold get() = targetString("llvmInlineThreshold")
+
     val runtimeDefinitions get() = targetList("runtimeDefinitions")
 }
 
@@ -85,6 +89,5 @@ interface WasmConfigurables : TargetableConfigurables, ClangFlags, LldFlags
 
 interface ZephyrConfigurables : TargetableConfigurables, ClangFlags {
     val boardSpecificClangFlags get() = targetList("boardSpecificClangFlags")
-    val targetCpu get() = targetString("targetCpu")
     val targetAbi get() = targetString("targetAbi")
 }