diff --git a/kotlin-native/backend.native/compiler/ir/backend.native/src/org/jetbrains/kotlin/backend/konan/llvm/LlvmAttributes.kt b/kotlin-native/backend.native/compiler/ir/backend.native/src/org/jetbrains/kotlin/backend/konan/llvm/LlvmAttributes.kt
index 4cd156f89d8..44fcea358dd 100644
--- a/kotlin-native/backend.native/compiler/ir/backend.native/src/org/jetbrains/kotlin/backend/konan/llvm/LlvmAttributes.kt
+++ b/kotlin-native/backend.native/compiler/ir/backend.native/src/org/jetbrains/kotlin/backend/konan/llvm/LlvmAttributes.kt
@@ -22,6 +22,7 @@ internal fun addLlvmFunctionWithDefaultAttributes(
         type: LLVMTypeRef
 ): LLVMValueRef = LLVMAddFunction(module, name, type)!!.also {
     addDefaultLlvmFunctionAttributes(context, it)
+    addTargetCpuAndFeaturesAttributes(context, it)
 }
 
 /**
@@ -35,6 +36,18 @@ private fun addDefaultLlvmFunctionAttributes(context: Context, llvmFunction: LLV
     }
 }
 
+/**
+ * Set target cpu and its features to make LLVM generate correct machine code.
+ */
+private fun addTargetCpuAndFeaturesAttributes(context: Context, llvmFunction: LLVMValueRef) {
+    context.config.platform.targetCpu?.let {
+        LLVMAddTargetDependentFunctionAttr(llvmFunction, "target-cpu", it)
+    }
+    context.config.platform.targetCpuFeatures?.let {
+        LLVMAddTargetDependentFunctionAttr(llvmFunction, "target-features", it)
+    }
+}
+
 internal fun addLlvmAttributesForKotlinFunction(context: Context, irFunction: IrFunction, llvmFunction: LLVMValueRef) {
     if (irFunction.returnType.isNothing()) {
         setFunctionNoReturn(llvmFunction)
diff --git a/kotlin-native/konan/konan.properties b/kotlin-native/konan/konan.properties
index 039e04511a2..c02b1b5af37 100644
--- a/kotlin-native/konan/konan.properties
+++ b/kotlin-native/konan/konan.properties
@@ -109,7 +109,7 @@ targetTriple.macos_x64 = x86_64-apple-macos
 targetSysRoot.macos_x64 = target-sysroot-xcode_12_5-macosx
 targetCpu.macos_x64 = core2
 targetCpuFeatures.macos_x64=+cx16,+cx8,+fxsr,+mmx,+sahf,+sse,+sse2,+sse3,+sse4.1,+ssse3,+x87
-clangFlags.macos_x64 = -cc1 -emit-obj -disable-llvm-passes -x ir -target-cpu $targetCpu.macos_x64
+clangFlags.macos_x64 = -cc1 -emit-obj -disable-llvm-passes -x ir
 clangNooptFlags.macos_x64 = -O1
 clangOptFlags.macos_x64 = -O3
 clangDebugFlags.macos_x64 = -O0
@@ -147,7 +147,7 @@ targetTriple.macos_arm64 = arm64-apple-macos
 targetSysRoot.macos_arm64 = target-sysroot-xcode_12_5-macosx
 # TODO: Check Clang behaviour.
 targetCpu.macos_arm64 = cyclone
-clangFlags.macos_arm64 = -cc1 -emit-obj -disable-llvm-passes -x ir -target-cpu $targetCpu.macos_arm64
+clangFlags.macos_arm64 = -cc1 -emit-obj -disable-llvm-passes -x ir
 clangNooptFlags.macos_arm64 = -O1
 clangOptFlags.macos_arm64 = -O3
 # See clangDebugFlags.ios_arm64
@@ -188,7 +188,7 @@ targetTriple.ios_arm32 = armv7-apple-ios
 # Shared with 64-bit version.
 targetSysRoot.ios_arm32 = target-sysroot-xcode_12_5-iphoneos
 targetCpu.ios_arm32 = generic
-clangFlags.ios_arm32 = -cc1 -emit-obj -disable-llvm-optzns -x ir -target-cpu $targetCpu.ios_arm32
+clangFlags.ios_arm32 = -cc1 -emit-obj -disable-llvm-optzns -x ir
 clangNooptFlags.ios_arm32 = -O1
 clangOptFlags.ios_arm32 = -O3
 clangDebugFlags.ios_arm32 = -O0
@@ -218,7 +218,7 @@ target-sysroot-xcode_12_5-iphoneos.default = \
 targetTriple.ios_arm64 = arm64-apple-ios
 targetSysRoot.ios_arm64 = target-sysroot-xcode_12_5-iphoneos
 targetCpu.ios_arm64 = cyclone
-clangFlags.ios_arm64 = -cc1 -emit-obj -disable-llvm-passes -x ir -target-cpu $targetCpu.ios_arm64
+clangFlags.ios_arm64 = -cc1 -emit-obj -disable-llvm-passes -x ir
 clangNooptFlags.ios_arm64 = -O1
 clangOptFlags.ios_arm64 = -O3
 # We force LLVM to use SelectionDAG because fast isel may miscompile sign extension.
@@ -247,7 +247,7 @@ target-sysroot-xcode_12_5-iphonesimulator.default = \
 targetTriple.ios_x64 = x86_64-apple-ios-simulator
 targetSysRoot.ios_x64 = target-sysroot-xcode_12_5-iphonesimulator
 targetCpu.ios_x64 = core2
-clangFlags.ios_x64 = -cc1 -emit-obj -disable-llvm-passes -x ir -target-cpu $targetCpu.ios_x64
+clangFlags.ios_x64 = -cc1 -emit-obj -disable-llvm-passes -x ir
 clangNooptFlags.ios_x64 = -O1
 clangOptFlags.ios_x64 = -O3
 clangDebugFlags.ios_x64 = -O0
@@ -273,7 +273,7 @@ target-sysroot-xcode_12_5-ios_simulator_arm64.default = \
 targetTriple.ios_simulator_arm64 = arm64-apple-ios-simulator
 targetSysRoot.ios_simulator_arm64 = target-sysroot-xcode_12_5-iphonesimulator
 targetCpu.ios_simulator_arm64 = cyclone
-clangFlags.ios_simulator_arm64 = -cc1 -emit-obj -disable-llvm-passes -x ir -target-cpu $targetCpu.ios_simulator_arm64
+clangFlags.ios_simulator_arm64 = -cc1 -emit-obj -disable-llvm-passes -x ir
 clangNooptFlags.ios_simulator_arm64 = -O1
 clangOptFlags.ios_simulator_arm64 = -O3
 clangDebugFlags.ios_simulator_arm64 = -O0 -mllvm -fast-isel=false -mllvm -global-isel=false
@@ -298,7 +298,7 @@ target-sysroot-xcode_12_5-appletvsimulator.default = \
 targetTriple.tvos_x64 = x86_64-apple-tvos-simulator
 targetSysRoot.tvos_x64 = target-sysroot-xcode_12_5-appletvsimulator
 targetCpu.tvos_x64 = core2
-clangFlags.tvos_x64 = -cc1 -emit-obj -disable-llvm-passes -x ir -target-cpu $targetCpu.tvos_x64
+clangFlags.tvos_x64 = -cc1 -emit-obj -disable-llvm-passes -x ir
 clangNooptFlags.tvos_x64 = -O1
 clangOptFlags.tvos_x64 = -O3
 clangDebugFlags.tvos_x64 = -O0
@@ -323,7 +323,7 @@ target-sysroot-xcode_12_5-tvos_simulator_arm64.default = \
 targetTriple.tvos_simulator_arm64 = arm64-apple-tvos-simulator
 targetSysRoot.tvos_simulator_arm64 = target-sysroot-xcode_12_5-appletvsimulator
 targetCpu.tvos_simulator_arm64 = cyclone
-clangFlags.tvos_simulator_arm64 = -cc1 -emit-obj -disable-llvm-passes -x ir -target-cpu $targetCpu.tvos_simulator_arm64
+clangFlags.tvos_simulator_arm64 = -cc1 -emit-obj -disable-llvm-passes -x ir
 clangNooptFlags.tvos_simulator_arm64 = -O1
 clangOptFlags.tvos_simulator_arm64 = -O3
 clangDebugFlags.tvos_simulator_arm64 = -O0 -mllvm -fast-isel=false -mllvm -global-isel=false
@@ -348,7 +348,7 @@ target-sysroot-xcode_12_5-appletvos.default = \
 targetTriple.tvos_arm64 = arm64-apple-tvos
 targetSysRoot.tvos_arm64 = target-sysroot-xcode_12_5-appletvos
 targetCpu.tvos_arm64 = cyclone
-clangFlags.tvos_arm64 = -cc1 -emit-obj -disable-llvm-passes -x ir -target-cpu $targetCpu.tvos_arm64
+clangFlags.tvos_arm64 = -cc1 -emit-obj -disable-llvm-passes -x ir
 clangNooptFlags.tvos_arm64 = -O1
 clangOptFlags.tvos_arm64 = -O3
 clangDebugFlags.tvos_arm64 = -O0 -mllvm -fast-isel=false -mllvm -global-isel=false
@@ -373,7 +373,7 @@ target-sysroot-xcode_12_5-watchos.default = \
 targetTriple.watchos_arm32 = armv7k-apple-watchos
 targetSysRoot.watchos_arm32 = target-sysroot-xcode_12_5-watchos
 targetCpu.watchos_arm32 = cortex-a7
-clangFlags.watchos_arm32 = -cc1 -emit-obj -disable-llvm-passes -x ir -target-cpu $targetCpu.watchos_arm32
+clangFlags.watchos_arm32 = -cc1 -emit-obj -disable-llvm-passes -x ir
 clangNooptFlags.watchos_arm32 = -O1
 clangOptFlags.watchos_arm32 = -O3
 clangDebugFlags.watchos_arm32 = -O0
@@ -396,9 +396,8 @@ targetToolchain.macos_arm64-watchos_arm64 = target-toolchain-xcode_12_5
 
 targetTriple.watchos_arm64 = arm64_32-apple-watchos
 targetSysRoot.watchos_arm64 = target-sysroot-xcode_12_5-watchos
+# TODO: Use -target-cpu=apple-s4 when we start generating bitcode for arm64_32.
 targetCpu.watchos_arm64 = cortex-a7
-# TODO: Use -target-cpu=apple-s4 after LLVM update.
-#  Note that we can't use $targetCpu.watchos_arm64 because bitcode is generated for watchos_arm32.
 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
@@ -426,12 +425,9 @@ target-sysroot-xcode_12_5-watchsimulator.default = \
 
 targetTriple.watchos_x86 = i386-apple-watchos-simulator
 targetSysRoot.watchos_x86 = target-sysroot-xcode_12_5-watchsimulator
-# -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.
+# Make sure that SSE2 is available by setting target-cpu to pentium4
 targetCpu.watchos_x86 = pentium4
-clangFlags.watchos_x86 = -cc1 -emit-obj -disable-llvm-passes -x ir -target-cpu $targetCpu.watchos_x86
+clangFlags.watchos_x86 = -cc1 -emit-obj -disable-llvm-passes -x ir
 clangNooptFlags.watchos_x86 = -O1
 clangOptFlags.watchos_x86 = -O3
 clangDebugFlags.watchos_x86 = -O0
@@ -456,7 +452,7 @@ target-sysroot-xcode_12_5-watchos_x64.default = \
 targetTriple.watchos_x64 = x86_64-apple-watchos-simulator
 targetSysRoot.watchos_x64 = target-sysroot-xcode_12_5-watchsimulator
 targetCpu.watchos_x64 = core2
-clangFlags.watchos_x64 = -cc1 -emit-obj -disable-llvm-passes -x ir -target-cpu $targetCpu.watchos_x64
+clangFlags.watchos_x64 = -cc1 -emit-obj -disable-llvm-passes -x ir
 clangNooptFlags.watchos_x64 = -O1
 clangOptFlags.watchos_x64 = -O3
 clangDebugFlags.watchos_x64 = -O0
@@ -481,7 +477,7 @@ target-sysroot-xcode_12_5-watchos_simulator_arm64.default = \
 targetTriple.watchos_simulator_arm64 = arm64-apple-watchos-simulator
 targetSysRoot.watchos_simulator_arm64 = target-sysroot-xcode_12_5-watchsimulator
 targetCpu.watchos_simulator_arm64 = cyclone
-clangFlags.watchos_simulator_arm64 = -cc1 -emit-obj -disable-llvm-passes -x ir -target-cpu $targetCpu.watchos_simulator_arm64
+clangFlags.watchos_simulator_arm64 = -cc1 -emit-obj -disable-llvm-passes -x ir
 clangNooptFlags.watchos_simulator_arm64 = -O1
 clangOptFlags.watchos_simulator_arm64 = -O3
 clangDebugFlags.watchos_simulator_arm64 = -O0 -mllvm -fast-isel=false -mllvm -global-isel=false
@@ -526,7 +522,7 @@ targetSysRoot.linux_x64 = $gccToolchain.linux_x64/x86_64-unknown-linux-gnu/sysro
 libGcc.linux_x64 = ../../lib/gcc/x86_64-unknown-linux-gnu/8.3.0
 targetCpu.linux_x64 = x86-64
 targetCpuFeatures.linux_x64 = +cx8,+fxsr,+mmx,+sse,+sse2,+x87
-clangFlags.linux_x64 = -cc1 -target-cpu $targetCpu.linux_x64 -emit-obj -disable-llvm-optzns -x ir \
+clangFlags.linux_x64 = -cc1 -emit-obj -disable-llvm-optzns -x ir \
    -ffunction-sections -fdata-sections
 clangNooptFlags.linux_x64 = -O1
 clangOptFlags.linux_x64 = -O3
@@ -593,7 +589,7 @@ targetCpu.linux_arm32_hfp = arm1136jf-s
 # Taken from clang-produced bitcode.
 # TODO: Not sure if we need so many features. Revise.
 targetCpuFeatures.linux_arm32_hfp = +armv6kz,+dsp,+fp64,+strict-align,+vfp2,+vfp2sp,-crypto,-d32,-fp-armv8,-fp-armv8d16,-fp-armv8d16sp,-fp-armv8sp,-fp16,-fp16fml,-fullfp16,-neon,-thumb-mode,-vfp3,-vfp3d16,-vfp3d16sp,-vfp3sp,-vfp4,-vfp4d16,-vfp4d16sp,-vfp4sp
-clangFlags.linux_arm32_hfp = -cc1 -target-cpu $targetCpu.linux_arm32_hfp -mfloat-abi hard -emit-obj -disable-llvm-optzns -x ir
+clangFlags.linux_arm32_hfp = -cc1 -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
@@ -649,7 +645,7 @@ linkerKonanFlags.linux_arm64 = -Bstatic -lstdc++ -Bdynamic -ldl -lm -lpthread \
 # targetSysroot-relative.
 libGcc.linux_arm64 = ../../lib/gcc/aarch64-unknown-linux-gnu/8.3.0
 targetCpu.linux_arm64 = cortex-a57
-clangFlags.linux_arm64 = -cc1 -target-cpu $targetCpu.linux_arm64 -emit-obj -disable-llvm-optzns -x ir
+clangFlags.linux_arm64 = -cc1 -emit-obj -disable-llvm-optzns -x ir
 clangNooptFlags.linux_arm64 = -O1
 clangOptFlags.linux_arm64 = -O3 -ffunction-sections
 clangDebugFlags.linux_arm64 = -O0
@@ -695,7 +691,8 @@ libGcc.linux_mips32 = ../../lib/gcc/mips-unknown-linux-gnu/8.3.0
 targetCpu.linux_mips32 = mips32r2
 # We generate a single binary with a big GOT. It causes problems for MIPS.
 # See https://gcc.gnu.org/onlinedocs/gcc/MIPS-Options.html about mxgot (aka -target-feature +xgot).
-clangFlags.linux_mips32 = -cc1 -emit-obj -disable-llvm-optzns -x ir -target-cpu $targetCpu.linux_mips32 -target-feature +xgot
+targetCpuFeatures.linux_mips32 = +xgot
+clangFlags.linux_mips32 = -cc1 -emit-obj -disable-llvm-optzns -x ir
 clangNooptFlags.linux_mips32 = -O1
 clangOptFlags.linux_mips32 = -O3 -ffunction-sections
 clangDebugFlags.linux_mips32 = -O0
@@ -734,7 +731,8 @@ linkerKonanFlags.linux_mipsel32 = -Bstatic -lstdc++ -Bdynamic -ldl -lm -lpthread
 # targetSysroot-relative.
 libGcc.linux_mipsel32 = ../../lib/gcc/mipsel-unknown-linux-gnu/8.3.0
 targetCpu.linux_mipsel32 = mips32r2
-clangFlags.linux_mipsel32 = -cc1 -emit-obj -disable-llvm-optzns -x ir -target-cpu $targetCpu.linux_mipsel32 -target-feature +xgot
+targetCpuFeatures.linux_mipsel32 = +xgot
+clangFlags.linux_mipsel32 = -cc1 -emit-obj -disable-llvm-optzns -x ir
 clangNooptFlags.linux_mipsel32 = -O1
 clangOptFlags.linux_mipsel32 = -O3 -ffunction-sections
 clangDebugFlags.linux_mipsel32 = -O0
@@ -773,7 +771,7 @@ dependencies.mingw_x64-android_arm32 = \
 targetTriple.android_arm32 = arm-unknown-linux-androideabi
 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
+clangFlags.android_arm32 = -cc1 -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
@@ -802,7 +800,7 @@ dependencies.mingw_x64-android_arm64 = \
 
 targetTriple.android_arm64 = aarch64-unknown-linux-android
 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
+clangFlags.android_arm64 = -cc1 -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
@@ -830,12 +828,9 @@ dependencies.mingw_x64-android_x86 = \
     target-toolchain-2-windows-android_ndk
 
 targetTriple.android_x86 = i686-unknown-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.
+# Make sure that SSE2 is available by setting target-cpu to pentium4
 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
+clangFlags.android_x86 = -cc1 -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
@@ -864,7 +859,7 @@ dependencies.mingw_x64-android_x64 = \
 
 targetTriple.android_x64 = x86_64-unknown-linux-android
 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
+clangFlags.android_x64 = -cc1 -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
@@ -902,7 +897,7 @@ targetCpuFeatures.mingw_x64 = +cx8,+fxsr,+mmx,+sse,+sse2,+x87
 # 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 $targetCpu.mingw_x64
+clangFlags.mingw_x64 = -cc1 -emit-obj -disable-llvm-passes -x ir -femulated-tls
 clangOptFlags.mingw_x64 = -O3 -ffunction-sections
 clangNooptFlags.mingw_x64 = -O1
 linkerNoDebugFlags.mingw_x64 = -Wl,-S
@@ -930,15 +925,12 @@ dependencies.macos_x64-mingw_x86 = \
 
 targetTriple.mingw_x86 = i686-pc-windows-gnu
 targetSysRoot.mingw_x86 = msys2-mingw-w64-i686-clang-llvm-lld-compiler_rt-8.0.1
+# Make sure that SSE2 is available by setting target-cpu to pentium4
+targetCpu.mingw_x86 = pentium4
 # 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 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.
-targetCpu.mingw_x86 = pentium4
-clangFlags.mingw_x86 = -cc1 -emit-obj -disable-llvm-passes -x ir -femulated-tls -target-cpu $targetCpu.mingw_x86
+clangFlags.mingw_x86 = -cc1 -emit-obj -disable-llvm-passes -x ir -femulated-tls
 clangOptFlags.mingw_x86 = -O3 -ffunction-sections
 clangNooptFlags.mingw_x86 = -O1
 linkerNoDebugFlags.mingw_x86 = -Wl,-S
@@ -975,7 +967,7 @@ dependencies.mingw_x64-wasm32 = \
 targetTriple.wasm32 = wasm32-unknown-unknown
 targetSysRoot.wasm32 = target-sysroot-4-embedded
 targetCpu.wasm32 = generic
-clangFlags.wasm32 = -cc1 -emit-obj -disable-llvm-optzns -x ir -target-cpu $targetCpu.wasm32
+clangFlags.wasm32 = -cc1 -emit-obj -disable-llvm-optzns -x ir
 clangNooptFlags.wasm32 = -O1
 clangOptFlags.wasm32 = -O3
 clangDebugFlags.wasm32 = -O0