[BUG] neon intrinsics fail to compile

[zchrissirhcz]

Description

The following code failed to compile when the template argument nc=4:

#include <stdint.h>
#include <arm_neon.h>

template<int32_t nc>
void test()
{
    float32x4_t tcurr = {1.f, 2.f, 3.f, 4.f};
    float32x4_t tnext = {5.f, 6.f, 7.f, 8.f};
    float32x4_t t_right;
    if (nc == 4) {
        t_right = tnext;
    } else {
        t_right = vextq_f32(tcurr, tnext, nc);
    }
}

int main()
{
    test<1>(); // ok
    test<3>(); // ok
    test<4>(); // compile error: argument value 4 is outside the valid range [0, 3]

    return 0;
}

compile script

#!/bin/bash

ANDROID_NDK=~/soft/android-ndk-r24-beta2
TOOLCHAIN=$ANDROID_NDK/build/cmake/android.toolchain.cmake

BUILD_DIR=android-arm64
mkdir -p $BUILD_DIR
cd $BUILD_DIR

cmake -G Ninja \
    -DCMAKE_TOOLCHAIN_FILE=$TOOLCHAIN \
    -DANDROID_LD=lld \
    -DANDROID_ABI="arm64-v8a" \
    -DANDROID_PLATFORM=android-24 \
    -DCMAKE_BUILD_TYPE=Release \
    ../..

#ninja
#cmake --build . --verbose
cmake --build .

cd ..

Error message

(base) zz@arcsoft-43% ./android-arm64-build.sh 
-- Android: Targeting API '24' with architecture 'arm64', ABI 'arm64-v8a', and processor 'aarch64'
-- Android: Selected unified Clang toolchain
-- The C compiler identification is Clang 14.0.0
-- The CXX compiler identification is Clang 14.0.0
-- Detecting C compiler ABI info
-- Detecting C compiler ABI info - done
-- Check for working C compiler: /home/zz/soft/android-ndk-r24-beta2/toolchains/llvm/prebuilt/linux-x86_64/bin/clang - skipped
-- Detecting C compile features
-- Detecting C compile features - done
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Check for working CXX compiler: /home/zz/soft/android-ndk-r24-beta2/toolchains/llvm/prebuilt/linux-x86_64/bin/clang++ - skipped
-- Detecting CXX compile features
-- Detecting CXX compile features - done
-- Configuring done
-- Generating done
CMake Warning:
  Manually-specified variables were not used by the project:

    ANDROID_LD

-- Build files have been written to: /home/zz/work/test/ndk_bug/build/android-arm64
[1/2] Building CXX object CMakeFiles/testbed.dir/main.cpp.o
FAILED: CMakeFiles/testbed.dir/main.cpp.o 
/home/zz/soft/android-ndk-r24-beta2/toolchains/llvm/prebuilt/linux-x86_64/bin/clang++ --target=aarch64-none-linux-android24 --sysroot=/home/zz/soft/android-ndk-r24-beta2/toolchains/llvm/prebuilt/linux-x86_64/sysroot   -DANDROID -fdata-sections -ffunction-sections -funwind-tables -fstack-protector-strong -no-canonical-prefixes -D_FORTIFY_SOURCE=2 -Wformat -Werror=format-security -fexceptions -frtti -stdlib=libc++ -O3 -DNDEBUG -fPIE -std=gnu++11 -MD -MT CMakeFiles/testbed.dir/main.cpp.o -MF CMakeFiles/testbed.dir/main.cpp.o.d -o CMakeFiles/testbed.dir/main.cpp.o -c /home/zz/work/test/ndk_bug/main.cpp
/home/zz/work/test/ndk_bug/main.cpp:13:19: error: argument value 4 is outside the valid range [0, 3]
        t_right = vextq_f32(tcurr, tnext, nc);
                  ^                       ~~
/home/zz/soft/android-ndk-r24-beta2/toolchains/llvm/prebuilt/linux-x86_64/lib64/clang/14.0.0/include/arm_neon.h:7103:25: note: expanded from macro 'vextq_f32'
  __ret = (float32x4_t) __builtin_neon_vextq_v((int8x16_t)__s0, (int8x16_t)__s1, __p2, 41); \
                        ^                                                        ~~~~
/home/zz/work/test/ndk_bug/main.cpp:21:5: note: in instantiation of function template specialization 'test<4>' requested here
    test<4>(); // compile error: argument value 4 is outside the valid range [0, 3]
    ^
1 error generated.
ninja: build stopped: subcommand failed.

Affected versions

r23, r24

Canary version

No response

Host OS

Linux

Host OS version

Ubuntu 20.04

Affected ABIs

arm64-v8a

Build system

CMake

Other build system

No response

minSdkVersion

24

Device API level

Not related to device API level.

[zchrissirhcz]

Tried the following Android NDK versions, all failed:

• r21b
• r21e
• r22b
• r23b
• r24-beta2

[Over17]

Your template specialization with <4> leads to this code emitted:

t_right = vextq_f32(tcurr, tnext, 4);

which is illegal (last f32 index in a 128-bit vector is 3) and leads to the above-mentioned error. The index range verification happens before the dead code elimination is kicked in (later on in opt stage).

Based on the above, looks like intended behaviour to me.

[DanAlbert]

Makes sense to me. Thanks for the diagnosis :)