Including <arm_neon.h> fails to compile in the sycl pass

[emtKazys]

Describe the bug

When building for ARM (cross compile or on target), including <arm_neon.h> generates compiler errors, but only in the sycl pass. Using #ifndef __SYCL_DEVICE_ONLY__ to excluding including that header will avoid the issue.

To reproduce

Example program:

// Neon.cpp
#include <arm_neon.h>

int main()
{
    return 0;
}

Command line which compiles:

clang++ neon.cpp

Command line which fails:

clang++ -fsycl neon.cpp

Example error messages:

In file included from neon.cpp:1:
/opt/intel/llvm/lib/clang/20/include/arm_neon.h:46:24: error: invalid vector element type 'poly8_t' (aka 'unsigned char')
   46 | typedef __attribute__((neon_polyvector_type(8))) poly8_t  poly8x8_t;
      |                        ^
/opt/intel/llvm/lib/clang/20/include/arm_neon.h:47:24: error: invalid vector element type 'poly8_t' (aka 'unsigned char')
   47 | typedef __attribute__((neon_polyvector_type(16))) poly8_t poly8x16_t;
      |                        ^
/opt/intel/llvm/lib/clang/20/include/arm_neon.h:48:24: error: invalid vector element type 'poly16_t' (aka 'unsigned short')
   48 | typedef __attribute__((neon_polyvector_type(4))) poly16_t poly16x4_t;
      |                        ^
/opt/intel/llvm/lib/clang/20/include/arm_neon.h:49:24: error: invalid vector element type 'poly16_t' (aka 'unsigned short')
   49 | typedef __attribute__((neon_polyvector_type(8))) poly16_t poly16x8_t;
      |                        ^
/opt/intel/llvm/lib/clang/20/include/arm_neon.h:50:24: error: invalid vector element type 'poly64_t' (aka 'unsigned long')
   50 | typedef __attribute__((neon_polyvector_type(1))) poly64_t poly64x1_t;
      |                        ^
/opt/intel/llvm/lib/clang/20/include/arm_neon.h:51:24: error: invalid vector element type 'poly64_t' (aka 'unsigned long')
   51 | typedef __attribute__((neon_polyvector_type(2))) poly64_t poly64x2_t;
      |                        ^
/opt/intel/llvm/lib/clang/20/include/arm_neon.h:7221:46: error: C-style cast from scalar 'poly8x8_t' (aka 'unsigned char') to vector 'int8x8_t' (vector of 8 'int8_t' values) of different size
 7221 |     __ret = (poly8x8_t)__builtin_neon_vadd_v((int8x8_t)__p0, (int8x8_t)__p1, 4);
      |                                              ^~~~~~~~~~~~~~
/opt/intel/llvm/lib/clang/20/include/arm_neon.h:7221:62: error: C-style cast from scalar 'poly8x8_t' (aka 'unsigned char') to vector 'int8x8_t' (vector of 8 'int8_t' values) of different size
 7221 |     __ret = (poly8x8_t)__builtin_neon_vadd_v((int8x8_t)__p0, (int8x8_t)__p1, 4);
      |                                                              ^~~~~~~~~~~~~~
/opt/intel/llvm/lib/clang/20/include/arm_neon.h:7252:47: error: C-style cast from scalar 'poly16x4_t' (aka 'unsigned short') to vector 'int8x8_t' (vector of 8 'int8_t' values) of different size
 7252 |     __ret = (poly16x4_t)__builtin_neon_vadd_v((int8x8_t)__p0, (int8x8_t)__p1, 5);
      |                                               ^~~~~~~~~~~~~~
/opt/intel/llvm/lib/clang/20/include/arm_neon.h:7252:63: error: C-style cast from scalar 'poly16x4_t' (aka 'unsigned short') to vector 'int8x8_t' (vector of 8 'int8_t' values) of different size
 7252 |     __ret = (poly16x4_t)__builtin_neon_vadd_v((int8x8_t)__p0, (int8x8_t)__p1, 5);
      |                                                               ^~~~~~~~~~~~~~
/opt/intel/llvm/lib/clang/20/include/arm_neon.h:7277:48: error: C-style cast from scalar 'poly8x16_t' (aka 'unsigned char') to vector 'int8x16_t' (vector of 16 'int8_t' values) of different size
 7277 |     __ret = (poly8x16_t)__builtin_neon_vaddq_v((int8x16_t)__p0, (int8x16_t)__p1, 36);
      |                                                ^~~~~~~~~~~~~~~
/opt/intel/llvm/lib/clang/20/include/arm_neon.h:7277:65: error: C-style cast from scalar 'poly8x16_t' (aka 'unsigned char') to vector 'int8x16_t' (vector of 16 'int8_t' values) of different size
 7277 |     __ret = (poly8x16_t)__builtin_neon_vaddq_v((int8x16_t)__p0, (int8x16_t)__p1, 36);
      |                                                                 ^~~~~~~~~~~~~~~
/opt/intel/llvm/lib/clang/20/include/arm_neon.h:7309:48: error: C-style cast from scalar 'poly64x2_t' (aka 'unsigned long') to vector 'int8x16_t' (vector of 16 'int8_t' values) of different size
 7309 |     __ret = (poly64x2_t)__builtin_neon_vaddq_v((int8x16_t)__p0, (int8x16_t)__p1, 38);
      |                                                ^~~~~~~~~~~~~~~
/opt/intel/llvm/lib/clang/20/include/arm_neon.h:7309:65: error: C-style cast from scalar 'poly64x2_t' (aka 'unsigned long') to vector 'int8x16_t' (vector of 16 'int8_t' values) of different size
 7309 |     __ret = (poly64x2_t)__builtin_neon_vaddq_v((int8x16_t)__p0, (int8x16_t)__p1, 38);
      |                                                                 ^~~~~~~~~~~~~~~
/opt/intel/llvm/lib/clang/20/include/arm_neon.h:7335:48: error: C-style cast from scalar 'poly16x8_t' (aka 'unsigned short') to vector 'int8x16_t' (vector of 16 'int8_t' values) of different size
 7335 |     __ret = (poly16x8_t)__builtin_neon_vaddq_v((int8x16_t)__p0, (int8x16_t)__p1, 37);
      |                                                ^~~~~~~~~~~~~~~
/opt/intel/llvm/lib/clang/20/include/arm_neon.h:7335:65: error: C-style cast from scalar 'poly16x8_t' (aka 'unsigned short') to vector 'int8x16_t' (vector of 16 'int8_t' values) of different size
 7335 |     __ret = (poly16x8_t)__builtin_neon_vaddq_v((int8x16_t)__p0, (int8x16_t)__p1, 37);
      |                                                                 ^~~~~~~~~~~~~~~
/opt/intel/llvm/lib/clang/20/include/arm_neon.h:8247:44: error: C-style cast from scalar 'poly8x8_t' (aka 'unsigned char') to vector 'int8x8_t' (vector of 8 'int8_t' values) of different size
 8247 |     )__builtin_neon_vbsl_v((int8x8_t)__p0, (int8x8_t)__p1, (int8x8_t)__p2, 4);
      |                                            ^~~~~~~~~~~~~~
/opt/intel/llvm/lib/clang/20/include/arm_neon.h:8247:60: error: C-style cast from scalar 'poly8x8_t' (aka 'unsigned char') to vector 'int8x8_t' (vector of 8 'int8_t' values) of different size
 8247 |     )__builtin_neon_vbsl_v((int8x8_t)__p0, (int8x8_t)__p1, (int8x8_t)__p2, 4);
      |                                                            ^~~~~~~~~~~~~~
/opt/intel/llvm/lib/clang/20/include/arm_neon.h:8276:44: error: C-style cast from scalar 'poly16x4_t' (aka 'unsigned short') to vector 'int8x8_t' (vector of 8 'int8_t' values) of different size
 8276 |     )__builtin_neon_vbsl_v((int8x8_t)__p0, (int8x8_t)__p1, (int8x8_t)__p2, 5);
      |                                            ^~~~~~~~~~~~~~
fatal error: too many errors emitted, stopping now [-ferror-limit=]

Environment

• OS: Ubuntu 22.04
• Target device and vendor: None
• DPC++ version: Custom build of Intel LLVM at v6.1.0
• Dependencies version: N/A

Additional context

No response

[AlexeySachkov]

Hi @emtKazys, thanks for the report

I assume that those errors happen during device compilation. That compilation pass targets HW-agnostic spir target which has no knowledge about any ARM-specific attributes, or built-ins, hence the errors.

Strictly speaking, we experience the same issue with regular x86, but it is handled by passing so-called aux-triple to the device compilation pass under the hood. That aux-triple makes the compiler aware of target-specific things even when we aim for spir.

My guess is that we don't set that aux-triple correctly when cross-compilation is happening, but that is just a guess.

If my initial assumption about this happening during device compilation pass is correct, then you can use #ifndef __SYCL_DEVICE_ONLY__ as a workaround. SYCL spec does not require us to support anything beyond what's in the spec and therefore any portable SYCL application won't use anything from arm_neon.h within device code. As such, you can carefully guard ARM-specific pieces of host code with the aforementioned macro so that they are skipped by the device compilation pass.

[Marcus-Forte]

I believe the issue that I posted happens for the same reason (the inclusion of arm_neon.h by Eigen), but in the newer clang-21 version.

Thanks to this post! I realised I just needed to prevent Eigen from including arm_neon.h. Defining EIGEN_ARM64_USE_SVE did it.

[emtKazys]

Hi @AlexeySachkov ,

Thanks for the reply. I agree that this is happening during the device specific pass, which is why using __SYCL_DEVICE_ONLY__ works.

As I understand your reply, you are suggesting that either;

1. I'm not setting aux-triple correctly, or
2. It's up to me to avoid including platform specific headers such as arm_neon.h.

Please clarify which is the more correct suggestions.

I don't really see 2 as being viable. I've distilled this down to a simple example, but in practice the issue comes up when including 3rd party libraries, like Eigen in @Marcus-Forte 's case. I've had this happen when including libglm and OpenCV headers. I've had to hack around the first and not use OpenCV in device code for the second, but this is limiting and requires case-by-case fixes.

Coming back to point 1, could you please explain how I might set aux-triple? I would have expected this is implicitly correct when building on an arm system. I'm also cross compiling in which case I do set --target=aarch64-linux-gnu and still have the same issue. Is this what you mean? I can't see any other clang options that may fit the bill.

Cheers.

[AlexeySachkov]

I'm not setting aux-triple correctly, or

aux-triple I'm talking about is an internal compiler option, it shouldn't be set explicitly by a user. I'm saying that our compiler sets it incorrectly for itself in a cross-compilation scenario.

UPD: I've just checked and it seems like we do set the aux-triple correctly (or at least we don't set it to x86 which is what I though happens here), so it must be some other compiler changes which are missing:

$ ./bin/clang++ -fsycl --target=aarch64-linux-gnu r.cpp -###
clang version 21.0.0git (https://github.com/intel/llvm 0a255cfb065e2c03edf6740d732ae1c4aafc8ecd)
Target: aarch64-unknown-linux-gnu
Thread model: posix
InstalledDir: intel/llvm/build/bin
Build config: +assertions
 "intel/llvm/build/bin/clang-21" "-cc1" "-triple" "spir64-unknown-unknown" "-aux-triple" "aarch64-unknown-linux-gnu" "-fsycl-is-device" "-fdeclare-spirv-
builtins" "-O2" "-mllvm" "-sycl-opt" "-fenable-sycl-dae" "-Wno-sycl-strict" "-fsycl-int-header=/tmp/r-header-a71258.h" "-fsycl-int-footer=/tmp/r-footer-052cf4.h" "-sycl-std=2020" "-ffin
e-grained-bitfield-accesses" "-fsycl-unique-prefix=uid697d81bddaec8a2b" "-D__SYCL_ANY_DEVICE_HAS_ANY_ASPECT__=1" "-Wspir-compat" "-fno-offload-use-alloca-addrspace-for-srets" "-emit-llv
m-bc" "-emit-llvm-uselists" "-disable-free" "-clear-ast-before-backend" "-main-file-name" "r.cpp" "-mrelocation-model" "static" "-mframe-pointer=all" "-fmath-errno" "-ffp-contract=on" "
-fno-rounding-math" "-mconstructor-aliases" "-aux-target-cpu" "generic" "-aux-target-feature" "+v8a" "-aux-target-feature" "+fp-armv8" "-aux-target-feature" "+neon" "-debugger-tuning=gd
b" "-fdebug-compilation-dir=intel/llvm/build" "-fcoverage-compilation-dir=intel/llvm/build" "-resource-dir" "/localdisk2/
asachkov/github.com/intel/llvm/build/lib/clang/21" "-internal-isystem" "intel/llvm/build/bin/../include/sycl/stl_wrappers" "-internal-isystem" "/localdis
k2/asachkov/github.com/intel/llvm/build/bin/../include" "-internal-isystem" "intel/llvm/build/bin/../include/sycl/stl_wrappers" "-internal-isystem" "/loc
aldisk2/asachkov/github.com/intel/llvm/build/bin/../include" "-internal-isystem" "intel/llvm/build/lib/clang/21/include" "-internal-isystem" "/usr/local/
include" "-internal-externc-isystem" "/include" "-internal-externc-isystem" "/usr/include" "-internal-isystem" "intel/llvm/build/lib/clang/21/include" "-
internal-isystem" "/usr/local/include" "-internal-externc-isystem" "/include" "-internal-externc-isystem" "/usr/include" "-std=c++17" "-fdeprecated-macro" "-ferror-limit" "19" "-fmessag
e-length=185" "-fgpu-rdc" "-fno-signed-char" "-fgnuc-version=4.2.1" "-fskip-odr-check-in-gmf" "-fcxx-exceptions" "-fexceptions" "-fcolor-diagnostics" "-faddrsig" "-D__GCC_HAVE_DWARF2_CF
I_ASM=1" "-o" "/tmp/r-da7d81.bc" "-x" "c++" "r.cpp"