lib/eal/arm/include/rte_vect.h fails to compile with clang14 for 32bit ARM

[Roger Melton (rmelton)]

Hey folks,

We are building DPDK with clang14 for a 32bit armv8-a based CPU and ran into a compile error with the following from lib/eal/arm/include/rte_vect.h:


#if (defined(RTE_ARCH_ARM) && defined(RTE_ARCH_32)) || \
(defined(RTE_ARCH_ARM64) && RTE_CC_IS_GNU<https://elixir.bootlin.com/dpdk/v24.11/C/ident/RTE_CC_IS_GNU> && (GCC_VERSION<https://elixir.bootlin.com/dpdk/v24.11/C/ident/GCC_VERSION> < 70000))
/* NEON intrinsic vcopyq_laneq_u32() is not supported in ARMv7-A(AArch32)
 * On AArch64, this intrinsic is supported since GCC version 7.
 */
static inline uint32x4_t
vcopyq_laneq_u32<https://elixir.bootlin.com/dpdk/v24.11/C/ident/vcopyq_laneq_u32>(uint32x4_t a, const int lane_a,
                 uint32x4_t b, const int lane_b)
{
        return vsetq_lane_u32(vgetq_lane_u32(b, lane_b), a, lane_a);
}
#endif

clang14 compile fails as follows:

In file included from ../../../../../../cisco-dpdk-upstream-arm-clang-fixes.git/lib/eal/common/eal_common_options.c:36:
../../../../../../cisco-dpdk-upstream-arm-clang-fixes.git/lib/eal/arm/include/rte_vect.h:80:24: error: argument to '__builtin_neon_vgetq_lane_i32' must be a constant integer
return vsetq_lane_u32(vgetq_lane_u32(b, lane_b), a, lane_a);
^ ~~~~~~
/auto/binos-tools/llvm14/llvm-14.0-p24/lib/clang/14.0.5/include/arm_neon.h:7697:22: note: expanded from macro 'vgetq_lane_u32'
__ret = (uint32_t) __builtin_neon_vgetq_lane_i32((int32x4_t)__s0, __p1); \
^ ~~~~
/auto/binos-tools/llvm14/llvm-14.0-p24/lib/clang/14.0.5/include/arm_neon.h:24148:19: note: expanded from macro 'vsetq_lane_u32'
uint32_t __s0 = __p0; \
^~~~
In file included from ../../../../../../cisco-dpdk-upstream-arm-clang-fixes.git/lib/eal/common/eal_common_options.c:36:
../../../../../../cisco-dpdk-upstream-arm-clang-fixes.git/lib/eal/arm/include/rte_vect.h:80:9: error: argument to '__builtin_neon_vsetq_lane_i32' must be a constant integer
return vsetq_lane_u32(vgetq_lane_u32(b, lane_b), a, lane_a);
^ ~~~~~~
/auto/binos-tools/llvm14/llvm-14.0-p24/lib/clang/14.0.5/include/arm_neon.h:24150:24: note: expanded from macro 'vsetq_lane_u32'
__ret = (uint32x4_t) __builtin_neon_vsetq_lane_i32(__s0, (int32x4_t)__s1, __p2); \
^ ~~~~
2 errors generated.


clang14 does appear to support the vcopyq_laneq_u32() intrinsic, s0 we want to skip the conditional implementation.

Two approaches I have tested to resolve the error are:

1) skip if building with clang:

#if !defined(__clang__) && ((defined(RTE_ARCH_ARM) && defined(RTE_ARCH_32)) || \
72 (defined(RTE_ARCH_ARM64) && RTE_CC_IS_GNU && (GCC_VERSION < 70000)))

2) skip if not building for ARMv7:

#if (defined(RTE_ARCH_ARMv7) && defined(RTE_ARCH_32)) || \
(defined(RTE_ARCH_ARM64) && RTE_CC_IS_GNU && (GCC_VERSION < 70000))

Both address our immediate problem, but may not be a appropriate for all cases.

Can anyone suggest the proper way to address this?  I'll be submitting an patch once I have a solution that is acceptable to the community.

Regards,
Roger





-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mails.dpdk.org/archives/dev/attachments/20241202/7f816cdb/attachment-0001.htm>