Missed autovectorization opportunity (aarch64, SVE, gather load)

[MatzeB]

Got a report of a simple loop that should autovectorize but does not do so on aarch64 (but does on x86 / AVX512). Repro:

#include <stdint.h>
#include <stdlib.h>

void noAutovec(uint32_t* __restrict ip, float* __restrict src, float* __restrict dst, size_t n) {
   //  If you encourage the compiler with the `#pragma` this does autovectorize.
   //  #pragma clang loop vectorize(enable)
    for (size_t i=0; i<n; ++i) {
      uint32_t idx = ip[i];
      dst[i] = src[idx];
    }
}

This vectorizes on x86 (clang -march=haswell -mavx512f -O3) but does not on aarch64 in my experiments (clang -target aarch64-redhat-linux-gnu -march=armv9-a+sve2+fp16).

Using #pragma clang loop vectorize(enable) makes vectorization work on aarch64. So this hints at the cost-model rejecting things (I assume vectorization should be beneficial when SVE is available).

(this mirrors meta T222824954 )

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

Author: Matthias Braun (MatzeB)

Got a report of a simple loop that should autovectorize but does not do so on aarch64 (but does on x86 / AVX512). Repro:

#include <stdint.h>
#include <stdlib.h>

void noAutovec(uint32_t* __restrict ip, float* __restrict src, float* __restrict dst, size_t n) {
   //  If you encourage the compiler with the `#pragma` this does autovectorize.
   //  #pragma clang loop vectorize(enable)
    for (size_t i=0; i<n; ++i) {
      uint32_t idx = ip[i];
      dst[i] = src[idx];
    }
}

This vectorizes on x86 (clang -march=haswell -mavx512f -O3) but does not on aarch64 in my experiments (clang -target aarch64-redhat-linux-gnu -march=armv9-a+sve2+fp16).

Using #pragma clang loop vectorize(enable) makes vectorization work on aarch64. So this hints at the cost-model rejecting things (I assume vectorization should be beneficial when SVE is available).

(this mirror meta T222824954 )

[davemgreen]

Gather/Scatter can be difficult, as they tend to be expensive operations (and can sometimes lead to more messy vectorization).

It will vectorize with a lower -sve-gather-overhead: https://godbolt.org/z/a1zsreejc
The default is 10. (X86 seems to have a GatherOverhead of 1024 without avx512! But is only 2 with).