LLVM ERROR: Cannot select: intrinsic %llvm.aarch64.neon.vcvtfp2fxs

[rscottmanley]

Getting an abort below when trying to compile for -mcpu=neoverse-v1 at 256 bits. I think it may have something to do with the const vector?

Compiling with llc bug.ll -mcpu=neoverse-v1 -aarch64-sve-vector-bits-min=256 -o bug.s

target triple = "arm64-linux"

; Function Attrs: mustprogress nofree noinline norecurse nosync nounwind null_pointer_is_valid willreturn
define dso_local void @crash_(i64* nocapture readonly %a, i64* nocapture writeonly %b) local_unnamed_addr #0 align 16 {
L.entry:
  %0 = bitcast i64* %a to <8 x float>*
  %1 = load <8 x float>, <8 x float>* %0, align 4
  %2 = fmul  <8 x float> %1, <float 4.096000e+03, float 4.096000e+03, float 4.096000e+03, float 4.096000e+03, float 4.096000e+03, float 4.096000e+03, float 4.096000e+03, float 4.096000e+03>
  %3 = fmul  <8 x float> %2, <float 4.096000e+03, float 4.096000e+03, float 4.096000e+03, float 4.096000e+03, float 4.096000e+03, float 4.096000e+03, float 4.096000e+03, float 4.096000e+03>
  %4 = fptosi <8 x float> %3 to <8 x i32>
  %5 = bitcast i64* %b to <8 x i32>*
  store <8 x i32> %4, <8 x i32>* %5, align 4
  %6 = getelementptr i64, i64* %a, i64 4
  %7 = bitcast i64* %6 to <8 x float>*
  %8 = load <8 x float>, <8 x float>* %7, align 4
  %9 = fmul  <8 x float> %8, <float 4.096000e+03, float 4.096000e+03, float 4.096000e+03, float 4.096000e+03, float 4.096000e+03, float 4.096000e+03, float 4.096000e+03, float 4.096000e+03>
  %10 = fmul  <8 x float> %9, <float 4.096000e+03, float 4.096000e+03, float 4.096000e+03, float 4.096000e+03, float 4.096000e+03, float 4.096000e+03, float 4.096000e+03, float 4.096000e+03>
  %11 = fptosi <8 x float> %10 to <8 x i32>
  %12 = getelementptr i64, i64* %b, i64 4
  %13 = bitcast i64* %12 to <8 x i32>*
  store <8 x i32> %11, <8 x i32>* %13, align 4
  %14 = getelementptr i64, i64* %a, i64 8
  %15 = bitcast i64* %14 to <8 x float>*
  %16 = load <8 x float>, <8 x float>* %15, align 4
  %17 = fmul  <8 x float> %16, <float 4.096000e+03, float 4.096000e+03, float 4.096000e+03, float 4.096000e+03, float 4.096000e+03, float 4.096000e+03, float 4.096000e+03, float 4.096000e+03>
  %18 = fmul  <8 x float> %17, <float 4.096000e+03, float 4.096000e+03, float 4.096000e+03, float 4.096000e+03, float 4.096000e+03, float 4.096000e+03, float 4.096000e+03, float 4.096000e+03>
  %19 = fptosi <8 x float> %18 to <8 x i32>
  %20 = getelementptr i64, i64* %b, i64 8
  %21 = bitcast i64* %20 to <8 x i32>*
  store <8 x i32> %19, <8 x i32>* %21, align 4
  ret void
}

attributes #0 = { mustprogress nofree noinline norecurse nosync nounwind null_pointer_is_valid willreturn "target-cpu"="neoverse-v1" }

[rscottmanley]

Not sure how to add people to notifications, apologies if doing this incorrectly. Seems weird I can't edit the bug after submission either.
@paulwalker-arm
@mcinally

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

[brads55]

Review to fix this: https://reviews.llvm.org/D119428

[rscottmanley]

Will this be fixed in LLVM 14?

[peterwaller-arm]

It should be, there is a backport request you can follow here: #53754