CODEWITHSUNDEEP
assemblyrust64-bitsimdauto-vectorization
Unlikus
Unlikus

Reputation: 1858

Rust auto vectorization of exp function

I am trying to get some good codegen for a Rust binary in x64 but it seems very hard to get simd operations for exp.

I have a simplified code example: https://godbolt.org/z/4MY34T1zj

The code should be easy to vectorize but I get scalar calls to glibc exp.

I can't show with godbold, but the same happens when I use rust libm and also link time optimizations.

What are the strategies to get vectorized codegen for floating point functions like exp, log, etc?

Edit: The LLVM-IR looks to me (who cannot really read llvm-ir) like it has all it needs to vectorize:

vector.body:
  %index = phi i64 [ 0, %vector.ph ], [ %index.next, %vector.body ]
  %0 = getelementptr inbounds [0 x double], ptr %x.0, i64 0, i64 %index
  %wide.load = load <4 x double>, ptr %0, align 8
  %1 = tail call <4 x double> @llvm.exp.v4f64(<4 x double> %wide.load)
  store <4 x double> %1, ptr %0, align 8
  %index.next = add nuw i64 %index, 4
  %2 = icmp eq i64 %index.next, %n.vec
  br i1 %2, label %middle.block, label %vector.body

So it seems to be a limitation of llvm

This seems relevant: https://reviews.llvm.org/D95373

Upvotes: 1

Views: 83

Answers (0)

Related Questions