llvm - 为什么 LLVM SIMD 不向量化此代码？

Question

我有以下 IR:

; ModuleID = 'vec.ir'
target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-apple-darwin15.3.0"

define void @patch(i64) {
entry:
  %1 = load float, float* inttoptr (i64 4388240000 to float*)
  %2 = load float, float* inttoptr (i64 4387644544 to float*)
  %3 = fadd float %1, %2
  %4 = load float, float* inttoptr (i64 4387729024 to float*)
  %5 = fadd float %1, %4
  %6 = load float, float* inttoptr (i64 4387730560 to float*)
  %7 = fadd float %1, %6
  %8 = load float, float* inttoptr (i64 4387513984 to float*)
  %9 = fadd float %1, %8
  store float %3, float* inttoptr (i64 4371309760 to float*)
  call void @__tickValue(i64 105553117467608, i64 %0)
  store float %5, float* inttoptr (i64 4371851456 to float*)
  call void @__tickValue(i64 105553117465688, i64 %0)
  store float %7, float* inttoptr (i64 4371574976 to float*)
  call void @__tickValue(i64 105553117465528, i64 %0)
  store float %9, float* inttoptr (i64 4371576512 to float*)
  call void @__tickValue(i64 105553117466648, i64 %0)
  ret void
}

declare void @__tickValue(i64, i64)

当我运行 /usr/local/opt/llvm/bin/opt -S -O3 vec.ir > vec-opt.ir 时，我得到:

; ModuleID = 'vec.ir'
target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-apple-darwin15.3.0"

define void @patch(i64) {
entry:
  %1 = load float, float* inttoptr (i64 4388240000 to float*), align 128
  %2 = load float, float* inttoptr (i64 4387644544 to float*), align 128
  %3 = fadd float %1, %2
  %4 = load float, float* inttoptr (i64 4387729024 to float*), align 128
  %5 = fadd float %1, %4
  %6 = load float, float* inttoptr (i64 4387730560 to float*), align 128
  %7 = fadd float %1, %6
  %8 = load float, float* inttoptr (i64 4387513984 to float*), align 128
  %9 = fadd float %1, %8
  store float %3, float* inttoptr (i64 4371309760 to float*), align 64
  tail call void @__tickValue(i64 105553117467608, i64 %0)
  store float %5, float* inttoptr (i64 4371851456 to float*), align 64
  tail call void @__tickValue(i64 105553117465688, i64 %0)
  store float %7, float* inttoptr (i64 4371574976 to float*), align 64
  tail call void @__tickValue(i64 105553117465528, i64 %0)
  store float %9, float* inttoptr (i64 4371576512 to float*), align 64
  tail call void @__tickValue(i64 105553117466648, i64 %0)
  ret void
}

declare void @__tickValue(i64, i64)

根据 http://llvm.org/docs/Vectorizers.html#the-slp-vectorizer ，我希望 fadd 指令可以合并。

我如何才能确定有关优化器未进行矢量化的原因的更多信息？

Answer 1

LLVM 具有内部成本模型，可以自动检测 SIMD 向量化是否有益。您可以通过 adding some flags to your build line 获得诊断信息

您还可以尝试通过添加 some directives 来“强制”矢量化在您的代码中。

如果你从 llvm IR 而不是源代码开始，你仍然有命令 line switches for opt