为什么'；LLVM SIMD是否将此代码矢量化？

我有以下资料：

; 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，我希望将

fadd

指令组合在一起

---

我如何确定优化器为什么不矢量化的更多信息？

LLVM有内部成本模型，可以自动检测SIMD矢量化是否有益。您可以通过以下方式获得诊断信息：

您还可以尝试通过添加代码来“强制”矢量化

如果您从llvm IR开始，而不是从源代码开始，那么您仍然有命令