[PATCH] D106353: [PowerPC] use lfiwax/lfiwzx for scalar_to_vector + load at PWR7

[ChenZheng via Phabricator via llvm-commits]

shchenz added a comment.

After more investigation, I think we don't need this patch.

`scalar_to_vector (scalar_load)` will put the scalar load result at the first element of the vector result and set other elements of the vector result as undef.

So it is always correct to convert `scalar_to_vector (scalar_load)` to a splat_load(all vector elements are the same). And we already improve the code gen for splat_load in https://reviews.llvm.org/D106555. I verified all the types in this patch v4i32/v4f32/v2i64_signext/v2i64_zeroext, it can be handled in patch https://reviews.llvm.org/D106555. We already recognize all of them as splat load.

The only unexpected case is the case `unadjusted_lxvwsx` in file `unadjusted_lxvwsx`. In this case, it has following patterns:

  t7: v4i8,ch = load<(load (s32) from %ir.0)> t0, t2, undef:i64

This will be legalized as:

  Legalizing node: t7: v4i8,ch = load<(load (s32) from %ir.0)> t0, t2, undef:i64
  Analyzing result type: v4i8
  Widen node result 0: t7: v4i8,ch = load<(load (s32) from %ir.0)> t0, t2, undef:i64
  
  Creating new node: t15: i32,ch = load<(load (s32) from %ir.0)> t0, t2, undef:i64
  Creating new node: t16: v4i32 = scalar_to_vector t15
  Creating new node: t17: v16i8 = bitcast t16

T


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D106353/new/

https://reviews.llvm.org/D106353