[PATCH] D88273: [SDAG] Do not convert undef to 0 when folding CONCAT/BUILD_VECTOR

[Krzysztof Parzyszek via Phabricator via llvm-commits]

kparzysz created this revision.
kparzysz added a reviewer: craig.topper.
Herald added a subscriber: hiraditya.
Herald added a project: LLVM.
kparzysz requested review of this revision.

When calling `DAG.getNode(ISD::CONCAT_VECTOR, ...)` with arguments that are `BUILD_VECTOR` and/or `undef`, the DAG will try to fold it into one large `BUILD_VECTOR`.  In the process of preparing the operands, it may perform sign- or zero-extension of all values having width less than the required one.  This includes `undef`s, which causes them to become zeros.  This conversion from `undef` to "def" is not necessary and may inhibit further optimizations.

Concrete example: the original code had this operation in it:

  v8i8 = BUILD_VECTOR t109, t109, t109, t109, t109, t109, t109, undef:i32

The operands were all of a legal type `i32`, `i8` as a scalar is not legal on this target.  The vector was then widened to `v128i8` by concatenating it with a number of `undef` vectors.  The folding described above broke up all these undef vectors into individual elements, and extended them to `i32`, which resulted in (unnecessarily) appending a number of 0s instead.


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D88273

Files:
  llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
  llvm/test/CodeGen/Hexagon/autohvx/isel-undef-not-zero.ll
  llvm/test/CodeGen/X86/vec-strict-cmp-sub128.ll

-------------- next part --------------
A non-text attachment was scrubbed...
Name: D88273.294210.patch
Type: text/x-patch
Size: 7639 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20200925/f460f0ff/attachment.bin>