[PATCH] D26066: [PowerPC] Improvements for BUILD_VECTOR Vol. 4

[Hal Finkel via llvm-commits]

hfinkel added a comment.

>   I think that keeping a clearly inferior code generation choice for a pattern when it can easily be improved is fairly hard to justify. 

I agree. Clearly there are limits (i.e. things that InstCombine should fix), but for things that are not clearly anti-canonical, having a high-quality backend means trying to do a good job on those inputs. There are lots of frontends, people using vector intrinsics, etc.

>   There is a combine of a build vector of int-to-fp nodes into an int-to-fp of a build vector node but not with the conversions going the other way around. So there's likely a good reason not to do this on some targets. My attempt to implement it there also causes failures in the ARM target which means that ARM somehow depends on this not happening. Again, rather than investigating whether this is a win on all targets (and whether ARM has a bug exposed by this), I felt that doing this for PPC where it is clearly a win is a better use of my time. 

Did you try adding the corresponding code as a target-specific DAGCombine?


Repository:
  rL LLVM

https://reviews.llvm.org/D26066