[llvm] [X86] Attempt to canonicalize vXf64 SHUFPD shuffle masks with undef elts to improve further folding (PR #116419)

[Mikhail Goncharov via llvm-commits]

metaflow wrote:

FYI I see that https://github.com/halide/Halide/blob/main/test/correctness/stage_strided_loads.cpp got flaky after this change and checks at llvm/lib/Target/X86/X86ISelLowering.cpp:9940 in unsigned int getSHUFPDImm(ArrayRef<int>): 0 <= FirstIndex && FirstIndex < 4 && "All undef shuffle mask"

https://github.com/llvm/llvm-project/pull/116419