[PATCH] D110926: [GlobalISel] Support vectors in LegalizerHelper::narrowScalarMul

[Jay Foad via Phabricator via llvm-commits]

foad added a comment.

The problem here is that having narrowed:

  %2:_(<2 x s64>) = G_UMULH %0:_, %1:_

into:

  %3:_(<2 x s32>), %4:_(<2 x s32>) = G_UNMERGE_VALUES %0:_(<2 x s64>)
  %5:_(<2 x s32>), %6:_(<2 x s32>) = G_UNMERGE_VALUES %1:_(<2 x s64>)
  %7:_(<2 x s32>) = G_MUL %3:_, %5:_
  %8:_(<2 x s32>) = G_MUL %4:_, %5:_
  %9:_(<2 x s32>) = G_MUL %3:_, %6:_
  %10:_(<2 x s32>) = G_UMULH %3:_, %5:_
  %11:_(<2 x s32>), %12:_(<2 x s1>) = G_UADDO %8:_, %9:_
  %13:_(<2 x s32>) = G_ZEXT %12:_(<2 x s1>)
  %14:_(<2 x s32>), %15:_(<2 x s1>) = G_UADDO %11:_, %10:_
  %16:_(<2 x s32>) = G_ZEXT %15:_(<2 x s1>)
  %17:_(<2 x s32>) = G_ADD %13:_, %16:_
  %18:_(<2 x s32>) = G_MUL %4:_, %6:_
  %19:_(<2 x s32>) = G_UMULH %4:_, %5:_
  %20:_(<2 x s32>) = G_UMULH %3:_, %6:_
  %21:_(<2 x s32>), %22:_(<2 x s1>) = G_UADDO %18:_, %19:_
  %23:_(<2 x s32>) = G_ZEXT %22:_(<2 x s1>)
  %24:_(<2 x s32>), %25:_(<2 x s1>) = G_UADDO %21:_, %20:_
  %26:_(<2 x s32>) = G_ZEXT %25:_(<2 x s1>)
  %27:_(<2 x s32>) = G_ADD %23:_, %26:_
  %28:_(<2 x s32>), %29:_(<2 x s1>) = G_UADDO %24:_, %17:_
  %30:_(<2 x s32>) = G_ZEXT %29:_(<2 x s1>)
  %31:_(<2 x s32>) = G_ADD %27:_, %30:_
  %32:_(<2 x s32>) = G_UMULH %4:_, %6:_
  %33:_(<2 x s32>) = G_ADD %32:_, %31:_

we need a way to merge %28:_(<2 x s32>) (a vector of the low parts of each element of the result) and %33:_(<2 x s32>) (same for the high parts) into a <2 x s64> result. I don't know how to do that. G_MERGE_VALUES only works on scalars, though I don't see why it couldn't be extended to work element-wise on vectors that all have the same length. And I can't see any existing helper functions in LegalizerHelper.cpp that do what I want either.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D110926/new/

https://reviews.llvm.org/D110926