[PATCH] D113852: [X86][Costmodel] `trunc v32i16 to v64i8` can appear after legalization, cost is same as for `trunc v32i16 to v32i8`

[Roman Lebedev via Phabricator via llvm-commits]

lebedev.ri added a comment.

In D113852#3131098 <https://reviews.llvm.org/D113852#3131098>, @RKSimon wrote:

> Have you been able to step through the mul fix/overflow cost calcs - I assume its the trunc <64 x i16> -> <64 x i8> cost calc that is causing this?

Thinking about it, i think this is working as expected,
e.g. `  %V64i8 = mul <64 x i8> undef, <i8 2, i8 4, i8 8, i8 16, i8 2, i8 4, i8 8, i8 16, i8 2, i8 4, i8 8, i8 16, i8 2, i8 4, i8 8, i8 16, i8 2, i8 4, i8 8, i8 16, i8 2, i8 4, i8 8, i8 16, i8 2, i8 4, i8 8, i8 16, i8 2, i8 4, i8 8, i8 16, i8 2, i8 4, i8 8, i8 16, i8 2, i8 4, i8 8, i8 16, i8 2, i8 4, i8 8, i8 16, i8 2, i8 4, i8 8, i8 16, i8 2, i8 4, i8 8, i8 16, i8 2, i8 4, i8 8, i8 16, i8 2, i8 4, i8 8, i8 16, i8 2, i8 4, i8 8, i8 16>`
queries `zext <64 x i8> to <64 x i16>` + `trunc <64 x i16> to <64 x i8>` costs,
and the patch adds said trunc cost, so we no longer fallback to scalarization cost,
which can be arbitrarily wrong. Not quite sure if anything is broken..


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D113852/new/

https://reviews.llvm.org/D113852