[PATCH] D52965: [X86][AVX] Improve vXi64 UITOFP vXf64/vXf32 support (P38226/PR38970)

[Simon Pilgrim via Phabricator via llvm-commits]

RKSimon created this revision.
RKSimon added reviewers: craig.topper, spatel, andreadb, efriedma, hfinkel.

An initial attempt to try and improve vXi64 UITOFP conversions:

vXi64-vXf64 - perform this as a true vectorization instead of (partially vectorized) scalar conversions by adding vector support to ExpandLegalINT_TO_FP)
vXi64-vXf32 - SSE customized versions of the ExpandLegalINT_TO_FP code, avoiding a lot of branches that were often poorly predicted

There's still room for improvement:

uitofp_4i64_to_4f64 - AVX1 codegen should be able to perform the vpsrlq xmm shifts as ymm (v8f32) shuffles
uitofp_Xi64_to_Xf32 - some of the BLENDV cases should be selected from the sign bit directly and not need a shift/comparison


Repository:
  rL LLVM

https://reviews.llvm.org/D52965

Files:
  lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
  lib/Target/X86/X86ISelLowering.cpp
  test/CodeGen/X86/vec_int_to_fp.ll

-------------- next part --------------
A non-text attachment was scrubbed...
Name: D52965.168577.patch
Type: text/x-patch
Size: 50142 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20181006/a6ac44e0/attachment-0001.bin>