[PATCH] D142084: [RFC][X86][MemFold] Upgrade the mechanism of auto-generated Memory Folding Table

[Bing Yu via Phabricator via llvm-commits]

yubing added inline comments.


================
Comment at: llvm/lib/Target/X86/X86MemFoldTables.inc:4009
   {X86::VMINSSZrr_Intkz, X86::VMINSSZrm_Intkz, TB_NO_REVERSE},
-  {X86::VMOVAPDZ128rrk, X86::VMOVAPDZ128rmk, TB_NO_REVERSE|TB_ALIGN_16},
-  {X86::VMOVAPDZ256rrk, X86::VMOVAPDZ256rmk, TB_NO_REVERSE|TB_ALIGN_32},
-  {X86::VMOVAPDZrrk, X86::VMOVAPDZrmk, TB_NO_REVERSE|TB_ALIGN_64},
-  {X86::VMOVAPSZ128rrk, X86::VMOVAPSZ128rmk, TB_NO_REVERSE|TB_ALIGN_16},
-  {X86::VMOVAPSZ256rrk, X86::VMOVAPSZ256rmk, TB_NO_REVERSE|TB_ALIGN_32},
-  {X86::VMOVAPSZrrk, X86::VMOVAPSZrmk, TB_NO_REVERSE|TB_ALIGN_64},
+  {X86::VMOVAPDZ128rrk, X86::VMOVAPDZ128rmk, TB_ALIGN_16},
+  {X86::VMOVAPDZ256rrk, X86::VMOVAPDZ256rmk, TB_ALIGN_32},
----------------
craig.topper wrote:
> Would this allow unfolding a VMOVAPDZ128rmk to VMOVAPDZ128rrk + VMOVAPDZ128rm? That would be a bug.
@craig.topper 
https://discourse.llvm.org/t/auto-generate-the-memory-folding-tables/61100/19
i remember we always fold whole load intrinsics into masked arithmetic. does it happen to masked movereg instruction as well?(only wholeload+ movereg => mask.load)

besides, do we need to add TB_NO_REVERSE for {X86::VMOVAPDZ128rr, X86::VMOVAPDZ128rm, TB_ALIGN_16}?


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D142084/new/

https://reviews.llvm.org/D142084