LowerSwitch: Avoid inserting NewDefault block
authorRuiling Song <ruiling.song@amd.com>
Tue, 12 Apr 2022 03:25:33 +0000 (11:25 +0800)
committerRuiling Song <ruiling.song@amd.com>
Thu, 14 Apr 2022 05:30:56 +0000 (13:30 +0800)
commit1e01f95057a702658a88879223586fde0122f038
tree066537953114bc05566a017264ebd0a5dd4d6cee
parent7c87d75d74f3c2943b286b239ec6ff96fc5109c7
LowerSwitch: Avoid inserting NewDefault block

The NewDefault was used to simplify the updating of PHI nodes, but it
causes some inefficiency for target that will run structurizer later. For
example, for a simple two-case switch, the extra NewDefault is causing
unstructured CFG like:

        O
       / \
      O   O
     / \ / \
    C1  ND C2
     \  |  /
      \ | /
        D

The change is to avoid the ND(NewDefault) block, that is we will get a
structured CFG for above example like:

        O
       / \
      /   \
     O     O
    / \   / \
   C1  \ /  C2
    \-> D <-/

The IR change introduced by this patch should be trivial to other targets,
so I am doing this unconditionally.

Fall-through among the cases will also cause unstructured CFG, but it need
more work and will be addressed in a separate change.

Reviewed by: arsenm

Differential Revision: https://reviews.llvm.org/D123607
llvm/lib/Transforms/Utils/LowerSwitch.cpp
llvm/test/CodeGen/AMDGPU/multilevel-break.ll
llvm/test/Transforms/FixIrreducible/switch.ll
llvm/test/Transforms/LowerSwitch/2014-06-23-PHIlowering.ll
llvm/test/Transforms/LowerSwitch/do-not-handle-impossible-values.ll
llvm/test/Transforms/LowerSwitch/feature.ll
llvm/test/Transforms/StructurizeCFG/interleaved-loop-order.ll
llvm/test/Transforms/UnifyLoopExits/switch.ll
llvm/test/Transforms/Util/lowerswitch.ll