Generalize MergeBlockIntoPredecessor. Replace uses of MergeBasicBlockIntoOnlyPred.
authorAlina Sbirlea <asbirlea@google.com>
Wed, 20 Jun 2018 22:01:04 +0000 (22:01 +0000)
committerAlina Sbirlea <asbirlea@google.com>
Wed, 20 Jun 2018 22:01:04 +0000 (22:01 +0000)
commitdfd14adeb04c79e6f763d491fc7d2fb3c3b6bc07
tree7029e176addeef1ae59e36818184a82926f9fba1
parent715ee079da4eb26fbe3c4c01cb8a3636d7a24667
Generalize MergeBlockIntoPredecessor. Replace uses of MergeBasicBlockIntoOnlyPred.

Summary:
Two utils methods have essentially the same functionality. This is an attempt to merge them into one.
1. lib/Transforms/Utils/Local.cpp : MergeBasicBlockIntoOnlyPred
2. lib/Transforms/Utils/BasicBlockUtils.cpp : MergeBlockIntoPredecessor

Prior to the patch:
1. MergeBasicBlockIntoOnlyPred
Updates either DomTree or DeferredDominance
Moves all instructions from Pred to BB, deletes Pred
Asserts BB has single predecessor
If address was taken, replace the block address with constant 1 (?)

2. MergeBlockIntoPredecessor
Updates DomTree, LoopInfo and MemoryDependenceResults
Moves all instruction from BB to Pred, deletes BB
Returns if doesn't have a single predecessor
Returns if BB's address was taken

After the patch:
Method 2. MergeBlockIntoPredecessor is attempting to become the new default:
Updates DomTree or DeferredDominance, and LoopInfo and MemoryDependenceResults
Moves all instruction from BB to Pred, deletes BB
Returns if doesn't have a single predecessor
Returns if BB's address was taken

Uses of MergeBasicBlockIntoOnlyPred that need to be replaced:

1. lib/Transforms/Scalar/LoopSimplifyCFG.cpp
Updated in this patch. No challenges.

2. lib/CodeGen/CodeGenPrepare.cpp
Updated in this patch.
  i. eliminateFallThrough is straightforward, but I added using a temporary array to avoid the iterator invalidation.
  ii. eliminateMostlyEmptyBlock(s) methods also now use a temporary array for blocks
Some interesting aspects:
  - Since Pred is not deleted (BB is), the entry block does not need updating.
  - The entry block was being updated with the deleted block in eliminateMostlyEmptyBlock. Added assert to make obvious that BB=SinglePred.
  - isMergingEmptyBlockProfitable assumes BB is the one to be deleted.
  - eliminateMostlyEmptyBlock(BB) does not delete BB on one path, it deletes its unique predecessor instead.
  - adding some test owner as subscribers for the interesting tests modified:
    test/CodeGen/X86/avx-cmp.ll
    test/CodeGen/AMDGPU/nested-loop-conditions.ll
    test/CodeGen/AMDGPU/si-annotate-cf.ll
    test/CodeGen/X86/hoist-spill.ll
    test/CodeGen/X86/2006-11-17-IllegalMove.ll

3. lib/Transforms/Scalar/JumpThreading.cpp
Not covered in this patch. It is the only use case using the DeferredDominance.
I would defer to Brian Rzycki to make this replacement.

Reviewers: chandlerc, spatel, davide, brzycki, bkramer, javed.absar

Subscribers: qcolombet, sanjoy, nemanjai, nhaehnle, jlebar, tpr, kbarton, RKSimon, wmi, arsenm, llvm-commits

Differential Revision: https://reviews.llvm.org/D48202

llvm-svn: 335183
34 files changed:
llvm/include/llvm/Transforms/Utils/BasicBlockUtils.h
llvm/lib/CodeGen/CodeGenPrepare.cpp
llvm/lib/Transforms/Scalar/LoopSimplifyCFG.cpp
llvm/lib/Transforms/Utils/BasicBlockUtils.cpp
llvm/test/CodeGen/AMDGPU/branch-relaxation.ll
llvm/test/CodeGen/AMDGPU/nested-loop-conditions.ll
llvm/test/CodeGen/AMDGPU/si-annotate-cf.ll
llvm/test/CodeGen/ARM/indirectbr.ll
llvm/test/CodeGen/PowerPC/memCmpUsedInZeroEqualityComparison.ll
llvm/test/CodeGen/PowerPC/memcmp-mergeexpand.ll
llvm/test/CodeGen/PowerPC/ppc-shrink-wrapping.ll
llvm/test/CodeGen/PowerPC/simplifyConstCmpToISEL.ll
llvm/test/CodeGen/Thumb2/2010-11-22-EpilogueBug.ll
llvm/test/CodeGen/Thumb2/thumb2-jtb.ll
llvm/test/CodeGen/X86/2006-11-17-IllegalMove.ll
llvm/test/CodeGen/X86/avx-cmp.ll
llvm/test/CodeGen/X86/avx-splat.ll
llvm/test/CodeGen/X86/avx2-vbroadcast.ll
llvm/test/CodeGen/X86/avx512-i1test.ll
llvm/test/CodeGen/X86/block-placement.ll
llvm/test/CodeGen/X86/hoist-spill.ll
llvm/test/CodeGen/X86/ins_subreg_coalesce-1.ll
llvm/test/CodeGen/X86/memcmp-mergeexpand.ll
llvm/test/CodeGen/X86/pr32108.ll
llvm/test/CodeGen/X86/setcc-lowering.ll
llvm/test/CodeGen/X86/split-store.ll
llvm/test/CodeGen/X86/tail-dup-merge-loop-headers.ll
llvm/test/DebugInfo/Generic/sunk-compare.ll
llvm/test/Transforms/CodeGenPrepare/X86/computedgoto.ll
llvm/test/Transforms/CodeGenPrepare/basic.ll
llvm/test/Transforms/LoopSimplifyCFG/scev.ll
llvm/test/Transforms/LoopStrengthReduce/ARM/2012-06-15-lsr-noaddrmode.ll
llvm/test/Transforms/LoopStrengthReduce/X86/2012-01-13-phielim.ll
llvm/test/Transforms/SimpleLoopUnswitch/trivial-unswitch-iteration.ll