[PATCH] D28522: Codegen: Make chains from lattice-shaped CFGs

[Kyle Butt via Phabricator via llvm-commits]

iteratee updated this revision to Diff 86937.
iteratee added a comment.

Looking over this, It feels like I can split the lattice portion out and put it in first. Even together they aren't very big, so If I can get some initial opinions on this before I start that, it would be appreciated.

I wanted to lay out lattice-type CFG's correctly even if the blocks are larger than what we would tail-duplicate.
To do that I worked out that for a lattice, we can compute which pair of edges forms the optimal fallthrough and use those edges. We don't really have to worry about CFG breaking, because with a lattice we can easily compute the optimal fallthrough pair and take it.

Lattice:
A Set of Predecessor blocks P that all have the same Successors S when |S| >= 2, |P| >=2 and S ∩ P = ∅
We can treat this as a graph optimization problem. There's a well known general algorithm, (the hungarian algorithm) but I just solved it for size 2 because that's the only size we really care about.

Thanks for taking a look, I'll try and get the plain lattice code separated from the tail duplication code tomorrow.


https://reviews.llvm.org/D28522

Files:
  lib/CodeGen/MachineBlockPlacement.cpp
  test/CodeGen/AArch64/branch-relax-cbz.ll
  test/CodeGen/AArch64/combine-comparisons-by-cse.ll
  test/CodeGen/AMDGPU/basic-branch.ll
  test/CodeGen/AMDGPU/branch-relaxation.ll
  test/CodeGen/AMDGPU/cf-loop-on-constant.ll
  test/CodeGen/AMDGPU/convergent-inlineasm.ll
  test/CodeGen/AMDGPU/salu-to-valu.ll
  test/CodeGen/ARM/2007-05-22-tailmerge-3.ll
  test/CodeGen/ARM/atomic-cmpxchg.ll
  test/CodeGen/ARM/fold-stack-adjust.ll
  test/CodeGen/PowerPC/tail-dup-break-cfg.ll
  test/CodeGen/PowerPC/tail-dup-layout.ll
  test/CodeGen/SPARC/sjlj.ll
  test/CodeGen/WebAssembly/mem-intrinsics.ll
  test/CodeGen/X86/block-placement.ll
  test/CodeGen/X86/bypass-slow-division-32.ll
  test/CodeGen/X86/sse1.ll
  test/CodeGen/X86/tail-dup-merge-loop-headers.ll
  test/CodeGen/X86/tail-dup-repeat.ll
  test/CodeGen/X86/tail-opts.ll
  test/CodeGen/X86/twoaddr-coalesce-3.ll
  test/CodeGen/X86/win-alloca-expander.ll

-------------- next part --------------
A non-text attachment was scrubbed...
Name: D28522.86937.patch
Type: text/x-patch
Size: 47107 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20170203/24af9687/attachment.bin>