[PATCH] D38594: [InlineCost] Tracking Values through PHI Nodes

[Haicheng Wu via Phabricator via llvm-commits]

haicheng created this revision.
Herald added a subscriber: javed.absar.
Herald added a reviewer: dberlin.

This patch fix this FIXME in visitPHI()

  // FIXME: We should potentially be tracking values through phi nodes,
  // especially when they collapse to a single value due to deleted CFG edges
  // during inlining.

To do this, I maintain two lists: dead blocks during inlining and a mapping of the blocks to their known successors.  For example,

  define i1 @outer9(i1 %cond) {             
    %C = call i1 @inner9(i32 0, i1 %cond)
    ret i1 %C
  }
    
  define i1 @inner9(i32 %cond1, i1 %cond2) {
  entry:                         
    switch i32 %cond1, label %exit [ i32 0, label %zero
                                     i32 1, label %one
                                     i32 2, label %two ]
    
  zero:
    br label %exit
  
  one:   
    br label %exit    
  
  two:
    br i1 %cond2, label %two_true, label %two_false
    
  two_true: 
    br label %exit
                
  two_false: 
    br label %exit
    
  exit:
    %phi = phi i32 [0, %zero], [1, %one], [2, %two_true], [2, %two_false], [-1, %entry] ; Simplified to 0
    %cmp = icmp eq i32 %phi, 0
    call void @pad()
    store i32 0, i32* @glbl
    ret i1 %cmp
  }

Blocks one, two, two_true, two_false are unreachable because of the deleted CFG edges.  The successor of block entry has to be block zero so block entry must not be the predecessor of block exit during inlining.  Thus, the phi node in block exit can be collapsed to a constant 0.

The algorithm used to find dead blocks is a simplified version from GVN::addDeadBlock().  Basically, if all the predecessors of a block are dead, this block is dead too.

This patch can reduce the inlinecost of the problematic callee of PR34173 <https://bugs.llvm.org/show_bug.cgi?id=34173> from 235 to 65.  The impact to spec is as follows

| Benchmark          | Code Size (%) | Perf (%)      |
|                    | (+ is bigger) | (+ is faster) |
| spec2000/gcc       | 0.05          | -0.39         |
| spec2006/gcc       | -0.04         | -0.03         |
| spec2006/h264ref   | -0.05         | 0.15          |
| spec2006/xalancbmk | 0.05          | -0.64         |
| spec2017/blender   | 0.02          | 0.54          |
| spec2017/gcc       | 0.04          | 0.03          |
| spec2017/imagick   | 0.01          | 1.03          |
| spec2017/perlbench | 0.01          | -0.04         |
| spec2017/xalancbmk | 0.04          | 0.29          |


Repository:
  rL LLVM

https://reviews.llvm.org/D38594

Files:
  lib/Analysis/InlineCost.cpp
  test/Transforms/Inline/AArch64/phi.ll

-------------- next part --------------
A non-text attachment was scrubbed...
Name: D38594.117838.patch
Type: text/x-patch
Size: 16429 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20171005/4c6599cb/attachment.bin>