[PATCH] D47371: [BPI] Apply invoke heuristic before loop branch heuristic

[Artur Pilipenko via Phabricator via llvm-commits]

apilipenko created this revision.
apilipenko added reviewers: chandlerc, vsk, skatkov.

Currently the loop branch heuristic is applied before the invoke heuristic which makes us overestimate the probability of the unwind destination of invokes inside loops. This in turn makes us grossly underestimate the frequencies of loops with invokes.

  loop:
    %i = phi i32 [ 0, %entry ], [ %i.next, %invoke.cont ]
    invoke void @foo() to label %invoke.cont unwind label %lpad
  
  invoke.cont:
    %i.next = add i32 %i, 1
    %cont = icmp ult i32 %i.next, %n
    br i1 %cont, label %loop, label %exit, !prof !{!"branch_weights", i32 9999, i32 1}



  -analyze -branch-prob: 
   ...
   edge loop -> invoke.cont probability is 0x7c000000 / 0x80000000 = 96.88% [HOT edge]
   edge loop -> lpad probability is 0x04000000 / 0x80000000 = 3.12%
   ...
  
  -analyze -block-freq:
   ...
   - loop: float = 31.901
   ...

Applying the invoke heuristic before the loop one makes the results more resonable:

  -analyze -branch-prob: 
   ...
   edge loop -> invoke.cont probability is 0x7ffff800 / 0x80000000 = 100.00% [HOT edge]
   edge loop -> lpad probability is 0x00000800 / 0x80000000 = 0.00%	
   ...
  
  -analyze -block-freq:
   ...
   - loop: float = 9905.6
   ...


Repository:
  rL LLVM

https://reviews.llvm.org/D47371

Files:
  lib/Analysis/BranchProbabilityInfo.cpp
  test/Analysis/BlockFrequencyInfo/loop_with_invoke.ll
  test/Analysis/BranchProbabilityInfo/loop.ll

-------------- next part --------------
A non-text attachment was scrubbed...
Name: D47371.148595.patch
Type: text/x-patch
Size: 3252 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20180525/f0a8429b/attachment.bin>