[llvm-bugs] [Bug 32984] New: Eliminate PHI nodes pass really slow at -O3

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=32984

            Bug ID: 32984
           Summary: Eliminate PHI nodes pass really slow at -O3
           Product: new-bugs
           Version: trunk
          Hardware: PC
                OS: Windows NT
            Status: NEW
          Severity: normal
          Priority: P
         Component: new bugs
          Assignee: unassignedbugs at nondot.org
          Reporter: douglas_yung at playstation.sony.com
                CC: llvm-bugs at lists.llvm.org

Created attachment 18424
  --> https://bugs.llvm.org/attachment.cgi?id=18424&action=edit
rtc.all.c

We have an internal test for which the compile time jumped from around 3
minutes to over 10 minutes after r295538. The attached test case, rtc.all.c,
when compiled with a Windows hosted compiler built from r295537 took 165
seconds to compile, while a Windows hosted compiler built from r295538 took 646
seconds. From a generated time report with the r295538 compiler, the culprit
seems to be the Eliminate PHI nodes for register allocation pass which took
almost 50% of the compilation time. In the generated time report for r295537,
this stage only took around 10% of the compilation time.

Here are the top 3 items in the time report when the attached file was built
using r295537:

  66.0040 ( 40.7%)   1.5912 ( 79.1%)  67.5952 ( 41.2%)  67.5629 ( 41.2%) 
Unroll loops
  16.5205 ( 10.2%)   0.0000 (  0.0%)  16.5205 ( 10.1%)  16.5549 ( 10.1%) 
Induction Variable Users
  15.0229 (  9.3%)   0.0000 (  0.0%)  15.0229 (  9.2%)  15.0129 (  9.1%) 
Eliminate PHI nodes for
register allocation

Compare that to the top 3 items in the time report when the attached file was
built using r295538:

  310.4420 ( 48.6%)   0.0624 (  1.0%)  310.5044 ( 48.2%)  310.5218 ( 48.2%) 
Eliminate PHI nodes
for register allocation
  128.4824 ( 20.1%)   3.9936 ( 61.5%)  132.4760 ( 20.5%)  132.7336 ( 20.6%) 
Unroll loops
  39.6867 (  6.2%)   0.0312 (  0.5%)  39.7179 (  6.2%)  39.7383 (  6.2%) 
Branch Probability Basic
Block Placement

To reproduce the slowdown, build the attached test case with a compiler >=
r295538, and compile with -O3.

-- 
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20170510/5a249923/attachment.html>