[Bug tree-optimization/101555] [12 Regression] Compile slowdown in tree PRE

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101555

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|---                         |12.0
           Keywords|                            |needs-bisection
            Summary|Compile slowdown in tree    |[12 Regression] Compile
                   |PRE                         |slowdown in tree PRE
           Assignee|unassigned at gcc dot gnu.org      |rguenth at gcc dot gnu.org
             Status|UNCONFIRMED                 |ASSIGNED
            Version|9.3.0                       |12.0
     Ever confirmed|0                           |1
   Last reconfirmed|                            |2021-07-22

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
Confirmed with g++-9 (SUSE Linux) 9.3.1 20200406 [revision
6db837a5288ee3ca5ec504fbd5a765817e556ac2]

 tree PRE                           : 645.41 ( 97%)   0.16 ( 34%) 645.60 ( 97%)
 177896 kB ( 15%)
 TOTAL                              : 665.14          0.47        665.73       
1154121 kB

with -fno-code-hoisting that get's down to

 tree PRE                           : 219.50 ( 93%)   0.04 ( 11%) 219.69 ( 93%)
  87813 kB ( 11%)
 TOTAL                              : 235.70          0.37        236.12       
 807306 kB

GCC 10.3.0 shows

 tree PRE                           : 267.62 ( 93%)   0.13 ( 32%) 267.69 ( 93%)
 140858 kB ( 13%)
 TOTAL                              : 288.92          0.41        289.37       
1122606 kB

GCC 11 does

 tree PRE                           : 251.41 ( 91%)   0.10 ( 23%) 251.71 ( 91%)
  137M ( 16%)
 TOTAL                              : 275.01          0.43        275.51       
  837M

Note that trunk regresses again:

 tree PRE                           : 312.12 ( 93%)   0.07 ( 15%) 312.46 ( 93%)
  137M ( 16%)
 TOTAL                              : 337.03          0.48        337.58       
  836M

bisection for that would be nice.

So part of this PR is a dup of PR97623, but there's still slowness of PRE
(as said elsewhere PRE is known quadratic).

I'm turning this PR into the GCC 12 regression that shows above.