[Bug rtl-optimization/126638] New: [17 Regression] Compile time slowdown in var-tracking since r17-2277-g6d975812860304

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=3D126638

            Bug ID: 126638
           Summary: [17 Regression] Compile time slowdown in var-tracking
                    since r17-2277-g6d975812860304
           Product: gcc
           Version: 17.0
            Status: UNCONFIRMED
          Keywords: compile-time-hog
          Severity: normal
          Priority: P3
         Component: rtl-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: pheeck at gcc dot gnu.org
                CC: law at gcc dot gnu.org, rguenth at gcc dot gnu.org
  Target Milestone: ---
              Host: aarch64-gnu-linux
            Target: aarch64-gnu-linux

Created attachment 65239
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=3D65239&action=3Dedit
The smaller testcase from pr39326, on which this slowdown occurs

This slowdown doesn't seem that interesting to me, but I wanted to let more
experienced people know about it.  It certainly would be nice to recover the
lost compile time but it is only 20% on a testcase that IIUC comes from
source-to-source compilation of an SPL program (I've not heard about the SPL
language nor was I able to find any information about it).  Therefore, this
seems like a minor issue to me, but I might be wrong.

----

Compilation of the smaller testcase from pr39326 with -O1 -g on Neoverse N1
machines slowed down by ~20% recently.  I'm attaching the testcase to this
report.  Here's a graph showing the slowdown:

https://lnt.opensuse.org/db_default/v4/CPP/graph?plot.0=3D686.621.8

Here is -ftime-report of a commit before the slowdown and after the slowdow=
n.=20
I removed lines with times under a second for brevity.  It looks like the e=
xtra
time is spent in var-tracking.

r17-2271-ge5af677005c7ab:
Time variable                                  wall           GGC
 phase opt and generate             :  70.85 (100%)   254M ( 93%)
 callgraph functions expansion      :  65.46 ( 92%)   223M ( 82%)
 callgraph ipa passes               :   5.31 (  7%)  9785k (  4%)
 alias stmt walking                 :  15.95 ( 22%)  2454k (  1%)
 tree DSE                           :   4.93 (  7%)     0  (  0%)
 tree loop invariant motion         :   2.79 (  4%)  3062k (  1%)
 integrated RA                      :   3.24 (  5%)    28M ( 11%)
 reload CSE regs                    :   2.05 (  3%)  1827k (  1%)
 variable tracking                  :   2.00 (  3%)  7839k (  3%)
 var-tracking dataflow              :  17.19 ( 24%)  1347k (  0%)
 var-tracking emit                  :  15.72 ( 22%)  1583k (  1%)
 TOTAL                              :  71.00          272M

r17-2297-g1f949c2f1ab60a:
Time variable                                  wall           GGC
 phase opt and generate             :  86.72 (100%)   256M ( 93%)
 callgraph functions expansion      :  81.29 ( 94%)   225M ( 82%)
 callgraph ipa passes               :   5.35 (  6%)  9785k (  3%)
 alias stmt walking                 :  16.06 ( 18%)  2454k (  1%)
 tree DSE                           :   4.96 (  6%)     0  (  0%)
 tree loop invariant motion         :   2.79 (  3%)  3062k (  1%)
 integrated RA                      :   3.28 (  4%)    28M ( 11%)
 reload CSE regs                    :   1.96 (  2%)  1827k (  1%)
 variable tracking                  :   1.94 (  2%)  7819k (  3%)
 var-tracking dataflow              :  26.27 ( 30%)  1666k (  1%)
 var-tracking emit                  :  22.35 ( 26%)  1838k (  1%)
 TOTAL                              :  86.88          274M


I bisected this to r17-2277-g6d975812860304

commit 6d97581286030449420dc4ee79612dd256067e2c
Author:     Jovan Dmitrovic <[email protected]>
AuthorDate: Thu Jul 9 07:31:17 2026 -0600
Commit:     Jeff Law <[email protected]>
CommitDate: Thu Jul 9 07:31:54 2026 -0600

    ivopts: Set register pressure cost to zero if there are enough register=
s.

    In function ivopts_estimate_reg_pressure, return zero if there are
    enough registers available so that spilling is not necessary.

    gcc/ChangeLog:

            * tree-ssa-loop-ivopts.cc (ivopts_estimate_reg_pressure):
            Return zero if there are enough registers to store new
            invariants and induction variables.

    Suggested-by:  Richard Biener <[email protected]>
    Signed-off-by: Jovan Dmitrovi=C4=87 <[email protected]>
    Signed-off-by: Djordje Todorovi=C4=87 <[email protected]>
    Signed-off-by: Chao-Ying Fu <[email protected]>
    Signed-off-by: Radosav Kruni=C4=87 <[email protected]>

I'm Ccing only Richard and Jeff, since the other people mentioned by the co=
mmit
don't have a gcc.gnu.org mail.  Bugzilla won't let me Cc them.=