[Bug c/101474] New: -fipa-icf generates worse code for identical function

[Bug c/101474] New: -fipa-icf generates worse code for identical function

[vegard.nossum at oracle dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101474

            Bug ID: 101474
           Summary: -fipa-icf generates worse code for identical function
           Product: gcc
           Version: 12.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
          Assignee: unassigned at gcc dot gnu.org
          Reporter: vegard.nossum at oracle dot com
  Target Milestone: ---

Created attachment 51166
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=51166&action=edit
Test case

I've come across a weird behaviour when using -fipa-icf, could maybe be related
to bug 80277.

Build with -O1 -fipa-icf and the second version of the identical function
actually has worse codegen than if you have just one of them there or than if
you hadn't passed -fipa-icf at all.

See example: https://godbolt.org/z/n8zz947aK

[Bug tree-optimization/101474] SRA sometimes produces worse code with inline functions (seen with -fipa-icf sometimes)

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101474

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
            Summary|-fipa-icf generates worse   |SRA sometimes produces
                   |code for identical function |worse code with inline
                   |                            |functions (seen with
                   |                            |-fipa-icf sometimes)
     Ever confirmed|0                           |1
   Last reconfirmed|                            |2021-07-16
             Status|UNCONFIRMED                 |NEW

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
This worse code generation has nothing to do with -fipa-icf really as shown by:
struct t {
    char x;
    int y;
};

void fn1(struct t *p, struct t v)
{
    *p = v;
}

void fn3(struct t *p, struct t v)
{
 fn1(p, v);
}

------ CUT ----
What is happening is SRA decides it is best to do a full scalar replacement for
the v that is inlined into fn3 (from fn1).

[Bug tree-optimization/101474] SRA sometimes produces worse code with inline functions (seen with -fipa-icf sometimes)

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101474

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jamborm at gcc dot gnu.org

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
Yes, when SRA decomposes a copy it leaves out padding which means we later fail
to apply store-merging and thus end up generating worse code for a simple
aggregate copy.  SRA assumes the total scalarization transform allow followup
optimizations, if this doesn't happen it's hardly a profitable transform.

[Bug tree-optimization/101474] SRA sometimes produces worse code with inline functions (seen with -fipa-icf sometimes)

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101474

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |RESOLVED
         Resolution|---                         |DUPLICATE

--- Comment #3 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
This is a dup of bug 96252.

*** This bug has been marked as a duplicate of bug 96252 ***

[Bug tree-optimization/101474] SRA sometimes produces worse code with inline functions (seen with -fipa-icf sometimes)

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101474

--- Comment #4 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Note ICF behavior and inlining is what PR 96252 is about.