[Bug lto/99559] New: LTO with -Os -ffreestanding -nodefaultlibs may generate call to memcpy, but fail linking even though memcpy is defined

[Bug lto/99559] New: LTO with -Os -ffreestanding -nodefaultlibs may generate call to memcpy, but fail linking even though memcpy is defined

[kabel at blackhole dot sk]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99559

            Bug ID: 99559
           Summary: LTO with -Os -ffreestanding -nodefaultlibs may
                    generate call to memcpy, but fail linking even though
                    memcpy is defined
           Product: gcc
           Version: 10.2.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: lto
          Assignee: unassigned at gcc dot gnu.org
          Reporter: kabel at blackhole dot sk
                CC: marxin at gcc dot gnu.org
  Target Milestone: ---

Created attachment 50372
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=50372&action=edit
test.c

When compiling for target where the assignment operator

  *dst = *src;

will generate a call to memcpy (for example aarch64), and when also using
compile flags

  -ffreestanding -Os -flto

and link flags

  -nodefaultlibs -nostartfiles

the linking will fail with

  undefined reference to `memcpy'

even if memcpy is defined in the source.

The memcpy is optimized away, even though it is used by the = operator. It
needs to be declared with __attribute__((used)) to not be optimized away.

What is interesting is that if the memcpy function is used explicitly, the
compilation may still fail with -Os, because the code will be inlined.

If memcpy is used explicitly multiple times, it may be generated with symbol
name
  memcpy.constprop.0.isra.0
or
  memcpy.isra.0
depending on whether the arguments passed are the same or different.

But the linking fill still fail, because the assignment operator will generate
a call to memcpy, and this symbol is not defined, only memcpy.isra.0 or
memcpy.constprop.0.isra.0 is.

I discovered this recently while trying to make U-Boot compile with LTO. We
have solved it by declaring "memcmp", "memset", "memcpy" and "memmove" with
attribute used, since according to GCC man page, calls to this functions may
still be done when -nodefaultlibs is used.

[Bug lto/99559] LTO with -Os -ffreestanding -nodefaultlibs may generate call to memcpy, but fail linking even though memcpy is defined

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99559

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |hubicka at gcc dot gnu.org,
                   |                            |jsm28 at gcc dot gnu.org

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Yeah, we do remove unused (as in unused from the source) memcpy implementations
early but then introduce uses ourselves.

I suppose we should avoid doing that with -ffreestanding, thus basically add
the used attribute on a select number of definitions automatically with this
option.

I wonder for example whether a static definition of memcpy would need to be
preserved since I think this isn't really special to LTO.

Joseph?  Honza?

[Bug lto/99559] LTO with -Os -ffreestanding -nodefaultlibs may generate call to memcpy, but fail linking even though memcpy is defined

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99559

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

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

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
This is a dup of bug 58203.

*** This bug has been marked as a duplicate of bug 58203 ***