Skip to content

GCC 15+ Issue with "-O2" Optimization #14

Description

@mesheets

With -O2 optimization enabled, GCC 15 and 16 have been shown to create corrupted compilations. Discussions on Debian bug and mailing lists:

Known working versions/options:

  • This issue has not been observed on GCC 14 or earlier.
  • If using GCC 15 but with the -O1 flag, the issue does not happen.

A minimalist code snippet (can be saved as a file conftest.cc):

/* If GCC 3.4.6 is compiled with GCC 15 and the -O2 flag,
 * compilation of this snippet by GCC 3.4.6 will fail.
 * If, however, GCC 3.4.6 is compiled with GCC 15 and the -O1 flag,
 * (or GCC 14 and earlier, even with the -O2 flag)
 * compilation of this snippet by GCC 3.4.6 succeeds.
 */
struct S { ~S(); };
void bar();
void foo()
{
  S s;
  bar();
}

Compilation Command:

/mnt/shared/src/GitHub.com/BrickBot/GNU-Legacy-Toolchain/build/suse/gcc3-O2/gcc3/gcc/xgcc -shared-libgcc -B/mnt/shared/src/GitHub.com/BrickBot/GNU-Legacy-Toolchain/build/suse/gcc3-O2/gcc3/gcc/ -nostdinc++ -L/mnt/shared/src/GitHub.com/BrickBot/GNU-Legacy-Toolchain/build/suse/gcc3-O2/gcc3/h8300-hitachi-coff/libstdc++-v3/src -L/mnt/shared/src/GitHub.com/BrickBot/GNU-Legacy-Toolchain/build/suse/gcc3-O2/gcc3/h8300-hitachi-coff/libstdc++-v3/src/.libs -nostdinc -B/mnt/shared/src/GitHub.com/BrickBot/GNU-Legacy-Toolchain/build/suse/gcc3-O2/gcc3/h8300-hitachi-coff/newlib/ -isystem /mnt/shared/src/GitHub.com/BrickBot/GNU-Legacy-Toolchain/build/suse/gcc3-O2/gcc3/h8300-hitachi-coff/newlib/targ-include -isystem /mnt/shared/src/GitHub.com/BrickBot/GNU-Legacy-Toolchain/sym-combined/gcc3/newlib/libc/include -B/opt/stow/h8300-hitachi-coff-gcc3/h8300-hitachi-coff/bin/ -B/opt/stow/h8300-hitachi-coff-gcc3/h8300-hitachi-coff/lib/ -isystem /opt/stow/h8300-hitachi-coff-gcc3/h8300-hitachi-coff/include -isystem /opt/stow/h8300-hitachi-coff-gcc3/h8300-hitachi-coff/sys-include -L/mnt/shared/src/GitHub.com/BrickBot/GNU-Legacy-Toolchain/build/suse/gcc3-O2/gcc3/ld -c -S -g -Os conftest.cc

 
The error message from the "GCC 15 -O2"-built GCC 3.4.6 with the debug decl.c (c.f. below) is as follows:

conftest.cc:1: error: Current decl type: void
…
conftest.cc:1: confused by earlier errors, bailing out

There could be other issues caused by GCC 15; this is just one that has been identified thus far. There are potentially at least two other cases noted on the Debian bug site (brickOS and sh4) [c.f. bug # 1121607].

 
Attached is that debug version of gcc-3.4.6/gcc/cp/decl.c that outputs extra info at the point of failure:
decl.c

Briefly, between the default: line and the my_friendly_assert (0, 20020917); line, this debug version adds the following (enum values are defined intree.h/tree.def):

	    error ("\
Current decl type: %T\n\
Found TREE_CODE:   %d\n\
ERROR_MARK:        %d\n\
IDENTIFIER_NODE:   %d\n\
TREE_LIST:         %d\n\
TREE_VEC:          %d\n\
BLOCK:             %d\n\
VOID_TYPE:         %d\n\
COND_EXPR:         %d\n\
BIT_NOT_EXPR:      %d\n\
ADDR_EXPR:         %d\n\
ARRAY_REF:         %d\n\
INDIRECT_REF:      %d\n\
CALL_EXPR:         %d\n\
TEMPLATE_ID_EXPR:  %d\n\
SCOPE_REF:         %d\n\
TYPE_DECL:         %d\n\
BASELINK:          %d\n\
TEMPLATE_DECL:     %d\n\
C++\n\
NAMESPACE_DECL:    %d\n\
",
decl,
TREE_CODE (decl),
ERROR_MARK,
IDENTIFIER_NODE,
TREE_LIST,
TREE_VEC,
BLOCK,
VOID_TYPE,
COND_EXPR,
BIT_NOT_EXPR,
ADDR_EXPR,
ARRAY_REF,
INDIRECT_REF,
CALL_EXPR,
TEMPLATE_ID_EXPR,
SCOPE_REF,
TYPE_DECL,
BASELINK,
TEMPLATE_DECL,
NAMESPACE_DECL );

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions