[Bug middle-end/108268] New: [13 Regression] Build failure on cross from powerpc64le-linux to nvptx-none

[Bug middle-end/108268] New: [13 Regression] Build failure on cross from powerpc64le-linux to nvptx-none

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108268

            Bug ID: 108268
           Summary: [13 Regression] Build failure on cross from
                    powerpc64le-linux to nvptx-none
           Product: gcc
           Version: 13.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: middle-end
          Assignee: unassigned at gcc dot gnu.org
          Reporter: jakub at gcc dot gnu.org
  Target Milestone: ---

../configure --disable-bootstrap --disable-sjlj-exceptions
--enable-newlib-io-long-long
--with-build-time-tools=/builddir/build/BUILD/gcc-13.0.0-20230102/obji/usr/nvptx-none/bin
--
target nvptx-none --enable-as-accelerator-for=ppc64le-redhat-linux
--enable-languages=c,c++,fortran,lto --prefix=/usr --mandir=/usr/share/man
--infodir=/usr/share/info --with-bugurl=
http://bugzilla.redhat.com/bugzilla --enable-checking=release
--with-system-zlib --with-gcc-major-version-only --without-isl
configured GCC ICEs on compilation of all C/C++ files.
The problem is that c_cpp_builtins function has &c_global_trees[0] value stored
in $r30
register, but when it calls
1525            int value = dwarf_reg_sizes_constant ();
the $r30 value is saved into stack frame and partially overwritten by buffer
overflow.
Before it is 0x0000000011939cb0 and after returning from that function
0x11930008.  This is because DWARF_FRAME_RETURN_COLUMN macro defaults to
DWARF_FRAME_REGISTERS, nvptx.h doesn't override it (it is a target without
register allocation nor debug info), dwarf_reg_sizes_constant allocates
343       poly_uint16 *sizes = XALLOCAVEC (poly_uint16, DWARF_FRAME_REGISTERS);
and then generate_dwarf_reg_sizes stores:
326       if (!init_state.wrote_return_column)
327         sizes[DWARF_FRAME_RETURN_COLUMN] = GET_MODE_SIZE (Pmode);

[Bug middle-end/108268] [13 Regression] Build failure on cross from powerpc64le-linux to nvptx-none

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108268

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|---                         |13.0
           Priority|P3                          |P1
                 CC|                            |fw at gcc dot gnu.org

[Bug middle-end/108268] [13 Regression] Build failure on cross from powerpc64le-linux to nvptx-none

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108268

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |build
           Severity|normal                      |blocker

[Bug middle-end/108268] [13 Regression] Build failure on cross from powerpc64le-linux to nvptx-none

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108268

--- Comment #1 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Looking at GCC 12, I think no target actually stored anything into that last
element
unless it was a return or alternate return column.
It is true that expand_builtin_init_dwarf_reg_sizes had:
  if (!init_state.wrote_return_column)
    init_return_column_size (mode, mem, DWARF_FRAME_RETURN_COLUMN);

#ifdef DWARF_ALT_FRAME_RETURN_COLUMN
  init_return_column_size (mode, mem, DWARF_ALT_FRAME_RETURN_COLUMN);
#endif
which is what the new gcc tries to preserve.
init_one_dwarf_reg_size had:
  const unsigned int dnum = DWARF_FRAME_REGNUM (regno);
  const unsigned int rnum = DWARF2_FRAME_REG_OUT (dnum, 1);
  const unsigned int dcol = DWARF_REG_TO_UNWIND_COLUMN (rnum);
...
  if (rnum >= DWARF_FRAME_REGISTERS)
    return;
though, and seems DWARF_REG_TO_UNWIND_COLUMN always uses the default definition
of
just returning its argument, so I think init_one_dwarf_reg_size would never
store to
the last element.  init_return_column_size can though.  And
TARGET_INIT_DWARF_REG_SIZES_EXTRA is only defined on msp430 and rs6000 and
neither writes at offsets >= DWARF_FRAME_REGISTERS.

So, I think the DWARF_FRAME_REGISTERS element should be considered into whether
everything is constant/same size only if it is DWARF_FRAME_RETURN_COLUMN or
DWARF_ATL_FRAME_RETURN_COLUMN.

[Bug middle-end/108268] [13 Regression] Build failure on cross from powerpc64le-linux to nvptx-none

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108268

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

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

--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
r13-4980-gd010efbfb89721bc9ca9e657e980ff0d8e6187af reverted the problematic
commit, so it should build now fine again.