[Bug debug/63572] [5 Regression] ICF breaks user debugging experience

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63572

--- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Better testcase, where ICF actually happens.

struct S { int a; int b; int c; };

__attribute__((noinline)) static int
f1 (struct S *x)
{
  static int u = 1;
  int g = x->a * 7;
  {
    static int v = 2;
    int h = x->b * 11;
    int i = x->c;
    return g + h + i;
  }
}

__attribute__((noinline)) static int
f2 (struct S *x)
{
  static int w = 3;
  int j = x->a * 7; int k = x->b * 11;
  {
    static int y = 4;
    int l = x->c;
    return j + k + l;
  }
}

__attribute__((noinline)) int f3 (struct S *x) { return f1 (x); }
__attribute__((noinline)) int f4 (struct S *x) { return f2 (x) + 1; }
__attribute__((noinline)) int f5 (struct S *x) { return f1 (x) + 2; }
__attribute__((noinline)) int f6 (struct S *x) { return f2 (x) + 3; }

int
main ()
{
  struct S s = { 1, 2, 3 };
  asm volatile ("" : : "r" (&s) : "memory");
  int a[4];
  a[0] = f3 (&s);
  a[1] = f4 (&s);
  a[2] = f5 (&s);
  a[3] = f6 (&s);
  asm volatile ("" : : "r" (a) : "memory");
  return 0;
}

As for .debug_line, the only solution which doesn't require any extensions
would be IMHO to put the icf_merged clone's DW_TAG_subprogram into its own
DW_TAG_partial_unit, import it into the DW_TAG_compile_unit where it is needed,
and use a different DW_AT_stmt_list offset in there, and emit part of
.debug_line (the one for the icf_merged aliases) manually into .debug_line by
the compiler and see whether the assembler will deal with it properly.

Anyway, the ideal user debugging experience IMHO with the above testcase is:
b f1 - debugger finds out that f1 and f2 subprograms have overlapping ranges,
puts a breakpoint into f1==f2 prologue, and when the breakpoint is hit,
unwinds, checks if from the backtrace it is possible using DW_TAG_GNU_call_site
figure out which of the functions has been called; if it is, depending on if it
is f1 or f2 either honors or ignores the breakpoint; if it isn't possible to
uniquely identify what the caller meant to call, honor the breakpoint.
To map instructions back to line info, blocks etc., again, check unwind
info/backtrace which function it is, if it is known which one it is, go into
.debug_line table corresponding to the CU/PU of the subprogram, otherwise pick
one.  Ditto for the
DW_TAG_subprogram/DW_TAG_lexical_block/DW_TAG_inlined_subroutine trees.

Does that sound like a plan?