[PATCH] Revert "gdb: change blockvector::contains() to handle blockvectors with "holes""

[PATCH] Revert "gdb: change blockvector::contains() to handle blockvectors with "holes""

[Jan Vrany]

This reverts commit cc1fc6af4150b19f9c4c70d0463ff498703fb637, since it
causes a number of regressions that seem not to be easily fixable.

The problem lies in existence of "freestanding" code, a code that is
part of a CU but does not have any block associated with it. Consider
following program:

    __asm__(
      ".type foo,@function   \n"
      "foo:                  \n"
      "  mov %rdi, %rax      \n"
      "  ret                 \n"
    );

    static int foo(int i);

    int main(int argc, char **argv) {
      return foo(argc);
    }

When compiled, the foo function has no block of itself:

    Blockvector:

    no map

    block #000, object at 0x55978957b510, 1 symbols in 0x1129..0x1148
     int main(int, char **); block object 0x55978957b380, 0x112d..0x1148 section .text
      block #001, object at 0x55978957b470 under 0x55978957b510, 2 symbols in 0x1129..0x1148
       typedef int int;
       typedef char char;
        block #002, object at 0x55978957b380 under 0x55978957b470, 2 symbols in 0x112d..0x1148, function main
         int argc; computed at runtime
         char **argv; computed at runtime

In this case lookup(0x1129) returns static block and, because of the
change in cc1fc6af4, contains(0x1129) which is wrong.

Such "freestanding" code is perhaps not common but it does exist,
especially in system code. In fact the regressions were at least in part
caused by such "freestanding" code in glibc (libc_sigaction.c).

The whole idea of commit cc1fc6af4 was to handle "holes" in CUs, a case
where one CU spans over multiple disjoint regions, possibly interleaved
with other CUs. Consider somewhat extreme case with two CUs:

    /* hole-1.c */
    int give_me_zero ();

    int
    main ()
    {
      return give_me_zero ();
    }

    /* hole-2.c */
        int __attribute__ ((section (".text_give_me_one"))) __attribute__((noinline))
    baz () { return 42; }

    __asm__(
      ".section  .text_give_me_one,\"ax\",@progbits\n"
      ".type foo,@function         \n"
      "foo:                        \n"
            "  mov %rdi, %rax      \n"
            "  ret                 \n"
            "  nop                 \n"
            "  nop                 \n"
            "  nop                 \n"
    );
    int __attribute__ ((section (".text_give_me_one"))) __attribute__((noinline))
    give_me_one ()
    {
      return 1;
    }

    __asm__(
      ".section  .text_give_me_zero,\"ax\",@progbits\n"
      "bar:                        \n"
            "  jmp give_me_one     \n"
            "  nop                 \n"
            "  nop                 \n"
            "  nop                 \n"
    );
    int __attribute__ ((section (".text_give_me_zero")))
    give_me_zero ()
    {
      extern int bar();
      return give_me_one() - 1;
    }

This when compiled with a carefully crafted linker script to force code
at certain positions, creates following layout:

   0x080000..0x080007   # "freestanding" bar from hole-2.c
   0x080008..0x080016   # give_me_zero() from hole-2.c
   0x080109..0x080114   # main from hole-1.c
   0xf00000..0xf0000b   # baz() from hole-2.c
   0xf0000b..0xf00011   # "freestanding" foo from hole-2.
   0xf0000b..0xf0001c   # gice_me_one() from hole-2.

The block vector for hole-1.c looks:

    Blockvector:

    no map

    block #000, object at 0x555a5d85fb90, 1 symbols in 0x80109..0x80114
     int main(void); block object 0x555a5d85faa0, 0x80109..0x80114 section .text
      block #001, object at 0x555a5d85faf0 under 0x555a5d85fb90, 1 symbols in 0x80109..0x80114
       typedef int int;
        block #002, object at 0x555a5d85faa0 under 0x555a5d85faf0, 0 symbols in 0x80109..0x80114, function main

And for hole-2.c:

    Blockvector:

    map
      0x0 -> 0x0
      0x80008 -> 0x555a5d85ff50
      0x80016 -> 0x0
      0xf00000 -> 0x555a5d860280
      0xf0000b -> 0x0
      0xf00012 -> 0x555a5d860110
      0xf0001d -> 0x0

    block #000, object at 0x555a5d8603b0, 3 symbols in 0x80008..0xf0001d
     int give_me_zero(void); block object 0x555a5d85ff50, 0x80008..0x80016 section .text
     int give_me_one(void); block object 0x555a5d860110, 0xf00012..0xf0001d section .text
     int baz(void); block object 0x555a5d860280, 0xf00000..0xf0000b section .text
      block #001, object at 0x555a5d8602d0 under 0x555a5d8603b0, 1 symbols in 0x80008..0xf0001d
       typedef int int;
        block #002, object at 0x555a5d85ff50 under 0x555a5d8602d0, 0 symbols in 0x80008..0x80016, function give_me_zero
        block #003, object at 0x555a5d860280 under 0x555a5d8602d0, 0 symbols in 0xf00000..0xf0000b, function baz
        block #004, object at 0x555a5d860110 under 0x555a5d8602d0, 0 symbols in 0xf00012..0xf0001d, function give_me_one

Note that despite the fact "freestanding" bar belongs to hole-2.c, the
corresponding CU's global and static blocks start at 0x80008! Looking
at DWARF for the second program, it looks like that the compiler (GCC 15)
did not record the presence of "freestanding" code:

    <0><71>: Abbrev Number: 1 (DW_TAG_compile_unit)
        <72>   DW_AT_producer    : (indirect string, offset: 0): GNU C23 15.2.0 -mtune=generic -march=x86-64 -g -fasynchronous-unwind-tables
        <76>   DW_AT_language    : 29   (C11)
        <77>   Unknown AT value: 90: 3
        <78>   Unknown AT value: 91: 0x31647
        <7c>   DW_AT_name        : (indirect line string, offset: 0x2d): hole-2.c
        <80>   DW_AT_comp_dir    : (indirect line string, offset: 0): test_programs
        <84>   DW_AT_ranges      : 0xc
        <88>   DW_AT_low_pc      : 0
        <90>   DW_AT_stmt_list   : 0x51

and corresponding part of .debug_aranges:

      Length:                   76
      Version:                  2
      Offset into .debug_info:  0x65
      Pointer Size:             8
      Segment Size:             0

        Address            Length
        0000000000f00000 000000000000000b
        0000000000f00012 000000000000000b
        0000000000080008 000000000000000e
        0000000000000000 0000000000000000

Thiago suggested to use minsymbols to tell whether or a CU contains
given address. I do not think this would work reliably as minsymbols do
no know to which CU they belong. In slightly more complicated case of
interleaved CUs it does not seem to be possible to tell for sure to which
one a given minsymbol belongs.

Moreover, Tom suggested that the comment in find_compunit_symtab_for_pc_sect
(which led to cc1fc6af4) may be outdated [2].

Given all that, I'm just reverting the change.

[1]: https://sourceware.org/bugzilla/show_bug.cgi?id=33679#c13
[2]: https://inbox.sourceware.org/gdb-patches/87cy6xzd3j.fsf@tromey.com/

Bug: https://sourceware.org/bugzilla/show_bug.cgi?id=33679
---
 gdb/block.c | 29 +----------------------------
 1 file changed, 1 insertion(+), 28 deletions(-)

diff --git a/gdb/block.c b/gdb/block.c
index 3d2c51cc554..e21580bcf63 100644
--- a/gdb/block.c
+++ b/gdb/block.c
@@ -864,34 +864,7 @@ blockvector::lookup (CORE_ADDR addr) const
 bool
 blockvector::contains (CORE_ADDR addr) const
 {
-  auto b = lookup (addr);
-  if (b == nullptr)
-    return false;
-
-  /* Handle the case that the blockvector has no address map but still has
-     "holes".  For example, consider the following blockvector:
-
-	B0    0x1000 - 0x4000   (global block)
-	B1    0x1000 - 0x4000   (static block)
-	 B3   0x1000 - 0x2000
-				(hole)
-	 B4   0x3000 - 0x4000
-
-     In this case, the above blockvector does not contain address 0x2500 but
-     lookup (0x2500) would return the blockvector's static block.
-
-     So here we check if the returned block is a static block and if yes, still
-     return false.  However, if the blockvector contains no blocks other than
-     the global and static blocks and ADDR falls into the static block,
-     conservatively return true.
-
-     See comment in find_compunit_symtab_for_pc_sect, symtab.c.
-
-     Also, note that if the blockvector in the above example would contain
-     an address map, then lookup (0x2500) would return NULL instead of
-     the static block.
-   */
-  return b != static_block () || num_blocks () == 2;
+  return lookup (addr) != nullptr;
 }
 
 /* See block.h.  */
-- 
2.51.0

[PING] Re: [PATCH] Revert "gdb: change blockvector::contains() to handle blockvectors with "holes""

[Jan Vraný]

Polite ping.

Thanks,
Jan

On Mon, 2025-12-08 at 18:07 +0000, Jan Vrany wrote:
> This reverts commit cc1fc6af4150b19f9c4c70d0463ff498703fb637, since it
> causes a number of regressions that seem not to be easily fixable.
> 
> The problem lies in existence of "freestanding" code, a code that is
> part of a CU but does not have any block associated with it. Consider
> following program:
> 
>     __asm__(
>       ".type foo,@function   \n"
>       "foo:                  \n"
>       "  mov %rdi, %rax      \n"
>       "  ret                 \n"
>     );
> 
>     static int foo(int i);
> 
>     int main(int argc, char **argv) {
>       return foo(argc);
>     }
> 
> When compiled, the foo function has no block of itself:
> 
>     Blockvector:
> 
>     no map
> 
>     block #000, object at 0x55978957b510, 1 symbols in 0x1129..0x1148
>      int main(int, char **); block object 0x55978957b380, 0x112d..0x1148 section .text
>       block #001, object at 0x55978957b470 under 0x55978957b510, 2 symbols in 0x1129..0x1148
>        typedef int int;
>        typedef char char;
>         block #002, object at 0x55978957b380 under 0x55978957b470, 2 symbols in 0x112d..0x1148, function main
>          int argc; computed at runtime
>          char **argv; computed at runtime
> 
> In this case lookup(0x1129) returns static block and, because of the
> change in cc1fc6af4, contains(0x1129) which is wrong.
> 
> Such "freestanding" code is perhaps not common but it does exist,
> especially in system code. In fact the regressions were at least in part
> caused by such "freestanding" code in glibc (libc_sigaction.c).
> 
> The whole idea of commit cc1fc6af4 was to handle "holes" in CUs, a case
> where one CU spans over multiple disjoint regions, possibly interleaved
> with other CUs. Consider somewhat extreme case with two CUs:
> 
>     /* hole-1.c */
>     int give_me_zero ();
> 
>     int
>     main ()
>     {
>       return give_me_zero ();
>     }
> 
>     /* hole-2.c */
>         int __attribute__ ((section (".text_give_me_one"))) __attribute__((noinline))
>     baz () { return 42; }
> 
>     __asm__(
>       ".section  .text_give_me_one,\"ax\",@progbits\n"
>       ".type foo,@function         \n"
>       "foo:                        \n"
>             "  mov %rdi, %rax      \n"
>             "  ret                 \n"
>             "  nop                 \n"
>             "  nop                 \n"
>             "  nop                 \n"
>     );
>     int __attribute__ ((section (".text_give_me_one"))) __attribute__((noinline))
>     give_me_one ()
>     {
>       return 1;
>     }
> 
>     __asm__(
>       ".section  .text_give_me_zero,\"ax\",@progbits\n"
>       "bar:                        \n"
>             "  jmp give_me_one     \n"
>             "  nop                 \n"
>             "  nop                 \n"
>             "  nop                 \n"
>     );
>     int __attribute__ ((section (".text_give_me_zero")))
>     give_me_zero ()
>     {
>       extern int bar();
>       return give_me_one() - 1;
>     }
> 
> This when compiled with a carefully crafted linker script to force code
> at certain positions, creates following layout:
> 
>    0x080000..0x080007   # "freestanding" bar from hole-2.c
>    0x080008..0x080016   # give_me_zero() from hole-2.c
>    0x080109..0x080114   # main from hole-1.c
>    0xf00000..0xf0000b   # baz() from hole-2.c
>    0xf0000b..0xf00011   # "freestanding" foo from hole-2.
>    0xf0000b..0xf0001c   # gice_me_one() from hole-2.
> 
> The block vector for hole-1.c looks:
> 
>     Blockvector:
> 
>     no map
> 
>     block #000, object at 0x555a5d85fb90, 1 symbols in 0x80109..0x80114
>      int main(void); block object 0x555a5d85faa0, 0x80109..0x80114 section .text
>       block #001, object at 0x555a5d85faf0 under 0x555a5d85fb90, 1 symbols in 0x80109..0x80114
>        typedef int int;
>         block #002, object at 0x555a5d85faa0 under 0x555a5d85faf0, 0 symbols in 0x80109..0x80114, function main
> 
> And for hole-2.c:
> 
>     Blockvector:
> 
>     map
>       0x0 -> 0x0
>       0x80008 -> 0x555a5d85ff50
>       0x80016 -> 0x0
>       0xf00000 -> 0x555a5d860280
>       0xf0000b -> 0x0
>       0xf00012 -> 0x555a5d860110
>       0xf0001d -> 0x0
> 
>     block #000, object at 0x555a5d8603b0, 3 symbols in 0x80008..0xf0001d
>      int give_me_zero(void); block object 0x555a5d85ff50, 0x80008..0x80016 section .text
>      int give_me_one(void); block object 0x555a5d860110, 0xf00012..0xf0001d section .text
>      int baz(void); block object 0x555a5d860280, 0xf00000..0xf0000b section .text
>       block #001, object at 0x555a5d8602d0 under 0x555a5d8603b0, 1 symbols in 0x80008..0xf0001d
>        typedef int int;
>         block #002, object at 0x555a5d85ff50 under 0x555a5d8602d0, 0 symbols in 0x80008..0x80016, function give_me_zero
>         block #003, object at 0x555a5d860280 under 0x555a5d8602d0, 0 symbols in 0xf00000..0xf0000b, function baz
>         block #004, object at 0x555a5d860110 under 0x555a5d8602d0, 0 symbols in 0xf00012..0xf0001d, function give_me_one
> 
> Note that despite the fact "freestanding" bar belongs to hole-2.c, the
> corresponding CU's global and static blocks start at 0x80008! Looking
> at DWARF for the second program, it looks like that the compiler (GCC 15)
> did not record the presence of "freestanding" code:
> 
>     <0><71>: Abbrev Number: 1 (DW_TAG_compile_unit)
>         <72>   DW_AT_producer    : (indirect string, offset: 0): GNU C23 15.2.0 -mtune=generic -march=x86-64 -g -fasynchronous-unwind-tables
>         <76>   DW_AT_language    : 29   (C11)
>         <77>   Unknown AT value: 90: 3
>         <78>   Unknown AT value: 91: 0x31647
>         <7c>   DW_AT_name        : (indirect line string, offset: 0x2d): hole-2.c
>         <80>   DW_AT_comp_dir    : (indirect line string, offset: 0): test_programs
>         <84>   DW_AT_ranges      : 0xc
>         <88>   DW_AT_low_pc      : 0
>         <90>   DW_AT_stmt_list   : 0x51
> 
> and corresponding part of .debug_aranges:
> 
>       Length:                   76
>       Version:                  2
>       Offset into .debug_info:  0x65
>       Pointer Size:             8
>       Segment Size:             0
> 
>         Address            Length
>         0000000000f00000 000000000000000b
>         0000000000f00012 000000000000000b
>         0000000000080008 000000000000000e
>         0000000000000000 0000000000000000
> 
> Thiago suggested to use minsymbols to tell whether or a CU contains
> given address. I do not think this would work reliably as minsymbols do
> no know to which CU they belong. In slightly more complicated case of
> interleaved CUs it does not seem to be possible to tell for sure to which
> one a given minsymbol belongs.
> 
> Moreover, Tom suggested that the comment in find_compunit_symtab_for_pc_sect
> (which led to cc1fc6af4) may be outdated [2].
> 
> Given all that, I'm just reverting the change.
> 
> [1]: https://sourceware.org/bugzilla/show_bug.cgi?id=33679#c13
> [2]: https://inbox.sourceware.org/gdb-patches/87cy6xzd3j.fsf@tromey.com/
> 
> Bug: https://sourceware.org/bugzilla/show_bug.cgi?id=33679
> ---
>  gdb/block.c | 29 +----------------------------
>  1 file changed, 1 insertion(+), 28 deletions(-)
> 
> diff --git a/gdb/block.c b/gdb/block.c
> index 3d2c51cc554..e21580bcf63 100644
> --- a/gdb/block.c
> +++ b/gdb/block.c
> @@ -864,34 +864,7 @@ blockvector::lookup (CORE_ADDR addr) const
>  bool
>  blockvector::contains (CORE_ADDR addr) const
>  {
> -  auto b = lookup (addr);
> -  if (b == nullptr)
> -    return false;
> -
> -  /* Handle the case that the blockvector has no address map but still has
> -     "holes".  For example, consider the following blockvector:
> -
> -	B0    0x1000 - 0x4000   (global block)
> -	B1    0x1000 - 0x4000   (static block)
> -	 B3   0x1000 - 0x2000
> -				(hole)
> -	 B4   0x3000 - 0x4000
> -
> -     In this case, the above blockvector does not contain address 0x2500 but
> -     lookup (0x2500) would return the blockvector's static block.
> -
> -     So here we check if the returned block is a static block and if yes, still
> -     return false.  However, if the blockvector contains no blocks other than
> -     the global and static blocks and ADDR falls into the static block,
> -     conservatively return true.
> -
> -     See comment in find_compunit_symtab_for_pc_sect, symtab.c.
> -
> -     Also, note that if the blockvector in the above example would contain
> -     an address map, then lookup (0x2500) would return NULL instead of
> -     the static block.
> -   */
> -  return b != static_block () || num_blocks () == 2;
> +  return lookup (addr) != nullptr;
>  }
>  
>  /* See block.h.  */

Re: [PATCH] Revert "gdb: change blockvector::contains() to handle blockvectors with "holes""

[Tom Tromey]

>>>>> "Jan" == Jan Vrany <jan.vrany@labware.com> writes:

Jan> This reverts commit cc1fc6af4150b19f9c4c70d0463ff498703fb637, since it
Jan> causes a number of regressions that seem not to be easily fixable.

I hate to revert this since I think it's a better direction for gdb.

I guess I think the 'lookup' and 'contains' methods should be
equivalent, and neither should ever return the global or static block --
only real code blocks.

Jan> The problem lies in existence of "freestanding" code, a code that is
Jan> part of a CU but does not have any block associated with it. Consider
Jan> following program:
...
Jan> When compiled, the foo function has no block of itself:

FWIW this is what I would expect.  It's just like any other case where a
function does not have debug info: no debug info mean no block.

Jan> Thiago suggested to use minsymbols to tell whether or a CU contains
Jan> given address. I do not think this would work reliably as minsymbols do
Jan> no know to which CU they belong. In slightly more complicated case of
Jan> interleaved CUs it does not seem to be possible to tell for sure to which
Jan> one a given minsymbol belongs.

The don't know the CU but at you should still be able to set a
breakpoint on them.

Anyway, I'd be willing to approve the backout on the basis that it
causes some regression.  But still I question whether it really makes
sense, and if it would be better instead to change the test.

Tom

Re: [PATCH] Revert "gdb: change blockvector::contains() to handle blockvectors with "holes""

[Jan Vraný]

On Mon, 2025-12-15 at 07:49 -0700, Tom Tromey wrote:
> > > > > > "Jan" == Jan Vrany <jan.vrany@labware.com> writes:
> 
> Jan> This reverts commit cc1fc6af4150b19f9c4c70d0463ff498703fb637, since it
> Jan> causes a number of regressions that seem not to be easily fixable.
> 
> I hate to revert this since I think it's a better direction for gdb.

Yeah, me too.
> 
> I guess I think the 'lookup' and 'contains' methods should be
> equivalent, and neither should ever return the global or static block --
> only real code blocks.
> 
> Jan> The problem lies in existence of "freestanding" code, a code that is
> Jan> part of a CU but does not have any block associated with it. Consider
> Jan> following program:
> ...
> Jan> When compiled, the foo function has no block of itself:
> 
> FWIW this is what I would expect.  It's just like any other case where a
> function does not have debug info: no debug info mean no block.
> 
> Jan> Thiago suggested to use minsymbols to tell whether or a CU contains
> Jan> given address. I do not think this would work reliably as minsymbols do
> Jan> no know to which CU they belong. In slightly more complicated case of
> Jan> interleaved CUs it does not seem to be possible to tell for sure to which
> Jan> one a given minsymbol belongs.
> 
> The don't know the CU but at you should still be able to set a
> breakpoint on them.
> 
> Anyway, I'd be willing to approve the backout on the basis that it
> causes some regression.  But still I question whether it really makes
> sense, and if it would be better instead to change the test.

We can change tests but I think GDB would still be broken - the problem is that
this change confuses unwinder in a way I do not really understand yet. So it seems to
me that if we keep this, we need "fix" some other part of GDB too. I spent some time
debugging this and tried again yesterday but so far the more I debug the less 
I understand :-(

I do not know what's the policy regarding regressions. If everyone can tolerate 
these regressions for longer, I can try to do some more debugging and see.

Jan

> 
> Tom

Re: [PATCH] Revert "gdb: change blockvector::contains() to handle blockvectors with "holes""

[Tom Tromey]

>>>>> "Jan" == Jan Vraný <Jan.Vrany@labware.com> writes:

Jan> I do not know what's the policy regarding regressions. If everyone can tolerate 
Jan> these regressions for longer, I can try to do some more debugging and see.

I suggest we land the backout and then investigate.
Like I said, I hate to do this, but I guess the bug seems big enough
that we shouldn't hold anything else up for this.

For the backout:
Approved-By: Tom Tromey <tom@tromey.com>

If you'd rather not look into the problems, that's totally fine, but
would you mind filing a bug with your findings?

thank you,
Tom

Re: [PATCH] Revert "gdb: change blockvector::contains() to handle blockvectors with "holes""

[Jan Vraný]

On Thu, 2025-12-18 at 11:55 -0700, Tom Tromey wrote:
> > > > > > "Jan" == Jan Vraný <Jan.Vrany@labware.com> writes:
> 
> Jan> I do not know what's the policy regarding regressions. If everyone can tolerate 
> Jan> these regressions for longer, I can try to do some more debugging and see.
> 
> I suggest we land the backout and then investigate.
> Like I said, I hate to do this, but I guess the bug seems big enough
> that we shouldn't hold anything else up for this.
> 
> For the backout:
> Approved-By: Tom Tromey <tom@tromey.com>

Thanks. Will run all tests again and push soon. 
> 
> If you'd rather not look into the problems, that's totally fine, but
> would you mind filing a bug with your findings?

I'm actually still looking into it when time allows. I do not like
not understanding what's going on. So far it looks like some weird 
interference with unwinding. I find that part of GDB particularly 
tricky :-) Of course I'll update the bug with my findings.

Jan

> 
> thank you,
> Tom