Re: pahole and LTO

From: Arnaldo Carvalho de Melo <hidden>
Date: 2020-01-10 17:29:42
Also in: bpf

On January 10, 2020 2:22:31 PM GMT-03:00, Stanislav Fomichev [off-list ref] wrote:

On 01/10, Arnaldo Carvalho de Melo wrote:

quoted

Em Fri, Jan 10, 2020 at 08:44:10AM -0800, Stanislav Fomichev

escreveu:

quoted

tl;dr - building the kernel with clang and lto breaks BTF

generation because

quoted

pahole doesn't seem to understand cross-cu references.

tl;dr response:

Yeah, so it may be the time to fix that, elfutils has interfaces for

it,

quoted

and the tools that come with it handle cross-cu references, so we

need

quoted

to study that and make pahole understand it.

Sure, we can definitely help with the implementation unless someone
is already actively working on it. Just wanted to make sure that's
a known problem.

From my (limited) looking at pahole sources, it seems that building
and index on the first pass and doing a second pass to resolve
cross-cu references is relatively easy to implement. Am I missing
anything? (not a dwarf expert in any sense).

Give it a try, please

And where do the patches for pahole go? I don't see any pahole patches
in bpf/netdev mailing lists.

Send it to me, cc
dwarves@vger.kernel.org

- Arnaldo

quoted

- Arnaldo

quoted

Can be reproduced with the following:
    $ cat a.c
    struct s;

    void f1() {}

    __attribute__((always_inline)) void f2(struct s *p)
    {
            if (p)
                    f1();
    }
    $ cat b.c
    struct s {
            int x;
    };

    void f2(struct s *p);

    int main()
    {
            struct s s = { 10 };
            f2(&s);
    }
    $ clang -fuse-ld=lld -flto {a,b}.c -g

    $ pahole a.out
    tag__recode_dwarf_type: couldn't find 0x3f type for 0x99

(inlined_subroutine)!

quoted

    lexblock__recode_dwarf_types: couldn't find 0x3f type for 0x99

(inlined_subroutine)!

quoted

    struct s {
            int                        x;                    /*

0     4 */

quoted

            /* size: 4, cachelines: 1, members: 1 */
            /* last cacheline: 4 bytes */
    };

From what I can tell, pahole internally loops over each cu and

resolves only

quoted

local references, while the dwarf spec (table 2.3) states the

following

quoted

about 'reference':
"Refers to one of the debugging information entries that describe

the program.

quoted

There are four types of reference. The first is an offset relative

to the

quoted

beginning of the compilation unit in which the reference occurs and

must

quoted

refer to an entry within that same compilation unit. The second

type of

quoted

reference is the offset of a debugging information entry in any

compilation

quoted

unit, including one different from the unit containing the

reference. The

quoted

third type of reference is an indirect reference to a type

definition using

quoted

an 8-byte signature for that type. The fourth type of reference is

a reference

quoted

from within the .debug_info section of the executable or shared

object file to

quoted

a debugging information entry in the .debug_info section of a

supplementary

quoted

object file."

In particular: "The second type of reference is the offset of a

debugging

quoted

information entry in any compilation unit, including one different

from the

quoted

unit containing the reference."


So the question is: is it a (known) issue? Is it something that's

ommitted

quoted

on purpose? Or it's not implemented because lto is not (yet) widely

used?

quoted


Here is the dwarf:

$ readelf --debug-dump=info a.out
Contents of the .debug_info section:

  Compilation Unit @ offset 0x0:
   Length:        0x44 (32-bit)
   Version:       4
   Abbrev Offset: 0x0
   Pointer Size:  8
 <0><b>: Abbrev Number: 1 (DW_TAG_compile_unit)
    <c>   DW_AT_producer    : (indirect string, offset: 0x11):

clang version 10.0.0 (https://github.com/llvm/llvm-project.git
5fe4679cc9cfb4941b766db07bf3cd928075d204)

quoted

    <10>   DW_AT_language    : 12	(ANSI C99)
    <12>   DW_AT_name        : (indirect string, offset: 0x0): a.c
    <16>   DW_AT_stmt_list   : 0x0
    <1a>   DW_AT_comp_dir    : (indirect string, offset: 0x7a):

/usr/local/google/home/sdf/tmp/lto

quoted

    <1e>   DW_AT_low_pc      : 0x201730
    <26>   DW_AT_high_pc     : 0x6
 <1><2a>: Abbrev Number: 2 (DW_TAG_subprogram)
    <2b>   DW_AT_low_pc      : 0x201730
    <33>   DW_AT_high_pc     : 0x6
    <37>   DW_AT_frame_base  : 1 byte block: 56 	(DW_OP_reg6 (rbp))
    <39>   DW_AT_name        : (indirect string, offset: 0xa4): f1
    <3d>   DW_AT_decl_file   : 1
    <3e>   DW_AT_decl_line   : 3
    <3f>   DW_AT_external    : 1
 <1><3f>: Abbrev Number: 3 (DW_TAG_subprogram)
    <40>   DW_AT_name        : (indirect string, offset: 0x4): f2
    <44>   DW_AT_decl_file   : 1
    <45>   DW_AT_decl_line   : 5
    <46>   DW_AT_prototyped  : 1
    <46>   DW_AT_external    : 1
    <46>   DW_AT_inline      : 1	(inlined)
 <1><47>: Abbrev Number: 0
  Compilation Unit @ offset 0x48:
   Length:        0x7f (32-bit)
   Version:       4
   Abbrev Offset: 0x0
   Pointer Size:  8
 <0><53>: Abbrev Number: 1 (DW_TAG_compile_unit)
    <54>   DW_AT_producer    : (indirect string, offset: 0x11):

clang version 10.0.0 (https://github.com/llvm/llvm-project.git
5fe4679cc9cfb4941b766db07bf3cd928075d204)

quoted

    <58>   DW_AT_language    : 12	(ANSI C99)
    <5a>   DW_AT_name        : (indirect string, offset: 0x7): b.c
    <5e>   DW_AT_stmt_list   : 0x3a
    <62>   DW_AT_comp_dir    : (indirect string, offset: 0x7a):

/usr/local/google/home/sdf/tmp/lto

quoted

    <66>   DW_AT_low_pc      : 0x201740
    <6e>   DW_AT_high_pc     : 0x1f
 <1><72>: Abbrev Number: 4 (DW_TAG_subprogram)
    <73>   DW_AT_low_pc      : 0x201740
    <7b>   DW_AT_high_pc     : 0x1f
    <7f>   DW_AT_frame_base  : 1 byte block: 56 	(DW_OP_reg6 (rbp))
    <81>   DW_AT_name        : (indirect string, offset: 0x9d):

main

quoted

    <85>   DW_AT_decl_file   : 1
    <86>   DW_AT_decl_line   : 7
    <87>   DW_AT_type        : <0xae>
    <8b>   DW_AT_external    : 1
 <2><8b>: Abbrev Number: 5 (DW_TAG_variable)
    <8c>   DW_AT_location    : 2 byte block: 91 78 	(DW_OP_fbreg:

-8)

quoted

    <8f>   DW_AT_name        : (indirect string, offset: 0xb): s
    <93>   DW_AT_decl_file   : 1
    <94>   DW_AT_decl_line   : 9
    <95>   DW_AT_type        : <0xb5>
 <2><99>: Abbrev Number: 6 (DW_TAG_inlined_subroutine)
    <9a>   DW_AT_abstract_origin: <0x3f>
    <9e>   DW_AT_low_pc      : 0x201752
    <a6>   DW_AT_high_pc     : 0x5
    <aa>   DW_AT_call_file   : 1
    <ab>   DW_AT_call_line   : 10
    <ac>   DW_AT_call_column : 
 <2><ad>: Abbrev Number: 0
 <1><ae>: Abbrev Number: 7 (DW_TAG_base_type)
    <af>   DW_AT_name        : (indirect string, offset: 0xd): int
    <b3>   DW_AT_encoding    : 5	(signed)
    <b4>   DW_AT_byte_size   : 4
 <1><b5>: Abbrev Number: 8 (DW_TAG_structure_type)
    <b6>   DW_AT_name        : (indirect string, offset: 0xb): s
    <ba>   DW_AT_byte_size   : 4
    <bb>   DW_AT_decl_file   : 1
    <bc>   DW_AT_decl_line   : 1
 <2><bd>: Abbrev Number: 9 (DW_TAG_member)
    <be>   DW_AT_name        : (indirect string, offset: 0xa2): x
    <c2>   DW_AT_type        : <0xae>
    <c6>   DW_AT_decl_file   : 1
    <c7>   DW_AT_decl_line   : 2
    <c8>   DW_AT_data_member_location: 0
 <2><c9>: Abbrev Number: 0
 <1><ca>: Abbrev Number: 0

-- 

- Arnaldo

`h`	back out one level
`j`	next message in thread
`k`	previous message in thread
`l`	drill in
`Esc`	close help / fold thread tree
`?`	toggle this help