Skip to contentSkip to conversation

dwarf_loader: encode char type as signed

PATCH In review Yonghong Song 3 participants 5 messages Aug 7, 2022 – Aug 10, 2022 BPF

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 dwarf_loader: encode char type as signed

Currently, the pahole treats 'char' or 'signed char' type as unsigned in BTF generation. The following is an example,

  $ cat t.c
  signed char a;
  char b;
  $ clang -O2 -g -c t.c
  $ pahole -JV t.o
  ...
  [1] INT signed char size=1 nr_bits=8 encoding=(none)
  [2] INT char size=1 nr_bits=8 encoding=(none)
In the above encoding '(none)' implies unsigned type.

But if the same program is compiled with bpf target,

  $ clang -target bpf -O2 -g -c t.c
  $ bpftool btf dump file t.o
  [1] INT 'signed char' size=1 bits_offset=0 nr_bits=8 encoding=SIGNED
  [2] VAR 'a' type_id=1, linkage=global
  [3] INT 'char' size=1 bits_offset=0 nr_bits=8 encoding=SIGNED
  [4] VAR 'b' type_id=3, linkage=global
  [5] DATASEC '.bss' size=0 vlen=2
          type_id=2 offset=0 size=1 (VAR 'a')
          type_id=4 offset=0 size=1 (VAR 'b')
the 'char' and 'signed char' are encoded as SIGNED integers.

Encode 'char' and 'signed char' as SIGNED should be a right to do and it will be consistent with bpf implementation.

With this patch,

  $ pahole -JV t.o
  ...
  [1] INT signed char size=1 nr_bits=8 encoding=SIGNED
  [2] INT char size=1 nr_bits=8 encoding=SIGNED
---
 dwarf_loader.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
dwarf_loader.c +1 -1
index d892bc3..c2ad2a0 100644
@@ -560,7 +560,7 @@ static struct base_type *base_type__new(Dwarf_Die *die, struct cu *cu, struct co
 		bt->bit_size = attr_numeric(die, DW_AT_byte_size) * 8;
 		uint64_t encoding = attr_numeric(die, DW_AT_encoding);
 		bt->is_bool = encoding == DW_ATE_boolean;
-		bt->is_signed = encoding == DW_ATE_signed;
+		bt->is_signed = (encoding == DW_ATE_signed) || (encoding == DW_ATE_signed_char);
 		bt->is_varargs = false;
 		bt->name_has_encoding = true;
 		bt->float_type = encoding_to_float_type(encoding);
Signature · 1 line
2.30.2
#1 lore raw
Quoted 28 lines · Yonghong Song

Currently, the pahole treats 'char' or 'signed char' type as unsigned in BTF generation. The following is an example,

  $ cat t.c
  signed char a;
  char b;
  $ clang -O2 -g -c t.c
  $ pahole -JV t.o
  ...
  [1] INT signed char size=1 nr_bits=8 encoding=(none)
  [2] INT char size=1 nr_bits=8 encoding=(none)
In the above encoding '(none)' implies unsigned type.

But if the same program is compiled with bpf target,

  $ clang -target bpf -O2 -g -c t.c
  $ bpftool btf dump file t.o
  [1] INT 'signed char' size=1 bits_offset=0 nr_bits=8 encoding=SIGNED
  [2] VAR 'a' type_id=1, linkage=global
  [3] INT 'char' size=1 bits_offset=0 nr_bits=8 encoding=SIGNED
  [4] VAR 'b' type_id=3, linkage=global
  [5] DATASEC '.bss' size=0 vlen=2
          type_id=2 offset=0 size=1 (VAR 'a')
          type_id=4 offset=0 size=1 (VAR 'b')
the 'char' and 'signed char' are encoded as SIGNED integers.

Encode 'char' and 'signed char' as SIGNED should be a right to do and it will be consistent with bpf implementation.

Quoted 8 lines

With this patch,

  $ pahole -JV t.o
  ...
  [1] INT signed char size=1 nr_bits=8 encoding=SIGNED
  [2] INT char size=1 nr_bits=8 encoding=SIGNED

---

LGTM.

Is there a plan to also add CHAR encoding bit?

Quoted 19 lines
 dwarf_loader.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
dwarf_loader.c +1 -2
index d892bc3..c2ad2a0 100644
@@ -560,7 +560,7 @@ static struct base_type *base_type__new(Dwarf_Die *die, struct cu *cu, struct co
                bt->bit_size = attr_numeric(die, DW_AT_byte_size) * 8;
                uint64_t encoding = attr_numeric(die, DW_AT_encoding);
                bt->is_bool = encoding == DW_ATE_boolean;
-               bt->is_signed = encoding == DW_ATE_signed;
+               bt->is_signed = (encoding == DW_ATE_signed) || (encoding == DW_ATE_signed_char);
                bt->is_varargs = false;
                bt->name_has_encoding = true;
                bt->float_type = encoding_to_float_type(encoding);
--

2.30.2

#2 lore raw
Quoted 36 lines · Andrii Nakryiko
Quoted 34 lines · Yonghong Song

Currently, the pahole treats 'char' or 'signed char' type as unsigned in BTF generation. The following is an example,

  $ cat t.c
  signed char a;
  char b;
  $ clang -O2 -g -c t.c
  $ pahole -JV t.o
  ...
  [1] INT signed char size=1 nr_bits=8 encoding=(none)
  [2] INT char size=1 nr_bits=8 encoding=(none)
In the above encoding '(none)' implies unsigned type.

But if the same program is compiled with bpf target,

  $ clang -target bpf -O2 -g -c t.c
  $ bpftool btf dump file t.o
  [1] INT 'signed char' size=1 bits_offset=0 nr_bits=8 encoding=SIGNED
  [2] VAR 'a' type_id=1, linkage=global
  [3] INT 'char' size=1 bits_offset=0 nr_bits=8 encoding=SIGNED
  [4] VAR 'b' type_id=3, linkage=global
  [5] DATASEC '.bss' size=0 vlen=2
          type_id=2 offset=0 size=1 (VAR 'a')
          type_id=4 offset=0 size=1 (VAR 'b')
the 'char' and 'signed char' are encoded as SIGNED integers.

Encode 'char' and 'signed char' as SIGNED should be a right to do and it will be consistent with bpf implementation.

With this patch,

  $ pahole -JV t.o
  ...
  [1] INT signed char size=1 nr_bits=8 encoding=SIGNED
  [2] INT char size=1 nr_bits=8 encoding=SIGNED
Quoted 4 lines
Quoted 2 lines

---

LGTM.

Quoted 23 lines

Is there a plan to also add CHAR encoding bit?

Quoted 19 lines
 dwarf_loader.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
dwarf_loader.c +1 -2
index d892bc3..c2ad2a0 100644
@@ -560,7 +560,7 @@ static struct base_type *base_type__new(Dwarf_Die *die, struct cu *cu, struct co
                bt->bit_size = attr_numeric(die, DW_AT_byte_size) * 8;
                uint64_t encoding = attr_numeric(die, DW_AT_encoding);
                bt->is_bool = encoding == DW_ATE_boolean;
-               bt->is_signed = encoding == DW_ATE_signed;
+               bt->is_signed = (encoding == DW_ATE_signed) || (encoding == DW_ATE_signed_char);
                bt->is_varargs = false;
                bt->name_has_encoding = true;
                bt->float_type = encoding_to_float_type(encoding);
--

2.30.2

#3 lore raw

Em Mon, Aug 08, 2022 at 03:52:38PM -0700, Andrii Nakryiko escreveu:

Quoted 38 lines
Quoted 36 lines · Andrii Nakryiko
Quoted 34 lines · Yonghong Song

Currently, the pahole treats 'char' or 'signed char' type as unsigned in BTF generation. The following is an example,

  $ cat t.c
  signed char a;
  char b;
  $ clang -O2 -g -c t.c
  $ pahole -JV t.o
  ...
  [1] INT signed char size=1 nr_bits=8 encoding=(none)
  [2] INT char size=1 nr_bits=8 encoding=(none)
In the above encoding '(none)' implies unsigned type.

But if the same program is compiled with bpf target,

  $ clang -target bpf -O2 -g -c t.c
  $ bpftool btf dump file t.o
  [1] INT 'signed char' size=1 bits_offset=0 nr_bits=8 encoding=SIGNED
  [2] VAR 'a' type_id=1, linkage=global
  [3] INT 'char' size=1 bits_offset=0 nr_bits=8 encoding=SIGNED
  [4] VAR 'b' type_id=3, linkage=global
  [5] DATASEC '.bss' size=0 vlen=2
          type_id=2 offset=0 size=1 (VAR 'a')
          type_id=4 offset=0 size=1 (VAR 'b')
the 'char' and 'signed char' are encoded as SIGNED integers.

Encode 'char' and 'signed char' as SIGNED should be a right to do and it will be consistent with bpf implementation.

With this patch,

  $ pahole -JV t.o
  ...
  [1] INT signed char size=1 nr_bits=8 encoding=SIGNED
  [2] INT char size=1 nr_bits=8 encoding=SIGNED
Quoted 6 lines
Quoted 4 lines
Quoted 2 lines

---

LGTM.

Thanks, tested before/after, applied.

Pushing out next for CI testing.

- Arnaldo

#4 lore raw
In reply to Andrii Nakryiko
Quoted 35 lines · Andrii Nakryiko
Quoted 34 lines · Yonghong Song

Currently, the pahole treats 'char' or 'signed char' type as unsigned in BTF generation. The following is an example,

   $ cat t.c
   signed char a;
   char b;
   $ clang -O2 -g -c t.c
   $ pahole -JV t.o
   ...
   [1] INT signed char size=1 nr_bits=8 encoding=(none)
   [2] INT char size=1 nr_bits=8 encoding=(none)
In the above encoding '(none)' implies unsigned type.

But if the same program is compiled with bpf target,

   $ clang -target bpf -O2 -g -c t.c
   $ bpftool btf dump file t.o
   [1] INT 'signed char' size=1 bits_offset=0 nr_bits=8 encoding=SIGNED
   [2] VAR 'a' type_id=1, linkage=global
   [3] INT 'char' size=1 bits_offset=0 nr_bits=8 encoding=SIGNED
   [4] VAR 'b' type_id=3, linkage=global
   [5] DATASEC '.bss' size=0 vlen=2
           type_id=2 offset=0 size=1 (VAR 'a')
           type_id=4 offset=0 size=1 (VAR 'b')
the 'char' and 'signed char' are encoded as SIGNED integers.

Encode 'char' and 'signed char' as SIGNED should be a right to do and it will be consistent with bpf implementation.

With this patch,

   $ pahole -JV t.o
   ...
   [1] INT signed char size=1 nr_bits=8 encoding=SIGNED
   [2] INT char size=1 nr_bits=8 encoding=SIGNED
Quoted 6 lines
Quoted 2 lines

---

LGTM.

Is there a plan to also add CHAR encoding bit?

Not yet... Need more thinking about use cases and backward comparability issues.

Quoted 21 lines
Quoted 21 lines

  dwarf_loader.c | 2 +-
  1 file changed, 1 insertion(+), 1 deletion(-)
dwarf_loader.c +1 -2
index d892bc3..c2ad2a0 100644
@@ -560,7 +560,7 @@ static struct base_type *base_type__new(Dwarf_Die *die, struct cu *cu, struct co
                 bt->bit_size = attr_numeric(die, DW_AT_byte_size) * 8;
                 uint64_t encoding = attr_numeric(die, DW_AT_encoding);
                 bt->is_bool = encoding == DW_ATE_boolean;
-               bt->is_signed = encoding == DW_ATE_signed;
+               bt->is_signed = (encoding == DW_ATE_signed) || (encoding == DW_ATE_signed_char);
                 bt->is_varargs = false;
                 bt->name_has_encoding = true;
                 bt->float_type = encoding_to_float_type(encoding);
--

2.30.2