Re: [PATCH 1/2] perf symbols: Ignore mapping symbols on aarch64

From: Victor Kamensky
Date: Wed Jan 14 2015 - 13:38:44 EST


On 14 January 2015 at 03:22, Will Deacon <will.deacon@xxxxxxx> wrote:
> On Tue, Jan 13, 2015 at 04:59:03PM +0000, Victor Kamensky wrote:
>> Aarch64 ELF files use mapping symbols with special names $x, $d
>> to identify regions of Aarch64 code (see Aarch64 ELF ABI - "ARM
>> IHI 0056B", section "4.5.4 Mapping symbols").
>>
>> The patch filters out these symbols at load time, similar to
>> "696b97a perf symbols: Ignore mapping symbols on ARM" changes
>> done for ARM before V8.
>>
>> Signed-off-by: Victor Kamensky <victor.kamensky@xxxxxxxxxx>
>> Cc: Peter Zijlstra <a.p.zijlstra@xxxxxxxxx>
>> Cc: Paul Mackerras <paulus@xxxxxxxxx>
>> Cc: Ingo Molnar <mingo@xxxxxxxxxx>
>> Cc: Arnaldo Carvalho de Melo <acme@xxxxxxxxxx>
>> Cc: Adrian Hunter <adrian.hunter@xxxxxxxxx>
>> Cc: Jiri Olsa <jolsa@xxxxxxxxxx>
>> Cc: Namhyung Kim <namhyung@xxxxxxxxxx>
>> Cc: Avi Kivity <avi@xxxxxxxxxxxxxxxxxxxx>
>> Cc: Masami Hiramatsu <masami.hiramatsu.pt@xxxxxxxxxxx>
>> Cc: Anton Blanchard <anton@xxxxxxxxx>
>> Cc: David Ahern <dsahern@xxxxxxxxx>
>> Cc: Will Deacon <will.deacon@xxxxxxx>
>> Cc: Dave Martin <Dave.Martin@xxxxxxx>
>> ---
>> tools/perf/util/symbol-elf.c | 8 ++++++++
>> 1 file changed, 8 insertions(+)
>>
>> diff --git a/tools/perf/util/symbol-elf.c b/tools/perf/util/symbol-elf.c
>> index 06fcd1b..1e188dd 100644
>> --- a/tools/perf/util/symbol-elf.c
>> +++ b/tools/perf/util/symbol-elf.c
>> @@ -862,6 +862,14 @@ int dso__load_sym(struct dso *dso, struct map *map,
>> !strcmp(elf_name, "$t"))
>> continue;
>> }
>> + /* Reject Aarch64 ELF "mapping symbols": these aren't unique and
>> + * don't identify functions, so will confuse the profile
>> + * output: */
>> + if (ehdr.e_machine == EM_AARCH64) {
>> + if (!strcmp(elf_name, "$x") ||
>> + !strcmp(elf_name, "$d"))
>> + continue;
>> + }
>
> Do we need to skip $x.* and $d.* too? I doubt GCC generates them, but they
> are permitted by the ELF ABI.

Fair enough. But I think it would need to be done for both EM_ARM and
EM_AARCH64. My above patch follows EM_ARM current case. Also it
seems that it would be quite rare case, as ABI suggests symbols
with dot notation should be used in manual asm where assembler does not
support multiple definitions of the same symbol.

Will something like the following suffice? I can add that as separate follow up
patch in this small series. Tested with explicit "$x.func1" symbol introduced in
manual asm.