Re: [PATCH v1 2/2] Strip build path directories in tools, xen and xen/arch/x86

[Jan Beulich <jbeulich@xxxxxxxx>]

On 19.03.2025 14:40, Marek Marczykowski-Górecki wrote:
> On Wed, Mar 19, 2025 at 01:43:59PM +0100, Jan Beulich wrote:
>> On 19.03.2025 12:58, Marek Marczykowski-Górecki wrote:
>>> On Wed, Mar 19, 2025 at 10:43:12AM +0100, Jan Beulich wrote:
>>>> On 19.03.2025 10:15, Jan Beulich wrote:
>>>>> On 18.03.2025 18:01, Marek Marczykowski-Górecki wrote:
>>>>>> --- a/xen/Makefile
>>>>>> +++ b/xen/Makefile
>>>>>> @@ -411,6 +411,8 @@ ifneq ($(CONFIG_CC_IS_CLANG),y)
>>>>>>  CFLAGS += -Wa,--strip-local-absolute
>>>>>>  endif
>>>>>>  
>>>>>> +$(call cc-option-add CFLAGS,CC,-ffile-prefix-map=$(XEN_ROOT)=.)
>>>>>
>>>>> This is lacking a comma:
>>>>>
>>>>> $(call cc-option-add,CFLAGS,CC,-ffile-prefix-map=$(XEN_ROOT)=.)
>>>>
>>>> And then, having tried the correct form (seeing the option then is passed
>>>> to the compiler), I can't spot any difference in the resulting
>>>> xen-syms.map. There were a few absolute paths there before (for
>>>> arch/x86/x86_64/kexec_reloc.S and arch/x86/acpi/wakeup_prot.S), and the
>>>> exact same ones are present afterwards.
>>>
>>> I'm not sure about xen-syms.map, it's about build path included in
>>> xen-syms. It appears at least once in .debug_str and once in
>>> .debug_line_str.
>>
>> In which case -fdebug-prefix-map= may suffice, which is available on
>> more compiler versions? And then only if DEBUG_INFO=y?
> 
> Oh, and xen.efi is full of build path. Binary on plain staging has 790
> occurrences. But there, -fdebug-prefix-map= also helps.
> 
> But also I don't think -fdebug-prefix-map= will be enough for tools, it
> looks like at least libxl has build path embedded in .rodata too.

And _all_ of them go away with -ffile-prefix-map=?

>>> But also, I see the patch lost a chunk during rebase (from before
>>> 4.17...), that adjusts XEN_ROOT to use $(realpath ...). That's the part
>>> even mentioned in the commit message...
>>>
>>> I'll send v2 shortly.
>>
>> Provided there's actually a need. I was in fact wondering whether this
>> was known to have significant effect prior to Anthony's work to make
>> out-of-tree builds possible (plus less related adjustments), but may
>> have lost most of its functionality since then (yet was still carried
>> forward for all the time).
> 
> There are clearly some build path embedding left. And
> -ffile-prefix-map=/-fdebug-prefix-map= doesn't work correctly with
> XEN_ROOT having xen/.. at the end.
> BTW, would it be acceptable to have this?
> 
>     $(call cc-option-add,CFLAGS,CC,-fdebug-prefix-map=$(realpath 
> $(XEN_ROOT))=.)
> 
> It may be less efficient (if make doesn't cache result),

What do you mean here? Variable evaluation depends solely on how we
use variables. I don't think there's any caching make does on its own?

As to $(realpath ...) - make 3.80 doesn't support that. We do provide a
fallback, but for that you need to use $(call realpath,...).

Jan

> but helps
> especially in tools, where XEN_ROOT is set in _a lot_ of places.
>