This is the mail archive of the mailing list for the binutils project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [PATCH] ARM: Add support for SHF_ARM_NOREAD section flag

On 08/01/16 13:33, Richard Earnshaw (lists) wrote:
On 08/01/16 13:00, Christophe MONAT wrote:
Hi Andre, Nick,

[Sorry that Mickael does not answer by himself, he was not subscribed to
the list so far -- now corrected].

On 01/07/16 19:33, Andre Vieira wrote:
Terry sent a patch upstream to handle the noread attribute in 2014:

Having seen this patch I believe the approach taken here to use section
names to represent the noread attribute in assembly is inferior to
Terry's approach.

For the GCC implementation of either an attribute or compile option for
execute-only we should not use section names to represent the noread
attribute, since for instance that means it can not be combined with
-ffunction-sections, or any other option that sets section names for

We disagree with that specific point : the section names that we emit
when gcc is using -ffunction-sections is in the form of:
which are perfectly caught and handled (it matters to us - we just
checked this).

I would like to rebase Terry's patch and make the necessary changes to
it, slightly different attribute name and so on, and use that instead of
this patch.

Would there be any objections to this?

The binutils patch that we contributed was in the perspective of
up-streaming the so-called PCROP support also in gcc also - the gcc
proposal is completed on our side, but still not public.

For the matter of marking the .text sections read-only, we tried the two
following strategies (and choose 2)):

1) keep the .text sections' names and emit the noread attribute in the
assembly (with the very same 'y' key) : we failed doing that because of
the specific treatment done by gas, that in the end *ignores* at some
final point the custom attributes on the pure .text name.
This implies that the .text sections in noread mode cannot be called
.text but must be called .text.something (which in your patch's tests
appears as, otherwise the noread attribute would not been have
accounted for).
In addition from the gcc standpoint, this emission forces to duplicate
emission code since the particular place we need to touch is hook-able,
but requires a complete duplication of the section attribute emission
(default_elf_asm_named_section) just to add two lines dealing with the
'y' case.

2) emit some .text.noread sections (that work nicely with
-ffunction-sections), without even requiring a section attribute change,
and delegating the treatment to gas that handles specifically those

This to write that your patch is certainly very good (just regretting
that it was not contributed at that time) but leads to uglier code in
gcc, which you may consider a problem.

Would it be possible to simply add the 'y' support from your patch, this
would not break our changes and fulfill your purposes ?

Just a quick comment, I don't currently have time to dig into this deeply.

My main concern is that requiring a specific template in the section
name feels like a cludge.  If gas/ld is not correctly dealing with the
section flags, then that is a bug that ought to be fixed properly.
Working around that by inventing new special meta-data is not helpful --
don't forget that other linkers besides GNU LD have to be able to
process these sections (and GNU LD has to process object files produced
by tools other than GCC/GAS).


Best regards,

So yes the unnamed, aka '.text', section can not be used as is because the assembler has a "default" .text with attributes 'ax' and when it merges with the .text from the incoming assembly it drops other attributes.

This however is not a problem for gcc as the plans for an execute-only option were to name any unnamed section that was noread and I don't mind using .text.noread as a default. However, my opinion is that that should not be the denominator for a 'noread' section. Instead it should really just be seen as a way to circumvent this property of unnamed sections.

I don't know why unnamed sections have this behavior though, I will try to dig that one up and see if maybe we ought to change the merging behavior of unnamed sections with different attributes.

I currently see three approaches we could take in the assembler to deal with unnamed sections with 'noread' ('y'): 1) merge all '.text' sections and if one of them is 'noread', all should be (not a big fan!); 2) rename, at assemble time, all '.text' sections that have the 'noread' attribute to '.text.noread'; 3) don't merge (current behavior), but warn about dropping the attribute, the code for this is already there, it just doesnt check for the NOREAD flag.

My favorite approach would be 3. That way the user would be hinted that he/she should name their 'noread' section.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]