Re: [PATCH 6/6] Compile sched_getaffinity.c with -fno-builtin-memset

["H.J. Lu" <hjl dot tools at gmail dot com>]

On Fri, Aug 14, 2015 at 8:46 AM, Joseph Myers <joseph@codesourcery.com> wrote:
> On Fri, 14 Aug 2015, H.J. Lu wrote:
>
>> On Fri, Aug 14, 2015 at 8:27 AM, Joseph Myers <joseph@codesourcery.com> wrote:
>> > On Fri, 14 Aug 2015, H.J. Lu wrote:
>> >
>> >> Since sched_getaffinity.c calls memset which may not be inlined, we
>> >> should compile it with -fno-builtin-memset so that the internal hidden
>> >> memset is called without PLT.
>> >>
>> >> OK for master?
>> >
>> > memset is called all over the place.  I don't like hardcoding special
>> > options for particular files based on calls to memset, and especially not
>> > based on what some particular compiler version happens to do with a
>> > particular call.  Why doesn't the call to libc_hidden_builtin_proto
>> > (memset) in include/string.h suffice?  Can you devise some set of
>> > declarations / toplevel asms to put in a header that will ensure memset
>> > calls go via a hidden alias (ideally with the compiler knowing it's
>> > hidden, not just the assembler / linker, so the compiler can do any
>> > optimizations based on it being an intra-library call)?
>>
>> libc_hidden_builtin_proto (memset)
>>
>> has no impact on memset inlined by compiler.
>
> Well, I see references to __GI_memset, not memset, in sched_getaffinity.os
> for both x86 and x86_64 (building with GCC 5).  So it seems to be working
> OK for me.  It's true some of those references use PLT-generating

Yes, it does reference __GI_memset, but via PLT, since GCC doesn't
know the libcall for the builtin memset is hidden.

> relocations, but those relocations shouldn't actually generate PLT entries
> when they are relocations against symbols not exported from the library -
> the linker should convert them to direct calls.  If they do generate PLT
> entries, it sounds like a linker optimization is missing.
>
> Apart from any missing linker optimizations, the compiler should know when

Linker isn't an issue.

> calls are to hidden functions as calls to those may be more efficient than
> calls that only the linker determines don't need a PLT entry.  On 32-bit
> x86, that means the compiler knowing it doesn't need to set %ebx up to

%ebx is the issue.

> point to the GOT as it would for a call through the PLT.  I don't know
> whether this gets optimized for built-in function calls using
> libc_hidden_builtin_proto, but if it doesn't, that seems like a missed
> optimization that should be addressed in the compiler not in glibc.
>

It is hard to say it is a GCC issue when we are kind of doing this behind
the back of GCC unless there is a way to tell GCC that the libcall of
a builtin function is hidden.

-- 
H.J.