Re: [PATCH 0/4] -ffreestanding/-fno-builtin-* patches

[Rasmus Villemoes <linux@rasmusvillemoes.dk>]

On 20/08/2020 19.56, Arvind Sankar wrote:
> On Thu, Aug 20, 2020 at 04:56:02PM +0200, Rasmus Villemoes wrote:
>> On 18/08/2020 23.41, Arvind Sankar wrote:
>>>
>>> Note that -fno-builtin-foo seems to mean slightly different things in
>>> clang and gcc. From experimentation, clang will neither optimize a call
>>> to foo, nor perform an optimization that introduces a call to foo. gcc
>>> will avoid optimizing calls to foo, but it can still generate new calls
>>> to foo while optimizing something else. Which means that
>>> -fno-builtin-{bcmp,stpcpy} only solves things for clang, not gcc. It's
>>> just that gcc doesn't seem to have implemented those optimizations.
>>>
>>
>> I think it's more than that. I've always read gcc's documentation
>>
>> '-fno-builtin'
>> '-fno-builtin-FUNCTION'
>>      Don't recognize built-in functions that do not begin with
>>      '__builtin_' as prefix. ...
>>
>>      GCC normally generates special code to handle certain built-in
>>      functions more efficiently; for instance, calls to 'alloca' may
>>      become single instructions which adjust the stack directly, and
>>      calls to 'memcpy' may become inline copy loops.
>>      ...
>>
>> to mean exactly that observed above and nothing more, i.e. that
>> -fno-builtin-foo merely means that gcc stops treating a call of a
>> function named foo to mean a call to a function implementing the
>> standard function by that name (and hence allows it to e.g. replace a
>> memcpy(d, s, 1) by byte load+store). It does not mean to prevent
>> emitting calls to foo, and I don't think it ever will - it's a bit sad
>> that clang has chosen to interpret these options differently.
> 
> That documentation is misleading, as it also goes on to say:
> "...nor can you change the behavior of the functions by linking with a
> different library"
> which implies that you _can_ change the behavior if you use the option,
> and which is what your "i.e." is saying as well.
> 
> My point is that this is not completely true: in gcc, foo by default is
> defined to be __builtin_foo, and -fno-builtin-foo simply removes this
> definition. So the effect is just that calls to foo in the original
> source will be left alone.

Yes, this is a much better way of putting it. And with -fbuiltin-foo in
effect, the compiler just needs to transform the code in some way as-if
the standard function by that name was called, which it can of course
decide to implement by emitting such a call, but it can also open-code
it - or synthesize it using other std functions.

> But in order for an optimization that introduces a new call to foo to be
> valid, foo _must_ have standard semantics: strchr(s,'\0') is not s +
> strlen(s) unless strlen has standard semantics.

Correct. So I agree that -fno-builtin-strlen should prevent the compiler
from generating calls to strlen() that don't appear in the code.

This is an oversight in
> gcc's optimizations: it converts to s + __builtin_strlen(s), which then
> (normally) becomes s + strlen(s).
> 
> Check out this horror: https://godbolt.org/z/a1r9fK
> 
> Clang will disable this optimization if -fno-builtin-strlen is
> specified.
>
> Clang's interpretation is more useful for embedded, since you can use
> -fno-builtin-foo and avoid calling __builtin_foo directly, and be
> guaranteed that there will be no calls to foo that you didn't write
> explicitly (outside of memcpy/memset/memcmp). In this case you are free
> to implement foo with non-standard semantics, or avoid implementing it
> altogether, and be reasonably confident that it will all work.

Yeah, except that the list of -fno-builtin-foo one would have to pass is
enourmous, so for targets with a somewhat wonky libc, I'd much rather be
able to do a blanket -fno-builtin, and then manually check their memcpy,
memset and memcmp implementations and opt back in with
-fbuiltin-mem{cpy,set,cmp} so that small constant-size memcpys do get
properly open-coded.

The advice in gcc's documentation of just #definining memcpy() to
__builtin_memcpy() doesn't work in the real world (for example it breaks
C++ code that uses std::memcpy(...)).

>> Thinking out load, it would be useful if both compilers grew
>>
>>   -fassume-provided-std-foo
>>
>> and
>>
>>   -fno-assume-provided-std-foo
>>
>> options to tell the compiler that a function named foo with standard
>> semantics can be assumed (or not) to be provided by the execution
>> environment; i.e. one half of what -f(no-)builtin-foo apparently does
>> for clang currently.
> 
> Not following: -fno-assume-provided-std-foo sounds like it would have
> exactly the same semantics as Clang's -fno-builtin-foo, except maybe in
> addition it should cause the compiler to error on seeing __builtin_foo
> if it can't implement that without calling foo.

Yeah, I think you've convinced me there's no use for a separate option
to prevent inventing calls to foo() - I was mostly thinking of it as a
way to avoid having to provide each and every libc function that may
have been half-way standardized at some point. But if one doesn't
provide, say, bcmp, the code base certainly doesn't use bcmp itself, so
one doesn't lose anything by just using -fno-builtin-bcmp; there are no
explicit bcmp() uses that fail to get optimized.

Rasmus