Inline assembly and value extensions

Inline assembly and value extensions

[Harald van Dijk]

Hi,

I am trying to port some x86-64 inline assembly to work properly for the 
x32 ABI and I am running into a small problem. The x32 ABI specifies 
that pointers are passed and returned zero-extended to 64 bits. When a 
pointer variable is defined by an inline assembly statement and then 
returned by a function, I do not see a way to inform GCC that the result 
is already zero-extended, that GCC does not need to zero-extend it again.

A silly example:

  void *return_a_pointer(void) {
    void *result;
    asm("movl $0x11223344, %%eax" : "=a"(result));
    return result;
  }

This function gets an extra "movl %eax, %eax" between the hand-written 
movl and the generated ret, which can be seen online at 
<https://godbolt.org/z/T8bGPo>. This extra movl is there to ensure the 
high bits of %rax are zero, but the initial movl already achieves that. 
How can I inform GCC that it does not need to emit that extra movl?

Likewise, is there an easy way to provide an inline assembly statement 
with a zero-extended pointer input? This one I am able to work around, 
as it is possible to instead of passing in a pointer value p, pass in an 
integer value (uint64_t)(uint32_t)p, but the workaround is kind of hard 
to read and I would like to avoid that if possible.

I looked the documentation for either relevant inline assembly 
constraints or relevant variable / type attributes, but was unable to 
find any. The most promising search result was the mode attribute, I was 
hoping it might be possible to give result a mode(DI) attribute, but the 
compiler rejects that.

Is there another approach that I can use instead?

Cheers,
Harald van Dijk

Re: Inline assembly and value extensions

[Harald van Dijk]

On 22/08/2020 16:30, Harald van Dijk wrote:
> Hi,
> 
> I am trying to port some x86-64 inline assembly to work properly for the 
> x32 ABI and I am running into a small problem. The x32 ABI specifies 
> that pointers are passed and returned zero-extended to 64 bits. When a 
> pointer variable is defined by an inline assembly statement and then 
> returned by a function, I do not see a way to inform GCC that the result 
> is already zero-extended, that GCC does not need to zero-extend it again.
> 
> A silly example:
> 
>   void *return_a_pointer(void) {
>     void *result;
>     asm("movl $0x11223344, %%eax" : "=a"(result));
>     return result;
>   }
> 
> This function gets an extra "movl %eax, %eax" between the hand-written 
> movl and the generated ret, which can be seen online at 
> <https://godbolt.org/z/T8bGPo>. This extra movl is there to ensure the 
> high bits of %rax are zero, but the initial movl already achieves that. 
> How can I inform GCC that it does not need to emit that extra movl?
> 
> Likewise, is there an easy way to provide an inline assembly statement 
> with a zero-extended pointer input? This one I am able to work around, 
> as it is possible to instead of passing in a pointer value p, pass in an 
> integer value (uint64_t)(uint32_t)p, but the workaround is kind of hard 
> to read and I would like to avoid that if possible.
> 
> I looked the documentation for either relevant inline assembly 
> constraints or relevant variable / type attributes, but was unable to 
> find any. The most promising search result was the mode attribute, I was 
> hoping it might be possible to give result a mode(DI) attribute, but the 
> compiler rejects that.

I have now found that forcing a different mode appears to be exactly how 
the zero-extension of arguments and return values is implemented: that 
is what ix86_promote_function_mode does.

The fact that this is not an option through variable attributes or 
inline assembly constraints looks like an unfortunate limitation of the 
inline assembly functionality, there is currently just no way to do what 
I am after. I very much hope to be proved wrong, but will try to just 
pick a workaround that does not look too bad.

> Is there another approach that I can use instead?
> 
> Cheers,
> Harald van Dijk

Re: Inline assembly and value extensions

[Segher Boessenkool]

On Sat, Oct 03, 2020 at 09:42:40PM +0100, Harald van Dijk via Gcc-help wrote:
> On 22/08/2020 16:30, Harald van Dijk wrote:
> >This function gets an extra "movl %eax, %eax" between the hand-written 
> >movl and the generated ret, which can be seen online at 
> ><https://godbolt.org/z/T8bGPo>. This extra movl is there to ensure the 
> >high bits of %rax are zero, but the initial movl already achieves that. 
> >How can I inform GCC that it does not need to emit that extra movl?

If your asm returns a 32-bit value, then GCC will not know what is in
the top 32 bits of the 64-bit register.

> >Likewise, is there an easy way to provide an inline assembly statement 
> >with a zero-extended pointer input? This one I am able to work around, 
> >as it is possible to instead of passing in a pointer value p, pass in an 
> >integer value (uint64_t)(uint32_t)p, but the workaround is kind of hard 
> >to read and I would like to avoid that if possible.

You can use much less chatty names for those very basic types (like u64
and u32), that makes it more readable.  Hiding what you do will not make
it more readable, that is just obfuscation, so using macros is not such
a good idea.  Inline asm is hard enough when you can see all there is
right in front of your eyes.

> >I looked the documentation for either relevant inline assembly 
> >constraints or relevant variable / type attributes, but was unable to 
> >find any. The most promising search result was the mode attribute, I was 
> >hoping it might be possible to give result a mode(DI) attribute, but the 
> >compiler rejects that.

Constraints just say which register (or memory addressed how, or what
kind of constnt).  The normal way to say something should have a certain
mode is by giving it a corresponding type in C (so SImode is "int" in C,
and DImode is "long long"; "long" is either, it depends on your ABI;
"u64" and "u32" should always be clear ;-) )

> I have now found that forcing a different mode appears to be exactly how 
> the zero-extension of arguments and return values is implemented: that 
> is what ix86_promote_function_mode does.
> 
> The fact that this is not an option through variable attributes or 
> inline assembly constraints looks like an unfortunate limitation of the 
> inline assembly functionality, there is currently just no way to do what 
> I am after. I very much hope to be proved wrong, but will try to just 
> pick a workaround that does not look too bad.
> 
> >Is there another approach that I can use instead?

You use a 64-bit expression (or preferably even a 64-bit variable).  The
same is true for outputs from the asm.

One way of making the code easier to read is to actually use 64-bit
variables for all these things in the asm, and then assign them from and
to the 32-bit things.

  int x;
  char *p;
  u64 xx = x;
  u64 pp = (u64)p;
  asm("smth %0,%1" : "+r"(pp), "+r"(xx));
  p = (char *)pp;
  x = xx;


Segher

Re: Inline assembly and value extensions

[Harald van Dijk]

On 05/10/2020 13:53, Segher Boessenkool wrote:
> On Sat, Oct 03, 2020 at 09:42:40PM +0100, Harald van Dijk via Gcc-help wrote:
>> On 22/08/2020 16:30, Harald van Dijk wrote:
>>> This function gets an extra "movl %eax, %eax" between the hand-written
>>> movl and the generated ret, which can be seen online at
>>> <https://godbolt.org/z/T8bGPo>. This extra movl is there to ensure the
>>> high bits of %rax are zero, but the initial movl already achieves that.
>>> How can I inform GCC that it does not need to emit that extra movl?
> 
> If your asm returns a 32-bit value, then GCC will not know what is in
> the top 32 bits of the 64-bit register.
> 
>>> Likewise, is there an easy way to provide an inline assembly statement
>>> with a zero-extended pointer input? This one I am able to work around,
>>> as it is possible to instead of passing in a pointer value p, pass in an
>>> integer value (uint64_t)(uint32_t)p, but the workaround is kind of hard
>>> to read and I would like to avoid that if possible.
> 
> You can use much less chatty names for those very basic types (like u64
> and u32), that makes it more readable.  Hiding what you do will not make
> it more readable, that is just obfuscation, so using macros is not such
> a good idea.  Inline asm is hard enough when you can see all there is
> right in front of your eyes.

I am not trying to hide anything. I am trying to pass in a pointer value 
in the same way that they are passed in all other cases. In the x32 ABI, 
the way pointer values are passed is by passing them in 64-bit registers 
with the high 32 bits zeroed, except in inline assembly. I'm trying to 
get the inline assembly to not be a special case.

>>> I looked the documentation for either relevant inline assembly
>>> constraints or relevant variable / type attributes, but was unable to
>>> find any. The most promising search result was the mode attribute, I was
>>> hoping it might be possible to give result a mode(DI) attribute, but the
>>> compiler rejects that.
> 
> Constraints just say which register (or memory addressed how, or what
> kind of constnt).  The normal way to say something should have a certain
> mode is by giving it a corresponding type in C (so SImode is "int" in C,
> and DImode is "long long"; "long" is either, it depends on your ABI;
> "u64" and "u32" should always be clear ;-) )

In the x32 ABI, pointers do not have a single mode. They are SImode, 
except when passed as parameters or returned, in which case they are 
DImode (see the ix86_promote_function_mode function I mentioned). This 
is really the source of the problems.

The value that I want to return is a pointer, and the assembly is 
returning a pointer. There should not be any need for extra code to 
convert the value returned from the assembly to the value returned from 
the function, since they are already the same type, but there is, 
because they are not the same mode.

>> I have now found that forcing a different mode appears to be exactly how
>> the zero-extension of arguments and return values is implemented: that
>> is what ix86_promote_function_mode does.
>>
>> The fact that this is not an option through variable attributes or
>> inline assembly constraints looks like an unfortunate limitation of the
>> inline assembly functionality, there is currently just no way to do what
>> I am after. I very much hope to be proved wrong, but will try to just
>> pick a workaround that does not look too bad.
>>
>>> Is there another approach that I can use instead?
> 
> You use a 64-bit expression (or preferably even a 64-bit variable).  The
> same is true for outputs from the asm.

That does not work for outputs though. Without informing GCC somehow 
that the high 32 bits of that 64-bit expression are zero, it will still 
emit an extra extension.

> One way of making the code easier to read is to actually use 64-bit
> variables for all these things in the asm, and then assign them from and
> to the 32-bit things.
> 
>    int x;
>    char *p;
>    u64 xx = x;
>    u64 pp = (u64)p;
>    asm("smth %0,%1" : "+r"(pp), "+r"(xx));
>    p = (char *)pp;
>    x = xx;

See <https://godbolt.org/z/Kx6d9q> for what happens when I modify the 
example I had provided to use a 64-bit variable like you suggested: that 
requires adding a cast in the return statement, and now it's that 
conversion that forces the extra "movl %eax, %eax". It cannot be avoided 
that way.

> Segher

Re: Inline assembly and value extensions

[Florian Weimer]

* Harald van Dijk via Gcc-help:

> In the x32 ABI, pointers do not have a single mode. They are SImode,
> except when passed as parameters or returned, in which case they are 
> DImode (see the ix86_promote_function_mode function I mentioned). This
> is really the source of the problems.

I think that GCC cannot optimize away unecessary sign and zero
extensions is also a contributing factor, no?

Thanks,
Florian
-- 
Red Hat GmbH, https://de.redhat.com/ , Registered seat: Grasbrunn,
Commercial register: Amtsgericht Muenchen, HRB 153243,
Managing Directors: Charles Cachera, Brian Klemm, Laurie Krebs, Michael O'Neill

Re: Inline assembly and value extensions

[Harald van Dijk]

On 05/10/2020 14:29, Florian Weimer wrote:
> * Harald van Dijk via Gcc-help:
> 
>> In the x32 ABI, pointers do not have a single mode. They are SImode,
>> except when passed as parameters or returned, in which case they are
>> DImode (see the ix86_promote_function_mode function I mentioned). This
>> is really the source of the problems.
> 
> I think that GCC cannot optimize away unecessary sign and zero
> extensions is also a contributing factor, no?

Hmm, yes, that is a good point too. An ugly way to try to inform GCC 
that the extension is unnecessary would be

   void *return_a_pointer(void) {
     unsigned long long result;
     asm("movl $0x11223344, %%eax" : "=a"(result));
     if (result & 0xFFFFFFFF00000000ULL)
       __builtin_unreachable();
     return (void *)result;
   }

This is not something I would want to have to write explicitly, but it 
does not help anyway at the moment, even in this case GCC emits the 
extra extension: <https://godbolt.org/z/vPb1Mr>. I think here the code 
does give enough information to GCC that it could in theory optimize it 
away.

(clang does not emit the extra extension, but that is because LLVM 
implements this ABI incorrectly, so not a fair comparison. I hope to be 
able to submit a patch to them for that at some later time.)