llvm dev - Apr 2016 - LTO renaming of constants with inline assembly

On Wed, Apr 6, 2016 at 10:46 AM, Peter Collingbourne <peter at pcc.me.uk>
wrote:
> I suspect that the right way to do promotion/renaming of this sort is to
> rename at the MC layer just before writing the symbol table to the object
> file.
>
I think that is too late - how would the symbols be distinguished in the
LTO case below after the IR is linked but before we renamed the duplicate?

Teresa

> Peter
>
> On Wed, Apr 6, 2016 at 10:37 AM, Teresa Johnson via llvm-dev <
> llvm-dev at lists.llvm.org> wrote:
>
>> I encountered an issue with ThinLTO handling of inline assembly, where
>> the inline assembly referenced a constant that was a local variable.
The
>> local var was renamed because it was promoted in ThinLTO mode, but the
>> inline assembly copy was not renamed and we ended up with an undef at
link
>> time.
>>
>> It looks like this is a general problem with inline assembly and LTO.
>> Wondering if it is a known issue. E.g. if I link in LTO mode two files
that
>> have inline assembly referencing local constants with the same name,
the
>> LTO linking will rename the second. However, the renaming doesn't
propagate
>> to inline assembly, resulting in the wrong output.
>>
>> For example, let's say we have two modules with inline assembly
that
>> writes a local constant var named "myvar" into the memory
pointed to by its
>> parameter "v", and a simple main that calls each function:
>>
>> $ cat inlineasm1.c
>> static const unsigned char __attribute__((used)) __attribute__
((aligned
>> (1))) myvar = 1;
>>
>> void foo(unsigned long int *v) {
>>   __asm__ volatile("movzbl     myvar(%%rip), %%eax\n\t"
>>                    "movq %%rax, %0\n\t"
>>                        : "=*m" (*v)
>>     :
>>     : "%eax"
>>     );
>> }
>>
>> $ cat inlineasm2.c
>> static const unsigned char __attribute__((used)) __attribute__
((aligned
>> (1))) myvar = 2;
>>
>> void bar(unsigned long int *v) {
>>   __asm__ volatile("movzbl     myvar(%%rip), %%eax\n\t"
>>                    "movq %%rax, %0\n\t"
>>                        : "=*m" (*v)
>>     :
>>     : "%eax"
>>     );
>> }
>>
>> $ cat inlineasm.c
>> #include <stdio.h>
>> extern void foo(unsigned long int *v);
>> extern void bar(unsigned long int *v);
>> int main() {
>>   unsigned long int f,b;
>>   foo(&f);
>>   bar(&b);
>>   printf("%lu %lu\n", f, b);
>> }
>>
>>
>> If compiled at -O2 (no LTO) this correctly prints out "1 2".
>>
>> However, when linked with LTO, the second copy of local
"myvar" is
>> renamed to "myvar.6". But the inline assembly which is still
hidden within
>> a call that hasn't been lowered, still references "myvar"
in that second
>> linked copy in bar(). The output is thus incorrect: "1 1" (or
"2 2" if the
>> bar() copy was linked first).
>>
>> Is this a known issue? Any ideas on how we could handle this?
>>
>> Thanks,
>> Teresa
>>
>> --
>> Teresa Johnson |  Software Engineer |  tejohnson at google.com |
>> 408-460-2413
>>
>> _______________________________________________
>> LLVM Developers mailing list
>> llvm-dev at lists.llvm.org
>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>>
>>
>
>
> --
> --
> Peter
>


-- 
Teresa Johnson |  Software Engineer |  tejohnson at google.com |  408-460-2413
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20160406/b8119199/attachment.html>

On Wed, Apr 6, 2016 at 10:49 AM, Teresa Johnson <tejohnson at google.com>
wrote:
>
>
> On Wed, Apr 6, 2016 at 10:46 AM, Peter Collingbourne <peter at
pcc.me.uk>
> wrote:
>
>> I suspect that the right way to do promotion/renaming of this sort is
to
>> rename at the MC layer just before writing the symbol table to the
object
>> file.
>>
>
> I think that is too late - how would the symbols be distinguished in the
> LTO case below after the IR is linked but before we renamed the duplicate?
>
Sorry, wasn't fully awake. I think we could do something along the lines of
the symbol renaming idea, but with directives that limit their scope to
inline asm blocks.

Specifically, we could teach the frontend to produce a mapping from symbol
names to globalvalues, for any internal names with the used attribute, and
attach that mapping to inline asm blocks.

For example, if myvar were renamed to myvar.6, the IR would look like this:

@myvar.6 = global i8 [...]

[...]

call asm("movzbl  myvar(%rip), ...", ..., "myvar")(..., i8*
@myvar.6)

The backend would produce assembly that would look like this:

.rename myvar, myvar.6
movzbl     myvar(%%rip), ...
.norename myvar

The .rename and .norename directives would delimit the scope of the
renaming.

Peter
>
> Teresa
>
>
>> Peter
>>
>> On Wed, Apr 6, 2016 at 10:37 AM, Teresa Johnson via llvm-dev <
>> llvm-dev at lists.llvm.org> wrote:
>>
>>> I encountered an issue with ThinLTO handling of inline assembly,
where
>>> the inline assembly referenced a constant that was a local
variable. The
>>> local var was renamed because it was promoted in ThinLTO mode, but
the
>>> inline assembly copy was not renamed and we ended up with an undef
at link
>>> time.
>>>
>>> It looks like this is a general problem with inline assembly and
LTO.
>>> Wondering if it is a known issue. E.g. if I link in LTO mode two
files that
>>> have inline assembly referencing local constants with the same
name, the
>>> LTO linking will rename the second. However, the renaming
doesn't propagate
>>> to inline assembly, resulting in the wrong output.
>>>
>>> For example, let's say we have two modules with inline assembly
that
>>> writes a local constant var named "myvar" into the memory
pointed to by its
>>> parameter "v", and a simple main that calls each
function:
>>>
>>> $ cat inlineasm1.c
>>> static const unsigned char __attribute__((used)) __attribute__
((aligned
>>> (1))) myvar = 1;
>>>
>>> void foo(unsigned long int *v) {
>>>   __asm__ volatile("movzbl     myvar(%%rip), %%eax\n\t"
>>>                    "movq %%rax, %0\n\t"
>>>                        : "=*m" (*v)
>>>     :
>>>     : "%eax"
>>>     );
>>> }
>>>
>>> $ cat inlineasm2.c
>>> static const unsigned char __attribute__((used)) __attribute__
((aligned
>>> (1))) myvar = 2;
>>>
>>> void bar(unsigned long int *v) {
>>>   __asm__ volatile("movzbl     myvar(%%rip), %%eax\n\t"
>>>                    "movq %%rax, %0\n\t"
>>>                        : "=*m" (*v)
>>>     :
>>>     : "%eax"
>>>     );
>>> }
>>>
>>> $ cat inlineasm.c
>>> #include <stdio.h>
>>> extern void foo(unsigned long int *v);
>>> extern void bar(unsigned long int *v);
>>> int main() {
>>>   unsigned long int f,b;
>>>   foo(&f);
>>>   bar(&b);
>>>   printf("%lu %lu\n", f, b);
>>> }
>>>
>>>
>>> If compiled at -O2 (no LTO) this correctly prints out "1
2".
>>>
>>> However, when linked with LTO, the second copy of local
"myvar" is
>>> renamed to "myvar.6". But the inline assembly which is
still hidden within
>>> a call that hasn't been lowered, still references
"myvar" in that second
>>> linked copy in bar(). The output is thus incorrect: "1 1"
(or "2 2" if the
>>> bar() copy was linked first).
>>>
>>> Is this a known issue? Any ideas on how we could handle this?
>>>
>>> Thanks,
>>> Teresa
>>>
>>> --
>>> Teresa Johnson |  Software Engineer |  tejohnson at google.com |
>>> 408-460-2413
>>>
>>> _______________________________________________
>>> LLVM Developers mailing list
>>> llvm-dev at lists.llvm.org
>>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>>>
>>>
>>
>>
>> --
>> --
>> Peter
>>
>
>
>
> --
> Teresa Johnson |  Software Engineer |  tejohnson at google.com |
> 408-460-2413
>


-- 
-- 
Peter
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20160406/2d61aa7f/attachment.html>

On Wed, Apr 6, 2016 at 11:16 AM, Peter Collingbourne <peter at pcc.me.uk>
wrote:
> On Wed, Apr 6, 2016 at 10:49 AM, Teresa Johnson <tejohnson at
google.com>
> wrote:
>
>>
>>
>> On Wed, Apr 6, 2016 at 10:46 AM, Peter Collingbourne <peter at
pcc.me.uk>
>> wrote:
>>
>>> I suspect that the right way to do promotion/renaming of this sort
is to
>>> rename at the MC layer just before writing the symbol table to the
object
>>> file.
>>>
>>
>> I think that is too late - how would the symbols be distinguished in
the
>> LTO case below after the IR is linked but before we renamed the
duplicate?
>>
>
> Sorry, wasn't fully awake. I think we could do something along the
lines
> of the symbol renaming idea, but with directives that limit their scope to
> inline asm blocks.
>
> Specifically, we could teach the frontend to produce a mapping from symbol
> names to globalvalues, for any internal names with the used attribute, and
> attach that mapping to inline asm blocks.
>
> For example, if myvar were renamed to myvar.6, the IR would look like this:
>
> @myvar.6 = global i8 [...]
>
> [...]
>
> call asm("movzbl  myvar(%rip), ...", ..., "myvar")(...,
i8* @myvar.6)
>
> The backend would produce assembly that would look like this:
>
> .rename myvar, myvar.6
> movzbl     myvar(%%rip), ...
> .norename myvar
>
> The .rename and .norename directives would delimit the scope of the
> renaming.
>
That's an interesting idea, thanks. Are .rename and .norename standard
directives? I did some web searches but couldn't find anything concrete on
them (I did find a .rename in some IBM Power documentation, but it seemed
to apply to string constants.

I think for ThinLTO purposes I will limit importing to/from modules with
inline assembly for now to avoid the issue.

Teresa

> Peter
>
>>
>> Teresa
>>
>>
>>> Peter
>>>
>>> On Wed, Apr 6, 2016 at 10:37 AM, Teresa Johnson via llvm-dev <
>>> llvm-dev at lists.llvm.org> wrote:
>>>
>>>> I encountered an issue with ThinLTO handling of inline
assembly, where
>>>> the inline assembly referenced a constant that was a local
variable. The
>>>> local var was renamed because it was promoted in ThinLTO mode,
but the
>>>> inline assembly copy was not renamed and we ended up with an
undef at link
>>>> time.
>>>>
>>>> It looks like this is a general problem with inline assembly
and LTO.
>>>> Wondering if it is a known issue. E.g. if I link in LTO mode
two files that
>>>> have inline assembly referencing local constants with the same
name, the
>>>> LTO linking will rename the second. However, the renaming
doesn't propagate
>>>> to inline assembly, resulting in the wrong output.
>>>>
>>>> For example, let's say we have two modules with inline
assembly that
>>>> writes a local constant var named "myvar" into the
memory pointed to by its
>>>> parameter "v", and a simple main that calls each
function:
>>>>
>>>> $ cat inlineasm1.c
>>>> static const unsigned char __attribute__((used)) __attribute__
>>>> ((aligned (1))) myvar = 1;
>>>>
>>>> void foo(unsigned long int *v) {
>>>>   __asm__ volatile("movzbl     myvar(%%rip),
%%eax\n\t"
>>>>                    "movq %%rax, %0\n\t"
>>>>                        : "=*m" (*v)
>>>>     :
>>>>     : "%eax"
>>>>     );
>>>> }
>>>>
>>>> $ cat inlineasm2.c
>>>> static const unsigned char __attribute__((used)) __attribute__
>>>> ((aligned (1))) myvar = 2;
>>>>
>>>> void bar(unsigned long int *v) {
>>>>   __asm__ volatile("movzbl     myvar(%%rip),
%%eax\n\t"
>>>>                    "movq %%rax, %0\n\t"
>>>>                        : "=*m" (*v)
>>>>     :
>>>>     : "%eax"
>>>>     );
>>>> }
>>>>
>>>> $ cat inlineasm.c
>>>> #include <stdio.h>
>>>> extern void foo(unsigned long int *v);
>>>> extern void bar(unsigned long int *v);
>>>> int main() {
>>>>   unsigned long int f,b;
>>>>   foo(&f);
>>>>   bar(&b);
>>>>   printf("%lu %lu\n", f, b);
>>>> }
>>>>
>>>>
>>>> If compiled at -O2 (no LTO) this correctly prints out "1
2".
>>>>
>>>> However, when linked with LTO, the second copy of local
"myvar" is
>>>> renamed to "myvar.6". But the inline assembly which
is still hidden within
>>>> a call that hasn't been lowered, still references
"myvar" in that second
>>>> linked copy in bar(). The output is thus incorrect: "1
1" (or "2 2" if the
>>>> bar() copy was linked first).
>>>>
>>>> Is this a known issue? Any ideas on how we could handle this?
>>>>
>>>> Thanks,
>>>> Teresa
>>>>
>>>> --
>>>> Teresa Johnson |  Software Engineer |  tejohnson at google.com
|
>>>> 408-460-2413
>>>>
>>>> _______________________________________________
>>>> LLVM Developers mailing list
>>>> llvm-dev at lists.llvm.org
>>>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>>>>
>>>>
>>>
>>>
>>> --
>>> --
>>> Peter
>>>
>>
>>
>>
>> --
>> Teresa Johnson |  Software Engineer |  tejohnson at google.com |
>> 408-460-2413
>>
>
>
>
> --
> --
> Peter
>


-- 
Teresa Johnson |  Software Engineer |  tejohnson at google.com |  408-460-2413
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20160406/ca175d83/attachment.html>