GCC needs __attribute__((returns_twice)) for SetJump

GCC needs __attribute__((returns_twice)) for SetJump

[Michael Zimmermann]

When compiling with any ARM toolchain and Os, registers can get
trashed when returning for the second time from SetJump because GCC
only handles this correctly when using standard names like 'setjmp' or
'getcontext'. When different names are used you have to use the
attribute 'returns_twice' to tell gcc to be extra careful.

example:
#define FN_NAME nonstandard_setjmp
extern int  FN_NAME(void*);

void jmp_buf_set(void *jmpb, void (*f)(void))
{
  if (!FN_NAME(jmpb))
    f();
}

this code produces this wrong code with Os:
00000000 <jmp_buf_set>:
   0: e92d4010 push {r4, lr}
   4: e1a04001 mov r4, r1
   8: ebfffffe bl 0 <nonstandard_setjmp>
   c: e3500000 cmp r0, #0
  10: 01a03004 moveq r3, r4
  14: 08bd4010 popeq {r4, lr}
  18: 012fff13 bxeq r3
  1c: e8bd4010 pop {r4, lr}
  20: e12fff1e bx lr

The generated code pushes backups of r4 and lr to the stack and then
saves all registers using nonstandard_setjmp.
Then it pops the stack and jumps to the function in r3 which is the
main problem because now the function can overwrite our register
backups on the stack.
When we return a second time from the call to nonstandard_setjmp, the
stack pointer has it's original(pushed) position and when the code
pops r4 and lr from the stack the values are not guaranteed to be the
same.

When using a standard name like setjmp or getcontext or adding
'__attribute__((returns_twice))' to nonstandard_setjmp's declaration
the code looks different:

00000000 <jmp_buf_set>:
   0: e92d4007 push {r0, r1, r2, lr}
   4: e58d1004 str r1, [sp, #4]
   8: ebfffffe bl 0 <setjmp>
   c: e3500000 cmp r0, #0
  10: 059d3004 ldreq r3, [sp, #4]
  14: 01a0e00f moveq lr, pc
  18: 012fff13 bxeq r3
  1c: e28dd00c add sp, sp, #12
  20: e49de004 pop {lr} ; (ldr lr, [sp], #4)
  24: e12fff1e bx lr

Here the problem is being solved by restoring r3 from the stack
without popping it.

I would have sent a patch but since there's no define for
'returns_twice' yet and I don't know how other compilers handle this I
want to discuss this first.

Thanks
Michael

Re: GCC needs __attribute__((returns_twice)) for SetJump

[Ard Biesheuvel]

On 8 December 2016 at 13:30, Michael Zimmermann
<sigmaepsilon92@gmail.com> wrote:
> When compiling with any ARM toolchain and Os, registers can get
> trashed when returning for the second time from SetJump because GCC
> only handles this correctly when using standard names like 'setjmp' or
> 'getcontext'. When different names are used you have to use the
> attribute 'returns_twice' to tell gcc to be extra careful.
>
> example:
> #define FN_NAME nonstandard_setjmp
> extern int  FN_NAME(void*);
>
> void jmp_buf_set(void *jmpb, void (*f)(void))
> {
>   if (!FN_NAME(jmpb))
>     f();
> }
>
> this code produces this wrong code with Os:
> 00000000 <jmp_buf_set>:
>    0: e92d4010 push {r4, lr}
>    4: e1a04001 mov r4, r1
>    8: ebfffffe bl 0 <nonstandard_setjmp>
>    c: e3500000 cmp r0, #0
>   10: 01a03004 moveq r3, r4
>   14: 08bd4010 popeq {r4, lr}
>   18: 012fff13 bxeq r3
>   1c: e8bd4010 pop {r4, lr}
>   20: e12fff1e bx lr
>
> The generated code pushes backups of r4 and lr to the stack and then
> saves all registers using nonstandard_setjmp.
> Then it pops the stack and jumps to the function in r3 which is the
> main problem because now the function can overwrite our register
> backups on the stack.
> When we return a second time from the call to nonstandard_setjmp, the
> stack pointer has it's original(pushed) position and when the code
> pops r4 and lr from the stack the values are not guaranteed to be the
> same.
>
> When using a standard name like setjmp or getcontext or adding
> '__attribute__((returns_twice))' to nonstandard_setjmp's declaration
> the code looks different:
>
> 00000000 <jmp_buf_set>:
>    0: e92d4007 push {r0, r1, r2, lr}
>    4: e58d1004 str r1, [sp, #4]
>    8: ebfffffe bl 0 <setjmp>
>    c: e3500000 cmp r0, #0
>   10: 059d3004 ldreq r3, [sp, #4]
>   14: 01a0e00f moveq lr, pc
>   18: 012fff13 bxeq r3
>   1c: e28dd00c add sp, sp, #12
>   20: e49de004 pop {lr} ; (ldr lr, [sp], #4)
>   24: e12fff1e bx lr
>
> Here the problem is being solved by restoring r3 from the stack
> without popping it.
>
> I would have sent a patch but since there's no define for
> 'returns_twice' yet and I don't know how other compilers handle this I
> want to discuss this first.
>

Well spotted!

This issue applies to all GCC supported architectures, not just ARM,
and so I think we need to solve this generically.

I have no idea how other toolchains deal with this, although I assume
Clang will support the same attribute

Re: GCC needs __attribute__((returns_twice)) for SetJump

[Michael Zimmermann]

I've tested it with clang 3.9.0 and it seems to support this attribute.
But what about MSVC and all the other compilers which can be found in
edk2's tools_def?

I wasn't able to find anything useful about twice-returning function
support for these compilers. So do they just never apply optimizations
which could break such functions or is there no way of fixing this?

On Thu, Dec 8, 2016 at 3:32 PM, Ard Biesheuvel
<ard.biesheuvel@linaro.org> wrote:
> On 8 December 2016 at 13:30, Michael Zimmermann
> <sigmaepsilon92@gmail.com> wrote:
>> When compiling with any ARM toolchain and Os, registers can get
>> trashed when returning for the second time from SetJump because GCC
>> only handles this correctly when using standard names like 'setjmp' or
>> 'getcontext'. When different names are used you have to use the
>> attribute 'returns_twice' to tell gcc to be extra careful.
>>
>> example:
>> #define FN_NAME nonstandard_setjmp
>> extern int  FN_NAME(void*);
>>
>> void jmp_buf_set(void *jmpb, void (*f)(void))
>> {
>>   if (!FN_NAME(jmpb))
>>     f();
>> }
>>
>> this code produces this wrong code with Os:
>> 00000000 <jmp_buf_set>:
>>    0: e92d4010 push {r4, lr}
>>    4: e1a04001 mov r4, r1
>>    8: ebfffffe bl 0 <nonstandard_setjmp>
>>    c: e3500000 cmp r0, #0
>>   10: 01a03004 moveq r3, r4
>>   14: 08bd4010 popeq {r4, lr}
>>   18: 012fff13 bxeq r3
>>   1c: e8bd4010 pop {r4, lr}
>>   20: e12fff1e bx lr
>>
>> The generated code pushes backups of r4 and lr to the stack and then
>> saves all registers using nonstandard_setjmp.
>> Then it pops the stack and jumps to the function in r3 which is the
>> main problem because now the function can overwrite our register
>> backups on the stack.
>> When we return a second time from the call to nonstandard_setjmp, the
>> stack pointer has it's original(pushed) position and when the code
>> pops r4 and lr from the stack the values are not guaranteed to be the
>> same.
>>
>> When using a standard name like setjmp or getcontext or adding
>> '__attribute__((returns_twice))' to nonstandard_setjmp's declaration
>> the code looks different:
>>
>> 00000000 <jmp_buf_set>:
>>    0: e92d4007 push {r0, r1, r2, lr}
>>    4: e58d1004 str r1, [sp, #4]
>>    8: ebfffffe bl 0 <setjmp>
>>    c: e3500000 cmp r0, #0
>>   10: 059d3004 ldreq r3, [sp, #4]
>>   14: 01a0e00f moveq lr, pc
>>   18: 012fff13 bxeq r3
>>   1c: e28dd00c add sp, sp, #12
>>   20: e49de004 pop {lr} ; (ldr lr, [sp], #4)
>>   24: e12fff1e bx lr
>>
>> Here the problem is being solved by restoring r3 from the stack
>> without popping it.
>>
>> I would have sent a patch but since there's no define for
>> 'returns_twice' yet and I don't know how other compilers handle this I
>> want to discuss this first.
>>
>
> Well spotted!
>
> This issue applies to all GCC supported architectures, not just ARM,
> and so I think we need to solve this generically.
>
> I have no idea how other toolchains deal with this, although I assume
> Clang will support the same attribute