Re: [Xenomai-core] Problem with gcc-4.5.1

[Gilles Chanteperdrix <gilles.chanteperdrix@xenomai.org>]

Anders Blomdell wrote:
> On 12/07/2010 12:51 PM, Gilles Chanteperdrix wrote:
>> Anders Blomdell wrote:
>>> When compiling Xenomai on Fedora-14 with gcc-4.5.1 [version 4.5.1
>>> 20100924 (Red Hat 4.5.1-4)], the loading of xeno_nucleus fails with the
>>> attached kernel OOPS, a notable difference between the 4.5.1 compiled
>>> version and a working one built with gcc-4.4.4 on the same system with
>>> the same configuration, sis tthat __rthal_x86_nodiv_ullimd is not
>>> inlined, is this anybody has seen before?
>> No, that is new, we need to see the disassembly of __rthal_x86_nodiv_ullimd
> 
> objdump -S:
> 
> static inline __attribute__((const)) unsigned long long
> __rthal_x86_nodiv_ullimd(const unsigned long long op,
> 			 const unsigned long long frac,
> 			 unsigned integ)
> {
>      e7a8:	55                   	push   %ebp
>      e7a9:	89 e5                	mov    %esp,%ebp
>      e7ab:	57                   	push   %edi
>      e7ac:	56                   	push   %esi
>      e7ad:	53                   	push   %ebx
>      e7ae:	83 ec 10             	sub    $0x10,%esp
>      e7b1:	8d 7d 08             	lea    0x8(%ebp),%edi
>      e7b4:	e8 fc ff ff ff       	call   e7b5 <__rthal_x86_nodiv_ullimd+0xd>
>      e7b9:	8b 1f                	mov    (%edi),%ebx
>      e7bb:	8b 4f 04             	mov    0x4(%edi),%ecx
> 	register unsigned rm __asm__("esi");
> 	register unsigned rh __asm__("edi");
> 	unsigned fracl, frach, opl, oph;
> 	register unsigned long long t;
> 
> 	__rthal_u64tou32(op, oph, opl);
>      e7be:	89 45 e8             	mov    %eax,-0x18(%ebp)
> 	__rthal_u64tou32(frac, frach, fracl);
>      e7c1:	89 5d f0             	mov    %ebx,-0x10(%ebp)
> 	register unsigned rm __asm__("esi");
> 	register unsigned rh __asm__("edi");
> 	unsigned fracl, frach, opl, oph;
> 	register unsigned long long t;
> 
> 	__rthal_u64tou32(op, oph, opl);
>      e7c4:	89 55 e4             	mov    %edx,-0x1c(%ebp)
> 	__rthal_u64tou32(frac, frach, fracl);
>      e7c7:	89 4d ec             	mov    %ecx,-0x14(%ebp)
> 
> 	__asm__ ("mov %[oph], %%eax\n\t"
>      e7ca:	8b 45 e4             	mov    -0x1c(%ebp),%eax
>      e7cd:	f7 65 ec             	mull   -0x14(%ebp)
>      e7d0:	89 c6                	mov    %eax,%esi
>      e7d2:	89 d7                	mov    %edx,%edi
>      e7d4:	8b 45 e8             	mov    -0x18(%ebp),%eax
>      e7d7:	f7 65 f0             	mull   -0x10(%ebp)
>      e7da:	89 d1                	mov    %edx,%ecx
>      e7dc:	d1 e0                	shl    %eax
>      e7de:	83 d1 00             	adc    $0x0,%ecx
>      e7e1:	83 d6 00             	adc    $0x0,%esi
>      e7e4:	83 d7 00             	adc    $0x0,%edi
>      e7e7:	8b 45 e4             	mov    -0x1c(%ebp),%eax
>      e7ea:	f7 65 f0             	mull   -0x10(%ebp)
>      e7ed:	01 c1                	add    %eax,%ecx
>      e7ef:	11 d6                	adc    %edx,%esi
>      e7f1:	83 d7 00             	adc    $0x0,%edi
>      e7f4:	8b 45 e8             	mov    -0x18(%ebp),%eax
>      e7f7:	f7 65 ec             	mull   -0x14(%ebp)
>      e7fa:	01 c1                	add    %eax,%ecx
>      e7fc:	11 d6                	adc    %edx,%esi
>      e7fe:	83 d7 00             	adc    $0x0,%edi
>      e801:	8b 45 e8             	mov    -0x18(%ebp),%eax
>      e804:	f7 67 08             	mull   0x8(%edi)

Problem is here: edi is used by gcc as if it contained an address
whereas it is used by the assembly for the computation. Should be marked
"early clobber". So,

in include/asm-x86/arith_32.h, replace:

: [rl]"=c"(rl), [rm]"=S"(rm), [rh]"=D"(rh), "=A"(t)

with:

: [rl]"=&c"(rl), [rm]"=&S"(rm), [rh]"=&D"(rh), "=&A"(t)


> But us I said, in the working version, the code seems to be inlined 
> everywhere. Should I send the two object modules as well (probably as a 
> private message?).

The code should work the same whatever gcc decides regarding inlining.
Whether we like gcc decision is a different issue. Note that there is an
option to get gcc to go back to the old behaviour (inlining as the
source command).

-- 
                                                                Gilles.