[PATCH] libxc-x86-64-fixes.patch

[PATCH] libxc-x86-64-fixes.patch

[Arun Sharma]

Build fixes for x86-64.

Signed-off-by: Nitin Kamble <nitin.a.kamble@intel.com>
Signed-off-by: Arun Sharma <arun.sharma@intel.com>

--- 1.4/tools/libxc/xc_ptrace.c	2005-03-20 07:07:46 -08:00
+++ edited/libxc/xc_ptrace.c	2005-03-29 15:18:27 -08:00
@@ -281,9 +281,9 @@
 		SET_PT_REGS(pt, ctxt.cpu_ctxt); 
 		memcpy(data, &pt, sizeof(elf_gregset_t));
 	} else if (request == PTRACE_GETFPREGS)
-	    memcpy(data, &ctxt.fpu_ctxt, sizeof(elf_fpregset_t));
+	    memcpy(data, &ctxt.fpu_ctxt, sizeof(ctxt.fpu_ctxt));
 	else /*if (request == PTRACE_GETFPXREGS)*/
-	    memcpy(data, &ctxt.fpu_ctxt, sizeof(elf_fpxregset_t));
+	    memcpy(data, &ctxt.fpu_ctxt, sizeof(ctxt.fpu_ctxt));
 	cr3 = ctxt.pt_base;
 	regs_valid = 1;
 	break;
--- 1.14/tools/libxc/xc_vmx_build.c	2005-03-20 03:53:25 -08:00
+++ edited/libxc/xc_vmx_build.c	2005-03-29 15:20:56 -08:00
@@ -478,10 +478,18 @@
 {
     int eax, ecx;
 
+#ifdef __i386__
     __asm__ __volatile__ ("pushl %%ebx; cpuid; popl %%ebx" 
 			  : "=a" (eax), "=c" (ecx) 
 			  : "0" (1) 
 			  : "dx");
+#elif defined __x86_64__
+    __asm__ __volatile__ ("pushq %%rbx; cpuid; popq %%rbx"
+                          : "=a" (eax), "=c" (ecx)
+                          : "0" (1)
+                          : "dx");
+#endif
+
     if (!(ecx & VMX_FEATURE_FLAG)) {
         return -1;
     }

Re: [PATCH] libxc-x86-64-fixes.patch

[Anthony Liguori]

Arun Sharma wrote:

> 
>+#ifdef __i386__
>     __asm__ __volatile__ ("pushl %%ebx; cpuid; popl %%ebx" 
> 			  : "=a" (eax), "=c" (ecx) 
> 			  : "0" (1) 
> 			  : "dx");
>+#elif defined __x86_64__
>+    __asm__ __volatile__ ("pushq %%rbx; cpuid; popq %%rbx"
>+                          : "=a" (eax), "=c" (ecx)
>+                          : "0" (1)
>+                          : "dx");
>+#endif
>+
>  
>
Does this mean the ebx-clobbering bug in gcc 3.4 also exists on x86-64 
(except it clobbers rbx instead)?

I really hate to see this end up a permanent part of the tree...  
perhaps we should add a Linux style cpuid() function:

static inline void cpuid(int op, int *eax, int *ebx, int *ecx, int *edx)
{
    int ax, bx, cx, dx;

#if defined __i386__ || defined __x86_64__
    __asm__("cpuid"
        : "=a" (ax),
          "=b" (bx),
          "=c" (cx),
          "=d" (dx)
        : "0" (op));
#else
#error cpuid not defined on current architecture
#endif

    if (eax) *eax = ax;
    if (ebx) *ebx = bx;
    if (ecx) *ecx = cx
    if (edx) *edx = dx;
}

This should take care of the ebx clobbering bug while also resulting in 
more shared code.

Just a thought.

Regards,
Anthony Liguori

Re: [PATCH] libxc-x86-64-fixes.patch

[Arun Sharma]

Anthony Liguori wrote:
> Arun Sharma wrote:
> 
>>
>> +#ifdef __i386__
>>     __asm__ __volatile__ ("pushl %%ebx; cpuid; popl %%ebx" 
>>               : "=a" (eax), "=c" (ecx)               : "0" (1) 
>>               : "dx");
>> +#elif defined __x86_64__
>> +    __asm__ __volatile__ ("pushq %%rbx; cpuid; popq %%rbx"
>> +                          : "=a" (eax), "=c" (ecx)
>> +                          : "0" (1)
>> +                          : "dx");
>> +#endif
>> +
>>  
>>
> Does this mean the ebx-clobbering bug in gcc 3.4 also exists on x86-64 
> (except it clobbers rbx instead)?
> 

There was no bug in pre gcc-3.4 generated code in the original version 
of vmx_identify(). The problem was with some other test case where the 
compiler generated code used %ebx within the inline assembly (which 
would be broken because cpuid clobbers %ebx and compiler didn't realize 
that).

gcc developers instead of making only those cases illegal, made even 
cases such as ours which look perfectly legal to me, illegal (because 
it's too much work for them to figure out what's legal and what's not).

> I really hate to see this end up a permanent part of the tree...  
> perhaps we should add a Linux style cpuid() function:
> 
> static inline void cpuid(int op, int *eax, int *ebx, int *ecx, int *edx)
> {
>    int ax, bx, cx, dx;
> 
> #if defined __i386__ || defined __x86_64__
>    __asm__("cpuid"
>        : "=a" (ax),
>          "=b" (bx),
>          "=c" (cx),
>          "=d" (dx)
>        : "0" (op));

As I pointed out in the earlier thread, the linux code in 
<asm/processor.h> is also broken when used in userland with -fPIC.

I used this particular case (<asm/processor.h>) as an argument with gcc 
developers on why they shouldn't be making code that was legal with an 
earlier version of gcc suddenly illegal. But I couldn't provide enough 
use cases to make a strong argument and since an easy (although ugly) 
workaround was available, I settled for it.

	-Arun