Re: [PATCH] x86/asm: improve how GEN_*_SUFFIXED_RMWcc() specify clobbers

From: Jan Beulich
Date: Thu Feb 22 2018 - 03:05:19 EST


>>> On 21.02.18 at 22:39, <keescook@xxxxxxxxxxxx> wrote:
> On Mon, Feb 19, 2018 at 6:49 AM, Jan Beulich <JBeulich@xxxxxxxx> wrote:
>> Commit df3405245a ("x86/asm: Add suffix macro for GEN_*_RMWcc()")
>> introduced "suffix" RMWcc operations, adding bogus clobber specifiers:
>> For one, on x86 there's no point explicitly clobbering "cc". In fact,
>
> Do you have more details on this? It seems from the GCC clobbers
> docs[1] that "cc" is needed for asm that changes the flags register.
> Given the explicit subl and decl being used for these macros, it seems
> needed to me.
>
>> with gcc properly fixed, this results in an overlap being detected by
>> the compiler between outputs and clobbers. Further more it seems bad
>
> Do you mean the flags register is already being included as an
> implicit clobber because there is an output of any kind? I can't find
> documentation that says this is true. If anything it looks like it
> could be "improved" from a full "cc" clobber to an output operand
> flag, like =@ccs.

As hpa has already said, "cc" has been automatically included as a
clobber forever on x86 (until the condition code outputs appeared,
at which point the clobber became added only if no such output
was present).

>> --- 4.16-rc2/arch/x86/include/asm/rmwcc.h
>> +++ 4.16-rc2-x86-rmwcc-clobbers/arch/x86/include/asm/rmwcc.h
>> @@ -2,8 +2,7 @@
>> #ifndef _ASM_X86_RMWcc
>> #define _ASM_X86_RMWcc
>>
>> -#define __CLOBBERS_MEM "memory"
>> -#define __CLOBBERS_MEM_CC_CX "memory", "cc", "cx"
>> +#define __CLOBBERS_MEM(clb...) "memory", ## clb
>
> This leaves a trailing comma in the non-cx case. I thought that caused
> me problems in the past, but maybe that's GCC version-specific?

No, it's formally specified to drop the comma. Note this is using a
gcc extension, not the C99 way of having a macro with variable
number of arguments (where - without again some gcc special
handling - this indeed would be a problem).

Jan