Re: [PATCH] take 2 of the tr-based current

[Manfred Spraul <manfred@colorfullife.com>]

[-- Attachment #1: Type: text/plain, Size: 1053 bytes --]

Anton Blanchard wrote:

>Hi,
>
>>No. const == never changes.
>>get_TR changes if a task calls schedule, and return on another cpu.
>>
>
>Yes, I found this exact problem on ppc64 where we would cache the
>processor data area across a schedule(). What was interesting was that
>__attribute__ ((pure)) was not enough to fix this.
>
>static inline struct Paca *get_paca(void) __attribute__ ((pure));
>static inline struct Paca *get_paca(void)
>{
>	struct Paca *rval;
>	__asm__ ("mfspr %0,0x113" : "=r" (rval));
>	return rval;
>}
>
>Alan Modra came to the rescue and found that gcc was optimising too much
>and since the function did not touch any global variables, it would
>upgrade the pure to const. This was on gcc 3.0.X.
>
But the function called schedule - mustn't gcc assume that schedule 
writes into global variables?
As far as I can see that sounds like a gcc bug.

Could you try how many get_cpu calls are generated by the attached testapp?
the i386 RH compilers generate correct code:
gcc-2.96-98: 4 calls.
gcc3-3.0.1-3: 2 calls.

--
    Manfred

[-- Attachment #2: test.c --]
[-- Type: application/msword, Size: 487 bytes --]