[U-Boot] [RFC PATCH v2] ARM: Avoid compiler optimization for usages of readb, writeb and friends.

[Albert ARIBAUD]

Le 21/12/2010 09:46, John Rigby a écrit :
> On Tue, Dec 21, 2010 at 1:35 AM, Dirk Behme<dirk.behme at googlemail.com>  wrote:
>>
>> (Resend with corrected broken example)
>>
>> On 21.12.2010 08:21, Albert ARIBAUD wrote:
>>> Hi Dirk,
>>>
>>> Le 21/12/2010 08:11, Dirk Behme a écrit :
>>>
>>>> But the issue with drivers/mtd/nand/omap_gpmc.c (i.e. the additional
>>>> ldrb    r3, [r3]) is still open? Has anybody tried to replace it with
>>>> a nop in the binary to be sure this is the root cause?
>>>
>>> Can you try and preprocess the C file for both the broken and working
>>> cases, then post the preprocessed C extract? Differences at the C level
>>> may help understanding differences at the asm level.
>>
>> gcc version 4.5.1 (Sourcery G++ Lite 2010.09-50)
>>
>> Work:
>> ====
>>
>> static void omap_nand_hwcontrol(struct mtd_info *mtd, int32_t cmd,
>>      uint32_t ctrl)
>> {
>>   register struct nand_chip *this = mtd->priv;
>>   ...
>>   if (cmd != -1)
>>
>>    (*(volatile unsigned char *)(this->IO_ADDR_W) = (cmd));
>> }
>>
>>         if (cmd != NAND_CMD_NONE)
>>    84:  e3710001        cmn     r1, #1
>>                 origwriteb(cmd, this->IO_ADDR_W);
>>    88:  15933004        ldrne   r3, [r3, #4]
>>    8c:  120110ff        andne   r1, r1, #255    ; 0xff
>>    90:  15c31000        strbne  r1, [r3]
>>    94:  e12fff1e        bx      lr
>>         ...
>>
>>
>> Broken:
>> ======
>>
>> static void omap_nand_hwcontrol(struct mtd_info *mtd, int32_t cmd,
>>      uint32_t ctrl)
>> {
>>   register struct nand_chip *this = mtd->priv;
>>
>> ...
>>
>>   if (cmd != -1)
>>    ({ __asm__ __volatile__ ("" : : : "memory"); (*(volatile unsigned
>> char *)(this->IO_ADDR_W) = (cmd)); });
>> }
>>
>>         if (cmd != NAND_CMD_NONE)
>>    84:  e3710001        cmn     r1, #1
>>    88:  012fff1e        bxeq    lr
>>                 writeb(cmd, this->IO_ADDR_W);
>>    8c:  e5933004        ldr     r3, [r3, #4]
>>    90:  e20110ff        and     r1, r1, #255    ; 0xff
>>    94:  e5c31000        strb    r1, [r3]
>>    98:  e5d33000        ldrb    r3, [r3]
>>    9c:  e12fff1e        bx      lr
>>
>>
>> The issue seems to be the additional 'ldrb      r3, [r3]' added by the
>> compiler in the broken version.
>>
>
> And I at your suggestion tried modifying the binary changing the extra
> ldrb to a nop and it works.

Seems like a compiler issue to me, as the preprocessed C source is the 
same for the register access and does not call for a re-read (that is 
what I wanted to see in the preprocessed version), yet the ASM sequence 
does the re-read.

Amicalement,
-- 
Albert.