On 2014-06-11, 1:17 PM, Chung-Lin Tang wrote:
> On 2014/6/11 下午 06:32, James Greenhalgh wrote:
>>
>> Hi,
>>
>> A recent change somewhere exposed a latent bug between LRA and the definition
>> of the movsi_compare0 pattern.
>>
>> This pattern ties the source and destination register of a set together
>> a (match_dup) and register constraints:
>>
>>     [(set (reg:CC CC_REGNUM)
>>      (compare:CC (match_operand:SI 1 "s_register_operand" "0,r")
>>                  (const_int 0)))
>>      (set (match_operand:SI 0 "s_register_operand" "=r,r")
>>      (match_dup 1))]
>>
>> This confuses LRA which expects the source and destination register of
>> a set to be different.
>>
>> reduced.c: In function '_IO_vfscanf_internal':
>> reduced.c:104:1: internal compiler error: in lra_create_copy, at lra.c:1512
>>       }
>>   ^
>> 0x8c3f9a lra_create_copy(int, int, int)
>>      /work/gcc-dev/src/gcc/gcc/lra.c:1512
>> 0x8e4ab0 process_bb_lives
>>      /work/gcc-dev/src/gcc/gcc/lra-lives.c:568
>> 0x8e4ab0 lra_create_live_ranges(bool)
>>      /work/gcc-dev/src/gcc/gcc/lra-lives.c:1019
>> 0x8c5a39 lra(_IO_FILE*)
>>      /work/gcc-dev/src/gcc/gcc/lra.c:2356
>> 0x873a96 do_reload
>>      /work/gcc-dev/src/gcc/gcc/ira.c:5415
>> 0x873a96 execute
>>      /work/gcc-dev/src/gcc/gcc/ira.c:5576
>> Please submit a full bug report,
>>
>> We can fix the pattern by moving away from match_dup and using register
>> tying with constraints consistently.
>>
>> I'm not entirely convinced that this is legitimate (my vague recollection is
>> that register tying should only be used to tie inputs to outputs).
>>
>> This has passed testing on a bunch of ARM targets, and fixes the build
>> issues I've been seeing.
> 
> Looking at this too, as an LRA exercise. I don't really think the
> pattern is wrong, rather LRA should just avoid creating the copy in this
> case; it's a result of operand constraining, after all.
> 
> Attached is the small LRA patch, pending testing. Vladimir should weight
> in on this.
> 

The patch is safe and ok.  Thanks for working on it, Chung-Lin.

>       * ira-lives.c (process_bb_lives): Skip creating copy during
>       insn sca when src/dest has constrained to same regno.
> 

Reply via email to