[rfc][ARM] Support for safepoint check based on memory protect rather than polling

[Andrew Dinn]

Oops, I missed a comment

On 14/05/12 18:01, Andrew Haley wrote:
>>  
>>  // If this is a backward branch, compile a safepoint check
>> @@ -4525,7 +4660,11 @@
>>    unsigned dest_taken = bci + offset;
>>  
>>    if (jinfo->bc_stackinfo[dest_taken] & BC_COMPILED) {
> 
> I think this is really the wrong test.  We should just be checking
> dest < bci.

Hmm, I guess so. Although this convention of checking the bc_stackinfo
flags is used throughout the compiler. I wonder if it the recursive
calls to Thumb2_codegen() don't invalidate your assumption -- I will check.

>> -    Thumb2_Safepoint(jinfo, stackdepth, bci);
>> +    // pass offset as positive so the safepoint code plant a forward
>> +    // skip over the test rather than doing an unconditional backwards
>> +    // branch. that allows the condition test to be planted by
>> +    // whatever followed this call
>> +    Thumb2_Safepoint(jinfo, stackdepth, bci, -dest_taken);
>>    }
>>  }

This one slipped through. The last argument should be -offset not
-dest_taken. Given the right program the original posted code causes an
error, as I found this very quickly when I tried to run javac. Having
patched it to pass -offset things now work as expected.

regards,


Andrew Dinn
-----------