ARM: More intrinsics

[pcpa at mandriva.com.br]

> On 03/15/2012 04:09 AM, pcpa at mandriva.com.br wrote:
>>> Xerxes noticed an inconsistency in my patch.  I've removed the unused
>>> masks, and redefined the condition in the atomic operations to use ETT
>>> (else, true, true) rather than the confused encoding I used before.
>>>
>>> So, it's now
>>>
>>> +      it(codebuf, COND_EQ, IT_MASK_ETT);
>>>
>>> instead of
>>>
>>> -      it(codebuf, COND_NE, IT_MASK_TEE);
>>
>>   I feel there is something bogus here, because the first condition
>> should be 'T' always, and the mask is based on bit 0 of the test
>> condition. Talking this because I wrote support for arm thumb jit
>> recently.
>
> Well, that's how the ARM asm describes it, but that's not what
> the encodings really do.  It's a shame they chose to describe it
> in that way.
>
>>> I had to do this because it's the only way the encodings make any
>>> sense.  There isn't really any encoding for TEE that works with
>>> COND_NE: it has to be expressed in the binary as COND_EQ:ETT.
>>> (If cond is true, don't do X, but do Y and Z.)
>>
>>   For easier viewing instead of pointing to some link, the code I use
>> to encode it* is:
>>
>> static void
>> _tcit(jit_state_t *_jit, unsigned int tc, int it)
>> {
>>     int		c;
>>     int		m;
>>     c = (tc >> 28) & 1;
>>     assert(!(tc & 0xfffffff) && tc != ARM_CC_NV);
>>     switch (it) {
>> 	case THUMB2_IT:		m =   1<<3; 			break;
>> 	case THUMB2_ITT:	m =  (c<<3)| (1<<2);		break;
>> 	case THUMB2_ITE:	m = (!c<<3)| (1<<2);		break;
>> 	case THUMB2_ITTT:	m =  (c<<3)| (c<<2)| (1<<1);	break;
>> 	case THUMB2_ITET:	m = (!c<<3)| (c<<2)| (1<<1);	break;
>> 	case THUMB2_ITTE:	m =  (c<<3)|(!c<<2)| (1<<1);	break;
>> 	case THUMB2_ITEE:	m = (!c<<3)|(!c<<2)| (1<<1);	break;
>> 	case THUMB2_ITTTT:	m =  (c<<3)| (c<<2)| (c<<1)|1;	break;
>> 	case THUMB2_ITETT:	m = (!c<<3)| (c<<2)| (c<<1)|1;	break;
>> 	case THUMB2_ITTET:	m =  (c<<3)|(!c<<2)| (c<<1)|1;	break;
>> 	case THUMB2_ITEET:	m = (!c<<3)|(!c<<2)| (c<<1)|1;	break;
>> 	case THUMB2_ITTTE:	m =  (c<<3)| (c<<2)|(!c<<1)|1;	break;
>> 	case THUMB2_ITETE:	m = (!c<<3)| (c<<2)|(!c<<1)|1;	break;
>> 	case THUMB2_ITTEE:	m =  (c<<3)|(!c<<2)|(!c<<1)|1;	break;
>> 	case THUMB2_ITEEE:	m = (!c<<3)|(!c<<2)|(!c<<1)|1;	break;
>> 	default:		abort();
>>     }
>>     assert(m && (tc != ARM_CC_AL || !(m & (m - 1))));
>>     is(0xbf00 | (tc >> 24) | m);
>> }
>
> LOL!  That's just the sort of thing I was trying to avoid!  :-)

  :-)

  Replying out of order.

> I'd love to use this, but I think I probably can't.  I could use this
> idea for just the couple of conditions I actually use, though, without
> requiring copyright assignment.

  I was just offering the idea for free, as I was under the impression
there was some confusion. The sample code is just verbatim "unrolled"
implementation of the description in the spec I used, and assertions
to ensure it does not generate undefined conditions.

> Thanks,
> Andrew.

Paulo