gcc 2.95.3/ARM problems comparing with negative
Toralf Lund
toralf@procaptura.com
Mon Nov 1 11:47:00 GMT 2004
Richard Earnshaw wrote:
>On Mon, 2004-11-01 at 09:52, Toralf Lund wrote:
>
>
>>I'm having a rather strange problem with an ARM code built with gcc
>>2.95.3. Maybe I'm missing something obvious, but it looks like the wrong
>>branch is taken after a comparison with a negative constant.
>>
>>I have the following test
>>
>> if(position<-ZSTP_MAX_POS)
>>
>>
>>where ZSTP_MAX_POS is #defined as 250000, and "position" is an argument
>>to the function containing the test. The problem is simply that if I
>>pass e.g. 1 for the position argument, the test evaluates to "true" -
>>according to the debugger, anyway - but surely 1 is not smaller than
>>-250000? Also, the following code is generated for the test:
>>
>> 8013458: e51b3014 ldr r3, [fp, -#20]
>> 801345c: e3e02a3d mvn r2, #249856 ; 0x3d000
>> 8013460: e2422090 sub r2, r2, #144 ; 0x90
>> 8013464: e1530002 cmp r3, r2
>> 8013468: 8a000017 bhi 80134cc <_SetZoomPos+0xc4>
>>
>>This is followed by the "test true" code, i.e. as far as I can tell, the
>>branch is supposed to be taken when the test is not true. But will it? I
>>would expect "bgt" instead of "bhi" for the branch instruction, I think.
>>
>>All this with gcc 2.95.3. Maybe I could upgrade, but it seems to me that
>>newer versions will actually generate less efficient code...
>>
>>
>
>bhi is a branch used for an unsigned comparison.
>
My point exactly...
> So I suspect your
>problem is the type of one of your arguments, most likely the
>declaration of position itself.
>
>
No, position is all right. But now that you mention it, ZSTP_MAX_POS is
actually set up via an indirect definition involving "sizeof", following
a recent update. I guess that makes ZSTP_MAX_POS an unsigned, and if I
use an "int" typecast, I do get the correct code. I still don't quite
get it, though. Shouldn't -ZSTP_MAX_POS implicitly be treated as a
signed in any case? I would at least expect either fully signed or fully
unsigned operation; the actual code here seems to be a mixture of bot.
- Toralf
------
Want more information? See the CrossGCC FAQ, http://www.objsw.com/CrossGCC/
Want to unsubscribe? Send a note to crossgcc-unsubscribe@sources.redhat.com
More information about the crossgcc
mailing list