Re: [PATCH v2, rs6000] Fix ICE on expand bcd<bcd_add_sub>_<code>_<mode> [PR100736]

[HAO CHEN GUI <guihaoc@linux.ibm.com>]

Segher,
  Does BCD comparison return false when either operand is invalid coding?
If yes, the result could be 3-way. We can check gt and eq bits for ge.
We still can't use crnot to only check lt bit as there could be invalid
coding.
  Also, do you think finite-math-only excludes invalid coding? Seems GCC
doesn't clear define it.

Thanks.
Gui Haochen


On 2/6/2022 上午 6:05, Segher Boessenkool wrote:
> Hi!
> 
> On Tue, May 31, 2022 at 06:56:00PM -0500, Segher Boessenkool wrote:
>> It's not clear to me how this can ever happen without finite_math_only?
>> The patch is safe, sure, but it may the real problem is elsewhere.
> 
> So, it is incorrect the RTL for our bcd{add,sub} insns uses CCFP at all.
> 
> CCFP stands for the result of a 4-way comparison, regular float
> comparison: lt gt eq un.  But bcdadd does not have an unordered at all.
> Instead, it has the result of a 3-way comparison (lt gt eq), and bit 3
> is set if an overflow happened -- but still exactly one of bits 0..2 is
> set then!  (If one of the inputs is an invalid number it sets bits 0..3
> to 0001 though.)
> 
> So it would be much more correct and sensible to use regular integer
> comparison results here, so, CC.
> 
> Does that fix the problem?
> 
> 
> Segher