On 2013/11/21 03:25 PM, Richard Henderson wrote: > On 11/21/2013 02:41 PM, Chung-Lin Tang wrote: >> I'm not saying we tolerate "wrong" RTL form, but rather that, it should >> be an issue of the RTL passes, not the backend. The backend should just >> be as much as possible, a "machine description". > > Matching non-canonical rtl does nothing but slow down the compiler and cause > confusion for the poor person that has to maintain the code later.
As I mentioned in the last mail, I will modify the constraints as suggested for the *norsi3 and EQ/NE cmp patterns in the final patch. Chung-Lin