Failure to optimise (a/b) and (a%b) into single __aeabi_idivmod call

Hi,

I am looking at best approach for http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43721 - Failure to optimise (a/b) and (a%b) into single __aeabi_idivmod call in ARM architecture

In sumary, the following c code results in __aeabi_idivmod() call and one __aeabi_idiv() call even though the former already calculates the quotient. int q = a / b; int r = a % b; return q + r;

My question is what would be the best way to handle it. As I see there are few options with some issues.

1. Handling in gimple level, try to reduce the operations to equivalent of this. We should do this for the targets without integer divide. {q, r} = a % b; Gimple assign stmts have only one lhs operation (?). Therefore, lhs has to be made 64bit to signify return values of R0 and R1 returned together. I am not too sure of any implications on other architectures here.

2. Handling in expand_divmod. Here, when we see a div or mod operation, we will have to do a linear search to see if there is a valid equivalent operation to combine. If we find one, we can generate __aeabi_idivmod() and cache the result for the equivalent operation. As I see, this can get messy and might not be acceptable.

3. An RTL pass to process and combine these library calls. Possibly using cse. I am still looking at this.

4. Ramana tried a prototype to do the same using target pattens. He has ruled this out. (if you want more info, please refer to at https://code.launchpad.net/~ramana/gcc-linaro/divmodsi4-experiments)

Any suggestion for best way to handle this?

Thanks, Kugan