Missed optimization: Reassociate multiplies by constants to shrink said constants

[Alcaro]

int a(int b, int c, int d, int e)
{
    return b*1000+c*100+d*10+e;
}
int a2(int b, int c, int d, int e)
{
    return ((b*10+c)*10+d)*10+e;
}

Expected: Same for both.

Actual: a2 is a series of lea, as expected, but former contains two imul.

GCC also misses this optimization, but MSVC performs it. https://godbolt.org/z/1fs9xhsdb

The missed optimization also shows up on ARM and RISC-V, where a loads three different constants, but a2 reuses the same one. https://godbolt.org/z/qzMEv1b8j https://godbolt.org/z/bTYzY4vGa (and GCC optimizes mul by 10 to some shifts and adds, but that's a separate missed optimization)

[pinskia]

I am not sure the ((b*10+c)*10+d)*10+e is better than b*1000+c*100+d*10+e in the case where you have a 2 or more issue pipeline. Even with the add+shift.

In the first case, it is linearized while in the 2nd case you can do with 3 adds.
linearization is a lot of the time worse.