The following two functions should be equal: unsigned f(int i, unsigned x) { unsigned y; if (i) y = 1024; else y = 1024*1024; return x/y ; } unsigned f1(int i, unsigned x) { unsigned y; if (i) y = x/ 1024; else y = x/(1024*1024); return y ; }
Confirmed.
*** Bug 30747 has been marked as a duplicate of this bug. ***
This actually undos some of what phi-opt might do and only should be done for a few operations. I think divide, multiple, shifts and rotates should be done and only with a constant operand. This has to be done late very close to expand even. I will be implementing this but might miss GCC 12 as I have a lot on my plate already for GCC 12.
(In reply to Andrew Pinski from comment #3) > This actually undos some of what phi-opt might do and only should be done > for a few operations. I think divide, multiple, shifts and rotates should > be done and only with a constant operand. This has to be done late very > close to expand even. > > I will be implementing this but might miss GCC 12 as I have a lot on my > plate already for GCC 12. It turns out this is still true. I had implemented the binary factoring out of operations for phiopt and end up with a failure with `gcc.dg/tree-prof/val-prof-1.c` which depends on phiopt not doing this. Note only one of the phi nodes needs to be constant for when we want to handle it. Oh I wonder if val-prof-1.c is still going to fail since we had originally: if (a == CST) b = CST/c; else b = a/c; And now we might have removed the if .. Woops. Anyways I will have to go and dig into that failure some more but this issue is now needed again too.
*** Bug 112608 has been marked as a duplicate of this bug. ***