[Bug rtl-optimization/92712] [8/9/10 Regression] Performance regression with assumed values

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92712

--- Comment #13 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Ok, so match.pd does the following transformations:
A*B+A -> (B+1)*A
A*B-A -> (B-1)*A
A-A*B -> (1-B)*A
A*B+A*C -> (B+C)*A
A*B-A*C -> (B-C)*A
(ignoring equivalent expressions with commutative + or *).
Let's consider for start just the simpler 3 cases.
For those, I believe the only cases where these introduce UB when it wasn't
before in two's complement world are:
A*B+A -> (B+1)*A with A==-1 && B==max
A*B+A -> (B+1)*A with A==0 && B==max
A*B-A -> (B-1)*A with A==0 && B==min
A-A*B -> (1-B)*A with A==-1 && B==-max
A-A*B -> (1-B)*A with A==0 && B==min
A-A*B -> (1-B)*A with A==0 && B==-max
Now, the current match.pd patterns disable this optimization if A might be 0 or
-1.
So, what we could do is: 1) for the A*B-A case, drop the must not be -1 check
2) add the checks for the ranges of the other operand, say for the first one it
is fine if we don't know anything about A (@0) if we know the B (@1) is not
maximum.
For the last 2, the question is if we can come up with something similar.