On 8/12/2026 10:52 AM, Jeffrey Law wrote:
On 8/10/2026 10:32 AM, Daniel Barboza wrote:
This new forwprop step is my attempt to implement Richi's suggestions from
v1 of this work [1] where he suggested to push things out of match.pd.
The idea is to simplify DIV/MOD into RSHIFT/BIT_AND ops in which the
divisor are pow2 integers in a PHI. E.g.:
phi_var = PHI <16,4>
_x = _y % phi_var
Can be turned into:
phi_var = PHI <15,3>
_x = _y & phi_var
As long as we know that _y is a positive number or '_x' is used just in
zero equality comparisons.
Most of 101179 use cases are solved by this change. PHI with 2+ args
are supported as long as every phi_arg meets the criteria.
Boostrapped and regression tested with x86_64, aarch64 and riscv64.
[1] https://gcc.gnu.org/pipermail/gcc-patches/2026-May/716303.html
PR tree-optimization/101179
gcc/ChangeLog:
* tree-ssa-forwprop.cc (simplify_phi_result_movdiv): New
forwprop step where MOD/DIV ops with pow2 divisors can be
simplified to BIT_AND/RSHIFT.
(pass_forwprop::execute): Call simplify_phi_result_movdiv.
gcc/testsuite/ChangeLog:
* gcc.dg/tree-ssa/pr101179.c: New test.
Can we really optimize the signed case when the result is used in a comparison
against zero? Consider something like:
y / (x ? 4 : 16) == 0
Where "y" is a small negative number, say -2. In the original form that
expression would be true. In the right shifted form it would be false.
Did I miss something here?
Nope, you are correct. For MOD we can do the bit_and transformation if the
dividend is positive or the MOD result is just used to compare with zero,
but for DIV this is only valid if the dividend is positive. I remember having
code for this check in earlier versions but didn't find them in the ML ...
I'll send a new version with the DIV restriction in place. Thanks,
Daniel
Jeff
