[range-op] Restrict division by power of 2 optimization to positive numbers.
authorAldy Hernandez <aldyh@redhat.com>
Mon, 7 Nov 2022 07:40:12 +0000 (08:40 +0100)
committerAldy Hernandez <aldyh@redhat.com>
Mon, 7 Nov 2022 11:41:41 +0000 (12:41 +0100)
commit3bff15c1c9fb3eb0bb042717e072476ec2d6d88c
tree1add755d5da016e9f67cb16be214c9cdad47cc07
parent071d00e0faabbd45449d2e83f207fca0f8e8ef68
[range-op] Restrict division by power of 2 optimization to positive numbers.

The problem here is that we are transforming a division by a power of
2 into a right shift, and using this to shift the maybe nonzero bits.
This gives the wrong result when the number being divided is negative.

In the testcase we are dividing this by 8:

[irange] int [-256, -255] NONZERO 0xffffff01

and coming up with:

[irange] int [-32, -31] NONZERO 0xffffffe0

The maybe nonzero bits are wrong as -31 has the LSB set (0xffffffe1)
whereas the bitmask says the lower 4 bits are off.

PR tree-optimization/107541

gcc/ChangeLog:

* range-op.cc (operator_div::fold_range): Restrict power of 2
optimization to positive numbers.

gcc/testsuite/ChangeLog:

* gcc.dg/tree-ssa/pr107541.c: New test.
gcc/range-op.cc
gcc/testsuite/gcc.dg/tree-ssa/pr107541.c [new file with mode: 0644]