On Tue, Nov 8, 2016 at 1:03 PM, Kyrill Tkachov
<kyrylo.tkac...@foss.arm.com> wrote:
> Hi all,
>
> There is an off-by-one error in the clear_bit_region helper in store merging
> in the case where it deals with
> multi-byte quantities starting at a non-zero bit offset. The particular
> input is
> {0xff, 0xff, 0xff} and we want to clear all bits except the least and most
> significant i.e. we want:
> {0x01, 0x00, 0x80} so it's called as clear_bit_region (input, 1, 22);
> This ends up clearing one more bit due to this bug. The patch fixes that.
> The last argument to clear_bit_region is the number of bits left to clear
> and since in the previous call we cleared
> BITS_PER_UNIT - start bits we should subtract exactly that amount from len
> when calculating the bits left to clear.
> This was uncovered when writing initial unit tests for these functions which
> are included in the followup patch.
>
> Bootstrapped and tested on aarch64 and x86_64 (the affected function is only
> called for little-endian code).
>
> Ok for trunk?Ok. Richard. > Thanks, > Kyrill > > 2016-11-08 Kyrylo Tkachov <kyrylo.tkac...@arm.com> > > PR tree-optimization/78234 > * gimple-ssa-store-merging.c (clear_bit_region): Fix off-by-one error > in start != 0 case.
