[Bug target/68651] New: [5/6 Regression][AArch64] Missing combination of shift-by-one with other arithmetic patterns with -mcpu=cortex-a53

[ktkachov at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68651

            Bug ID: 68651
           Summary: [5/6 Regression][AArch64] Missing combination of
                    shift-by-one with other arithmetic patterns with
                    -mcpu=cortex-a53
           Product: gcc
           Version: 6.0
            Status: UNCONFIRMED
          Keywords: missed-optimization
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: ktkachov at gcc dot gnu.org
  Target Milestone: ---
            Target: aarch64*

Consider the testcases:

int
foo (int x)
{
  return (x * 2) & 65535;
}

int
bar (int x, int y)
{
  return (x * 2) | y;
}

With GCC 4.9 for aarch64 with -O2 -mcpu=cortex-a53 we generated:
foo:
        ubfiz   w0, w0, 1, 15
        ret

bar:
        orr     w0, w1, w0, lsl 1
        ret

whereas with GCC 5 and trunk we fail to combine:
foo:
        add     w0, w0, w0
        uxth    w0, w0
        ret

bar:
        add     w0, w0, w0
        orr     w0, w0, w1
        ret

This is because of updated costs information for cortex-a53 which leads to
x * 2 being expanded as x + x rather than x << 1 and combine can't merge the
PLUS form.

This can be fixed simply by mutating the pattern during recog_for_combine in a
similar way to change_zero_ext