[Bug tree-optimization/128130] New: [17 Regression] arm Cortex-M: adjacent word loads are no longer paired into ldrd at -O2

[yuki.tanaka at defaultcase dot de via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=128130

            Bug ID: 128130
           Summary: [17 Regression] arm Cortex-M: adjacent word loads are
                    no longer paired into ldrd at -O2
           Product: gcc
           Version: 17.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: yuki.tanaka at defaultcase dot de
  Target Milestone: ---

int a(int *p)
{
  return p[0] + p[1] + p[2] + p[3];
}

Compiler Explorer (GCC 16.1 vs. 17):

With -O2 -mcpu=cortex-m33 -mfloat-abi=soft.
ARM GCC 16.1:
a:
        ldrd    r3, r1, [r0]
        add     r3, r3, r1
        ldrd    r2, r0, [r0, #8]
        add     r3, r3, r2
        add     r0, r0, r3
        bx      lr

ARM GCC 17:
a:
        ldr     r1, [r0, #4]
        ldr     r2, [r0]
        ldrd    ip, r3, [r0, #8]
        add     r0, r2, ip
        add     r3, r3, r1
        add     r0, r0, r3
        bx      lr

GCC 17 adds p[0] + p[2] and p[1] + p[3] instead of p[0] + p[1] and p[2] + p[3],
so the loads are no longer paired. -fno-tree-slp-vectorize gives the GCC 16.1
code.

Possibly related to r17-2654-gfb0838bc107 ("Support two-lane vector BB
reductions without target support").