[Bug target/58621] New: With -fsection-anchors, a superfluous 'add' is performed

["b.grayson at samsung dot com" <gcc-bugzilla at gcc dot gnu dot org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58621

            Bug ID: 58621
           Summary: With -fsection-anchors, a superfluous 'add' is
                    performed
           Product: gcc
           Version: 4.9.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: b.grayson at samsung dot com
            Target: AArch64

In some scenarios, the code emitted with -fsection-anchors is suboptimal.  Here
is the generated assembly for some code, compiled with -O3 -fsection-anchors
-fno-common:

int a;
char b;

int foo() {
    return a+b;
}


0000000000000000 <foo>:
   0:   90000000        adrp    x0, 0 <foo>
   4:   91000002        add     x2, x0, #0x0
   8:   39400001        ldrb    w1, [x0]
   c:   b9400440        ldr     w0, [x2,#4]
  10:   0b000020        add     w0, w1, w0
  14:   d65f03c0        ret

Here is the source assembly:

foo:
        adrp    x0, .LANCHOR0
        add     x2, x0, :lo12:.LANCHOR0
        ldrb    w1, [x0,#:lo12:.LANCHOR0]
        ldr     w0, [x2,4]
        add     w0, w1, w0
        ret

Note the add.  It is computing the address of variable b, and using that as the
section anchor, so effectively the ldr is using the address x0 + #0 + #4.

If the ldr instead used #:lo12:.LANCHOR0 + 4, it would eliminate the extra add
instruction:

foo:
        adrp    x0, .LANCHOR0
        ldrb    w1, [x0,#:lo12:.LANCHOR0]
        ldr     w0, [x2,#:lo12:.LANCHOR0+4]
        add     w0, w1, w0
        ret

Of course, the offset must be small enough to not cross a page boundary from
.LANCHOR0's page address, but that restriction is already in there that the
offset from address of the chosen section anchor to address of 'a' must be less
than a page.

(Yes, in this case the add looks amazingly redundant since it is adding 0, but
I've seen other cases where the ANCHOR is not at offset 0, in which case we
have an add plus an offsetted load, where it could be done with just an
offsetted load.)