[Bug middle-end/88085] New: User alignments on var decls not respected if smaller than type alignment

[krebbel at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88085

            Bug ID: 88085
           Summary: User alignments on var decls not respected if smaller
                    than type alignment
           Product: gcc
           Version: 9.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: middle-end
          Assignee: unassigned at gcc dot gnu.org
          Reporter: krebbel at gcc dot gnu.org
  Target Milestone: ---

typedef int __attribute__((vector_size(16))) v4si; 

v4si a4 __attribute__((aligned(4))); 

void 
foo (v4si a) 
{ 
 a4 += a; 
}

-O3 -dP

MEM_ALIGN for variable a4 is 128 bits although it is emitted with an alignment
requirement of just 4: .comm a4,16,4

It works when moving the aligned attribute over to the typedef instead.

foo:
#(insn:TI 6 3 7 2 (set (reg:V4SI 20 xmm0 [85])
#        (plus:V4SI (reg:V4SI 20 xmm0 [86])
#            (mem/c:V4SI (symbol_ref:DI ("a4") [flags 0x2] <var_decl
0x7ff6d222eab0 a4>) [1 a4+0 S16 A128]))) "t.c":8:6 3122 {*addv4si3}
#     (expr_list:REG_EQUIV (mem/c:V4SI (symbol_ref:DI ("a4") [flags 0x2]
<var_decl 0x7ff6d222eab0 a4>) [1 a4+0 S16 A128])
#        (nil)))
        paddd   a4(%rip), %xmm0 # 6     [c=12 l=8]  *addv4si3/0
#(insn:TI 7 6 18 2 (set (mem/c:V4SI (symbol_ref:DI ("a4") [flags 0x2] <var_decl
0x7ff6d222eab0 a4>) [1 a4+0 S16 A128])
#        (reg:V4SI 20 xmm0 [85])) "t.c":8:6 1198 {movv4si_internal}
#     (expr_list:REG_DEAD (reg:V4SI 20 xmm0 [85])
#        (nil)))
        movaps  %xmm0, a4(%rip) # 7     [c=4 l=7]  movv4si_internal/3
#(jump_insn:TI 14 18 15 2 (simple_return) "t.c":9:1 688
{simple_return_internal}
#     (nil)
# -> simple_return)
        ret             # 14    [c=0 l=1]  simple_return_internal
        .size   foo, .-foo
        .comm   a4,16,4
        .ident  "GCC: (GNU) 9.0.0 20181114 (experimental)"



set_mem_attributes_minus_bitpos in emit-rtl.c:

This code sets the alignment field based on the type alignment:

      /* ??? If T is a type, respecting mode alignment may *also* be wrong
         e.g. if the type carries an alignment attribute.  Should we be
         able to simply always use TYPE_ALIGN?  */
    }

  /* We can set the alignment from the type if we are making an object or if
     this is an INDIRECT_REF.  */
  if (objectp || TREE_CODE (t) == INDIRECT_REF)
    attrs.align = MAX (attrs.align, TYPE_ALIGN (type));


This code later would use the alignment in the aligned attribute (in obj_align)
but ignores it if it is lower than the alignment we already have?! As expected
setting bigger alignments on the var decl works.

      /* If this is an indirect reference, record it.  */
      else if (TREE_CODE (t) == MEM_REF 
               || TREE_CODE (t) == TARGET_MEM_REF)
        {
          attrs.expr = t;
          attrs.offset_known_p = true;
          attrs.offset = 0;
          apply_bitpos = bitpos;
        }

      /* Compute the alignment.  */
      unsigned int obj_align;
      unsigned HOST_WIDE_INT obj_bitpos;
      get_object_alignment_1 (t, &obj_align, &obj_bitpos);
      unsigned int diff_align = known_alignment (obj_bitpos - bitpos);
      if (diff_align != 0)
        obj_align = MIN (obj_align, diff_align);
      attrs.align = MAX (attrs.align, obj_align);
    }