[Bug middle-end/70424] New: [4.9/5/6 Regression] Pointer derived from integer gets reduced alignment

["amonakov at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70424

            Bug ID: 70424
           Summary: [4.9/5/6 Regression] Pointer derived from integer gets
                    reduced alignment
           Product: gcc
           Version: 4.9.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: middle-end
          Assignee: unassigned at gcc dot gnu.org
          Reporter: amonakov at gcc dot gnu.org
  Target Milestone: ---

int f(long a)
{
  int *p=(int*)(a<<1);
  //asm("" : "+r"(p));
  return *p;
}

Starting from 4.9, in the above example GCC assumes that *p is aligned to 16
bits (on 4.8 and earlier, to 32 bits, like normal int*). This causes the load
to be torn in two on strict-alignment targets; using -O0 or uncommenting the
asm restores old behavior (one 32-bit load). This change seems unintended.

On x86_64 it's visible on RTL level (note A32->A16 change):

gcc-4.8.0 -S t.c -Os -o- -dP

#(insn:TI 7 3 15 2 (set (reg:SI 0 ax [orig:66 *p_3 ] [66])
#        (mem:SI (plus:DI (reg/v:DI 5 di [orig:63 a ] [63])
#                (reg/v:DI 5 di [orig:63 a ] [63])) [2 *p_3+0 S4 A32]))
align.c:5 89 {*movsi_internal}
#     (expr_list:REG_DEAD (reg/v:DI 5 di [orig:63 a ] [63])
#        (nil)))
        movl    (%rdi,%rdi), %eax       # 7     *movsi_internal/1       [length
= 3]

gcc-4.9.2 -S t.c -Os -o- -dP

#(insn:TI 7 3 13 2 (set (reg:SI 0 ax [orig:90 *p_3 ] [90])
#        (mem:SI (plus:DI (reg/v:DI 5 di [orig:87 a ] [87])
#                (reg/v:DI 5 di [orig:87 a ] [87])) [2 *p_3+0 S4 A16]))
align.c:5 90 {*movsi_internal}
#     (expr_list:REG_DEAD (reg/v:DI 5 di [orig:87 a ] [87])
#        (nil)))
        movl    (%rdi,%rdi), %eax       # 7     *movsi_internal/1       [length
= 3]