[Bug c/57356] New: gcc-4.8: SSE2 instructions generated with '-mno-sse2'

["thutt at vmware dot com" <gcc-bugzilla at gcc dot gnu dot org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57356

            Bug ID: 57356
           Summary: gcc-4.8: SSE2 instructions generated with '-mno-sse2'
           Product: gcc
           Version: unknown
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
          Assignee: unassigned at gcc dot gnu.org
          Reporter: thutt at vmware dot com

The following example shows a defect in gcc 4.8 when using the
'-mno-sse2' command line option: SSE2 instructions are still
generated.

Compile with:

  gcc-4.8 -m64 -O1 -mno-sse2 -c -o /tmp/gungla.o /tmp/gungla.c

Sample Code:

  typedef struct s128 {
      char x[16];
  } s128;

  typedef struct wrapper {
      s128 elem;
  } wrapper;

  void test2(s128 *);

  void
  test(wrapper **p, int *num)
  {
      int   i;
      s128 *a = &(*p)->elem;
      s128  b = *a;

      for (i = 1; i < *num; i++) {
         test2(a);
         if (i == 10) {
            *a = b;
            break;
         }
      }
      if (p) {
         *p = 0;
      }
  }

We have not been able to simplify this code any further.

  Disassembly of section .text:

  0000000000000000 <test>:
     0:    41 56                    push   %r14
     2:    41 55                    push   %r13
     4:    41 54                    push   %r12
     6:    55                       push   %rbp
     7:    53                       push   %rbx
     8:    48 83 ec 10              sub    $0x10,%rsp
     c:    49 89 fe                 mov    %rdi,%r14
     f:    48 89 f5                 mov    %rsi,%rbp
    12:    4c 8b 2f                 mov    (%rdi),%r13
--> 15:    f3 41 0f 6f 45 00        movdqu 0x0(%r13),%xmm0
--> 1b:    66 0f 7f 04 24           movdqa %xmm0,(%rsp)
    20:    83 3e 01                 cmpl   $0x1,(%rsi)
    23:    7f 32                    jg     57 <test+0x57>
    25:    eb 22                    jmp    49 <test+0x49>
    27:    4c 89 e7                 mov    %r12,%rdi
    2a:    e8 00 00 00 00           callq  2f <test+0x2f>
              2b: R_X86_64_PC32    test2-0x4
    2f:    83 fb 0a                 cmp    $0xa,%ebx
    32:    75 0d                    jne    41 <test+0x41>
    34:    66 0f 6f 0c 24           movdqa (%rsp),%xmm1
    39:    f3 41 0f 7f 4d 00        movdqu %xmm1,0x0(%r13)
    3f:    eb 08                    jmp    49 <test+0x49>
    41:    83 c3 01                 add    $0x1,%ebx
    44:    39 5d 00                 cmp    %ebx,0x0(%rbp)
    47:    7f de                    jg     27 <test+0x27>
    49:    4d 85 f6                 test   %r14,%r14
    4c:    74 1b                    je     69 <test+0x69>
    4e:    49 c7 06 00 00 00 00     movq   $0x0,(%r14)
    55:    eb 12                    jmp    69 <test+0x69>
    57:    4d 89 ec                 mov    %r13,%r12
    5a:    4c 89 ef                 mov    %r13,%rdi
    5d:    e8 00 00 00 00           callq  62 <test+0x62>
              5e: R_X86_64_PC32    test2-0x4
    62:    bb 01 00 00 00           mov    $0x1,%ebx
    67:    eb d8                    jmp    41 <test+0x41>
    69:    48 83 c4 10              add    $0x10,%rsp
    6d:    5b                       pop    %rbx
    6e:    5d                       pop    %rbp
    6f:    41 5c                    pop    %r12
    71:    41 5d                    pop    %r13
    73:    41 5e                    pop    %r14
    75:    c3                       retq

One oddity is that using '-mno-sse', the offending instructions are
not generated.  (Maybe this implies that the instructions are simply
misclassified as SSE, rather than SSE2?)

gcc-4.7 does not exhibit the problem with the same code.

Can you folks please confirm, or refute, if gcc-4.7 suffers from the
same defect?

Possibly related to bug 46716.