[Bug rtl-optimization/42586] New: load-modify-store on x86 should be single instruction

["andi-gcc at firstfloor dot org" <gcc-bugzilla at gcc dot gnu dot org>]

Continuation from 42585

>From one of the examples of
http://embed.cs.utah.edu/embarrassing/dec_09/harvest/gcc-head_llvm-gcc-head/

struct _fat_ptr
{
  unsigned char *curr;
  unsigned char *base;
  unsigned char *last_plus_one;
};
int Cyc_string_ungetc (int ignore, struct _fat_ptr *sptr);
int
Cyc_string_ungetc (int ignore, struct _fat_ptr *sptr)
{
  struct _fat_ptr *_T0;
  struct _fat_ptr *_T1;
  struct _fat_ptr _T2;
  int _T3;
  struct _fat_ptr _ans;
  int _change;

  {
    _T0 = sptr;
    _T1 = sptr;
    _T2 = *sptr;
    _T3 = -1;
    _ans = _T2;
    _change = -1;
    _ans.curr += 4294967295U;
    *sptr = _ans;
    return (0);
  }
}

when compiled with -O2 -m32 on 4.5.0 20091219 generates

Cyc_string_ungetc:
        subl    $32, %esp
        movl    40(%esp), %eax
        movl    (%eax), %edx
        subl    $1, %edx
        movl    %edx, (%eax)
        xorl    %eax, %eax
        addl    $32, %esp
        ret

Apart from the useless stack frame manipulation it should use

       subl $1,(%eax) 

not load-modify-store as recommended in the Intel/AMD optimization
manuals for any modern CPU.

I experimented with different -mtune=s
(thinking it was maybe optimizing for Pentium5 where this made sense), but that
didn't help (apart from turning the subl into a decl). The code snippet
above is for -mtune=generic -m32


-- 
           Summary: load-modify-store on x86 should be single instruction
           Product: gcc
           Version: unknown
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: rtl-optimization
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: andi-gcc at firstfloor dot org
  GCC host triplet: x86_64-linux
GCC target triplet: x86_64-linux -m32


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42586