[Bug middle-end/53090] New: suboptimal ivopt

[xinliangli at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53090

             Bug #: 53090
           Summary: suboptimal ivopt
    Classification: Unclassified
           Product: gcc
           Version: 4.8.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: middle-end
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: xinliangli@gmail.com


Compiling the attached benchmark code with trunk gcc, the code generated for
the hot memory swap loop (line 60) is very inefficient : both icc and llvm use
two ivs and generate a tight loop with 9 instructions, but gcc decides to use 3
ivs, and the loop exit testing code is wierd and inefficient -- it ends up
produce a loop with 11 instructions.

#define XCH(x,y)    { Aint t_mp; t_mp=(x); (x)=(y); (y)=t_mp; }

    for( i=1, j=k-1 ; i<j ; ++i, --j ) {
        XCH(perm[i], perm[j])
    }

The tight version:


.LBB0_11:                               # %for.body57.i
                                        #   Parent Loop BB0_1 Depth=1
                                        #     Parent Loop BB0_9 Depth=2
                                        # =>    This Inner Loop Header: Depth=3
    movl    (%rbx,%rdi,4), %ebp
    movl    (%rbx,%rsi,4), %eax
    movl    %eax, (%rbx,%rdi,4)
    movl    %ebp, (%rbx,%rsi,4)
    decq    %rsi
    incq    %rdi
    cmpl    %edx, %edi
    leal    -1(%rdx), %edx
    jl    .LBB0_11


The gcc version:

.L18:
    movl    (%rdx), %edi
    addl    $1, %ecx
    movl    (%rsi), %eax
    movl    %eax, (%rdx)
    addq    $4, %rdx
    movl    %edi, (%rsi)
    movl    %r8d, %edi
    subq    $4, %rsi
    subl    %ecx, %edi
    cmpl    %edi, %ecx
    jl    .L18


However gcc is doing the right thing when applied on the extracted test case:

#define XCH(x,y) { int t_mp; t_mp=(x); (x)=(y); (y)=t_mp; }

void foo (int *perm, int k)
{
 int i,j;
 for( i=1, j=k-1 ; i<j ; ++i, --j ) {
    XCH(perm[i], perm[j])
 }
}