liveness analysis with stack slots

[Maurizio Vitale <mav at cuma dot polymath-solutions dot lan>]

Good day,
     let me first apologize if this is the wrong group, gcc.devel
might have been a better choice.

I'm writing a library for fixed-point arithmetic using expression
templates. I have properties of expressions that sometimes can be
computed completely at compile time. GCC builds the expression in
memory (on the stack). Values are correctly computed at compile time
and the code for building the expression is completely inlined
(somewhat to my surprise, given the complexity of the machinery behind
it).
What baffles me is why GCC (and the Intel compiler) doesn't recognize
that the expression is totally dead. Without going into to many
details (I can add info once people steer me in the direction they
need to go for understanding the issue), what I get from the following
fragment:

      typedef typename number_of_trait<target_traits>::type target_number;
      typedef typename number_of_trait<source_traits>::type source_number;
      MARK ("B meta");
      int shift = meta_eval (q_shift, target_number(target), source_number(source));
      MARK ("E meta");
      std:: cout << "shift would be " << shift << "\n";

[q_shift is an expression defined elsewhere as right(_2)-right(_1) where
_1 would be bound to target and _2 bound to source]

The resulting assembly is:

        #### B meta
#NO_APP
        movl    (%r13), %eax
        movq    %rsp, (%rsp)
        movl    %eax, 144(%rsp)
        movl    %eax, 128(%rsp)
        movl    %eax, 160(%rsp)
        movzwl  (%r15), %eax
        movw    %ax, 192(%rsp)
        movw    %ax, 176(%rsp)
        movw    %ax, 208(%rsp)
        leaq    208(%rsp), %rax
        movq    %rax, 8(%rsp)
        leaq    160(%rsp), %rax
        movq    %rax, 16(%rsp)
#APP
        #### E meta
#NO_APP
        movl    $.LC339, %esi
        movl    $_ZSt4cout, %edi
        call    _ZStlsISt11char_traitsIcEERSt13basic_ostreamIcT_ES5_PKc
        movl    $.LC340, %esi
        movl    $_ZSt4cout, %edi
        call    _ZStlsISt11char_traitsIcEERSt13basic_ostreamIcT_ES5_PKc
        movl    $1, %esi                                       #<== compile time result of q_shift = right(_2)-right(_1)
        movq    %rax, %rdi
        call    _ZNSolsEi

The entire portion between '#### B meta' and '#### E meta' is the
building of the expression. At least at the C++ source level
everything in there is dead.

Can somebody shed some light on this and hint at why GCC (and the
Intel compiler) decide that they cannot remove that code?

I remember a few years ago GCC had problems with liveness analysis w/
stack slots. Is that still the case?

Thanks a lot,

       Maurizio