[Bug rtl-optimization/35775] collapsing popping args for tail calls at -Os

[tkoenig at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=35775

--- Comment #3 from Thomas Koenig <tkoenig at gcc dot gnu.org> ---
We now get

$ cat tailcall.c 
void bar(void);
void baz(void);

void foo(int a)
{
  if (a)
    bar();
  else
    baz();
}
$ gcc -S -Os tailcall.c 
$ cat tailcall.s 
        .file   "tailcall.c"
        .text
        .globl  foo
        .type   foo, @function
foo:
.LFB0:
        .cfi_startproc
        testl   %edi, %edi
        je      .L2
        jmp     bar
.L2:
        jmp     baz
        .cfi_endproc
.LFE0:
        .size   foo, .-foo
        .ident  "GCC: (GNU) 9.0.1 20190414 (experimental)"
        .section        .note.GNU-stack,"",@progbits

so the original test case is fixed.

With a slightly more complex test case, we now get

$ cat tailcall2.s 
        .file   "tailcall2.c"
        .text
        .globl  foo
        .type   foo, @function
foo:
.LFB0:
        .cfi_startproc
        pushq   %rcx
        .cfi_def_cfa_offset 16
        call    gargle
        testl   %eax, %eax
        je      .L2
        popq    %rdx
        .cfi_remember_state
        .cfi_def_cfa_offset 8
        jmp     bar
.L2:
        .cfi_restore_state
        popq    %rax
        .cfi_def_cfa_offset 8
        jmp     baz
        .cfi_endproc
.LFE0:
        .size   foo, .-foo
        .ident  "GCC: (GNU) 9.0.1 20190414 (experimental)"
        .section        .note.GNU-stack,"",@progbits

so the problem is still present in this case (using two different
registers seems strange, but it should not matter).