Re: [PATCH] expr.cc: avoid unexpected side effects in expand_expr_divmod optimization

["Jose E. Marchesi" <jose.marchesi@oracle.com>]

Hi Jakub.

> On Thu, Dec 08, 2022 at 11:59:44AM +0100, Jose E. Marchesi via Gcc-patches wrote:
>> gcc/ChangeLog
>> 
>> 	* expr.cc (expand_expr_divmod): Avoid side-effects of trying
>> 	sequences involving funcalls in optimization.
>
> That looks wrong.
> The globals for mentioned calls just shouldn't be emitted during expansion,
> especially if it is bigger annoyance than just having some extra symbols
> in the symbol table.
> expand_expr_divmod is definitely not the only place where something is
> expanded and later not used, lots of other places in the expander do that,
> and more importantly, there are over 80 optimization passes after expansion,
> many of them can remove code determined to be dead, and while lots of dead
> code is removed in GIMPLE optimizations already, definitely not all.
> So, rather than add hacks for this in a single spot, much better is to emit
> the globals only for stuff that is actually needed (so during final or
> immediately before it).

Yeah I see the point.

The culprit of the leadked .global seems to be a call to
assemble_external_libcall in emit_library_call_value_1:

expand_expr_divmod
  expand_divmod -> This will result in libcall
   sign_expand_divmod
     emit_library_call_value
       emit_library_call_value_1
         ...
         /* If this machine requires an external definition for library
            functions, write one out.  */
         assemble_external_libcall (fun);
         ...

The documented purpose of assemble_external_libcall is, as stated in
output.h, to "Assemble a string constant".

So, it seems to me that emit_library_call_value should not assemble
anything, since it is used by expand functions whose expansions may be
eventually discarded.

However, simply removing that call to assemble_external_libcall makes
.global declarations to not be emitted even when the funcall is actually
emitted in final:

For:

  int foo(unsigned int len)
  {
    return ((long)len) * 234 / 5;
  }

we get:

    .file   "foo.c"
    .text
    <------------- NO .global __divdi3
    .align  3
    .global foo
    .type   foo, @function
  foo:
    mov32   %r1,%r1
    mov     %r2,5
    mul     %r1,234
    call    __divdi3
    exit
  .size   foo, .-foo
  .ident  "GCC: (GNU) 13.0.0 20221207 (experimental)"

Note that BPF lacks signed division instructions.

So, I guess the right fix would be to call assemble_external_libcall
during final?  The `.global FOO' directive would be generated
immediately before the call sequence, but I guess that would be ok.

WDYT?