[Bug tree-optimization/104582] [11/12 Regression] Unoptimal code for __negdi2 (and others) from libgcc2 due to unwanted vectorization

["rguenth at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104582

--- Comment #9 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Jakub Jelinek from comment #8)
> Just trying a dumb microbenchmark:
> struct S { unsigned long a, b; } s;
> 
> __attribute__((noipa)) void
> foo (unsigned long a, unsigned long b)
> {
>   s.a = a;
>   s.b = b;
> }
> 
> int
> main ()
> {
>   int i;
>   for (i = 0; i < 1000000000; i++)
>     foo (42, 43);
>   return 0;
> }
> the GCC 11 vs. GCC 12 code:
> -	movq	%rdi, s(%rip)
> -	movq	%rsi, s+8(%rip)
> +	movq	%rdi, %xmm0
> +	movq	%rsi, %xmm1
> +	punpcklqdq	%xmm1, %xmm0
> +	movaps	%xmm0, s(%rip)
> seems to be exactly the same speed (on i9-7960X) and the GCC 11 code is 7
> bytes smaller.

The GCC 12 code is 30% slower on Zen 2 (the gpr -> xmm move is comparatively
more costly there).  As said we fail to account for that.  But as I said
the cost is not there if it's

struct S { unsigned long a, b; } s;

__attribute__((noipa)) void
foo (unsigned long *a, unsigned long *b)
{
  unsigned long a_ = *a;
  unsigned long b_ = *b;
  s.a = a_;
  s.b = b_;
}

which vectorizes to

        movq    (%rdi), %xmm0
        movhps  (%rsi), %xmm0
        movaps  %xmm0, s(%rip)
        ret

which is _smaller_ than the scalar code.  So it's important to be able
to distinguish those cases.  The above is also

a__3 1 times scalar_store costs 12 in body
b__5 1 times scalar_store costs 12 in body
a__3 1 times vector_store costs 12 in body
<unknown> 1 times vec_construct costs 8 in prologue