[PATCH] Treat tail calls as scheduling barriers (PR rtl-optimization/48721)

[Vladimir Makarov]

On 11/30/2011 12:54 PM, Jakub Jelinek wrote:
> Hi!
>
> As discussed in the PR, I think tail calls are best handled as full
> scheduling barriers, scheduling tail calls before other insns effectively
> means we remove those other insns, but isn't dead insn removal the job
> of other passes instead?
> In any case, tail calls need to come up after all restoration of
> call saved register, on this testcase GCC wanted to schedule a tail
> call before r12 restoring insn.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
>
> 2011-11-30  Jakub Jelinek<jakub@redhat.com>
>
> 	PR rtl-optimization/48721
> 	* sched-deps.c (sched_analyze_insn): For SIBLING_CALL_P set
> 	reg_pending_barrier to TRUE_BARRIER.
>
> 	* gcc.target/i386/pr48721.c: New test.
>
> --- gcc/sched-deps.c.jj	2011-11-28 17:58:04.000000000 +0100
> +++ gcc/sched-deps.c	2011-11-30 14:10:55.117267854 +0100
> @@ -2873,7 +2873,11 @@ sched_analyze_insn (struct deps_desc *de
>   	  else
>   	    sched_analyze_2 (deps, XEXP (link, 0), insn);
>   	}
> -      if (find_reg_note (insn, REG_SETJMP, NULL))
> +      /* Don't schedule anything after a tail call, tail call needs
> +	 to use at least all call-saved registers.  */
> +      if (SIBLING_CALL_P (insn))
> +	reg_pending_barrier = TRUE_BARRIER;
> +      else if (find_reg_note (insn, REG_SETJMP, NULL))
>   	reg_pending_barrier = MOVE_BARRIER;
>       }
>
> --- gcc/testsuite/gcc.target/i386/pr48721.c.jj	2011-11-30 14:14:26.290628842 +0100
> +++ gcc/testsuite/gcc.target/i386/pr48721.c	2011-11-30 12:24:06.000000000 +0100
> @@ -0,0 +1,51 @@
> +/* PR rtl-optimization/48721 */
> +/* { dg-do compile } */
> +/* { dg-options "-O -foptimize-sibling-calls -fsched2-use-superblocks -fschedule-insns2 -mtune=core2" } */
> +
> +extern unsigned char a[];
> +extern int b[], d[], e[], f[], g[], *h[], m[], *n[], o[];
> +extern char c[];
> +
> +struct S
> +{
> +  unsigned char s1;
> +  int s2, s3, s4, s5, s6, s7, s8;
> +};
> +
> +__attribute__((noinline, noclone)) int
> +foo (int x)
> +{
> +  return 0;
> +}
> +
> +int
> +bar (int x, struct S *y)
> +{
> +  int z;
> +  switch (x)
> +    {
> +    case 1:
> +    case 2:
> +      {
> +	int t2, t4, t5, t6, t7, t8;
> +	z = o[y->s8 * 6];
> +	t8 = *n[m[x] * 5];
> +	t4 = *h[y->s7];
> +	t7 = z;
> +	z = g[f[x] + y->s6];
> +	t6 = e[y->s5];
> +	t5 = d[c[x] + y->s3 * 17];
> +	if (z)
> +	  t2 = b[z];
> +	if (a[z] != y->s1)
> +	  return foo (x);
> +	y->s8 = t8;
> +	y->s4 = t4;
> +	y->s7 = t7;
> +	y->s6 = t6;
> +	y->s5 = t5;
> +	y->s2 = t2;
> +      }
> +    }
> +  return 0;
> +}
>
Ok, thanks.