ifcvt: Fix up noce_convert_multiple_sets [PR106590]

  Hi!

The following testcase is miscompiled on x86_64-linux.
The problem is in the noce_convert_multiple_sets optimization.
We essentially have:
if (g == 1)
  {
    g = 1;
    f = 23;
  }
else
  {
    g = 2;
    f = 20;
  }
and for each insn try to create a conditional move sequence.
There is code to detect overlap with the regs used in the condition
and the destinations, so we actually try to construct:
tmp_g = g == 1 ? 1 : 2;
f = g == 1 ? 23 : 20;
g = tmp_g;
which is fine.  But, we actually try to create two different
conditional move sequences in each case, seq1 with the whole
(eq (reg/v:HI 82 [ g ]) (const_int 1 [0x1]))
condition and seq2 with cc_cmp
(eq (reg:CCZ 17 flags) (const_int 0 [0]))
to rely on the earlier present comparison.  In each case, we
compare the rtx costs and choose the cheaper sequence (seq1 if both
have the same cost).
The problem is that with the skylake tuning,
tmp_g = g == 1 ? 1 : 2;
is actually expanded as
tmp_g = (g == 1) + 1;
in seq1 (which clobbers (reg 17 flags)) and as a cmov in seq2
(which doesn't).  The tuning says both have the same cost, so we
pick seq1.  Next we check sequences for
f = g == 1 ? 23 : 20; and here the seq2 cmov is cheaper, but it
uses (reg 17 flags) which has been clobbered earlier.

The following patch fixes that by detecting if we in the chosen
sequence clobber some register mentioned in cc_cmp or rev_cc_cmp,
and if yes, arranges for only seq1 (i.e. sequences that emit the
comparison itself) to be used after that.

Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?

2022-08-15  Jakub Jelinek  <jakub@redhat.com>

	PR rtl-optimization/106590
	* ifcvt.cc (check_for_cc_cmp_clobbers): New function.
	(noce_convert_multiple_sets_1): If SEQ sets or clobbers any regs
	mentioned in cc_cmp or rev_cc_cmp, don't consider seq2 for any
	further conditional moves.

	* gcc.dg/torture/pr106590.c: New test.


	Jakub
  

On Mon, 15 Aug 2022, Jakub Jelinek wrote:

> Hi!
> 
> The following testcase is miscompiled on x86_64-linux.
> The problem is in the noce_convert_multiple_sets optimization.
> We essentially have:
> if (g == 1)
>   {
>     g = 1;
>     f = 23;
>   }
> else
>   {
>     g = 2;
>     f = 20;
>   }
> and for each insn try to create a conditional move sequence.
> There is code to detect overlap with the regs used in the condition
> and the destinations, so we actually try to construct:
> tmp_g = g == 1 ? 1 : 2;
> f = g == 1 ? 23 : 20;
> g = tmp_g;
> which is fine.  But, we actually try to create two different
> conditional move sequences in each case, seq1 with the whole
> (eq (reg/v:HI 82 [ g ]) (const_int 1 [0x1]))
> condition and seq2 with cc_cmp
> (eq (reg:CCZ 17 flags) (const_int 0 [0]))
> to rely on the earlier present comparison.  In each case, we
> compare the rtx costs and choose the cheaper sequence (seq1 if both
> have the same cost).
> The problem is that with the skylake tuning,
> tmp_g = g == 1 ? 1 : 2;
> is actually expanded as
> tmp_g = (g == 1) + 1;
> in seq1 (which clobbers (reg 17 flags)) and as a cmov in seq2
> (which doesn't).  The tuning says both have the same cost, so we
> pick seq1.  Next we check sequences for
> f = g == 1 ? 23 : 20; and here the seq2 cmov is cheaper, but it
> uses (reg 17 flags) which has been clobbered earlier.
> 
> The following patch fixes that by detecting if we in the chosen
> sequence clobber some register mentioned in cc_cmp or rev_cc_cmp,
> and if yes, arranges for only seq1 (i.e. sequences that emit the
> comparison itself) to be used after that.
> 
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?

OK.

Thanks,
Richard.

> 2022-08-15  Jakub Jelinek  <jakub@redhat.com>
> 
> 	PR rtl-optimization/106590
> 	* ifcvt.cc (check_for_cc_cmp_clobbers): New function.
> 	(noce_convert_multiple_sets_1): If SEQ sets or clobbers any regs
> 	mentioned in cc_cmp or rev_cc_cmp, don't consider seq2 for any
> 	further conditional moves.
> 
> 	* gcc.dg/torture/pr106590.c: New test.
> 
> --- gcc/ifcvt.cc.jj	2022-07-26 10:32:23.000000000 +0200
> +++ gcc/ifcvt.cc	2022-08-12 16:31:45.348151269 +0200
> @@ -3369,6 +3369,20 @@ noce_convert_multiple_sets (struct noce_
>    return TRUE;
>  }
>  
> +/* Helper function for noce_convert_multiple_sets_1.  If store to
> +   DEST can affect P[0] or P[1], clear P[0].  Called via note_stores.  */
> +
> +static void
> +check_for_cc_cmp_clobbers (rtx dest, const_rtx, void *p0)
> +{
> +  rtx *p = (rtx *) p0;
> +  if (p[0] == NULL_RTX)
> +    return;
> +  if (reg_overlap_mentioned_p (dest, p[0])
> +      || (p[1] && reg_overlap_mentioned_p (dest, p[1])))
> +    p[0] = NULL_RTX;
> +}
> +
>  /* This goes through all relevant insns of IF_INFO->then_bb and tries to
>     create conditional moves.  In case a simple move sufficis the insn
>     should be listed in NEED_NO_CMOV.  The rewired-src cases should be
> @@ -3519,7 +3533,7 @@ noce_convert_multiple_sets_1 (struct noc
>  
>  	 as min/max and emit an insn, accordingly.  */
>        unsigned cost1 = 0, cost2 = 0;
> -      rtx_insn *seq, *seq1, *seq2;
> +      rtx_insn *seq, *seq1, *seq2 = NULL;
>        rtx temp_dest = NULL_RTX, temp_dest1 = NULL_RTX, temp_dest2 = NULL_RTX;
>        bool read_comparison = false;
>  
> @@ -3531,9 +3545,10 @@ noce_convert_multiple_sets_1 (struct noc
>  	 as well.  This allows the backend to emit a cmov directly without
>  	 creating an additional compare for each.  If successful, costing
>  	 is easier and this sequence is usually preferred.  */
> -      seq2 = try_emit_cmove_seq (if_info, temp, cond,
> -				 new_val, old_val, need_cmov,
> -				 &cost2, &temp_dest2, cc_cmp, rev_cc_cmp);
> +      if (cc_cmp)
> +	seq2 = try_emit_cmove_seq (if_info, temp, cond,
> +				   new_val, old_val, need_cmov,
> +				   &cost2, &temp_dest2, cc_cmp, rev_cc_cmp);
>  
>        /* The backend might have created a sequence that uses the
>  	 condition.  Check this.  */
> @@ -3588,6 +3603,24 @@ noce_convert_multiple_sets_1 (struct noc
>  	  return FALSE;
>  	}
>  
> +      if (cc_cmp)
> +	{
> +	  /* Check if SEQ can clobber registers mentioned in
> +	     cc_cmp and/or rev_cc_cmp.  If yes, we need to use
> +	     only seq1 from that point on.  */
> +	  rtx cc_cmp_pair[2] = { cc_cmp, rev_cc_cmp };
> +	  for (walk = seq; walk; walk = NEXT_INSN (walk))
> +	    {
> +	      note_stores (walk, check_for_cc_cmp_clobbers, cc_cmp_pair);
> +	      if (cc_cmp_pair[0] == NULL_RTX)
> +		{
> +		  cc_cmp = NULL_RTX;
> +		  rev_cc_cmp = NULL_RTX;
> +		  break;
> +		}
> +	    }
> +	}
> +
>        /* End the sub sequence and emit to the main sequence.  */
>        emit_insn (seq);
>  
> --- gcc/testsuite/gcc.dg/torture/pr106590.c.jj	2022-08-12 16:39:33.931965959 +0200
> +++ gcc/testsuite/gcc.dg/torture/pr106590.c	2022-08-12 16:39:17.752179521 +0200
> @@ -0,0 +1,75 @@
> +/* PR rtl-optimization/106590 } */
> +/* { dg-do run } */
> +/* { dg-additional-options "-mtune=skylake" { target { i?86-*-* x86_64-*-* } } } */
> +
> +typedef struct A { short a; } A;
> +typedef A *B;
> +typedef struct C { int c, d; } C;
> +typedef C *D;
> +
> +B
> +foo (void)
> +{
> +  static A r = { .a = 1 };
> +  return &r;
> +}
> +
> +D
> +bar (void)
> +{
> +  static C r = { .c = 1, .d = 23 };
> +  return &r;
> +}
> +
> +static inline int __attribute__((always_inline))
> +baz (short a)
> +{
> +  int e = 1, f;
> +  short g;
> +  D h;
> +
> +  switch (a)
> +    {
> +    case 1:
> +      f = 23;
> +      g = 1;
> +      break;
> +    case 2:
> +      f = 20;
> +      g = 2;
> +      break;
> +    }
> +
> +  h = bar ();
> +
> +  if (h->d != f || h->c != g)
> +    __builtin_abort ();
> +  return e;
> +}
> +
> +int
> +qux (void)
> +{
> +  B i = foo ();
> +  int e = 1;
> +
> +  switch (i->a)
> +    {
> +    case 1:
> +    case 2:
> +      e = baz (i->a);
> +      break;
> +    case 3:
> +      e = 0;
> +      break;
> +    }
> +
> +  return e;
> +}
> +
> +int
> +main ()
> +{
> +  qux ();
> +  return 0;
> +}
> 
> 	Jakub
> 
>