[PATCH] Fix (X >> C1) & C2 folding (PR tree-optimization/61158)

[Richard Biener]

On Thu, 15 May 2014, Jakub Jelinek wrote:

> Hi!
> 
> fold_binary_loc for (X >> C1) & C2, if X is zero extended narrower
> value, decreases prec, but if the shift count is bigger
> than the narrower prec, we then attempt to zerobits <<= negative_value.
> 
> In that case the result is necessarily zero though, all possibly non-zero
> bits are shifted away, so this patch fixes this case by making sure zerobits
> is all ones in that case, which results in (X, 0) folding a few lines below.
> 
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk/4.9/4.8?

Ok.

Thanks,
Richard.

> 2014-05-15  Jakub Jelinek  <jakub@redhat.com>
> 
> 	PR tree-optimization/61158
> 	* fold-const.c (fold_binary_loc): If X is zero-extended and
> 	shiftc >= prec, make sure zerobits is all ones instead of
> 	invoking undefined behavior.
> 
> 	* gcc.dg/pr61158.c: New test.
> 
> --- gcc/fold-const.c.jj	2014-05-14 09:46:07.000000000 +0200
> +++ gcc/fold-const.c	2014-05-14 11:35:19.593730226 +0200
> @@ -11972,11 +11972,17 @@ fold_binary_loc (location_t loc,
>  		      /* See if we can shorten the right shift.  */
>  		      if (shiftc < prec)
>  			shift_type = inner_type;
> +		      /* Otherwise X >> C1 is all zeros, so we'll optimize
> +			 it into (X, 0) later on by making sure zerobits
> +			 is all ones.  */
>  		    }
>  		}
>  	      zerobits = ~(unsigned HOST_WIDE_INT) 0;
> -	      zerobits >>= HOST_BITS_PER_WIDE_INT - shiftc;
> -	      zerobits <<= prec - shiftc;
> +	      if (shiftc < prec)
> +		{
> +		  zerobits >>= HOST_BITS_PER_WIDE_INT - shiftc;
> +		  zerobits <<= prec - shiftc;
> +		}
>  	      /* For arithmetic shift if sign bit could be set, zerobits
>  		 can contain actually sign bits, so no transformation is
>  		 possible, unless MASK masks them all away.  In that
> @@ -11994,7 +12000,7 @@ fold_binary_loc (location_t loc,
>  	  /* ((X << 16) & 0xff00) is (X, 0).  */
>  	  if ((mask & zerobits) == mask)
>  	    return omit_one_operand_loc (loc, type,
> -				     build_int_cst (type, 0), arg0);
> +					 build_int_cst (type, 0), arg0);
>  
>  	  newmask = mask | zerobits;
>  	  if (newmask != mask && (newmask & (newmask + 1)) == 0)
> --- gcc/testsuite/gcc.dg/pr61158.c.jj	2014-05-14 12:18:06.066817887 +0200
> +++ gcc/testsuite/gcc.dg/pr61158.c	2014-05-14 12:18:48.046598622 +0200
> @@ -0,0 +1,12 @@
> +/* PR tree-optimization/61158 */
> +/* { dg-do compile } */
> +/* { dg-options "-O2 -fdump-tree-original" } */
> +
> +unsigned long long
> +foo (unsigned int x)
> +{
> +  return ((unsigned long long) x & 0x00ff000000000000ULL) >> 40;
> +}
> +
> +/* { dg-final { scan-tree-dump "return 0;" "original" { target { ilp32 || lp64 } } } } */
> +/* { dg-final { cleanup-tree-dump "original" } } */
> 
> 	Jakub
> 
> 

-- 
Richard Biener <rguenther@suse.de>
SUSE / SUSE Labs
SUSE LINUX Products GmbH - Nuernberg - AG Nuernberg - HRB 16746
GF: Jeff Hawn, Jennifer Guild, Felix Imend"orffer