Re: [PATCH, RFT s390x] Bug in truncation simplifcation

[Adam Nemet <anemet at sonic dot net>]

I write:
> I found this bug while working on my previous patch.
>
> If MODE is the mode of the truncation and OP is the operand then in
> order to see whether the bits outside of MODE have the same value as
> the new sign bit, the number of leading bits we need to check is:
>
>   GET_MODE_BITSIZE (GET_MODE (OP)) - GET_MODE_BITSIZE (MODE) + 1
>
> Instead the code checks GET_MODE_BITSIZE (mode) + 1 bits which happens
> to work if the new mode is half the size of the original mode.
>
> Unfortunately, I couldn't sufficiently exercise the new code.  I
> bootstrapped and tested on i686-pc-linux-gnu.  However on x86,
> TRUNCATEs only occur before an LSHIFTRT and a MULT where this
> transformation is disabled specifically.  I tried testing with a
> h8300-elf toolchain but the build fails in newlib.  Finally, I built
> an s390x-ibm-linux-gnu cc1 and compared the assembly output for
> gcc.c-torture/exercise and for GCC preprocessed files.  There is no
> change in the output before and after the patch and the code does
> trigger a few times.

After this week's changes from Roger
(http://gcc.gnu.org/ml/gcc-patches/2006-03/msg00187.html) the
middle-end started generating TRUNCATEs for TRULY_NOOP_TRUNCATION
targets as well.

I checked and with freshly updated sources my changes do trigger now
on x86_64-pc-linux-gnu.  So I retested and bootstrapped my patch on
x86_64-pc-linux-gnu.  Testing revealed no problems.

Also note that RTH has agreed that this was a bug
(http://gcc.gnu.org/ml/gcc-patches/2006-03/msg00137.html).

OK to install?

Adam

>
> 	* simplify-rtx.c (simplify_unary_operation_1) <TRUNCATE>: When
> 	trying to remove TRUNCATE check if all bits outside the new mode
> 	are identical to the sign bit.
>
> Index: simplify-rtx.c
> ===================================================================
> --- simplify-rtx.c	(revision 111573)
> +++ simplify-rtx.c	(working copy)
> @@ -640,7 +640,8 @@ simplify_unary_operation_1 (enum rtx_cod
>        if ((TRULY_NOOP_TRUNCATION (GET_MODE_BITSIZE (mode),
>  				 GET_MODE_BITSIZE (GET_MODE (op)))
>  	   ? (num_sign_bit_copies (op, GET_MODE (op))
> -	      >= (unsigned int) (GET_MODE_BITSIZE (mode) + 1))
> +	      > (unsigned int) (GET_MODE_BITSIZE (GET_MODE (op))
> +				- GET_MODE_BITSIZE (mode)))
>  	   : truncated_to_mode (mode, op))
>  	  && ! (GET_CODE (op) == LSHIFTRT
>  		&& GET_CODE (XEXP (op, 0)) == MULT))