This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [PATCH] Fix PR39715 (Thumb pessimization)
- From: Adam Nemet <anemet at caviumnetworks dot com>
- To: Paolo Bonzini <paolo dot bonzini at gmail dot com>
- Cc: gcc-patches at gcc dot gnu dot org, Richard Earnshaw <rearnsha at arm dot com>
- Date: Sat, 01 Aug 2009 17:13:59 -0700
- Subject: Re: [PATCH] Fix PR39715 (Thumb pessimization)
- References: <1249161859-12993-2-git-send-email-bonzini@gnu.org>
Hi Paolo,
Paolo Bonzini <bonzini@gnu.org> writes:
> Index: gcc/simplify-rtx.c
> ===================================================================
> --- gcc/simplify-rtx.c (branch cond-optab2)
> +++ gcc/simplify-rtx.c (working copy)
> @@ -3990,6 +3990,49 @@ simplify_relational_operation_1 (enum rt
> simplify_gen_binary (XOR, cmp_mode,
> XEXP (op0, 1), op1));
>
> + /* Look for useless extensions of a subreg, if the extension
> + is from the same mode as the subreg. */
> + if (op0code == SIGN_EXTEND
> + && GET_CODE (XEXP (op0, 0)) == SUBREG
> + && subreg_lowpart_p (XEXP (op0, 0))
> + && GET_MODE (op0) == GET_MODE (SUBREG_REG (XEXP (op0, 0)))
> + && num_sign_bit_copies (SUBREG_REG (XEXP (op0, 0)), GET_MODE (op0))
> + > (unsigned) (GET_MODE_BITSIZE (GET_MODE (op0))
> + - GET_MODE_BITSIZE (GET_MODE (XEXP (op0, 0)))))
> + return simplify_gen_relational (code, mode, cmp_mode,
> + SUBREG_REG (XEXP (op0, 0)), op1);
> +
> + if (GET_CODE (op1) == SIGN_EXTEND
> + && GET_CODE (XEXP (op1, 0)) == SUBREG
> + && subreg_lowpart_p (XEXP (op0, 0))
> + && GET_MODE (op1) == GET_MODE (SUBREG_REG (XEXP (op1, 0)))
> + && num_sign_bit_copies (SUBREG_REG (XEXP (op1, 0)), GET_MODE (op1))
> + > (unsigned) (GET_MODE_BITSIZE (GET_MODE (op1))
> + - GET_MODE_BITSIZE (GET_MODE (XEXP (op1, 0)))))
> + return simplify_gen_relational (code, mode, cmp_mode,
> + op0, SUBREG_REG (XEXP (op1, 0)));
These seem to be true even without the enclosing comparison:
(sign_extend:W (subreg:N (reg:W X) <low>))
->
(reg:W X) if (reg:W X) is already sign-extended from N to W.
Doesn't my changes to simplify_shift_const_1 in combine fix this? Or I
guess the problem is that since you create these in simplify-rtx we
never actually see this in the combiner proper.
Then maybe we should do this in simplify_unary_operation_1
<SIGN_EXTEND>? It's bad that we'd have this in two places but in order
to share code we would have to expand and rebuild compounds in
simplify-rtx too which does not seem like a good idea.
Adam