This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [Patch] MIPS: Optimize sync_compare_and_swap for const zero.


David Daney <ddaney@avtrex.com> writes:
> This patch fixes a missed optimization problem with the 
> sync_compare_and_swap insns I recently added.
>
> For a constant zero compare value we can directly compare against $0, 
> instead of loading zero into a register for the compare.  This typically 
> saves one instruction and reduces register usage.
>
> Tested on mipsel-linux-gnu with bootstrap of all default languages 
> yielding no regressions.
>
> OK to commit?
>
> 2007-10-02  David Daney  <ddaney@avtrex.com>
>
>     * config/mips/mips.md (sync_compare_and_swap<mode>): Handle compare
>     against constant zero.
>     * config/mips/mips.h (MIPS_COMPARE_AND_SWAP):  Handle constant zero
>     operand.

Thanks for catching this.

> Index: gcc/config/mips/mips.md
> ===================================================================
> --- gcc/config/mips/mips.md	(revision 128829)
> +++ gcc/config/mips/mips.md	(working copy)
> @@ -4331,7 +4331,7 @@ (define_insn "sync_compare_and_swap<mode
>    [(set (match_operand:GPR 0 "register_operand" "=&d,&d")
>  	(match_operand:GPR 1 "memory_operand" "+R,R"))
>     (set (match_dup 1)
> -	(unspec_volatile:GPR [(match_operand:GPR 2 "register_operand" "d,d")
> +	(unspec_volatile:GPR [(match_operand:GPR 2 "reg_or_0_operand" "Jd,Jd")
>  			      (match_operand:GPR 3 "arith_operand" "I,d")]
>  	 UNSPEC_COMPARE_AND_SWAP))]
>    "GENERATE_LL_SC"

Silly minor nit, but please use dJ rather than Jd for consistency with
other patterns.  OK with that change; no need to retest.

Thanks,
Richard


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]