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]

ping^3 [PATCH] Subreg handling in reload_inner_reg_of_subreg() and push_reload()


On Mon, Jul 30, 2007 at 06:10:23PM +0200, Rask Ingemann Lambertsen wrote:
>    In <URL:http://gcc.gnu.org/ml/gcc/2007-05/msg00505.html> and
> <URL:http://gcc.gnu.org/ml/gcc-patches/2007-05/msg01570.html> (which was
> briefly discussed) I described a case where reload is given HImode subreg of
> a multiword SImode hard reg. Because the HImode subreg is made of two QImode
> hard regs rather than one HImode hard reg, reload decides to write to the
> whole SImode hard reg instead of just the HImode subreg on the assumption
> that the subreg is "problematic", as a comment puts it. In fact, the rest of
> reload handles such a subreg just fine with no special treatment. This patch
> adjusts the checks in reload_inner_reg_of_subreg() and push_reload()
> accordingly.
> 
>    This and the other two reload patches in this patch set were bootstrapped
> and tested together on x86_64-unknown-linux-gnu with no new failures. I also
> built and tested cross compilers for arm-unknown-elf, cris-axis-elf,
> m32c-unknown-elf, mipsisa64-unknown-elf, sh-unknown-elf and v850-unknown-elf
> with no new failures.
> 
>    Ok for trunk?
> 
> 2007-07-30  Rask Ingemann Lambertsen  <rask@sygehus.dk>
> 
> 	* reload.c (reload_inner_reg_of_subreg): No special handling is
> 	  needed if a subreg consists of more hard regs than words.
> 	  (push_reload): Likewise.
> 
> Index: gcc/reload.c
> ===================================================================
> --- gcc/reload.c	(revision 126653)
> +++ gcc/reload.c	(working copy)
> @@ -809,13 +809,13 @@ reload_inner_reg_of_subreg (rtx x, enum 
>      return 1;
>  
>    /* If the outer part is a word or smaller, INNER larger than a
> -     word and the number of regs for INNER is not the same as the
> +     word and the number of regs for INNER is smaller than the
>       number of words in INNER, then INNER will need reloading.  */
>    return (GET_MODE_SIZE (mode) <= UNITS_PER_WORD
>  	  && output
>  	  && GET_MODE_SIZE (GET_MODE (inner)) > UNITS_PER_WORD
>  	  && ((GET_MODE_SIZE (GET_MODE (inner)) / UNITS_PER_WORD)
> -	      != (int) hard_regno_nregs[REGNO (inner)][GET_MODE (inner)]));
> +	      > (int) hard_regno_nregs[REGNO (inner)][GET_MODE (inner)]));
>  }
>  
>  /* Return nonzero if IN can be reloaded into REGNO with mode MODE without
> @@ -1035,8 +1035,8 @@ push_reload (rtx in, rtx out, rtx *inloc
>  		       > UNITS_PER_WORD)
>  		   && ((GET_MODE_SIZE (GET_MODE (SUBREG_REG (in)))
>  			/ UNITS_PER_WORD)
> -		       != (int) hard_regno_nregs[REGNO (SUBREG_REG (in))]
> -						[GET_MODE (SUBREG_REG (in))]))
> +		       > (int) hard_regno_nregs[REGNO (SUBREG_REG (in))]
> +					       [GET_MODE (SUBREG_REG (in))]))
>  		  || ! HARD_REGNO_MODE_OK (subreg_regno (in), inmode)))
>  	  || (secondary_reload_class (1, class, inmode, in) != NO_REGS
>  	      && (secondary_reload_class (1, class, GET_MODE (SUBREG_REG (in)),
> @@ -1128,8 +1128,8 @@ push_reload (rtx in, rtx out, rtx *inloc
>  		       > UNITS_PER_WORD)
>  		   && ((GET_MODE_SIZE (GET_MODE (SUBREG_REG (out)))
>  			/ UNITS_PER_WORD)
> -		       != (int) hard_regno_nregs[REGNO (SUBREG_REG (out))]
> -						[GET_MODE (SUBREG_REG (out))]))
> +		       > (int) hard_regno_nregs[REGNO (SUBREG_REG (out))]
> +					       [GET_MODE (SUBREG_REG (out))]))
>  		  || ! HARD_REGNO_MODE_OK (subreg_regno (out), outmode)))
>  	  || (secondary_reload_class (0, class, outmode, out) != NO_REGS
>  	      && (secondary_reload_class (0, class, GET_MODE (SUBREG_REG (out)),
> 
> -- 
> Rask Ingemann Lambertsen


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