[PATCH][i386] Implement ix86_emit_swdivsf more efficiently

[PATCH][i386] Implement ix86_emit_swdivsf more efficiently

[Richard Guenther]

This rewrites the iteration step of swdivsf to be more register
efficient (two registers instead of four, no load of a FP constant).
This matches how ICC emits the rcp sequence and causes no overall loss
of precision (Micha might still remember the exact details).  The patch is
fallout of the work trying to fix PR47989.

Bootstrapped and tested on x86_64-unknown-linux-gnu, ok for 4.7?

Thanks,
Richard.

2011-03-14  Richard Guenther  <rguenther@suse.de>

	* config/i386/i386.c (ix86_emit_swdivsf): Implement more
	efficiently.

Index: trunk/gcc/config/i386/i386.c
===================================================================
--- trunk.orig/gcc/config/i386/i386.c	2011-03-09 11:52:21.000000000 +0100
+++ trunk/gcc/config/i386/i386.c	2011-03-10 15:43:47.000000000 +0100
@@ -31747,38 +31747,38 @@ void ix86_emit_i387_log1p (rtx op0, rtx
 
 void ix86_emit_swdivsf (rtx res, rtx a, rtx b, enum machine_mode mode)
 {
-  rtx x0, x1, e0, e1, two;
+  rtx x0, x1, e0, e1;
 
   x0 = gen_reg_rtx (mode);
   e0 = gen_reg_rtx (mode);
   e1 = gen_reg_rtx (mode);
   x1 = gen_reg_rtx (mode);
 
-  two = CONST_DOUBLE_FROM_REAL_VALUE (dconst2, SFmode);
-
-  if (VECTOR_MODE_P (mode))
-    two = ix86_build_const_vector (mode, true, two);
-
-  two = force_reg (mode, two);
-
-  /* a / b = a * rcp(b) * (2.0 - b * rcp(b)) */
+  /* a / b = a * ((rcp(b) + rcp(b)) - (b * rcp(b) * rcp (b))) */
 
   /* x0 = rcp(b) estimate */
   emit_insn (gen_rtx_SET (VOIDmode, x0,
 			  gen_rtx_UNSPEC (mode, gen_rtvec (1, b),
 					  UNSPEC_RCP)));
-  /* e0 = x0 * a */
+  /* e0 = x0 * b */
   emit_insn (gen_rtx_SET (VOIDmode, e0,
-			  gen_rtx_MULT (mode, x0, a)));
-  /* e1 = x0 * b */
-  emit_insn (gen_rtx_SET (VOIDmode, e1,
 			  gen_rtx_MULT (mode, x0, b)));
-  /* x1 = 2. - e1 */
+
+  /* e0 = x0 * e0 */
+  emit_insn (gen_rtx_SET (VOIDmode, e0,
+			  gen_rtx_MULT (mode, x0, e0)));
+
+  /* e1 = x0 + x0 */
+  emit_insn (gen_rtx_SET (VOIDmode, e1,
+			  gen_rtx_PLUS (mode, x0, x0)));
+
+  /* x1 = e1 - e0 */
   emit_insn (gen_rtx_SET (VOIDmode, x1,
-			  gen_rtx_MINUS (mode, two, e1)));
-  /* res = e0 * x1 */
+			  gen_rtx_MINUS (mode, e1, e0)));
+
+  /* res = a * x1 */
   emit_insn (gen_rtx_SET (VOIDmode, res,
-			  gen_rtx_MULT (mode, e0, x1)));
+			  gen_rtx_MULT (mode, a, x1)));
 }
 
 /* Output code to perform a Newton-Rhapson approximation of a

Re: [PATCH][i386] Implement ix86_emit_swdivsf more efficiently

[Michael Matz]

Hi,

On Mon, 14 Mar 2011, Richard Guenther wrote:

> This rewrites the iteration step of swdivsf to be more register 
> efficient (two registers instead of four, no load of a FP constant). 
> This matches how ICC emits the rcp sequence and causes no overall loss 
> of precision (Micha might still remember the exact details).

I haven't done a full error analysis for the intermediate rounding steps, 
but merely a statistical analysis for a subset of dividends and the full 
set of divisors.  On AMD and Intel processors (that matters because rcpss 
accuracy is different on both) the sum of all absolute errors between the 
quotient as from divss and the quotients from either our old and the new 
method is better for the new method.  The max error is 2ulps in each case.


Ciao,
Michael.

Re: [PATCH][i386] Implement ix86_emit_swdivsf more efficiently

[Uros Bizjak]

Hello!

> This rewrites the iteration step of swdivsf to be more register
> efficient (two registers instead of four, no load of a FP constant).
> This matches how ICC emits the rcp sequence and causes no overall loss
> of precision (Micha might still remember the exact details).  The patch is
> fallout of the work trying to fix PR47989.
>
> Bootstrapped and tested on x86_64-unknown-linux-gnu, ok for 4.7?
>
> Thanks,
> Richard.
>
> 2011-03-14  Richard Guenther  <rguenther@suse.de>
>
> 	* config/i386/i386.c (ix86_emit_swdivsf): Implement more
> 	efficiently.

OK for 4.7.

Thanks,
Uros.