about summary refs log tree commit diff
path: root/sysdeps/powerpc/powerpc64/fpu/s_llroundf.S
diff options
context:
space:
mode:
Diffstat (limited to 'sysdeps/powerpc/powerpc64/fpu/s_llroundf.S')
-rw-r--r--sysdeps/powerpc/powerpc64/fpu/s_llroundf.S53
1 files changed, 37 insertions, 16 deletions
diff --git a/sysdeps/powerpc/powerpc64/fpu/s_llroundf.S b/sysdeps/powerpc/powerpc64/fpu/s_llroundf.S
index bbbd05492e..a211879393 100644
--- a/sysdeps/powerpc/powerpc64/fpu/s_llroundf.S
+++ b/sysdeps/powerpc/powerpc64/fpu/s_llroundf.S
@@ -1,5 +1,5 @@
 /* llroundf function.  PowerPC64 version.
-   Copyright (C) 2004, 2006 Free Software Foundation, Inc.
+   Copyright (C) 2004, 2006, 2007 Free Software Foundation, Inc.
    This file is part of the GNU C Library.
 
    The GNU C Library is free software; you can redistribute it and/or
@@ -20,8 +20,8 @@
 #include <sysdep.h>
 
 	.section	".toc","aw"
-.LC0:	/* -0.0 */
-	.tc FD_00000000_0[TC],0x0000000000000000
+.LC0:	/* 2^23 */
+	.tc FD_41600000_0[TC],0x4160000000000000
 .LC1:	/* 0.5 */
 	.tc FD_3fe00000_0[TC],0x3fe0000000000000
 	.section	".text"
@@ -34,24 +34,45 @@
    tie, choose the one that is even (least significant bit o).". 
    So we can't use the PowerPC "round to Nearest" mode. Instead we set
    "round toward Zero" mode and round by adding +-0.5 before rounding
-   to the integer value.  */
+   to the integer value.
+
+   It is necessary to detect when x is (+-)0x1.fffffffffffffp-2
+   because adding +-0.5 in this case will cause an erroneous shift,
+   carry and round.  We simply return 0 if 0.5 > x > -0.5.  Likewise
+   if x is and odd number between +-(2^23 and 2^24-1) a shift and
+   carry will erroneously round if biased with +-0.5.  Therefore if x
+   is greater/less than +-2^23 we don't need to bias the number with
+   +-0.5.  */
 
 ENTRY (__llroundf)
 	CALL_MCOUNT 0
-	lfd	fp12,.LC0@toc(2)
-	lfd	fp10,.LC1@toc(2)
-	fcmpu	cr6,fp1,fp12	/* if (x < 0.0)  */
-	fsubs	fp3,fp1,fp10	/* x-= 0.5;  */
-	ble-	cr6,.L9
-	fadds	fp3,fp1,fp10	/* x+= 0.5;  */
-.L9:
-	fctidz	fp2,fp3		/* Convert To Integer DW round toward 0.  */
-	stfd	fp2,-16(r1)
-	nop	/* Insure the following load is in a different dispatch group */
-	nop	/* to avoid pipe stall on POWER4&5.  */
+	lfd	fp9,.LC0@toc(2)	/* Load 2^23 into fpr9.  */
+	lfd	fp10,.LC1@toc(2)/* Load 0.5 into fpr10.  */
+	fabs	fp2,fp1		/* Get the absolute value of x.  */
+	fsub	fp12,fp10,fp10	/* Compute 0.0 into fp12.  */
+	fcmpu	cr6,fp2,fp10	/* if |x| < 0.5  */
+	fcmpu	cr4,fp2,fp9	/* if |x| >= 2^23  */
+	fcmpu	cr3,fp1,fp12	/* x is negative? x < 0.0  */
+	blt-	cr6,.Lretzero	/* 0.5 > x < -0.5 so just return 0.  */
+	bge-	cr4,.Lnobias	/* 2^23 > x < -2^23 just convert with no bias.  */
+	fadd	fp3,fp2,fp10	/* |x|+=0.5 bias to prepare to round.  */
+	bge	cr3,.Lconvert	/* x is positive so don't negate x.  */
+	fnabs	fp3,fp3		/* -(|x|+=0.5)  */
+.Lconvert:
+	fctidz	fp4,fp3		/* Convert to Integer double word round toward 0.  */
+	stfd	fp4,-16(r1)
+	nop
+	nop
 	nop
-	ld	r3,-16(r1)
+	ld	r3,-16(r1)	/* Load return as integer.  */
+.Lout:
 	blr
+.Lretzero:			/* 0.5 > x > -0.5  */
+	li	r3,0		/* return 0.  */
+	b	.Lout
+.Lnobias:
+	fmr	fp3,fp1
+	b	.Lconvert
 	END (__llroundf)
 
 strong_alias (__llroundf, __lroundf)