This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [PATCH] Fix ARM NAN fraction bits
- From: "Maciej W. Rozycki" <macro at codesourcery dot com>
- To: Joey Ye <joey dot ye at arm dot com>, "Joseph S. Myers" <joseph at codesourcery dot com>
- Cc: <libc-alpha at sourceware dot org>, <gcc-patches at gcc dot gnu dot org>
- Date: Fri, 16 May 2014 17:57:51 +0100
- Subject: Re: [PATCH] Fix ARM NAN fraction bits
- Authentication-results: sourceware.org; auth=none
- References: <000b01cf3367$439c5280$cad4f780$ at arm dot com>
On Thu, 27 Feb 2014, Joey Ye wrote:
> Current ARM soft-float implementation is violating the RTABI
> (http://infocenter.arm.com/help/topic/com.arm.doc.ihi0043d/IHI0043D_rtabi.pd
> f) Section 4.1.1.1:
>
> When not otherwise specified by IEEE 754, the result on an invalid operation
> should be the quiet NaN bit pattern with only the most significant bit of
> the significand set, and all other significand bits zero.
>
> This patch fixes it by setting _FP_NANFRAC_* to zero.
>
> Ran make check test with âmfloat-abi=soft. No regression.
>
> OK to checkin?
>
> 2014-02-27 Joey Ye <joey.ye@arm.com>
> * sysdeps/arm/soft-fp/sfp-machine.h
> ÂÂ(_FP_NANFRAC_S, _FP_NANFRAC_D, _FP_NANFRAC_Q):
> Â Set to zero.
>
>
> diff --git a/sysdeps/arm/soft-fp/sfp-machine.h
> b/sysdeps/arm/soft-fp/sfp-machine.h
> index 52a08b5..32697fe 100644
> --- a/sysdeps/arm/soft-fp/sfp-machine.h
> +++ b/sysdeps/arm/soft-fp/sfp-machine.h
> @@ -21,9 +21,9 @@
> #define _FP_DIV_MEAT_D(R,X,Y)ÂÂÂÂÂÂÂÂÂ _FP_DIV_MEAT_2_udiv(D,R,X,Y)
> #define _FP_DIV_MEAT_Q(R,X,Y)ÂÂÂÂÂÂÂÂÂ _FP_DIV_MEAT_4_udiv(Q,R,X,Y)
>
> -#define _FP_NANFRAC_SÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ ((_FP_QNANBIT_S << 1) - 1)
> -#define _FP_NANFRAC_DÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ ((_FP_QNANBIT_D << 1) - 1),
> -1
> -#define _FP_NANFRAC_QÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ ((_FP_QNANBIT_Q << 1) - 1),
> -1, -1, -1
> +#define _FP_NANFRAC_SÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ 0
> +#define _FP_NANFRAC_DÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ 0, 0
> +#define _FP_NANFRAC_QÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ 0, 0, 0, 0
> #define _FP_NANSIGN_SÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ 0
> #define _FP_NANSIGN_DÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ 0
> #define _FP_NANSIGN_QÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ 0
This did regrettably, when propagated to libgcc, regress
gcc.dg/torture/builtin-math-7.c on soft-fp arm-eabi targets, currently
ARMv6-M (`-march=armv6-m -mthumb') only. This is because these NANFRAC
macros have now no bits set and as a result when used to construct a NaN
in the semi-raw mode, they build an infinity instead. Consequently
operations such as (Inf - Inf) now produce Inf rather than NaN. The
change worked for the original test case, because division is made in the
canonical mode, where the quiet bit is set separately, from the fp class.
Here's a fix making code match the commit description quoted above, that
is set the most significant bit of the significand. This is also what
targets similar in this respect do.
OK to apply? OK for libgcc (against libgcc/config/arm/sfp-machine.h), in
particular for GCC 4.8 and 4.9?
2014-05-16 Maciej W. Rozycki <macro@codesourcery.com>
PR libgcc/60166
* sysdeps/arm/soft-fp/sfp-machine.h (_FP_NANFRAC_S, _FP_NANFRAC_D)
(_FP_NANSIGN_Q): Set the quiet bit.
Maciej
glibc-soft-fp-arm-nanfrac.diff
Index: glibc-fsf-trunk-quilt/sysdeps/arm/soft-fp/sfp-machine.h
===================================================================
--- glibc-fsf-trunk-quilt.orig/sysdeps/arm/soft-fp/sfp-machine.h 2014-05-16 03:25:52.000000000 +0100
+++ glibc-fsf-trunk-quilt/sysdeps/arm/soft-fp/sfp-machine.h 2014-05-16 03:31:34.451805339 +0100
@@ -21,9 +21,9 @@
#define _FP_DIV_MEAT_D(R,X,Y) _FP_DIV_MEAT_2_udiv(D,R,X,Y)
#define _FP_DIV_MEAT_Q(R,X,Y) _FP_DIV_MEAT_4_udiv(Q,R,X,Y)
-#define _FP_NANFRAC_S 0
-#define _FP_NANFRAC_D 0, 0
-#define _FP_NANFRAC_Q 0, 0, 0, 0
+#define _FP_NANFRAC_S _FP_QNANBIT_S
+#define _FP_NANFRAC_D _FP_QNANBIT_D, 0
+#define _FP_NANFRAC_Q _FP_QNANBIT_Q, 0, 0, 0
#define _FP_NANSIGN_S 0
#define _FP_NANSIGN_D 0
#define _FP_NANSIGN_Q 0