This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [PATCH][GCC][PATCHv3] Improve fpclassify w.r.t IEEE like numbers in GIMPLE.
- From: Joseph Myers <joseph at codesourcery dot com>
- To: Tamar Christina <Tamar dot Christina at arm dot com>
- Cc: Jeff Law <law at redhat dot com>, GCC Patches <gcc-patches at gcc dot gnu dot org>, Wilco Dijkstra <Wilco dot Dijkstra at arm dot com>, "rguenther at suse dot de" <rguenther at suse dot de>, Michael Meissner <meissner at linux dot vnet dot ibm dot com>, nd <nd at arm dot com>
- Date: Thu, 19 Jan 2017 14:46:06 +0000
- Subject: Re: [PATCH][GCC][PATCHv3] Improve fpclassify w.r.t IEEE like numbers in GIMPLE.
- Authentication-results: sourceware.org; auth=none
- References: <HE1PR0801MB2027D1CA7B116B16DD1E5641FFBB0@HE1PR0801MB2027.eurprd08.prod.outlook.com> <alpine.DEB.2.20.1611112157270.32183@digraph.polyomino.org.uk> <VI1PR0801MB2031689EAA3643B992C12E0AFFB60@VI1PR0801MB2031.eurprd08.prod.outlook.com> <alpine.DEB.2.20.1611241821520.21096@digraph.polyomino.org.uk> <VI1PR0801MB203191C5CA00160F3DAC7E40FF890@VI1PR0801MB2031.eurprd08.prod.outlook.com> <bad2845d-1944-d792-d4d1-91eac63eb73c@redhat.com> <VI1PR0801MB20311CC45C387326291E49F1FF9D0@VI1PR0801MB2031.eurprd08.prod.outlook.com> <alpine.DEB.2.20.1612151858040.2106@digraph.polyomino.org.uk> <VI1PR0801MB20313CD2538258D3AAA60F5DFF910@VI1PR0801MB2031.eurprd08.prod.outlook.com> <72535306-7d9c-bc2d-b7a3-06360633cfef@redhat.com> <alpine.DEB.2.20.1701181641540.1980@digraph.polyomino.org.uk>,<alpine.DEB.2.20.1701181703580.1980@digraph.polyomino.org.uk> <VI1PR0801MB2031DC35E0B3ADDF40EB5A0AFF7E0@VI1PR0801MB2031.eurprd08.prod.outlook.com>
On Thu, 19 Jan 2017, Tamar Christina wrote:
> > Also, I don't think the call to perform_ibm_extended_fixups in
> > is_subnormal is correct. Subnormal for IBM long double is *not* the same
> > as subnormal double high part. Likewise it's incorrect in is_normal as
> > well.
>
> The calls to is_zero and is_subnormal were incorrect indeed. I've
> corrected them by not calling the fixup code and to instead make sure it
> falls through into the old fp based code which did normal floating point
> operations on the number. This is the same code as was before in
> fpclassify so it should work.
For is_zero it's fine to test based on the high part for IBM long double;
an IBM long double is (zero, infinite, NaN, finite) if and only if the
high part is. The problem is the different threshold between normal and
subnormal.
> As for is_normal, the code is almost identical as the code that used to
> be in fold_builtin_interclass_mathfn in BUILT_IN_ISNORMAL, with the
> exception that I don't check <= max_value but instead < inifity, so I
> can reuse the same constant.
The old code set orig_arg before converting IBM long double to double.
Your code sets it after the conversion. The old code set min_exp based on
a string set from REAL_MODE_FORMAT (orig_mode)->emin - 1; your code uses
the adjusted mode. Both of those are incorrect for IBM long double.
--
Joseph S. Myers
joseph@codesourcery.com