Re: Anyone knows this issue? std/math.d(2759): Error: number '0x1p-1024' is not representable
On 20 Jun 2015 11:35, Martin Nowak via Digitalmars-d digitalmars-d@puremagic.com wrote: On 06/20/2015 09:52 AM, Walter Bright wrote: Which C runtime are you using? The math functions in C runtimes are often inadequate. So we now call the host's C library to perform constant folding/CTFE of exp? This is a point for Iain's proposal to use a high precision floating point implementation in the compiler to avoid machine/platform differences. I thought dmd used the C library implementations for all intrinsics via CTFE, not just exp? In any case, isn't the problem you are seeing related to the conversion of string to float? Which again dmd is at the mercy of the C library implementation to handle correctly.
Re: Anyone knows this issue? std/math.d(2759): Error: number '0x1p-1024' is not representable
On 06/19/2015 06:54 PM, Andrei Alexandrescu wrote: This is happening on my older CentOS as well, it's a problem in the C library. I suggest we just version those tests. -- Andrei Even more worrisome, I don't understand why dmd is compiling the unittests at all. Might be related to https://issues.dlang.org/show_bug.cgi?id=14431 or https://issues.dlang.org/show_bug.cgi?id=14508.
Re: Anyone knows this issue? std/math.d(2759): Error: number '0x1p-1024' is not representable
On 06/20/2015 09:52 AM, Walter Bright wrote: Which C runtime are you using? The math functions in C runtimes are often inadequate. So we now call the host's C library to perform constant folding/CTFE of exp? This is a point for Iain's proposal to use a high precision floating point implementation in the compiler to avoid machine/platform differences.
Re: Anyone knows this issue? std/math.d(2759): Error: number '0x1p-1024' is not representable
On 06/20/2015 02:00 PM, Martin Nowak wrote: On 06/20/2015 12:06 PM, Walter Bright wrote: On 6/20/2015 3:06 AM, Walter Bright wrote: Apparently we need to roll our own version of strtod() and put it in Port. We already do our own strtold(), so it should be straightforward. Thanks, I'll work on a fix. I can't use the strtold_dm from the backend, b/c it's not available for the other compilers, and AFAIU licensing doesn't allow me to move it to the frontend.
Re: Anyone knows this issue? std/math.d(2759): Error: number '0x1p-1024' is not representable
On 6/19/2015 9:18 AM, Martin Nowak wrote: I'm getting this error while trying to build the relase. std/math.d(2759): Error: number '0x1p-1024' is not representable I'm getting that while compiling a -m32 (X86) static release phobos library, but I can't reproduce the error using the same command. https://github.com/D-Programming-Language/phobos/blob/abb21ab9bcd3857f1567bfde11c2d5d385f5caf0/std/math.d#L2759 Don't even know why the unittests are compiled. Which C runtime are you using? The math functions in C runtimes are often inadequate.
Re: Anyone knows this issue? std/math.d(2759): Error: number '0x1p-1024' is not representable
On 06/20/2015 09:52 AM, Walter Bright wrote: Which C runtime are you using? The math functions in C runtimes are often inadequate. Debian 7.4, there is nothing I can do to avoid glibc.
Re: Anyone knows this issue? std/math.d(2759): Error: number '0x1p-1024' is not representable
On 6/20/2015 3:06 AM, Walter Bright wrote: Apparently we need to roll our own version of strtod() and put it in Port. We already do our own strtold(), so it should be straightforward.
Re: Anyone knows this issue? std/math.d(2759): Error: number '0x1p-1024' is not representable
On 06/20/2015 12:06 PM, Walter Bright wrote: On 6/20/2015 3:06 AM, Walter Bright wrote: Apparently we need to roll our own version of strtod() and put it in Port. We already do our own strtold(), so it should be straightforward. Thanks, I'll work on a fix.
Re: Anyone knows this issue? std/math.d(2759): Error: number '0x1p-1024' is not representable
On 06/20/2015 01:35 PM, Iain Buclaw via Digitalmars-d wrote: Can we make this a boolean check removing the need for me to explicitly set errno? The function returns a float, so errno seems like the better choice.
Re: Anyone knows this issue? std/math.d(2759): Error: number '0x1p-1024' is not representable
On 6/20/2015 2:48 AM, Iain Buclaw via Digitalmars-d wrote:
On 20 Jun 2015 11:35, Martin Nowak via Digitalmars-d
digitalmars-d@puremagic.com mailto:digitalmars-d@puremagic.com wrote:
On 06/20/2015 09:52 AM, Walter Bright wrote:
Which C runtime are you using? The math functions in C runtimes are
often inadequate.
So we now call the host's C library to perform constant folding/CTFE of exp?
This is a point for Iain's proposal to use a high precision floating
point implementation in the compiler to avoid machine/platform differences.
I thought dmd used the C library implementations for all intrinsics via CTFE,
not just exp?
In any case, isn't the problem you are seeing related to the conversion of
string to float? Which again dmd is at the mercy of the C library implementation
to handle correctly.
The code is the following, in lexer.c:
errno = 0;
(void)Port::strtod((char *)stringbuffer.data, NULL);
if (errno == ERANGE)
{
const char *suffix = (result == TOKfloat32v || result == TOKimaginary32v) ?
f : ;
error(scanloc, number '%s%s' is not representable, (char
*)stringbuffer.data, suffix);
}
The point of the Port:: code is to correct for inadequacies of various C
standard library implementations, in this case Debian 7.4.
Apparently we need to roll our own version of strtod() and put it in Port.
Re: Anyone knows this issue? std/math.d(2759): Error: number '0x1p-1024' is not representable
On 20 June 2015 at 12:06, Walter Bright via Digitalmars-d digitalmars-d@puremagic.com wrote: On 6/20/2015 2:48 AM, Iain Buclaw via Digitalmars-d wrote: On 20 Jun 2015 11:35, Martin Nowak via Digitalmars-d digitalmars-d@puremagic.com mailto:digitalmars-d@puremagic.com wrote: On 06/20/2015 09:52 AM, Walter Bright wrote: Which C runtime are you using? The math functions in C runtimes are often inadequate. So we now call the host's C library to perform constant folding/CTFE of exp? This is a point for Iain's proposal to use a high precision floating point implementation in the compiler to avoid machine/platform differences. I thought dmd used the C library implementations for all intrinsics via CTFE, not just exp? In any case, isn't the problem you are seeing related to the conversion of string to float? Which again dmd is at the mercy of the C library implementation to handle correctly. The code is the following, in lexer.c: errno = 0; (void)Port::strtod((char *)stringbuffer.data, NULL); if (errno == ERANGE) Ah yes, that brings back memories of fixing GDC to set errno = ERANGE if casting the result from 160bits down returns Infinity. Can we make this a boolean check removing the need for me to explicitly set errno?
Re: Anyone knows this issue? std/math.d(2759): Error: number '0x1p-1024' is not representable
On 6/20/15 2:30 AM, Martin Nowak wrote: On 06/19/2015 06:54 PM, Andrei Alexandrescu wrote: This is happening on my older CentOS as well, it's a problem in the C library. I suggest we just version those tests. -- Andrei Even more worrisome, I don't understand why dmd is compiling the unittests at all. I think it's just tokenizing them that triggers the errors. -- Andrei
Re: Anyone knows this issue? std/math.d(2759): Error: number '0x1p-1024' is not representable
On 6/20/2015 5:16 AM, Martin Nowak wrote: On 06/20/2015 02:00 PM, Martin Nowak wrote: On 06/20/2015 12:06 PM, Walter Bright wrote: On 6/20/2015 3:06 AM, Walter Bright wrote: Apparently we need to roll our own version of strtod() and put it in Port. We already do our own strtold(), so it should be straightforward. Thanks, I'll work on a fix. I can't use the strtold_dm from the backend, b/c it's not available for the other compilers, and AFAIU licensing doesn't allow me to move it to the frontend. Ironically, this just appeared: https://www.reddit.com/r/programming/comments/3aif1a/visual_c_strtod_still_broken/
Re: Anyone knows this issue? std/math.d(2759): Error: number '0x1p-1024' is not representable
On 20 June 2015 at 14:03, Martin Nowak via Digitalmars-d digitalmars-d@puremagic.com wrote: On 06/20/2015 01:35 PM, Iain Buclaw via Digitalmars-d wrote: Can we make this a boolean check removing the need for me to explicitly set errno? The function returns a float, so errno seems like the better choice. It's at this point which I say it's a longdouble for GDC. Incase I need to remind people that there are no uses of native FP whatsoever in my compiler. :-)
Re: Anyone knows this issue? std/math.d(2759): Error: number '0x1p-1024' is not representable
On 6/20/2015 5:16 AM, Martin Nowak wrote: I can't use the strtold_dm from the backend, b/c it's not available for the other compilers, and AFAIU licensing doesn't allow me to move it to the frontend. Hmm. Perhaps google for similar code that would be available? Or submit a bug report to the Debian folks?
Re: Anyone knows this issue? std/math.d(2759): Error: number '0x1p-1024' is not representable
On Saturday, 20 June 2015 at 17:04:48 UTC, Walter Bright wrote: On 6/20/2015 5:16 AM, Martin Nowak wrote: I can't use the strtold_dm from the backend, b/c it's not available for the other compilers, and AFAIU licensing doesn't allow me to move it to the frontend. Hmm. Perhaps google for similar code that would be available? Might work, as the issue dates back to 2013, I'd say we skip it for this release and fix it independently.
Re: Anyone knows this issue? std/math.d(2759): Error: number '0x1p-1024' is not representable
On Friday, 19 June 2015 at 16:19:05 UTC, Martin Nowak wrote: I'm getting this error while trying to build the relase. std/math.d(2759): Error: number '0x1p-1024' is not representable I'm getting that while compiling a -m32 (X86) static release phobos library, but I can't reproduce the error using the same command. https://github.com/D-Programming-Language/phobos/blob/abb21ab9bcd3857f1567bfde11c2d5d385f5caf0/std/math.d#L2759 Don't even know why the unittests are compiled. See http://forum.dlang.org/post/mjjfjr$nlk$1...@digitalmars.com and http://forum.dlang.org/post/fbkyfjfofwaoitaxd...@forum.dlang.org
Re: Anyone knows this issue? std/math.d(2759): Error: number '0x1p-1024' is not representable
On 6/19/15 9:18 AM, Martin Nowak wrote: I'm getting this error while trying to build the relase. std/math.d(2759): Error: number '0x1p-1024' is not representable I'm getting that while compiling a -m32 (X86) static release phobos library, but I can't reproduce the error using the same command. https://github.com/D-Programming-Language/phobos/blob/abb21ab9bcd3857f1567bfde11c2d5d385f5caf0/std/math.d#L2759 Don't even know why the unittests are compiled. This is happening on my older CentOS as well, it's a problem in the C library. I suggest we just version those tests. -- Andrei
Anyone knows this issue? std/math.d(2759): Error: number '0x1p-1024' is not representable
I'm getting this error while trying to build the relase. std/math.d(2759): Error: number '0x1p-1024' is not representable I'm getting that while compiling a -m32 (X86) static release phobos library, but I can't reproduce the error using the same command. https://github.com/D-Programming-Language/phobos/blob/abb21ab9bcd3857f1567bfde11c2d5d385f5caf0/std/math.d#L2759 Don't even know why the unittests are compiled.
Re: Anyone knows this issue? std/math.d(2759): Error: number '0x1p-1024' is not representable
On 6/19/15 9:34 AM, John Colvin wrote: On Friday, 19 June 2015 at 16:19:05 UTC, Martin Nowak wrote: I'm getting this error while trying to build the relase. std/math.d(2759): Error: number '0x1p-1024' is not representable I'm getting that while compiling a -m32 (X86) static release phobos library, but I can't reproduce the error using the same command. https://github.com/D-Programming-Language/phobos/blob/abb21ab9bcd3857f1567bfde11c2d5d385f5caf0/std/math.d#L2759 Don't even know why the unittests are compiled. See http://forum.dlang.org/post/mjjfjr$nlk$1...@digitalmars.com and http://forum.dlang.org/post/fbkyfjfofwaoitaxd...@forum.dlang.org Links taken from the new forum interface seem useless. Neither of these go nowhere relevant (nor does the one posted by Martin). I'm using the vertical split layout if that matters. In related news: the fact that I need to point to the post I want to see with Ctrl-Up and Ctrl-Down, and THEN press Enter like an idiot to see it is definitely not working out for me. I understand it saves bandwidth, but if most bandwidth savings come from nobody using the interface then it's no gain to anyone. Please have the message viewer follow the list focus. cc Vladimir Thanks, Andrei
Re: Anyone knows this issue? std/math.d(2759): Error: number '0x1p-1024' is not representable
On Friday, 19 June 2015 at 16:59:20 UTC, Andrei Alexandrescu wrote: On 6/19/15 9:34 AM, John Colvin wrote: On Friday, 19 June 2015 at 16:19:05 UTC, Martin Nowak wrote: [...] See http://forum.dlang.org/post/mjjfjr$nlk$1...@digitalmars.com and http://forum.dlang.org/post/fbkyfjfofwaoitaxd...@forum.dlang.org Links taken from the new forum interface seem useless. Neither of these go nowhere relevant (nor does the one posted by Martin). I'm using the vertical split layout if that matters. This seems to be specific to the vertical-split layout. Works fine in the other 3 layouts.
