Eric Blake [EMAIL PROTECTED] wrote:
According to Eric Blake on 3/30/2008 4:03 PM:
| And just now, I tested cygwin with -O2, and got:
|
| ../../m4/tests/test-strtod.c:149: assertion failed
| ../../m4/tests/test-strtod.c:221: assertion failed
|
| which looks like rounding errors in computing
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
According to Jim Meyering on 4/1/2008 1:02 AM:
| Wow - gnulib's version of strtod really IS inaccurate. The buildbot is
| still showing failures, since 5 * pow (10.0, -1) != 0.5. Any suggestions
| on how to make this test pass? Would it be
Eric Blake ebb9 at byu.net writes:
Done as follows. FLT_EPSILON represents a 2-ulp error for 0.5;
Actually, FLT_EPSILON represents a 2-ulp error for 0.5f, and unless
FLT_EVAL_METHOD0, it's waaay more than 2-ulp of the double 0.5. I also
botched FLT_MIN instead of DBL_MIN, now fixed.
--
Eric Blake wrote:
* doc/posix-headers/math.texi (math.h): Document this.
It's not only Solaris 8 which lacks it. I'm updating the doc.
Also, I prefer not to libel a particular compiler as broken. If we label
every compiler or platform which has a bug as broken, not only every system
will have
Eric Blake wrote:
* tests/test-math.c (main): Test it.
This test fails on IRIX 6.5 with CC=cc -O. The reason is that it inlines
the 'd == d' test. Here is the machine code that the compiler generates:
.entmain
.globl main
main:# 0x0
.frame $sp, 0, $31
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
According to Bruno Haible on 4/1/2008 7:05 PM:
| While I like that you collected the many instances of the function which
| returns a NaN into a single place, I don't like two things here:
|
| 1) POSIX says that math.h defines the macro NAN, but
Eric Blake wrote:
| 1) POSIX says that math.h defines the macro NAN, but also says that it
| expands into a constant expression. We cannot guarantee that. Therefore
| is it better to define the macro or not?
Well, I did document in doc/posix-functions/math.texi that NAN might not
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
According to Bruno Haible on 4/1/2008 7:45 PM:
| Were you planning on backing out the math.in.h changes?
|
| It depends on your vote... If you say no, then no.
Let's leave it in for now: I'm brave enough to make m4 1.4.11 the litmus
test. If
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
According to Bruno Haible on 3/30/2008 6:11 PM:
| Find attached the results on a dozen of platforms.
Thanks.
|
| I integrated these test results into test-strtod.c as comments (since we
don't
| yet have a matrix like in m4/printf.m4), using a
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
According to Eric Blake on 3/31/2008 6:24 AM:
| | - You seem to have seen problems on Solaris 8 that I did not see.
| What can
| | be the reason? Did you use gnulib's signbit and isnan functions? I
| used
| | the ones from the system, or
Eric Blake [EMAIL PROTECTED] wrote:
According to Jim Meyering on 3/30/2008 8:49 AM:
...
Ahh. That makes sense - I was going off the line numbers in gnulib, not
in my bootstrapped copy of coreutils. Maybe the same bootstrap process
that adds those two lines should add a third line?
#line 1
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
According to Jim Meyering on 3/31/2008 11:24 AM:
|
| Actually, strtod is doing the right thing: it produces -0,
| with both the libc version and the gnulib-supplied function.
|
| The trouble is that signbit(result) returns INT_MIN,
| while
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
According to Eric Blake on 3/31/2008 1:22 PM:
| | The trouble is that signbit(result) returns INT_MIN,
| | while signbit(-0.0) returns 1. Both seem to be allowed, so
| | how about this change? It solves the problem for me.
|
| Good catch. Yes,
Eric Blake [EMAIL PROTECTED] wrote:
According to Jim Meyering on 3/31/2008 11:24 AM:
|
| Actually, strtod is doing the right thing: it produces -0,
| with both the libc version and the gnulib-supplied function.
|
| The trouble is that signbit(result) returns INT_MIN,
| while signbit(-0.0)
Eric Blake wrote:
should we raise this as a gcc bug, that when it
does constant-folding optimization of signbit at compile-time
(signbit(-0.0) = 1), it results in a different value than when the macro
is used on a runtime value (signbit(result) = mask off all but the sign bit)?
I don't think
Eric Blake wrote:
The problem was that Solaris 8 does not define NAN anywhere, so the
test failures I saw in the NaN section of test-strtod.c were because
Solaris was correctly parsing NaN, but the test thought that NaN was not
possible.
So, the '#ifdef NAN' needs to go. It's less portable
Eric Blake wrote:
* tests/test-strtod.c: New file.
The file is using C99 syntax; Solaris cc does not grok it. This converts it
to C89 syntax.
2008-03-31 Bruno Haible [EMAIL PROTECTED]
* tests/test-strtod.c (main): Don't use C99 syntax.
--- tests/test-strtod.c.orig2008-04-01
Eric Blake wrote:
Even more thanks! I'm adding the following, which does some more
interpretation of your results.
Thanks. Somehow now I get slightly different results on OSF/1 4.0.
Updated:
2008-03-31 Bruno Haible [EMAIL PROTECTED]
* tests/test-strtod.c (main): Update results for
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
According to Bruno Haible on 3/31/2008 4:04 PM:
| Eric Blake wrote:
| The problem was that Solaris 8 does not define NAN anywhere, so the
| test failures I saw in the NaN section of test-strtod.c were because
| Solaris was correctly parsing NaN, but
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
According to Eric Blake on 3/30/2008 4:03 PM:
| And just now, I tested cygwin with -O2, and got:
|
| ../../m4/tests/test-strtod.c:149: assertion failed
| ../../m4/tests/test-strtod.c:221: assertion failed
|
| which looks like rounding errors in
Eric Blake [EMAIL PROTECTED] wrote:
...
Therefore, I'm asking for opinions - should I apply this patch as is, to
Seems fine that you've applied it, except for a new failure
in coreutils' running of the gnulib strtod test. E.g.,
test-strtod.c:544: assertion failed
test-strtod.c:592:
Jim Meyering [EMAIL PROTECTED] wrote:
Eric Blake [EMAIL PROTECTED] wrote:
...
Therefore, I'm asking for opinions - should I apply this patch as is, to
Seems fine that you've applied it, except for a new failure
in coreutils' running of the gnulib strtod test. E.g.,
The buildbot is also
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
According to Jim Meyering on 3/30/2008 7:31 AM:
|
| The buildbot is also showing a new compile failure:
|
| http://buildbot.proulx.com:9000/
|
| cc1: warnings being treated as errors
Thanks for the report; I'm applying this:
- --
Don't work too
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
According to Jim Meyering on 3/30/2008 7:28 AM:
| Eric Blake [EMAIL PROTECTED] wrote:
| ...
| Therefore, I'm asking for opinions - should I apply this patch as is, to
|
| Seems fine that you've applied it, except for a new failure
| in coreutils'
Eric Blake [EMAIL PROTECTED] wrote:
According to Jim Meyering on 3/30/2008 7:28 AM:
| Eric Blake [EMAIL PROTECTED] wrote:
| ...
| Therefore, I'm asking for opinions - should I apply this patch as is, to
|
| Seems fine that you've applied it, except for a new failure
| in coreutils' running
Hi Eric,
Of all the
systems I tried, I did not find a single one that was fully C99 compliant.
I'd like to document the flaws across more platforms. Do you have a set of
test programs, one snippet for each bug you found so far, that I could run
on a variety of systems? I could try to extract
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
According to Bruno Haible on 3/30/2008 2:44 PM:
| Hi Eric,
|
| Of all the
| systems I tried, I did not find a single one that was fully C99 compliant.
|
| I'd like to document the flaws across more platforms. Do you have a set of
| test programs, one
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
According to Jim Meyering on 3/30/2008 8:49 AM:
|
| In my copy of test-strtod.c, they're all comparing sign bits:
|
| ASSERT (signbit (result1) != signbit (result2));
The default value of NaN (both sign bit and other bits) are enough of a
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
According to Eric Blake on 3/30/2008 4:03 PM:
| I'm wondering if it is a gcc -O2 bug? I'm not quite sure what
| to do here, without access to a machine with those failures. Can you step
| through the replacement function with a debugger?
Or maybe
Eric Blake wrote:
test-strtod.c is all I had; but I intentionally made it catch all errors
rather than aborting on the first
Ah, I see. Sorry, I had not seen this detail.
Find attached the results on a dozen of platforms.
I integrated these test results into test-strtod.c as comments (since
Oops, the shown patch in the last mail was incomplete. See
http://git.sv.gnu.org/gitweb/?p=gnulib.git;a=commitdiff;h=b442459322dbef0bcb547770565b05724e98b47d
for the full patch.
(Sorry, a 'git' pitfall. I had used 'git add' on one file and not on the other.
What is the command to get the union of
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
According to Bruno Haible on 3/30/2008 6:18 PM:
| Oops, the shown patch in the last mail was incomplete. See
|
http://git.sv.gnu.org/gitweb/?p=gnulib.git;a=commitdiff;h=b442459322dbef0bcb547770565b05724e98b47d
| for the full patch.
|
| (Sorry, a
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
According to Eric Blake on 3/29/2008 3:29 PM:
| Therefore, I'm asking for opinions - should I apply this patch as is, to
| at least guarantee 'inf' and 'nan' parsing on all platforms, even though
| it can cause regressions on the accuracy of the
33 matches
Mail list logo
