Re: [PATCH] utimens: fix broken m4 test
Eric Blake wrote: > * m4/utimens.m4 (gl_UTIMENS): Include correct headers. ... > diff --git a/m4/utimens.m4 b/m4/utimens.m4 ... > +#include Thanks. That fixed it for me, too.
Re: icc bug on test-frexp-nolibm
On 11/03/2010 05:10 PM, Eric Blake wrote:
> I'm running out of time today to get to the bottom of this, so if anyone
> else wants to help out, great. If not, I'll see if I can resume testing
> tomorrow and get a minimal test case that exposes the compiler bug, as
> well as try and isolate ways to work around it.
>
> I'm on a Linux x86_64 machine with icc (ICC) 10.0 20070426.
>
> I hate it when a heisenbug disappears because of the mere act of adding
> debugging hooks. :(
I've reduced it to a minimal testcase:
$ cat foo.c
#define _GNU_SOURCE 1
#include
#include
double zero;
double bar (void) { return -zero; }
int main (void)
{
assert (!signbit (zero));
assert (signbit (bar ()));
return 0;
}
$ icc -g -o foo foo.c
$ ./foo
$ icc-o foo foo.c
$ ./foo
foo: foo.c:9: main: Assertion `(( sizeof( bar () ) > sizeof( double )) ?
__signbitl( (long double)(bar ()) ) : (( sizeof( bar () ) == sizeof(
float )) ? __signbitf( (float)(bar ()) ) : __signbit( (double)(bar ())
)))' failed.
Aborted (core dumped)
$
Looks like icc with debugging computes -0.0 correctly, but without -g it
makes a mistaken optimization. Comparing the assembly, I see:
with -g:
Dump of assembler code for function bar:
0x004005f8 <+0>: push %rbp
0x004005f9 <+1>: mov%rsp,%rbp
0x004005fc <+4>: sub$0x10,%rsp
0x00400600 <+8>: fldl 0x20087a(%rip)# 0x600e80
0x00400606 <+14>:fchs
0x00400608 <+16>:fstpl -0x10(%rbp)
0x0040060b <+19>:movsd -0x10(%rbp),%xmm0
0x00400610 <+24>:leaveq
0x00400611 <+25>:retq
without -g:
Dump of assembler code for function bar:
0x00400b3a <+0>: pxor %xmm0,%xmm0
0x00400b3e <+4>: subsd 0x203542(%rip),%xmm0#
0x604088
0x00400b46 <+12>:retq
0x00400b47 <+13>:nop
Gnulib's testsuite already has several workarounds for computing -0.0 in
spite of buggy compilers, so I think it's just a matter of making all
the tests uniformly use those same workarounds.
In particular, I tested that this works for icc regardless of whether
CFLAGS includes -g (borrowing from test-ceill.c's workarounds for -0.0L
as my starting point):
#if defined __hpux || defined __sgi || defined __ICC
static double
compute_minus_zero (void)
{
return -DBL_MIN * DBL_MIN;
}
# define minus_zero compute_minus_zero ()
#else
double minus_zero = -0.0;
#endif
--
Eric Blake ebl...@redhat.com+1-801-349-2682
Libvirt virtualization library http://libvirt.org
signature.asc
Description: OpenPGP digital signature
[PATCH 2/2] ceil, floor, round, trunc: enhance tests of -0
* tests/test-ceilf1.c (main): Ensure correct sign of result.
* tests/test-ceill.c (main): Likewise.
* tests/test-floorf1.c (main): Likewise.
* tests/test-floorl.c (main): Likewise.
* tests/test-round1.c (main): Likewise.
* tests/test-roundf1.c (main): Likewise.
* tests/test-roundl.c (main): Likewise.
* tests/test-trunc1.c (main): Likewise.
* tests/test-truncf1.c (main): Likewise.
* tests/test-truncl.c (main): Likewise.
Signed-off-by: Eric Blake
---
In my minus_zero cleanups, I noticed that these tests were rather
weak; it's possible that there are implementations that return 0.0
when they should be returning -0.0, so let's test that explicitly.
The new checks should work even on systems that lack -0.0 (if there
are any; I'm not convinced that SGI really lacks -0.0 on MIPS, so
much as has a buggy compiler that doesn't expose it well, but I
haven't recently logged on to an Irix machine to test this).
ChangeLog| 12
modules/ceilf-tests |1 +
tests/test-ceilf1.c |2 ++
tests/test-ceill.c |2 ++
tests/test-floorf1.c |2 ++
tests/test-floorl.c |2 ++
tests/test-round1.c |2 ++
tests/test-roundf1.c |2 ++
tests/test-roundl.c |2 ++
tests/test-trunc1.c |2 ++
tests/test-truncf1.c |2 ++
tests/test-truncl.c |2 ++
12 files changed, 33 insertions(+), 0 deletions(-)
diff --git a/ChangeLog b/ChangeLog
index 9aca76a..01a4b39 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,5 +1,17 @@
2010-11-04 Eric Blake
+ ceil, floor, round, trunc: enhance tests of -0
+ * tests/test-ceilf1.c (main): Ensure correct sign of result.
+ * tests/test-ceill.c (main): Likewise.
+ * tests/test-floorf1.c (main): Likewise.
+ * tests/test-floorl.c (main): Likewise.
+ * tests/test-round1.c (main): Likewise.
+ * tests/test-roundf1.c (main): Likewise.
+ * tests/test-roundl.c (main): Likewise.
+ * tests/test-trunc1.c (main): Likewise.
+ * tests/test-truncf1.c (main): Likewise.
+ * tests/test-truncl.c (main): Likewise.
+
frexp, tests: work around ICC bug with -zero
* m4/frexp.m4 (gl_FUNC_FREXP_WORKS): Compute -0.0 in a way that
works with more compilers.
diff --git a/modules/ceilf-tests b/modules/ceilf-tests
index 040fd1b..a0472a7 100644
--- a/modules/ceilf-tests
+++ b/modules/ceilf-tests
@@ -9,6 +9,7 @@ tests/macros.h
Depends-on:
float
isnanf-nolibm
+signbit
stdbool
stdint
diff --git a/tests/test-ceilf1.c b/tests/test-ceilf1.c
index 86bb49e..4e38247 100644
--- a/tests/test-ceilf1.c
+++ b/tests/test-ceilf1.c
@@ -33,7 +33,9 @@ main ()
{
/* Zero. */
ASSERT (ceilf (0.0f) == 0.0f);
+ ASSERT (!signbit (ceilf (0.0f)));
ASSERT (ceilf (minus_zerof) == 0.0f);
+ ASSERT (!!signbit (minus_zerof) == !!signbit (ceilf (minus_zerof)));
/* Positive numbers. */
ASSERT (ceilf (0.3f) == 1.0f);
ASSERT (ceilf (0.7f) == 1.0f);
diff --git a/tests/test-ceill.c b/tests/test-ceill.c
index cac45b7..f627a9e 100644
--- a/tests/test-ceill.c
+++ b/tests/test-ceill.c
@@ -40,7 +40,9 @@ main ()
/* Zero. */
ASSERT (ceill (0.0L) == 0.0L);
+ ASSERT (!signbit (ceill (0.0L)));
ASSERT (ceill (minus_zerol) == 0.0L);
+ ASSERT (!!signbit (minus_zerol) == !!signbit (ceill (minus_zerol)));
/* Positive numbers. */
ASSERT (ceill (0.3L) == 1.0L);
ASSERT (ceill (0.7L) == 1.0L);
diff --git a/tests/test-floorf1.c b/tests/test-floorf1.c
index 39261d8..f0f7716 100644
--- a/tests/test-floorf1.c
+++ b/tests/test-floorf1.c
@@ -33,7 +33,9 @@ main ()
{
/* Zero. */
ASSERT (floorf (0.0f) == 0.0f);
+ ASSERT (!signbit (floorf (0.0f)));
ASSERT (floorf (minus_zerof) == 0.0f);
+ ASSERT (!!signbit (minus_zerof) == !!signbit (floorf (minus_zerof)));
/* Positive numbers. */
ASSERT (floorf (0.3f) == 0.0f);
ASSERT (floorf (0.7f) == 0.0f);
diff --git a/tests/test-floorl.c b/tests/test-floorl.c
index 84717f8..2675a9e 100644
--- a/tests/test-floorl.c
+++ b/tests/test-floorl.c
@@ -40,7 +40,9 @@ main ()
/* Zero. */
ASSERT (floorl (0.0L) == 0.0L);
+ ASSERT (!signbit (floorl (0.0L)));
ASSERT (floorl (minus_zerol) == 0.0L);
+ ASSERT (!!signbit (minus_zerol) == !!signbit (floorl (minus_zerol)));
/* Positive numbers. */
ASSERT (floorl (0.3L) == 0.0L);
ASSERT (floorl (0.7L) == 0.0L);
diff --git a/tests/test-round1.c b/tests/test-round1.c
index 86fd1de..4711a74 100644
--- a/tests/test-round1.c
+++ b/tests/test-round1.c
@@ -35,7 +35,9 @@ main ()
{
/* Zero. */
ASSERT (round (0.0) == 0.0);
+ ASSERT (!signbit (round (0.0)));
ASSERT (round (minus_zerod) == 0.0);
+ ASSERT (!!signbit (minus_zerod) == !!signbit (round (minus_zerod)));
/* Positive numbers. */
ASSERT (round (0.3) == 0.0);
ASSERT (round (0.5) == 1.0);
diff --git a/tests/test-roundf1.c b/tests/test-roundf1.c
index 33609bb..354dcf0 100644
--- a/tests/test-roundf1.c
+++ b/tests/test-roundf1.c
@@ -35,7 +35,9 @@ main ()
{
/* Zero. */
ASSERT (roundf (0.0f) == 0.0f);
+ ASSERT (!sig
[PATCH 1/2] frexp, tests: work around ICC bug with -zero
* m4/frexp.m4 (gl_FUNC_FREXP_WORKS): Compute -0.0 in a way that works with more compilers. * tests/minus-zero.h: New file. * modules/ceilf-tests (Files): Include it. * modules/ceill-tests (Files): Likewise. * modules/floorf-tests (Files): Likewise. * modules/floorl-tests (Files): Likewise. * modules/frexp-nolibm-tests (Files): Likewise. * modules/frexp-tests (Files): Likewise. * modules/frexpl-nolibm-tests (Files): Likewise. * modules/frexpl-tests (Files): Likewise. * modules/isnan-tests (Files): Likewise. * modules/isnand-nolibm-tests (Files): Likewise. * modules/isnand-tests (Files): Likewise. * modules/isnanf-nolibm-tests (Files): Likewise. * modules/isnanf-tests (Files): Likewise. * modules/isnanl-nolibm-tests (Files): Likewise. * modules/isnanl-tests (Files): Likewise. * modules/round-tests (Files): Likewise. * modules/roundf-tests (Files): Likewise. * modules/roundl-tests (Files): Likewise. * modules/ldexpl-tests (Files): Likewise. * modules/signbit-tests (Files): Likewise. * modules/snprintf-posix-tests (Files): Likewise. * modules/sprintf-posix-tests (Files): Likewise. * modules/strtod-tests (Files): Likewise. * modules/trunc-tests (Files): Likewise. * modules/truncf-tests (Files): Likewise. * modules/truncl-tests (Files): Likewise. * modules/vsnprintf-posix-tests (Files): Likewise. * modules/vsprintf-posix-tests (Files): Likewise. * modules/vasnprintf-posix-tests (Files): Likewise. * modules/vasprintf-posix-tests (Files): Likewise. * tests/test-ceilf1.c (main): Use it. * tests/test-ceill.c (main): Likewise. * tests/test-floorf1.c (main): Likewise. * tests/test-floorl.c (main): Likewise. * tests/test-frexp.c (main): Likewise. * tests/test-frexpl.c (main): Likewise. * tests/test-isnan.c (main): Likewise. * tests/test-isnand.h (main): Likewise. * tests/test-isnanf.h (main): Likewise. * tests/test-isnanl.h (main): Likewise. * tests/test-ldexpl.c (main): Likewise. * tests/test-round.c (main): Likewise. * tests/test-roundf.c (main): Likewise. * tests/test-roundl.c (main): Likewise. * tests/test-signbit.c (test_signbitf, test_signbitd) (test_signbitl): Likewise. * tests/test-snprintf-posix.h (test_function): Likewise. * tests/test-sprintf-posix.h (test_function): Likewise. * tests/test-strtod.c (main): Likewise. * tests/test-trunc1.c (main): Likewise. * tests/test-truncf1.c (main): Likewise. * tests/test-truncl.c (main): Likewise. Signed-off-by: Eric Blake --- Refactors everything so that future compiler workarounds necessary to get at -0.0 can be limited to a single test file (.m4 files may need more efforts, though). Tested that all of these tests still pass with gcc; there are still some ICC failures, but I think they are due to NaN issues and not -0.0 issues (at any rate, I'll probably have more patches coming up later). ChangeLog | 59 + m4/frexp.m4| 18 -- modules/ceilf-tests|1 + modules/ceill-tests|1 + modules/floorf-tests |1 + modules/floorl-tests |1 + modules/frexp-nolibm-tests |2 + modules/frexp-tests|1 + modules/frexpl-nolibm-tests|2 + modules/frexpl-tests |1 + modules/isnan-tests|1 + modules/isnand-nolibm-tests|1 + modules/isnand-tests |1 + modules/isnanf-nolibm-tests|1 + modules/isnanf-tests |1 + modules/isnanl-nolibm-tests|1 + modules/isnanl-tests |1 + modules/ldexpl-tests |1 + modules/round-tests|1 + modules/roundf-tests |1 + modules/roundl-tests |1 + modules/signbit-tests |1 + modules/snprintf-posix-tests |1 + modules/sprintf-posix-tests|1 + modules/strtod-tests |1 + modules/trunc-tests|1 + modules/truncf-tests |1 + modules/truncl-tests |1 + modules/vasnprintf-posix-tests |1 + modules/vasprintf-posix-tests |1 + modules/vsnprintf-posix-tests |1 + modules/vsprintf-posix-tests |1 + tests/minus-zero.h | 70 tests/test-ceilf1.c|7 +--- tests/test-ceill.c | 19 +-- tests/test-floorf1.c |7 +--- tests/test-floorl.c| 19 +-- tests/test-frexp.c |7 +--- tests/test-frexpl.c| 19 +-- tests/test-isnan.c | 29 ++--- tests/test-isnand.h|7 +--- tests/test-isnanf.h|7 +--- tests/test-isnanl.h| 19 +-- tests/test-ldexpl.c| 19 +-- tests/test-round1.c|7 +--- tests/test-roundf1.c |7 +--- tests/test-roundl.c| 19 +-- tests/test-signbit.c | 35 tests/test-snprintf-posix.h| 33 --
[PATCH] isnanl: work around icc bug
* lib/isnan.c (FUNC): Compute run-time NaN under ICC as well.
Signed-off-by: Eric Blake
---
On a Linux x86_64 system where isnan() fails to recognize pseudo-nan
long doubles, ICC 10.1 fails to compile isnan.c with:
../../gllib/isnan.c(133): error: floating-point operation result is out of range
static memory_double nan = { L_(0.0) / L_(0.0) };
^
../../gllib/isnan.c(134): error: floating-point operation result is out of range
static DOUBLE plus_inf = L_(1.0) / L_(0.0);
^
../../gllib/isnan.c(135): error: floating-point operation result is out of range
static DOUBLE minus_inf = -L_(1.0) / L_(0.0);
ChangeLog |5 -
lib/isnan.c |4 ++--
2 files changed, 6 insertions(+), 3 deletions(-)
diff --git a/ChangeLog b/ChangeLog
index 68c0a75..7f1a824 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,4 +1,7 @@
-2010-11-03 Eric Blake
+2010-11-04 Eric Blake
+
+ isnan: work around icc bug
+ * lib/isnan.c (FUNC): Compute run-time NaN under ICC as well.
tests: fix compiler warnings
* tests/test-getopt.h (test_getopt): Fix condition.
diff --git a/lib/isnan.c b/lib/isnan.c
index ddad7fb..d183ac4 100644
--- a/lib/isnan.c
+++ b/lib/isnan.c
@@ -118,8 +118,8 @@ FUNC (DOUBLE x)
/* Be careful to not do any floating-point operation on x, such as x == x,
because x may be a signaling NaN. */
# if defined __TINYC__ || defined __SUNPRO_C || defined __DECC \
- || (defined __sgi && !defined __GNUC__)
- /* The Sun C 5.0 compilers and the Compaq (ex-DEC) 6.4 compilers don't
+ || (defined __sgi && !defined __GNUC__) || defined __ICC
+ /* The Sun C 5.0, Intel ICC 10.0, and Compaq (ex-DEC) 6.4 compilers don't
recognize the initializers as constant expressions. The latter compiler
also fails when constant-folding 0.0 / 0.0 even when constant-folding is
not required. The SGI MIPSpro C compiler complains about "floating-point
--
1.7.3.2
Re: [PATCH 2/2] ceil, floor, round, trunc: enhance tests of -0
Eric Blake writes: > In my minus_zero cleanups, I noticed that these tests were rather > weak; it's possible that there are implementations that return 0.0 > when they should be returning -0.0, so let's test that explicitly. > The new checks should work even on systems that lack -0.0 (if there > are any; I'm not convinced that SGI really lacks -0.0 on MIPS, so > much as has a buggy compiler that doesn't expose it well, but I > haven't recently logged on to an Irix machine to test this). This looks good to me, for the "round" modules that I maintain. -- Ben Pfaff http://benpfaff.org
Re: [PATCH 1/2] frexp, tests: work around ICC bug with -zero
Eric Blake writes: > Refactors everything so that future compiler workarounds necessary > to get at -0.0 can be limited to a single test file (.m4 files may > need more efforts, though). Tested that all of these tests still > pass with gcc; there are still some ICC failures, but I think they > are due to NaN issues and not -0.0 issues (at any rate, I'll probably > have more patches coming up later). This looks fine to me for the "round" and "isnan" modules that I maintain. -- Ben Pfaff http://benpfaff.org
