Fix x86/x86_64 expm1l spurious underflow exceptions (bug 16539)
Commit Message
This patch fixes bug 16539, spurious underflow exceptions from x86 /
x86-64 expm1l. The problem is that the computation of a base-2
exponent with extra precision involves spurious underflows for
arguments that are small but not subnormal, so a check is added to
just return the argument in those cases. (If the argument *is*
subnormal, underflowing is correct and the existing code will always
underflow, so it suffices to keep using the existing code in that
case; some expm1 implementations have a bug (bug 16353) with missing
underflow exceptions, but I don't think there's such a bug in this
particular version.)
Tested x86_64 and x86; no ulps updates needed.
(auto-libm-test-out diffs omitted below.)
2014-06-24 Joseph Myers <joseph@codesourcery.com>
[BZ #16539]
* sysdeps/i386/fpu/e_expl.S (IEEE754_EXPL) [USE_AS_EXPM1L]: Just
return the argument for normal arguments with exponent below -64.
* sysdeps/x86_64/fpu/e_expl.S (IEEE754_EXPL) [USE_AS_EXPM1L]:
Likewise.
* math/auto-libm-test-in: Add another test of expm1.
* math/auto-libm-test-out: Regenerated.
Comments
On 06/24/2014 10:47 PM, Joseph S. Myers wrote:
> This patch fixes bug 16539, spurious underflow exceptions from x86 /
> x86-64 expm1l. The problem is that the computation of a base-2
> exponent with extra precision involves spurious underflows for
> arguments that are small but not subnormal, so a check is added to
> just return the argument in those cases. (If the argument *is*
> subnormal, underflowing is correct and the existing code will always
> underflow, so it suffices to keep using the existing code in that
> case; some expm1 implementations have a bug (bug 16353) with missing
> underflow exceptions, but I don't think there's such a bug in this
> particular version.)
Thanks,
Andreas
Hi Joseph,
on S390 test-ldouble fails after this patch:
Failure: expm1 (0x4.0000000000000028p-16384): Exception "Underflow" set
Failure: expm1_downward (0x4.0000000000000028p-16384): Exception
"Underflow" set
Failure: expm1_towardzero (0x4.0000000000000028p-16384): Exception
"Underflow" set
Failure: expm1_upward (0x4.0000000000000028p-16384): Exception
"Underflow" set
But I haven´t debugged it yet.
Bye
On 06/24/2014 10:47 PM, Joseph S. Myers wrote:
> This patch fixes bug 16539, spurious underflow exceptions from x86 /
> x86-64 expm1l. The problem is that the computation of a base-2
> exponent with extra precision involves spurious underflows for
> arguments that are small but not subnormal, so a check is added to
> just return the argument in those cases. (If the argument *is*
> subnormal, underflowing is correct and the existing code will always
> underflow, so it suffices to keep using the existing code in that
> case; some expm1 implementations have a bug (bug 16353) with missing
> underflow exceptions, but I don't think there's such a bug in this
> particular version.)
>
> Tested x86_64 and x86; no ulps updates needed.
>
> (auto-libm-test-out diffs omitted below.)
>
> 2014-06-24 Joseph Myers <joseph@codesourcery.com>
>
> [BZ #16539]
> * sysdeps/i386/fpu/e_expl.S (IEEE754_EXPL) [USE_AS_EXPM1L]: Just
> return the argument for normal arguments with exponent below -64.
> * sysdeps/x86_64/fpu/e_expl.S (IEEE754_EXPL) [USE_AS_EXPM1L]:
> Likewise.
> * math/auto-libm-test-in: Add another test of expm1.
> * math/auto-libm-test-out: Regenerated.
>
> diff --git a/math/auto-libm-test-in b/math/auto-libm-test-in
> index 4eaa013..0826825 100644
> --- a/math/auto-libm-test-in
> +++ b/math/auto-libm-test-in
> @@ -960,6 +960,8 @@ expm1 0x1p-64
> expm1 -0x1p-64
> expm1 0x1p-100
> expm1 -0x1p-100
> +# Bug 16353: underflow exception may be missing
> +expm1 0x4.0000000000000028p-16384 missing-underflow
>
> fma 1.0 2.0 3.0
> fma 1.25 0.75 0.0625
> diff --git a/sysdeps/i386/fpu/e_expl.S b/sysdeps/i386/fpu/e_expl.S
> index eb4086b..c7e4373 100644
> --- a/sysdeps/i386/fpu/e_expl.S
> +++ b/sysdeps/i386/fpu/e_expl.S
> @@ -108,6 +108,16 @@ ENTRY(IEEE754_EXPL)
> andb %ah, %dh
> cmpb $0x40, %dh
> je 2f
> +
> + /* Test for arguments that are small but not subnormal. */
> + movzwl 4+8(%esp), %eax
> + andl $0x7fff, %eax
> + cmpl $0x3fbf, %eax
> + jge 3f
> + /* Argument's exponent below -64; avoid spurious underflow if
> + normal. */
> + cmpl $0x0001, %eax
> + jge 2f
> #else
> movzwl 4+8(%esp), %eax
> andl $0x7fff, %eax
> diff --git a/sysdeps/x86_64/fpu/e_expl.S b/sysdeps/x86_64/fpu/e_expl.S
> index 338527b..0ebe388 100644
> --- a/sysdeps/x86_64/fpu/e_expl.S
> +++ b/sysdeps/x86_64/fpu/e_expl.S
> @@ -105,6 +105,16 @@ ENTRY(IEEE754_EXPL)
> andb %ah, %dh
> cmpb $0x40, %dh
> je 2f
> +
> + /* Test for arguments that are small but not subnormal. */
> + movzwl 8+8(%rsp), %eax
> + andl $0x7fff, %eax
> + cmpl $0x3fbf, %eax
> + jge 3f
> + /* Argument's exponent below -64; avoid spurious underflow if
> + normal. */
> + cmpl $0x0001, %eax
> + jge 2f
> #else
> movzwl 8+8(%rsp), %eax
> andl $0x7fff, %eax
>
@@ -960,6 +960,8 @@ expm1 0x1p-64
expm1 -0x1p-64
expm1 0x1p-100
expm1 -0x1p-100
+# Bug 16353: underflow exception may be missing
+expm1 0x4.0000000000000028p-16384 missing-underflow
fma 1.0 2.0 3.0
fma 1.25 0.75 0.0625
@@ -108,6 +108,16 @@ ENTRY(IEEE754_EXPL)
andb %ah, %dh
cmpb $0x40, %dh
je 2f
+
+ /* Test for arguments that are small but not subnormal. */
+ movzwl 4+8(%esp), %eax
+ andl $0x7fff, %eax
+ cmpl $0x3fbf, %eax
+ jge 3f
+ /* Argument's exponent below -64; avoid spurious underflow if
+ normal. */
+ cmpl $0x0001, %eax
+ jge 2f
#else
movzwl 4+8(%esp), %eax
andl $0x7fff, %eax
@@ -105,6 +105,16 @@ ENTRY(IEEE754_EXPL)
andb %ah, %dh
cmpb $0x40, %dh
je 2f
+
+ /* Test for arguments that are small but not subnormal. */
+ movzwl 8+8(%rsp), %eax
+ andl $0x7fff, %eax
+ cmpl $0x3fbf, %eax
+ jge 3f
+ /* Argument's exponent below -64; avoid spurious underflow if
+ normal. */
+ cmpl $0x0001, %eax
+ jge 2f
#else
movzwl 8+8(%rsp), %eax
andl $0x7fff, %eax