[COMMITTED] tilegx: enable wordsize-64 support for ieee745 dbl-64.
Commit Message
On 12/23/2014 3:47 PM, Chris Metcalf wrote:
> On 12/23/2014 3:06 PM, Joseph Myers wrote:
>> I take it that the ABI on tilegx32 allows the lround alias to llround
>
> No, it doesn't - I missed this possibility. I assume the best fix is to
> provide an lround override in sysdeps/tile/tilegx/tilegx32/s_lround.c
> that invokes llround() and then sign-extends the result?
>
> But, why doesn't MIPS itself use this approach?
>
> Thanks for catching this!
Actually, it is, of course, a little trickier than that, since
we have to prevent s_llround.c from setting up the alias for
lround() as well. But I think it may make sense to use global
conditional logic to do this. The current users of dbl-64/wordsize-64
are aarch64, alpha, sparc64, and x86_64. So as far as I know, the
question is whether the forthcoming ILP32 version of aarch64 would
benefit from this hack (probably yes) and whether the existing x32
version of x86_64 would (not clear to me, but plausibly better to clean
up the high bits of the returned value). If it's OK to just add
ILP32-conditional logic, we can do use the appended diff.
If not, the simplest thing for me to do would be to just add s_lround.c
and s_llround.c in sysdeps/tile/tilegx/tilegx32 that use the default
versions of dbl-64 and not worry about the slight performance
improvement from using the full 64-bit register size. Or, we could
come up with some way to use a conditional in this code: for example,
just add a random #ifdef, #define it in a tilegx-specific override
of s_llround.c, and #include the standard version explicitly. But
this runs counter to the glibc rule of avoiding #ifdef. I'm not sure what
the cleanest path would be if the use of "#ifdef _LP64" isn't OK.
Comments
On Tue, Dec 23, 2014 at 1:11 PM, Chris Metcalf <cmetcalf@ezchip.com> wrote:
> On 12/23/2014 3:47 PM, Chris Metcalf wrote:
>>
>> On 12/23/2014 3:06 PM, Joseph Myers wrote:
>>>
>>> I take it that the ABI on tilegx32 allows the lround alias to llround
>>
>>
>> No, it doesn't - I missed this possibility. I assume the best fix is to
>> provide an lround override in sysdeps/tile/tilegx/tilegx32/s_lround.c
>> that invokes llround() and then sign-extends the result?
>>
>> But, why doesn't MIPS itself use this approach?
>>
>> Thanks for catching this!
>
>
> Actually, it is, of course, a little trickier than that, since
> we have to prevent s_llround.c from setting up the alias for
> lround() as well. But I think it may make sense to use global
> conditional logic to do this. The current users of dbl-64/wordsize-64
> are aarch64, alpha, sparc64, and x86_64. So as far as I know, the
> question is whether the forthcoming ILP32 version of aarch64 would
> benefit from this hack (probably yes) and whether the existing x32
> version of x86_64 would (not clear to me, but plausibly better to clean
> up the high bits of the returned value). If it's OK to just add
> ILP32-conditional logic, we can do use the appended diff.
>
> If not, the simplest thing for me to do would be to just add s_lround.c
> and s_llround.c in sysdeps/tile/tilegx/tilegx32 that use the default
> versions of dbl-64 and not worry about the slight performance
> improvement from using the full 64-bit register size. Or, we could
> come up with some way to use a conditional in this code: for example,
> just add a random #ifdef, #define it in a tilegx-specific override
> of s_llround.c, and #include the standard version explicitly. But
> this runs counter to the glibc rule of avoiding #ifdef. I'm not sure what
> the cleanest path would be if the use of "#ifdef _LP64" isn't OK.
>
> --- a/sysdeps/ieee754/dbl-64/wordsize-64/s_llround.c
> +++ b/sysdeps/ieee754/dbl-64/wordsize-64/s_llround.c
> @@ -64,16 +64,31 @@ __llround (double x)
>
> weak_alias (__llround, llround)
> #ifdef NO_LONG_DOUBLE
> -strong_alias (__llround, __lroundl)
> -weak_alias (__llround, lroundl)
> +strong_alias (__llround, __llroundl)
> +weak_alias (__llround, llroundl)
> #endif
>
> -/* long has the same width as long long on 64-bit machines. */
> +/* long has the same width as long long on 64-bit machines.
> + If building for ILP32 on a machine with 64-bit registers, however,
> + do an explicit cast on the return value instead. */
> #undef lround
> #undef __lround
> +#ifdef _LP64
> strong_alias (__llround, __lround)
> weak_alias (__llround, lround)
> -#ifdef NO_LONG_DOUBLE
> -strong_alias (__llround, __llroundl)
> -weak_alias (__llround, llroundl)
> +# ifdef NO_LONG_DOUBLE
> +strong_alias (__llround, __lroundl)
> +weak_alias (__llround, lroundl)
> +# endif
> +#else
> +long int
> +__lround (double x)
> +{
> + return __llround (x);
> +}
> +weak_alias (__lround, lround)
> +# ifdef NO_LONG_DOUBLE
> +strong_alias (__lround, __lroundl)
> +weak_alias (__lround, lroundl)
> +# endif
> #endif
>
Please make sure that x32 version is unchanged where
lround is an alias of llround.
@@ -64,16 +64,31 @@ __llround (double x)
weak_alias (__llround, llround)
#ifdef NO_LONG_DOUBLE
-strong_alias (__llround, __lroundl)
-weak_alias (__llround, lroundl)
+strong_alias (__llround, __llroundl)
+weak_alias (__llround, llroundl)
#endif
-/* long has the same width as long long on 64-bit machines. */
+/* long has the same width as long long on 64-bit machines.
+ If building for ILP32 on a machine with 64-bit registers, however,
+ do an explicit cast on the return value instead. */
#undef lround
#undef __lround
+#ifdef _LP64
strong_alias (__llround, __lround)
weak_alias (__llround, lround)
-#ifdef NO_LONG_DOUBLE
-strong_alias (__llround, __llroundl)
-weak_alias (__llround, llroundl)
+# ifdef NO_LONG_DOUBLE
+strong_alias (__llround, __lroundl)
+weak_alias (__llround, lroundl)
+# endif
+#else
+long int
+__lround (double x)
+{
+ return __llround (x);
+}
+weak_alias (__lround, lround)
+# ifdef NO_LONG_DOUBLE
+strong_alias (__lround, __lroundl)
+weak_alias (__lround, lroundl)
+# endif
#endif