Patchwork Sparc exp(), expf() performance improvement

login
register
mail settings
Submitter Patrick McGehearty
Date July 31, 2017, 7:39 p.m.
Message ID <1501529969-96949-1-git-send-email-patrick.mcgehearty@oracle.com>
Download mbox | patch
Permalink /patch/21827/
State New
Headers show

Comments

David Miller - July 31, 2017, 7:47 p.m.
From: Patrick McGehearty <patrick.mcgehearty@oracle.com>
Date: Mon, 31 Jul 2017 15:39:29 -0400

> This PATCH is intended to improve exp() and expf() performance on Sparc.
> These changes will only be active on Sparc platforms and only for
> those platforms that support HWCAP_SPARC_CRYPTO (niagara4 and later).

Can you explain which instructions exactly help make the compiled
C code for exp() and expf() faster instead of being vague like
this?

Wouldn't the new C code you are adding be faster on other CPUs as
well, even without gcc generating instructions for Niagara 4 and
later?

Thank you.
Carlos O'Donell - July 31, 2017, 8:12 p.m.
On 07/31/2017 03:47 PM, David Miller wrote:
> From: Patrick McGehearty <patrick.mcgehearty@oracle.com>
> Date: Mon, 31 Jul 2017 15:39:29 -0400
> 
>> This PATCH is intended to improve exp() and expf() performance on Sparc.
>> These changes will only be active on Sparc platforms and only for
>> those platforms that support HWCAP_SPARC_CRYPTO (niagara4 and later).
> 
> Can you explain which instructions exactly help make the compiled
> C code for exp() and expf() faster instead of being vague like
> this?
> 
> Wouldn't the new C code you are adding be faster on other CPUs as
> well, even without gcc generating instructions for Niagara 4 and
> later?


... I would also like to see the results of the glibc microbenchmark
*before* and *after* the patches. We have *specific* microbenchmarks
for lots of math functions.
Joseph Myers - July 31, 2017, 9:06 p.m.
On Mon, 31 Jul 2017, Patrick McGehearty wrote:

>     * sysdeps/sparc/configure.ac: manage -mcpu=niagara4 test
>     * sysdeps/sparc/configure: Regenerate

As far as I can see, niagara4 support was added to GCC in r178554 
(2011-09-05), so would have been in GCC 4.7.  The minimum GCC version for 
building glibc is 4.9.  So this configure test, and any associated 
conditionals in Makefiles or elsewhere, should not be necessary.

> diff --git a/sysdeps/sparc/fpu/libm_endian.h b/sysdeps/sparc/fpu/libm_endian.h
> new file mode 100644
> index 0000000..a49abcd
> --- /dev/null
> +++ b/sysdeps/sparc/fpu/libm_endian.h
> @@ -0,0 +1,32 @@
> +/*
> +   Contributed by Oracle Inc.
> +   Copyright (C) 2017 Free Software Foundation, Inc.

All new files should start with a descriptive comment on their first line, 
before the copyright notice.  "Contributed by" is no longer used; 
significant contributions can be mentioned in the NEWS file (as well as in 
contrib.texi).

> +extern double exp (double);

No such declaration should be needed, since you're including <math.h>.

> +extern double __ieee754_exp (double);

And this should come through math_private.h.

> +/*
> + * For i = 0, ..., 66,
> + *   TBL2[2*i] is a double precision number near (i+1)*2^-6, and

We don't use leading '*' on each line of a comment.

> +static const double zC[] = {
> +  0.5,
> +  4.61662413084468283841e+01,	/* 0x40471547, 0x652b82fe */
> +  2.16608493865351192653e-02,	/* 0x3f962e42, 0xfee00000 */
> +  5.96317165397058656257e-12,	/* 0x3d9a39ef, 0x35793c76 */

Please put a comment on each file-scope variable or function explaining 
its semantics.

If you're commenting exact representations of floating-point numbers, I'd 
advise writing them as C99 hex float constants instead of decimal (and 
then omitting those comments) - generally I think hex float constants are 
a good idea for any constants that have been computed to have some 
property (polynomial coefficients, etc.) and can't be expected to be 
human-readable.

> +#ifndef DBL_MIN
> +#define DBL_MIN 2.2250738585072014E-308
> +#endif

Just include <float.h>.  No such #ifndef is then needed (and we use 
indentation inside #if, "# define" etc.).

> +#define	half		zC[0]
> +#define	invln2_32	zC[1]

It would be better to make each of these into a separate static const 
variable with its own comment rather than defining an array full of 
unrelated values.

Much the same comments apply to the float code.
Patrick McGehearty - July 31, 2017, 9:06 p.m.
Sparc has a significant performance issue with RAW (read after write).
That is, if a value is stored to a particular address and then read from 
that
address before the store has reached L2 cache, a pipeline hiccup occurs
and a 30+ cycle delay is seen. Most commonly this issue is seen in the case
of register spill/fills, but it also occurs when a value in an integer 
register
to stored to a temporary in memory and then loaded to a floating point 
register.
The int to fp and fp to int operations are common in exp() algorithms due
to cracking the exponent from the mantissa to determine which special
case to use in handling particular input data ranges.

Starting with Niagara4 (T4), direct int to fp and fp to int transfer 
instructions
were added, avoiding this performance issue. If we compile for any Sparc
platform instead of T4 and later, we can't use the direct transfers.
Note that T4 was first introduced in 2011, meaning most current
Sparc/Linux platforms will have this support.

For comparison, recent x86 chips from Intel have thrown enough HW at
the RAW issue to not have any delays when a read-after-write occurs.

The new algorithm is significantly different from the existing 
sysdeps/ieee754 algorithm.
The new algorithm matches the one used by the Solaris/Studio libm exp(), 
expf() code.
My effort was involved in porting (with Oracle corporate permission), not
algorithm construction.

It seems likely that this code could be faster on other CPUs, but I've 
only tested it on Sparc
as that's the machines I have ready access to. The advantage may be much 
less on other platforms.

- patrick


On 7/31/2017 2:47 PM, David Miller wrote:
> From: Patrick McGehearty <patrick.mcgehearty@oracle.com>
> Date: Mon, 31 Jul 2017 15:39:29 -0400
>
>> This PATCH is intended to improve exp() and expf() performance on Sparc.
>> These changes will only be active on Sparc platforms and only for
>> those platforms that support HWCAP_SPARC_CRYPTO (niagara4 and later).
> Can you explain which instructions exactly help make the compiled
> C code for exp() and expf() faster instead of being vague like
> this?
>
> Wouldn't the new C code you are adding be faster on other CPUs as
> well, even without gcc generating instructions for Niagara 4 and
> later?
>
> Thank you.
David Miller - July 31, 2017, 9:21 p.m.
From: Patrick McGehearty <patrick.mcgehearty@oracle.com>
Date: Mon, 31 Jul 2017 16:06:44 -0500

> Sparc has a significant performance issue with RAW (read after write).
> That is, if a value is stored to a particular address and then read
> from that
> address before the store has reached L2 cache, a pipeline hiccup
> occurs
> and a 30+ cycle delay is seen. Most commonly this issue is seen in the
> case
> of register spill/fills, but it also occurs when a value in an integer
> register
> to stored to a temporary in memory and then loaded to a floating point
> register.
> The int to fp and fp to int operations are common in exp() algorithms
> due
> to cracking the exponent from the mantissa to determine which special
> case to use in handling particular input data ranges.
> 
> Starting with Niagara4 (T4), direct int to fp and fp to int transfer
> instructions
> were added, avoiding this performance issue. If we compile for any
> Sparc
> platform instead of T4 and later, we can't use the direct transfers.
> Note that T4 was first introduced in 2011, meaning most current
> Sparc/Linux platforms will have this support.
> 
> For comparison, recent x86 chips from Intel have thrown enough HW at
> the RAW issue to not have any delays when a read-after-write occurs.
> 
> The new algorithm is significantly different from the existing
> sysdeps/ieee754 algorithm.
> The new algorithm matches the one used by the Solaris/Studio libm
> exp(), expf() code.
> My effort was involved in porting (with Oracle corporate permission),
> not
> algorithm construction.
> 
> It seems likely that this code could be faster on other CPUs, but I've
> only tested it on Sparc
> as that's the machines I have ready access to. The advantage may be
> much less on other platforms.

You miss my point.

You are doing two _completely_ different things here.

First, you could simply build the existing exp() and expf() C code in
glibc with niagara4.  In fact, if this float<-->int move instruction
helps so much, you probably want to build the entire math library
this way with appropriate ifunc hooks.  Not just exp/expf.

Second, you could then introduce the new C code implementation of exp
and expf functions and:

1) See if it is faster on other sparc cpus.

2) Ask other glibc developers to test whether it is faster on
   non-sparc cpus as well.

Making both changes and only targetting post-niagara4 cpus is
completely the wrong way to go about this.
Patrick McGehearty - July 31, 2017, 9:30 p.m.
On 7/31/2017 3:12 PM, Carlos O'Donell wrote:
> On 07/31/2017 03:47 PM, David Miller wrote:
>> From: Patrick McGehearty <patrick.mcgehearty@oracle.com>
>> Date: Mon, 31 Jul 2017 15:39:29 -0400
>>
>>> This PATCH is intended to improve exp() and expf() performance on Sparc.
>>> These changes will only be active on Sparc platforms and only for
>>> those platforms that support HWCAP_SPARC_CRYPTO (niagara4 and later).
>> Can you explain which instructions exactly help make the compiled
>> C code for exp() and expf() faster instead of being vague like
>> this?
>>
>> Wouldn't the new C code you are adding be faster on other CPUs as
>> well, even without gcc generating instructions for Niagara 4 and
>> later?
>
> ... I would also like to see the results of the glibc microbenchmark
> *before* and *after* the patches. We have *specific* microbenchmarks
> for lots of math functions.
>

I'm assuming you are referring to the results of running
"make bench". I found some exp results in benchtests/bench.out
On my test machine (a single core VM in a Sparc S7 running at 4.3GHz):

ieee754  (before)
   "exp": {
    "": {
     "duration": 4.34656e+10,
     "iterations": 8.224e+06,
     "max": 16550.3,
     "min": 400.426,
     "mean": 5285.21
    },

new sparc code (after)
   "exp": {
    "": {
     "duration": 4.25365e+10,
     "iterations": 6.07034e+08,
     "max": 183.446,
     "min": 27.095,
     "mean": 70.0726
    },

I have to say that the ratio of 5285/70 = 75x speedup seems way
too optimistic for my new code. I have not investigated the reason
for the apparently super slow max value.

I wrote my own standalone tests which tested a variety of different 
values for x
with a repeat factor of 500 (i.e. time the computation of each value 500 
times).
A sample of results:
x= 10  ieee754 exp(x) = 172 nsec; new exp(x) = 37 nsec (or 51 nsec 
without t4 optimizations)
x=-10  ieee754 exp(x) = 172 nsec; new exp(x) = 37 nsec
x=-0.9 ieee754 exp(x) = 172 nsec; new exp(x) = 19 nsec
x= 0.9 ieee754 exp(x) = 172 nsec; new exp(x) = 19 nsec
x= 0   ieee754 exp(x) = 116 nsec; new exp(x) =  8 nsec

The expf() is around 200 nsec for ieee754.
The new expf() time is typically around 12-13 nsec/call.

- patrick
Joseph Myers - July 31, 2017, 9:54 p.m.
On Mon, 31 Jul 2017, Patrick McGehearty wrote:

> I have to say that the ratio of 5285/70 = 75x speedup seems way
> too optimistic for my new code. I have not investigated the reason
> for the apparently super slow max value.

That would be the code that tries to ensure correctly rounded results 
(only for round-to-even) in cases where the mathematical result is very 
close to a half-way value.

There is no need for functions such as exp to be correctly rounded, only 
functions such as fma and sqrt that are fully bound to IEEE 754 
operations.  (TS 18661-4 reserves names such as crexp for 
correctly-rounded implementations, which need to be correctly rounded, 
with correct exceptions including "inexact", in all rounding modes.)  
Thus in fact the slow cases could be removed - *provided* an analysis of 
the existing checks for when to use them shows that the fast cases still 
have a sufficiently small error bound (it's possible some slow cases may 
still be needed, if the errors from the fast case can be too large for 
some inputs).
Carlos O'Donell - July 31, 2017, 11:22 p.m.
On 07/31/2017 05:30 PM, Patrick McGehearty wrote:
> On 7/31/2017 3:12 PM, Carlos O'Donell wrote:
>> On 07/31/2017 03:47 PM, David Miller wrote:
>>> From: Patrick McGehearty <patrick.mcgehearty@oracle.com>
>>> Date: Mon, 31 Jul 2017 15:39:29 -0400
>>>
>>>> This PATCH is intended to improve exp() and expf() performance on Sparc.
>>>> These changes will only be active on Sparc platforms and only for
>>>> those platforms that support HWCAP_SPARC_CRYPTO (niagara4 and later).
>>> Can you explain which instructions exactly help make the compiled
>>> C code for exp() and expf() faster instead of being vague like
>>> this?
>>>
>>> Wouldn't the new C code you are adding be faster on other CPUs as
>>> well, even without gcc generating instructions for Niagara 4 and
>>> later?
>>
>> ... I would also like to see the results of the glibc microbenchmark
>> *before* and *after* the patches. We have *specific* microbenchmarks
>> for lots of math functions.
>>
> 
> I'm assuming you are referring to the results of running
> "make bench". I found some exp results in benchtests/bench.out
> On my test machine (a single core VM in a Sparc S7 running at 4.3GHz):

My apologies, I realize I was being too harsh in my last email.
We look forward to more posts from Oracle to help out with SPARC.

Yes, this is exactly what you need to do, run `make bench` to get
data before and after the changes.

> ieee754  (before)
>   "exp": {
>    "": {
>     "duration": 4.34656e+10,
>     "iterations": 8.224e+06,
>     "max": 16550.3,
>     "min": 400.426,
>     "mean": 5285.21
>    },
> 
> new sparc code (after)
>   "exp": {
>    "": {
>     "duration": 4.25365e+10,
>     "iterations": 6.07034e+08,
>     "max": 183.446,
>     "min": 27.095,
>     "mean": 70.0726
>    },
> 
> I have to say that the ratio of 5285/70 = 75x speedup seems way
> too optimistic for my new code. I have not investigated the reason
> for the apparently super slow max value.

It looks like you've made significant performance gains for the input
values being tested by the microbenchmark.

One reason you might see 75x is that you've eliminated a slow path
that used to go through multi-precision arithmetic?

> I wrote my own standalone tests which tested a variety of different values for x
> with a repeat factor of 500 (i.e. time the computation of each value 500 times).
> A sample of results:
> x= 10  ieee754 exp(x) = 172 nsec; new exp(x) = 37 nsec (or 51 nsec without t4 optimizations)
> x=-10  ieee754 exp(x) = 172 nsec; new exp(x) = 37 nsec
> x=-0.9 ieee754 exp(x) = 172 nsec; new exp(x) = 19 nsec
> x= 0.9 ieee754 exp(x) = 172 nsec; new exp(x) = 19 nsec
> x= 0   ieee754 exp(x) = 116 nsec; new exp(x) =  8 nsec
> 
> The expf() is around 200 nsec for ieee754.
> The new expf() time is typically around 12-13 nsec/call.

Perhaps we need more indicative inputs added to benchtests/exp-inputs?
Patrick McGehearty - Aug. 1, 2017, 4:06 p.m.
On 7/31/2017 4:21 PM, David Miller wrote:
> From: Patrick McGehearty <patrick.mcgehearty@oracle.com>
> Date: Mon, 31 Jul 2017 16:06:44 -0500
>
>> Sparc has a significant performance issue with RAW (read after write).
>> That is, if a value is stored to a particular address and then read
>> from that
>> address before the store has reached L2 cache, a pipeline hiccup
>> occurs
>> and a 30+ cycle delay is seen. Most commonly this issue is seen in the
>> case
>> of register spill/fills, but it also occurs when a value in an integer
>> register
>> to stored to a temporary in memory and then loaded to a floating point
>> register.
>> The int to fp and fp to int operations are common in exp() algorithms
>> due
>> to cracking the exponent from the mantissa to determine which special
>> case to use in handling particular input data ranges.
>>
>> Starting with Niagara4 (T4), direct int to fp and fp to int transfer
>> instructions
>> were added, avoiding this performance issue. If we compile for any
>> Sparc
>> platform instead of T4 and later, we can't use the direct transfers.
>> Note that T4 was first introduced in 2011, meaning most current
>> Sparc/Linux platforms will have this support.
>>
>> For comparison, recent x86 chips from Intel have thrown enough HW at
>> the RAW issue to not have any delays when a read-after-write occurs.
>>
>> The new algorithm is significantly different from the existing
>> sysdeps/ieee754 algorithm.
>> The new algorithm matches the one used by the Solaris/Studio libm
>> exp(), expf() code.
>> My effort was involved in porting (with Oracle corporate permission),
>> not
>> algorithm construction.
>>
>> It seems likely that this code could be faster on other CPUs, but I've
>> only tested it on Sparc
>> as that's the machines I have ready access to. The advantage may be
>> much less on other platforms.
> You miss my point.
>
> You are doing two _completely_ different things here.
>
> First, you could simply build the existing exp() and expf() C code in
> glibc with niagara4.  In fact, if this float<-->int move instruction
> helps so much, you probably want to build the entire math library
> this way with appropriate ifunc hooks.  Not just exp/expf.
>
> Second, you could then introduce the new C code implementation of exp
> and expf functions and:
>
> 1) See if it is faster on other sparc cpus.
>
> 2) Ask other glibc developers to test whether it is faster on
>     non-sparc cpus as well.
>
> Making both changes and only targetting post-niagara4 cpus is
> completely the wrong way to go about this.

I'm preparing to do a trial run on -mcpu=niagara4 for glibc.
I'll report back on any interesting differences for make bench
with/without -mcpu=niagara4 for the current sourceware tree.

I will note from my point of view, this project is focused only
on exp() and expf() as Sparc/Solaris/Studio showed dramatically
better performance on those specific functions. There are a few
other functions which run faster on Sparc/Solaris/Studio, but
nothing like the performance difference for exp() and expf().

- patrick
Patrick McGehearty - Aug. 1, 2017, 4:33 p.m.
On 7/31/2017 6:22 PM, Carlos O'Donell wrote:
> On 07/31/2017 05:30 PM, Patrick McGehearty wrote:
>> On 7/31/2017 3:12 PM, Carlos O'Donell wrote:
>>> On 07/31/2017 03:47 PM, David Miller wrote:
>>>> From: Patrick McGehearty <patrick.mcgehearty@oracle.com>
>>>> Date: Mon, 31 Jul 2017 15:39:29 -0400
>>>>
>>>>> This PATCH is intended to improve exp() and expf() performance on Sparc.
>>>>> These changes will only be active on Sparc platforms and only for
>>>>> those platforms that support HWCAP_SPARC_CRYPTO (niagara4 and later).
>>>> Can you explain which instructions exactly help make the compiled
>>>> C code for exp() and expf() faster instead of being vague like
>>>> this?
>>>>
>>>> Wouldn't the new C code you are adding be faster on other CPUs as
>>>> well, even without gcc generating instructions for Niagara 4 and
>>>> later?
>>> ... I would also like to see the results of the glibc microbenchmark
>>> *before* and *after* the patches. We have *specific* microbenchmarks
>>> for lots of math functions.
>>>
>> I'm assuming you are referring to the results of running
>> "make bench". I found some exp results in benchtests/bench.out
>> On my test machine (a single core VM in a Sparc S7 running at 4.3GHz):
> My apologies, I realize I was being too harsh in my last email.
> We look forward to more posts from Oracle to help out with SPARC.
>
> Yes, this is exactly what you need to do, run `make bench` to get
> data before and after the changes.
>
>> ieee754  (before)
>>    "exp": {
>>     "": {
>>      "duration": 4.34656e+10,
>>      "iterations": 8.224e+06,
>>      "max": 16550.3,
>>      "min": 400.426,
>>      "mean": 5285.21
>>     },
>>
>> new sparc code (after)
>>    "exp": {
>>     "": {
>>      "duration": 4.25365e+10,
>>      "iterations": 6.07034e+08,
>>      "max": 183.446,
>>      "min": 27.095,
>>      "mean": 70.0726
>>     },
>>
>> I have to say that the ratio of 5285/70 = 75x speedup seems way
>> too optimistic for my new code. I have not investigated the reason
>> for the apparently super slow max value.
> It looks like you've made significant performance gains for the input
> values being tested by the microbenchmark.
>
> One reason you might see 75x is that you've eliminated a slow path
> that used to go through multi-precision arithmetic?
>
>> I wrote my own standalone tests which tested a variety of different values for x
>> with a repeat factor of 500 (i.e. time the computation of each value 500 times).
>> A sample of results:
>> x= 10  ieee754 exp(x) = 172 nsec; new exp(x) = 37 nsec (or 51 nsec without t4 optimizations)
>> x=-10  ieee754 exp(x) = 172 nsec; new exp(x) = 37 nsec
>> x=-0.9 ieee754 exp(x) = 172 nsec; new exp(x) = 19 nsec
>> x= 0.9 ieee754 exp(x) = 172 nsec; new exp(x) = 19 nsec
>> x= 0   ieee754 exp(x) = 116 nsec; new exp(x) =  8 nsec
>>
>> The expf() is around 200 nsec for ieee754.
>> The new expf() time is typically around 12-13 nsec/call.
> Perhaps we need more indicative inputs added to benchtests/exp-inputs?
>
Thank you for explaining the source of the very slow paths in the 
current code.
It bothers me to have unusually good results from a code change
when I don't have a clue about the reason.

Yes, I can see value in adding a wider range of values for
the inputs to exp.  Some larger and some negative values
would be more representative. Also, by adding more values,
the unusual values would be less heavily weighted in the performance
averages. Of course, changes to benchtests/exp-inputs would
be a separate patch from the code I'm proposing.

- patrick
Adhemerval Zanella Netto - Aug. 1, 2017, 6:27 p.m.
On 01/08/2017 13:06, Patrick McGehearty wrote:
> On 7/31/2017 4:21 PM, David Miller wrote:
>> From: Patrick McGehearty <patrick.mcgehearty@oracle.com>
>> Date: Mon, 31 Jul 2017 16:06:44 -0500
>>
>>> Sparc has a significant performance issue with RAW (read after write).
>>> That is, if a value is stored to a particular address and then read
>>> from that
>>> address before the store has reached L2 cache, a pipeline hiccup
>>> occurs
>>> and a 30+ cycle delay is seen. Most commonly this issue is seen in the
>>> case
>>> of register spill/fills, but it also occurs when a value in an integer
>>> register
>>> to stored to a temporary in memory and then loaded to a floating point
>>> register.
>>> The int to fp and fp to int operations are common in exp() algorithms
>>> due
>>> to cracking the exponent from the mantissa to determine which special
>>> case to use in handling particular input data ranges.
>>>
>>> Starting with Niagara4 (T4), direct int to fp and fp to int transfer
>>> instructions
>>> were added, avoiding this performance issue. If we compile for any
>>> Sparc
>>> platform instead of T4 and later, we can't use the direct transfers.
>>> Note that T4 was first introduced in 2011, meaning most current
>>> Sparc/Linux platforms will have this support.
>>>
>>> For comparison, recent x86 chips from Intel have thrown enough HW at
>>> the RAW issue to not have any delays when a read-after-write occurs.
>>>
>>> The new algorithm is significantly different from the existing
>>> sysdeps/ieee754 algorithm.
>>> The new algorithm matches the one used by the Solaris/Studio libm
>>> exp(), expf() code.
>>> My effort was involved in porting (with Oracle corporate permission),
>>> not
>>> algorithm construction.
>>>
>>> It seems likely that this code could be faster on other CPUs, but I've
>>> only tested it on Sparc
>>> as that's the machines I have ready access to. The advantage may be
>>> much less on other platforms.
>> You miss my point.
>>
>> You are doing two _completely_ different things here.
>>
>> First, you could simply build the existing exp() and expf() C code in
>> glibc with niagara4.  In fact, if this float<-->int move instruction
>> helps so much, you probably want to build the entire math library
>> this way with appropriate ifunc hooks.  Not just exp/expf.
>>
>> Second, you could then introduce the new C code implementation of exp
>> and expf functions and:
>>
>> 1) See if it is faster on other sparc cpus.
>>
>> 2) Ask other glibc developers to test whether it is faster on
>>     non-sparc cpus as well.
>>
>> Making both changes and only targetting post-niagara4 cpus is
>> completely the wrong way to go about this.
> 
> I'm preparing to do a trial run on -mcpu=niagara4 for glibc.
> I'll report back on any interesting differences for make bench
> with/without -mcpu=niagara4 for the current sourceware tree.
> 
> I will note from my point of view, this project is focused only
> on exp() and expf() as Sparc/Solaris/Studio showed dramatically
> better performance on those specific functions. There are a few
> other functions which run faster on Sparc/Solaris/Studio, but
> nothing like the performance difference for exp() and expf().
> 
> - patrick
> 

I agree with David, we should refrain of adding even more platform
specific assembly optimization where a default C code could be as
good as and also improve generic performance on other platforms as
well.

The problem you specific is very similar to the one on POWER before POWER8,
where floating pointer to integer transfer issues a load-hit-store that
increases latency.  I tried to mitigate this on sin/cos by tweaking the 
internal code using a hackish hooks (commit 77a2a8b4a19f0), but currently
I am convinced that a new algorithm for single float exp, sin, cos (and
probably others) is in fact a better solution.

In fact, some architecture already went on this path, although by 
providing specific assembly implementation of algorithms that could be
provided in a more C generic way for a possible default one.  For instance,
power8 [1], x86_64 [2], and i686 sse2 [3] seems to use same algorithms 
(based on initial code comments).  You could use this as base as well to 
evaluate which one shows better performance and precision.

We also has a similar discussion for AArch64 proposal of sinf/cosf 
optimization [4].

[1] sysdeps/powerpc/powerpc64/power8/fpu/e_expf.S
[2] sysdeps/x86_64/fpu/e_expf.S
[3] sysdeps/i386/i686/fpu/multiarch/e_expf-sse2.S
[4] https://sourceware.org/ml/libc-alpha/2017-06/msg00503.html
Patrick McGehearty - Aug. 2, 2017, 7:53 p.m.
My report on -mcpu=niagara4 for benchtests/math (and make check)
See below for results and recommendations.

On 8/1/2017 11:06 AM, Patrick McGehearty wrote:
> On 7/31/2017 4:21 PM, David Miller wrote:
>> You miss my point.
>>
>> You are doing two _completely_ different things here.
>>
>> First, you could simply build the existing exp() and expf() C code in
>> glibc with niagara4.  In fact, if this float<-->int move instruction
>> helps so much, you probably want to build the entire math library
>> this way with appropriate ifunc hooks.  Not just exp/expf.
>>
>> Second, you could then introduce the new C code implementation of exp
>> and expf functions and:
>>
>> 1) See if it is faster on other sparc cpus.
>>
>> 2) Ask other glibc developers to test whether it is faster on
>>     non-sparc cpus as well.
>>
>> Making both changes and only targetting post-niagara4 cpus is
>> completely the wrong way to go about this.
>
> I'm preparing to do a trial run on -mcpu=niagara4 for glibc.
> I'll report back on any interesting differences for make bench
> with/without -mcpu=niagara4 for the current sourceware tree.
>
> I will note from my point of view, this project is focused only
> on exp() and expf() as Sparc/Solaris/Studio showed dramatically
> better performance on those specific functions. There are a few
> other functions which run faster on Sparc/Solaris/Studio, but
> nothing like the performance difference for exp() and expf().
>
> - patrick
>
Benchtests comparison  based on Aug 1, 2017 glibc
using -mcpu=niagara4 on an S7

Niagara4 provides new features, including direct fma support, direct
transfers between fp and int registers, and new types of branch
instructions.

Roughly 40% of tests showed more than 10% change.
I've grouped them by type of test.

Mean execution time (in nsec/call)
Fcn       Base     -mcpu  % improvement
exp       5318      5109       4%
log        534       343      56%
log2       182       157      16%
powf       511       396      29%

acos      1148       730      57%
asin      1120       691      62%
asinh      313       249      26%
atan      1199       915      31%
atanh      242       211      15%
cos       1527      1199      27%
cosh       304       265      15%
sin       1562      1194      31%
sincos    1671      1295      29%
sinh       573       523      10%
tan       1365      1098      24%
tanh       311       223      39%

fmin        14.6      17.6   -17% <<
fmax        14.6      17.6   -19% <<
fminf       14.6      18.1   -19% <<
fmaxf       14.6      18.2   -20% <<
ffsll       32        22      45%
(ffsll: find first bit set in a word)

I suspect fmin/fmax/fminf/fmaxf regression may have placed a branch
instruction in a branch delay slot which causes a pipeline hiccup, or
something similar but I have not investigated specifics for any of
these tests.

I also found that using -mcpu=niagara4 caused warnings to be
generated for math/e_sqrt.c. That's because -mcpu=niagara4 defines
__FP_FAST_FMA which affects the definition of EMULV and DLA_FMS.
Either e_sqrt.c needs to be revised to avoid the warnings or
-mcpu=niagara4 needs to not be used on that routine.

Also, make check got the following new failures when using -mcpu=niagara4.
FAIL: math/test-float-cpow
FAIL: math/test-float-finite-cpow
FAIL: math/test-ifloat-cpow
(above 3 identical single value, 5ulp error instead of 4 ulp)
FAIL: math/test-float-ctanh
FAIL: math/test-float-finite-ctanh
FAIL: math/test-ifloat-ctanh
(above 3 identical single value, 2ulp error instead of 1 ulp)
These may be due to the use of FMA or other optimizations due
to -mcpu=niagara4. They have not been investigated in depth.

I'll note that the tests are not exhaustive. Specifically, only powf,
fminf, and fmaxf are included as 32-bit precision test. Complex
functions are not tested. Some other common functions such as expm1
and log10 are not present. That's not a criticism, just an
observation that more tests are needed before a blanket change of
default behavior should be recommended.

I concur that it would be beneficial to use -mcpu=niagara4 for
selected libm functions, but do not believe it should be default on
without more extensive testing (and likely more compiler tuning). I'd
like to move forward on that issue, but those tasks are beyond the
scope of my current assignment. I will put them on our internal list
of desirable enhancements but can't make any commitments as to whether
resources will be assigned to making them happen.

My goal/assignment is for exp() and expf() on Linux to match Solaris
libm performance. exp() is included in libmicro() and in one or more
of the new SPECcpu2017 applications, making it highly visible when
doing benchmark comparisons between systems.
Even without -mcpu=niagara4, the new code is much faster than the
ieee754 code, but to achieve full parity, the -mcpu=niagara4 switch is
needed. I believe that switch should be only used selectively on
functions where a clear benefit is shown. Also, for those functions, a
fall-back version that does not use -mcpu=niagara4 should be included
and controlled by ifunc to allow a Sparc-based glibc to run on any
Sparc platform even if it is built on a newer Sparc platform.
The approach I've implemented supports both of those requirements.

I can remove the Makefile code to use -mcpu=niagara4 iff the compiler
supports it. That will mean any attempt to build glibc on Sparc with
an older version of gcc will fail.  Internal to Oracle, some Oracle SW
still depends on gcc 4.4.7, meaning I'll maintain two different
versions, one for Oracle Linux/Sparc and one for external use.

respectfully submitted by Patrick McGehearty
David Miller - Aug. 2, 2017, 11:19 p.m.
From: Patrick McGehearty <patrick.mcgehearty@oracle.com>
Date: Wed, 2 Aug 2017 14:53:19 -0500

> I concur that it would be beneficial to use -mcpu=niagara4 for
> selected libm functions, but do not believe it should be default on
> without more extensive testing (and likely more compiler tuning).

This is precisely what has been suggested, have an niagara4 IFUNC
vector for math functions that benefit greatly from the fp<-->int
register move instructions.
Patrick McGehearty - Aug. 2, 2017, 11:35 p.m.
Apologies for the many earlier versions of this email.
My fingers forgot I was in a mailer not an editor and
every time I typed "^S" to save my temporary version,
the mailer did a send.

This is the intended version of this note.

- patrick


On 7/31/2017 4:06 PM, Joseph Myers wrote:
> On Mon, 31 Jul 2017, Patrick McGehearty wrote:
>
>>      * sysdeps/sparc/configure.ac: manage -mcpu=niagara4 test
>>      * sysdeps/sparc/configure: Regenerate
> As far as I can see, niagara4 support was added to GCC in r178554
> (2011-09-05), so would have been in GCC 4.7.  The minimum GCC version for
> building glibc is 4.9.  So this configure test, and any associated
> conditionals in Makefiles or elsewhere, should not be necessary.
Internally I need to support some Oracle Linux releases which use gcc 4.4.7.
Putting the test for -mcpu=niagara4 support allowed me to use one set of 
changes for
both the Oracle and sourceware versions.
I'll simplify my patch for sourceware, omitting the test.

Same applies to the use of
#ifndef DBL_MIN
as DBL_MIN was not defined in older glibc/gcc trees.
I'll remove the DBL_MIN def.


>
>> diff --git a/sysdeps/sparc/fpu/libm_endian.h b/sysdeps/sparc/fpu/libm_endian.h
>> new file mode 100644
>> index 0000000..a49abcd
>> --- /dev/null
>> +++ b/sysdeps/sparc/fpu/libm_endian.h
>> @@ -0,0 +1,32 @@
>> +/*
>> +   Contributed by Oracle Inc.
>> +   Copyright (C) 2017 Free Software Foundation, Inc.
> All new files should start with a descriptive comment on their first line,
> before the copyright notice.  "Contributed by" is no longer used;
> significant contributions can be mentioned in the NEWS file (as well as in
> contrib.texi).
Ok, removed.  Will say "EXP function - Compute double precision 
exponential" for the comment.

>
>> +extern double exp (double);
> No such declaration should be needed, since you're including <math.h>.
Ok, removed.
>
>> +extern double __ieee754_exp (double);
> And this should come through math_private.h.
I'm not finding it in any math_private.h that I've checked.
Do you mean I should put it in sysdeps/sparc/fpu/math_private.h?


>
>> +/*
>> + * For i = 0, ..., 66,
>> + *   TBL2[2*i] is a double precision number near (i+1)*2^-6, and
> We don't use leading '*' on each line of a comment.
Fixed.
>> +static const double zC[] = {
>> +  0.5,
>> +  4.61662413084468283841e+01,	/* 0x40471547, 0x652b82fe */
>> +  2.16608493865351192653e-02,	/* 0x3f962e42, 0xfee00000 */
>> +  5.96317165397058656257e-12,	/* 0x3d9a39ef, 0x35793c76 */
> Please put a comment on each file-scope variable or function explaining
> its semantics.
>
> If you're commenting exact representations of floating-point numbers, I'd
> advise writing them as C99 hex float constants instead of decimal (and
> then omitting those comments) - generally I think hex float constants are
> a good idea for any constants that have been computed to have some
> property (polynomial coefficients, etc.) and can't be expected to be
> human-readable.

I agree that C99 hex float constants would be better than decimal, but 
I'm reluctant to
change the formatting of any of the existing constants as it seems an 
easy opportunity to
introduce a possible typo. Even something as simple as transposed digits 
in one table
entry could slip by even extended testing procedures.

On the other hand, I can readily change the zC table into the following:
static const half       0.5;
static const invln2_32  4.61662413084468283841e+01;  /* 0x40471547, 
0x652b82fe */
...
and add comments to the limit of my understanding of the algorithm.
I am porting this code, not the original developer.
I will make those changes.


>
>> +#ifndef DBL_MIN
>> +#define DBL_MIN 2.2250738585072014E-308
>> +#endif
> Just include <float.h>.  No such #ifndef is then needed (and we use
> indentation inside #if, "# define" etc.).
will remove for sourceware version of code.
>> +#define	half		zC[0]
>> +#define	invln2_32	zC[1]
> It would be better to make each of these into a separate static const
> variable with its own comment rather than defining an array full of
> unrelated values.
will do.

> Much the same comments apply to the float code.
>
I will make similar changes to float.
Joseph Myers - Aug. 3, 2017, 11:17 a.m.
On Thu, 3 Aug 2017, Patrick McGehearty wrote:

> > > +extern double __ieee754_exp (double);
> > And this should come through math_private.h.
> I'm not finding it in any math_private.h that I've checked.
> Do you mean I should put it in sysdeps/sparc/fpu/math_private.h?

No.  sysdeps/generic/math_private.h includes math_private_calls.h for each 
floating-point type.  Just #include <math_private.h> and let that arrange 
for the appropriate declarations to be visible.

> > If you're commenting exact representations of floating-point numbers, I'd
> > advise writing them as C99 hex float constants instead of decimal (and
> > then omitting those comments) - generally I think hex float constants are
> > a good idea for any constants that have been computed to have some
> > property (polynomial coefficients, etc.) and can't be expected to be
> > human-readable.
> 
> I agree that C99 hex float constants would be better than decimal, but 
> I'm reluctant to change the formatting of any of the existing constants 
> as it seems an easy opportunity to introduce a possible typo. Even 
> something as simple as transposed digits in one table entry could slip 
> by even extended testing procedures.

All you need to do is compare the generated object files before and after 
the change and make sure they are identical, so the constants have the 
same values in decimal and hex float.
Patrick McGehearty - Aug. 3, 2017, 4:50 p.m.
On 8/2/2017 6:19 PM, David Miller wrote:
> From: Patrick McGehearty <patrick.mcgehearty@oracle.com>
> Date: Wed, 2 Aug 2017 14:53:19 -0500
>
>> I concur that it would be beneficial to use -mcpu=niagara4 for
>> selected libm functions, but do not believe it should be default on
>> without more extensive testing (and likely more compiler tuning).
> This is precisely what has been suggested, have an niagara4 IFUNC
> vector for math functions that benefit greatly from the fp<-->int
> register move instructions.

I'm new to linux glibc development and have been using existing examples 
to suggest
how I should do things like use ifunc. If there is a different approach 
that is recommended,
I'm willing to use it but will need some hints/instruction on what is.

My current approach of using
sparc_libm_ifunc (exp, hwcap & HWCAP_SPARC_CRYPTO ?
                   __exp_niagara4 : __ieee754_exp);
in sysdeps/sparc/sparc64/fpu/multiarch/e_exp-niagara4.c matches the 
approach used in
sysdeps/sparc/sparc64/fpu/multiarch/s_ceil.c
sysdeps/sparc/sparc64/fpu/multiarch/s_ceilf.c
sysdeps/sparc/sparc64/fpu/multiarch/s_floor.c
sysdeps/sparc/sparc64/fpu/multiarch/s_floorf.c
sysdeps/sparc/sparc64/fpu/multiarch/s_fma.c
sysdeps/sparc/sparc64/fpu/multiarch/s_fmaf.c
sysdeps/sparc/sparc64/fpu/multiarch/s_trunc.c
sysdeps/sparc/sparc64/fpu/multiarch/s_truncf.c

I'll note that simply using -mcpu=niagara4 on the existing ieee754 exp() 
code only
provides a 5% performance gain. Due to other assignments, I will not 
have time/resources
to redo other libm functions at this time.

- patrick
Adhemerval Zanella Netto - Aug. 3, 2017, 6:43 p.m.
On 03/08/2017 13:50, Patrick McGehearty wrote:
> On 8/2/2017 6:19 PM, David Miller wrote:
>> From: Patrick McGehearty <patrick.mcgehearty@oracle.com>
>> Date: Wed, 2 Aug 2017 14:53:19 -0500
>>
>>> I concur that it would be beneficial to use -mcpu=niagara4 for
>>> selected libm functions, but do not believe it should be default on
>>> without more extensive testing (and likely more compiler tuning).
>> This is precisely what has been suggested, have an niagara4 IFUNC
>> vector for math functions that benefit greatly from the fp<-->int
>> register move instructions.
> 
> I'm new to linux glibc development and have been using existing examples to suggest
> how I should do things like use ifunc. If there is a different approach that is recommended,
> I'm willing to use it but will need some hints/instruction on what is.
> 
> My current approach of using
> sparc_libm_ifunc (exp, hwcap & HWCAP_SPARC_CRYPTO ?
>                   __exp_niagara4 : __ieee754_exp);
> in sysdeps/sparc/sparc64/fpu/multiarch/e_exp-niagara4.c matches the approach used in
> sysdeps/sparc/sparc64/fpu/multiarch/s_ceil.c
> sysdeps/sparc/sparc64/fpu/multiarch/s_ceilf.c
> sysdeps/sparc/sparc64/fpu/multiarch/s_floor.c
> sysdeps/sparc/sparc64/fpu/multiarch/s_floorf.c
> sysdeps/sparc/sparc64/fpu/multiarch/s_fma.c
> sysdeps/sparc/sparc64/fpu/multiarch/s_fmaf.c
> sysdeps/sparc/sparc64/fpu/multiarch/s_trunc.c
> sysdeps/sparc/sparc64/fpu/multiarch/s_truncf.c
> 
> I'll note that simply using -mcpu=niagara4 on the existing ieee754 exp() code only
> provides a 5% performance gain. Due to other assignments, I will not have time/resources
> to redo other libm functions at this time.
> 
> - patrick
> 

SPARC code is still using a old way to define ifunc code which replicates the
logic already present on include/libc-symbols.h.  I would suggest you to use
the newer ifunc macro definition for new code:

---
#include <math.h>
#include <init-arch.h>

extern double __ieee754_exp_niagara (double);
extern double __ieee754_exp (double);

libm_ifunc (__ieee754_exp,
            hwcap & HWCAP_SPARC_CRYPTO
	       ? __exp_niagara4 : __ieee754_exp
strong_alias (__ieee754_exp, __exp_finite)
---

You will need also to provide a init-arch.h file, as for other architectures
(x86, aarch64, powerpc) which I think should do basically:

---
#include <ldsodefs.h>

#define INIT_ARCH() \
  unsigned long int hwcap = __GLRO(dl_hwcap); 
---

I would also advise against to include the default implementation on current
ifunc definition file (such as sparc s_ceil.c for instance). I would recommend
create a default file, for instance sysdeps/sparc/sparc64/fpu/multiarch/e_exp-default.c,
and then include the default exp function.
Patrick McGehearty - Aug. 4, 2017, 12:16 a.m.
I want to thank David Miller, Carlos O'Donell, Joseph Myers, and 
Adhemerval Zanella
for their helpful comments on improving the patch I'm proposing for
Sparc exp(), expf().

I will be going on vacation after tomorrow, returning after the USA 
Solar eclipse (Aug 21).
As it is generally considered 'bad form' to submit code to a shared 
project and then
disappear for a time, I plan to work on this patch tomorrow but hold off 
my next PATCH
submission until I get back.

Just sending this message so you all won't think I've disappeared into 
the ether.  :-)

- patrick

Patch

This PATCH is intended to improve exp() and expf() performance on Sparc.
These changes will only be active on Sparc platforms and only for
those platforms that support HWCAP_SPARC_CRYPTO (niagara4 and later).

Typical Performance Gains for new exp() code for Sparc

exp()
With -mcpu=niagara4,    8x to 14x (depending on input value)
Without -mcpu=niagara4, 5x to  8x (depending on input value)

expf()
With -mcpu=niagara4,    16x
Without -mcpu=niagara4, 15x

Some portion of these speedups are due to taking advantage of
specific features new to the niagara4 and later architectures.

Glibc correctness tests for exp() and expf() were run. Within the
test suite 3 input values were found to cause 1 bit differences (ulp)
when "FE_TONEAREST" rounding mode is set. No differences were
seen for the tested values for the other rounding modes.
Typical example:
exp(-0x1.760cd2p+0)  (-1.46113312244415283203125)
 new code:    2.31973271630014299393707e-01   0x1.db14cd799387ap-3
 old code:    2.31973271630014271638132e-01   0x1.db14cd7993879p-3
    exp    =  2.31973271630014285508337 (high precision)
Old delta: off by 0.49 ulp
New delta: off by 0.51 ulp
This is a difference of 0.02 ulp which is well within the 1 ulp requirement
although the test function reports the difference as a failure.

For expf(), no differences in the test suite were seen in
FE_TONEAREST, FE_DOWNWARD, FE_TOWARDZERO. Two differences were
seen with FE_UPWARD. For both cases, the precise value was less
than 0.00005 less than the value reported by the new code.
The difference of the old and new was 1 ulp.
Typical example:
Failure: Test: exp_upward (-0x1p-20)
Result:
 new code:  9.999990463256e-01 0x1.ffffe0p-1
 old code:  9.999991059303e-01 0x1.ffffe2p-1
precise     9.999990463261e-01 (high precision) (using bc -l at scale=60)

Changes to be committed:
    (use "git reset HEAD <file>..." to unstage)

    * sysdeps/sparc/configure.ac: manage -mcpu=niagara4 test
    * sysdeps/sparc/configure: Regenerate
    * sysdeps/sparc/fpu/libm_endian.h: New file
    * sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/Makefile:
        add e_exp*-niagara4.c functions and -mcpu=niagara4 option
    * sysdeps/sparc/sparc64/fpu/multiarch/Makefile: Likewise

    * sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/e_exp-niagara4.c: New file
    * sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/e_expf-niagara4.c: New file
    * sysdeps/sparc/sparc64/fpu/multiarch/e_exp-niagara4.c: New file
    * sysdeps/sparc/sparc64/fpu/multiarch/e_expf-niagara4.c: New file

---
 sysdeps/sparc/configure                            |   24 +
 sysdeps/sparc/configure.ac                         |   12 +
 sysdeps/sparc/fpu/libm_endian.h                    |   32 ++
 .../sparc/sparc32/sparcv9/fpu/multiarch/Makefile   |    9 +-
 .../sparc32/sparcv9/fpu/multiarch/e_exp-niagara4.c |    1 +
 .../sparcv9/fpu/multiarch/e_expf-niagara4.c        |    1 +
 sysdeps/sparc/sparc64/fpu/multiarch/Makefile       |    9 +-
 .../sparc/sparc64/fpu/multiarch/e_exp-niagara4.c   |  433 +++++++++++++++++++
 .../sparc/sparc64/fpu/multiarch/e_expf-niagara4.c  |  445 ++++++++++++++++++++
 9 files changed, 964 insertions(+), 2 deletions(-)
 create mode 100644 sysdeps/sparc/fpu/libm_endian.h
 create mode 100644 sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/e_exp-niagara4.c
 create mode 100644 sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/e_expf-niagara4.c
 create mode 100644 sysdeps/sparc/sparc64/fpu/multiarch/e_exp-niagara4.c
 create mode 100644 sysdeps/sparc/sparc64/fpu/multiarch/e_expf-niagara4.c

diff --git a/sysdeps/sparc/configure b/sysdeps/sparc/configure
index 90a86f6..cf4d567 100644
--- a/sysdeps/sparc/configure
+++ b/sysdeps/sparc/configure
@@ -81,3 +81,27 @@  if test $libc_cv_sparc_gcc_gotdata = yes; then
   $as_echo "#define PI_STATIC_AND_HIDDEN 1" >>confdefs.h
 
 fi
+
+# Check for GCC support of -mcpu=niagara4
+{ $as_echo "$as_me:${as_lineno-$LINENO}: checking for niagara4 compiler support" >&5
+$as_echo_n "checking for niagara4 compiler support... " >&6; }
+if ${libc_cv_sparc_niagara4+:} false; then :
+  $as_echo_n "(cached) " >&6
+else
+  echo > conftest.S
+if { ac_try='${CC-cc} -mcpu=niagara4 conftest.S -S'
+  { { eval echo "\"\$as_me\":${as_lineno-$LINENO}: \"$ac_try\""; } >&5
+  (eval $ac_try) 2>&5
+  ac_status=$?
+  $as_echo "$as_me:${as_lineno-$LINENO}: \$? = $ac_status" >&5
+  test $ac_status = 0; }; }; then
+  libc_cv_sparc_niagara4=yes
+else
+  libc_cv_sparc_niagara4=no
+fi
+rm -f conftest*
+fi
+{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $libc_cv_sparc_niagara4" >&5
+$as_echo "$libc_cv_sparc_niagara4" >&6; }
+config_vars="$config_vars
+sparc_n4flags = $libc_cv_sparc_niagara4"
diff --git a/sysdeps/sparc/configure.ac b/sysdeps/sparc/configure.ac
index 982077c..f9f5667 100644
--- a/sysdeps/sparc/configure.ac
+++ b/sysdeps/sparc/configure.ac
@@ -57,3 +57,15 @@  fi
 if test $libc_cv_sparc_gcc_gotdata = yes; then
   AC_DEFINE(PI_STATIC_AND_HIDDEN)
 fi
+
+# Check for GCC support of -mcpu=niagara4
+AC_CACHE_CHECK(for niagara4 compiler support, libc_cv_sparc_niagara4, [dnl
+echo > conftest.S
+dnl
+if AC_TRY_COMMAND([${CC-cc} -mcpu=niagara4 conftest.S -S]); then
+  libc_cv_sparc_niagara4=yes
+else
+  libc_cv_sparc_niagara4=no
+fi
+rm -f conftest*])
+LIBC_CONFIG_VAR([sparc_n4flags], [$libc_cv_sparc_niagara4])
diff --git a/sysdeps/sparc/fpu/libm_endian.h b/sysdeps/sparc/fpu/libm_endian.h
new file mode 100644
index 0000000..a49abcd
--- /dev/null
+++ b/sysdeps/sparc/fpu/libm_endian.h
@@ -0,0 +1,32 @@ 
+/*
+   Contributed by Oracle Inc.
+   Copyright (C) 2017 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#ifndef _LIBM_ENDIAN_H
+#define	_LIBM_ENDIAN_H
+
+#define	HIWORD		0
+#define	LOWORD		1
+#define	HIXWORD		0		/* index of int containing exponent */
+#define	XSGNMSK		0x80000000	/* exponent bit mask within the int */
+/* XBIASED_EXP(x) must be an int expression; so ~0x80000000 is no good */
+#define	XBIASED_EXP(x)	((((int *)&x)[HIXWORD] & 0x7fffffff) >> 16)
+#define	ISZEROL(x)	(((((int *)&x)[0] & ~XSGNMSK) | ((int *)&x)[1] | \
+				((int *)&x)[2] | ((int *)&x)[3]) == 0)
+
+#endif	/* !defined(_LIBM_ENDIAN_H) */
diff --git a/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/Makefile b/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/Makefile
index 322e300..2d5bc0b 100644
--- a/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/Makefile
+++ b/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/Makefile
@@ -1,4 +1,5 @@ 
 ifeq ($(subdir),math)
+libm-sysdep_routines += e_exp-niagara4 e_expf-niagara4
 ifeq ($(have-as-vis3),yes)
 libm-sysdep_routines += m_copysignf-vis3 m_copysign-vis3 s_fabs-vis3 \
 			s_fabsf-vis3 s_llrintf-vis3 s_llrint-vis3 \
@@ -7,8 +8,14 @@  libm-sysdep_routines += m_copysignf-vis3 m_copysign-vis3 s_fabs-vis3 \
 			s_fmaf-vis3 s_fma-vis3 s_nearbyint-vis3 \
 			s_nearbyintf-vis3 s_fdimf-vis3 s_fdim-vis3
 sysdep_routines += s_copysignf-vis3 s_copysign-vis3
-
 CFLAGS-s_fdimf-vis3.c += -Wa,-Av9d -mvis3
 CFLAGS-s_fdim-vis3.c += -Wa,-Av9d -mvis3
 endif
+ifeq ($(sparc_n4flags),yes)
+N4_CFLAGS = -mcpu=niagara4
+else
+N4_CFLAGS =
+endif
+CFLAGS-e_exp-niagara4.c += ${N4_CFLAGS}
+CFLAGS-e_expf-niagara4.c += ${N4_CFLAGS}
 endif
diff --git a/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/e_exp-niagara4.c b/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/e_exp-niagara4.c
new file mode 100644
index 0000000..fc29e48
--- /dev/null
+++ b/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/e_exp-niagara4.c
@@ -0,0 +1 @@ 
+#include <sparc64/fpu/multiarch/e_exp-niagara4.c>
diff --git a/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/e_expf-niagara4.c b/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/e_expf-niagara4.c
new file mode 100644
index 0000000..2cd92c2
--- /dev/null
+++ b/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/e_expf-niagara4.c
@@ -0,0 +1 @@ 
+#include <sparc64/fpu/multiarch/e_expf-niagara4.c>
diff --git a/sysdeps/sparc/sparc64/fpu/multiarch/Makefile b/sysdeps/sparc/sparc64/fpu/multiarch/Makefile
index 03a271d..6379424 100644
--- a/sysdeps/sparc/sparc64/fpu/multiarch/Makefile
+++ b/sysdeps/sparc/sparc64/fpu/multiarch/Makefile
@@ -1,4 +1,5 @@ 
 ifeq ($(subdir),math)
+libm-sysdep_routines +=  e_exp-niagara4 e_expf-niagara4
 ifeq ($(have-as-vis3),yes)
 libm-sysdep_routines += m_signbitf-vis3 m_signbit-vis3 m_finitef-vis3 \
 			m_finite-vis3 m_isinff-vis3 m_isinf-vis3 \
@@ -11,7 +12,6 @@  libm-sysdep_routines += m_signbitf-vis3 m_signbit-vis3 m_finitef-vis3 \
 sysdep_routines += s_signbitf-vis3 s_signbit-vis3 s_finitef-vis3 \
 		   s_finite-vis3 s_isinff-vis3 s_isinf-vis3 \
 		   s_isnanf-vis3 s_isnan-vis3
-
 CFLAGS-s_ceilf-vis3.c += -Wa,-Av9d -mvis3
 CFLAGS-s_ceil-vis3.c += -Wa,-Av9d -mvis3
 CFLAGS-s_floorf-vis3.c += -Wa,-Av9d -mvis3
@@ -19,4 +19,11 @@  CFLAGS-s_floor-vis3.c += -Wa,-Av9d -mvis3
 CFLAGS-s_truncf-vis3.c += -Wa,-Av9d -mvis3
 CFLAGS-s_trunc-vis3.c += -Wa,-Av9d -mvis3
 endif
+ifeq ($(sparc_n4flags),yes)
+N4_CFLAGS = -mcpu=niagara4
+else
+N4_CFLAGS =
+endif
+CFLAGS-e_exp-niagara4.c += ${N4_CFLAGS}
+CFLAGS-e_expf-niagara4.c += ${N4_CFLAGS}
 endif
diff --git a/sysdeps/sparc/sparc64/fpu/multiarch/e_exp-niagara4.c b/sysdeps/sparc/sparc64/fpu/multiarch/e_exp-niagara4.c
new file mode 100644
index 0000000..a179668
--- /dev/null
+++ b/sysdeps/sparc/sparc64/fpu/multiarch/e_exp-niagara4.c
@@ -0,0 +1,433 @@ 
+/* EXP function
+   Contributed by Oracle Inc.
+   Copyright (C) 2017 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+/*
+ * exp(x)
+ * Hybrid algorithm of Peter Tang's Table driven method (for large
+ * arguments) and an accurate table (for small arguments).
+ * Written by K.C. Ng, November 1988.
+ * Method (large arguments):
+ *	1. Argument Reduction: given the input x, find r and integer k
+ *	   and j such that
+ *	             x = (k+j/32)*(ln2) + r,  |r| <= (1/64)*ln2
+ *
+ *	2. exp(x) = 2^k * (2^(j/32) + 2^(j/32)*expm1(r))
+ *	   a. expm1(r) is approximated by a polynomial:
+ *	      expm1(r) ~ r + t1*r^2 + t2*r^3 + ... + t5*r^6
+ *	      Here t1 = 1/2 exactly.
+ *	   b. 2^(j/32) is represented to twice double precision
+ *	      as TBL[2j]+TBL[2j+1].
+ *
+ * Note: If divide were fast enough, we could use another approximation
+ *	 in 2.a:
+ *	      expm1(r) ~ (2r)/(2-R), R = r - r^2*(t1 + t2*r^2)
+ *	      (for the same t1 and t2 as above)
+ *
+ * Special cases:
+ *	exp(INF) is INF, exp(NaN) is NaN;
+ *	exp(-INF)=  0;
+ *	for finite argument, only exp(0)=1 is exact.
+ *
+ * Accuracy:
+ *	According to an error analysis, the error is always less than
+ *	an ulp (unit in the last place).  The largest errors observed
+ *	are less than 0.55 ulp for normal results and less than 0.75 ulp
+ *	for subnormal results.
+ *
+ * Misc. info.
+ *	For IEEE double
+ *		if x >  7.09782712893383973096e+02 then exp(x) overflow
+ *		if x < -7.45133219101941108420e+02 then exp(x) underflow
+ */
+
+/* #pragma weak exp = __exp_niagara4 */
+#include <sparc-ifunc.h>
+#include "libm_endian.h"
+#include <math.h>
+#include <math_private.h>
+#include <errno.h>
+
+extern double exp (double);
+extern double __exp_niagara4 (double);
+extern double __ieee754_exp (double);
+
+sparc_libm_ifunc (exp, hwcap & HWCAP_SPARC_CRYPTO ?
+		  __exp_niagara4 : __ieee754_exp);
+
+
+static const double TBL[] = {
+  1.00000000000000000000e+00, 0.00000000000000000000e+00,
+  1.02189714865411662714e+00, 5.10922502897344389359e-17,
+  1.04427378242741375480e+00, 8.55188970553796365958e-17,
+  1.06714040067682369717e+00, -7.89985396684158212226e-17,
+  1.09050773266525768967e+00, -3.04678207981247114697e-17,
+  1.11438674259589243221e+00, 1.04102784568455709549e-16,
+  1.13878863475669156458e+00, 8.91281267602540777782e-17,
+  1.16372485877757747552e+00, 3.82920483692409349872e-17,
+  1.18920711500272102690e+00, 3.98201523146564611098e-17,
+  1.21524735998046895524e+00, -7.71263069268148813091e-17,
+  1.24185781207348400201e+00, 4.65802759183693679123e-17,
+  1.26905095719173321989e+00, 2.66793213134218609523e-18,
+  1.29683955465100964055e+00, 2.53825027948883149593e-17,
+  1.32523664315974132322e+00, -2.85873121003886075697e-17,
+  1.35425554693689265129e+00, 7.70094837980298946162e-17,
+  1.38390988196383202258e+00, -6.77051165879478628716e-17,
+  1.41421356237309514547e+00, -9.66729331345291345105e-17,
+  1.44518080697704665027e+00, -3.02375813499398731940e-17,
+  1.47682614593949934623e+00, -3.48399455689279579579e-17,
+  1.50916442759342284141e+00, -1.01645532775429503911e-16,
+  1.54221082540794074411e+00, 7.94983480969762085616e-17,
+  1.57598084510788649659e+00, -1.01369164712783039808e-17,
+  1.61049033194925428347e+00, 2.47071925697978878522e-17,
+  1.64575547815396494578e+00, -1.01256799136747726038e-16,
+  1.68179283050742900407e+00, 8.19901002058149652013e-17,
+  1.71861929812247793414e+00, -1.85138041826311098821e-17,
+  1.75625216037329945351e+00, 2.96014069544887330703e-17,
+  1.79470907500310716820e+00, 1.82274584279120867698e-17,
+  1.83400808640934243066e+00, 3.28310722424562658722e-17,
+  1.87416763411029996256e+00, -6.12276341300414256164e-17,
+  1.91520656139714740007e+00, -1.06199460561959626376e-16,
+  1.95714412417540017941e+00, 8.96076779103666776760e-17,
+};
+
+/*
+ * For i = 0, ..., 66,
+ *   TBL2[2*i] is a double precision number near (i+1)*2^-6, and
+ *   TBL2[2*i+1] = exp(TBL2[2*i]) to within a relative error less
+ *   than 2^-60.
+ *
+ * For i = 67, ..., 133,
+ *   TBL2[2*i] is a double precision number near -(i+1)*2^-6, and
+ *   TBL2[2*i+1] = exp(TBL2[2*i]) to within a relative error less
+ *   than 2^-60.
+ */
+static const double TBL2[] = {
+  1.56249999999984491572e-02, 1.01574770858668417262e+00,
+  3.12499999999998716305e-02, 1.03174340749910253834e+00,
+  4.68750000000011102230e-02, 1.04799100201663386578e+00,
+  6.24999999999990632493e-02, 1.06449445891785843266e+00,
+  7.81249999999999444888e-02, 1.08125780744903954300e+00,
+  9.37500000000013322676e-02, 1.09828514030782731226e+00,
+  1.09375000000001346145e-01, 1.11558061464248226002e+00,
+  1.24999999999999417133e-01, 1.13314845306682565607e+00,
+  1.40624999999995337063e-01, 1.15099294469117108264e+00,
+  1.56249999999996141975e-01, 1.16911844616949989195e+00,
+  1.71874999999992894573e-01, 1.18752938276309216725e+00,
+  1.87500000000000888178e-01, 1.20623024942098178158e+00,
+  2.03124999999361649516e-01, 1.22522561187652545556e+00,
+  2.18750000000000416334e-01, 1.24452010776609567344e+00,
+  2.34375000000003524958e-01, 1.26411844775347081971e+00,
+  2.50000000000006328271e-01, 1.28402541668774961003e+00,
+  2.65624999999982791543e-01, 1.30424587476761533189e+00,
+  2.81249999999993727240e-01, 1.32478475872885725906e+00,
+  2.96875000000003275158e-01, 1.34564708304941493822e+00,
+  3.12500000000002886580e-01, 1.36683794117380030819e+00,
+  3.28124999999993394173e-01, 1.38836250675661765364e+00,
+  3.43749999999998612221e-01, 1.41022603492570874906e+00,
+  3.59374999999992450483e-01, 1.43243386356506730017e+00,
+  3.74999999999991395772e-01, 1.45499141461818881638e+00,
+  3.90624999999997613020e-01, 1.47790419541173490003e+00,
+  4.06249999999991895372e-01, 1.50117780000011058483e+00,
+  4.21874999999996613820e-01, 1.52481791053132154090e+00,
+  4.37500000000004607426e-01, 1.54883029863414023453e+00,
+  4.53125000000004274359e-01, 1.57322082682725961078e+00,
+  4.68750000000008326673e-01, 1.59799544995064657371e+00,
+  4.84374999999985456078e-01, 1.62316021661928200359e+00,
+  4.99999999999997335465e-01, 1.64872127070012375327e+00,
+  5.15625000000000222045e-01, 1.67468485281178436352e+00,
+  5.31250000000003441691e-01, 1.70105730184840653330e+00,
+  5.46874999999999111822e-01, 1.72784505652716169344e+00,
+  5.62499999999999333866e-01, 1.75505465696029738787e+00,
+  5.78124999999993338662e-01, 1.78269274625180318417e+00,
+  5.93749999999999666933e-01, 1.81076607211938656050e+00,
+  6.09375000000003441691e-01, 1.83928148854178719063e+00,
+  6.24999999999995559108e-01, 1.86824595743221411048e+00,
+  6.40625000000009103829e-01, 1.89766655033813602671e+00,
+  6.56249999999993782751e-01, 1.92755045016753268072e+00,
+  6.71875000000002109424e-01, 1.95790495294292221651e+00,
+  6.87499999999992450483e-01, 1.98873746958227681780e+00,
+  7.03125000000004996004e-01, 2.02005552770870666635e+00,
+  7.18750000000007105427e-01, 2.05186677348799140219e+00,
+  7.34375000000008770762e-01, 2.08417897349558689513e+00,
+  7.49999999999983901766e-01, 2.11700001661264058939e+00,
+  7.65624999999997002398e-01, 2.15033791595229351046e+00,
+  7.81250000000005884182e-01, 2.18420081081563077774e+00,
+  7.96874999999991451283e-01, 2.21859696867912603579e+00,
+  8.12500000000000000000e-01, 2.25353478721320854561e+00,
+  8.28125000000008215650e-01, 2.28902279633221983346e+00,
+  8.43749999999997890576e-01, 2.32506966027711614586e+00,
+  8.59374999999999444888e-01, 2.36168417973090827289e+00,
+  8.75000000000003219647e-01, 2.39887529396710563745e+00,
+  8.90625000000013433699e-01, 2.43665208303232461162e+00,
+  9.06249999999980571097e-01, 2.47502376996297712708e+00,
+  9.21874999999984456878e-01, 2.51399972303748420188e+00,
+  9.37500000000001887379e-01, 2.55358945806293169412e+00,
+  9.53125000000003330669e-01, 2.59380264069854327147e+00,
+  9.68749999999989119814e-01, 2.63464908881560244680e+00,
+  9.84374999999997890576e-01, 2.67613877489447116176e+00,
+  1.00000000000001154632e+00, 2.71828182845907662113e+00,
+  1.01562499999999333866e+00, 2.76108853855008318234e+00,
+  1.03124999999995980993e+00, 2.80456935623711389738e+00,
+  1.04687499999999933387e+00, 2.84873489717039740654e+00,
+  -1.56249999999999514277e-02, 9.84496437005408453480e-01,
+  -3.12499999999955972718e-02, 9.69233234476348348707e-01,
+  -4.68749999999993824384e-02, 9.54206665969188905230e-01,
+  -6.24999999999976130205e-02, 9.39413062813478028090e-01,
+  -7.81249999999989314103e-02, 9.24848813216205822840e-01,
+  -9.37499999999995975442e-02, 9.10510361380034494161e-01,
+  -1.09374999999998584466e-01, 8.96394206635151680196e-01,
+  -1.24999999999998556710e-01, 8.82496902584596676355e-01,
+  -1.40624999999999361622e-01, 8.68815056262843721235e-01,
+  -1.56249999999999111822e-01, 8.55345327307423297647e-01,
+  -1.71874999999924144012e-01, 8.42084427143446223596e-01,
+  -1.87499999999996752598e-01, 8.29029118180403035154e-01,
+  -2.03124999999988037347e-01, 8.16176213022349550386e-01,
+  -2.18749999999995947686e-01, 8.03522573689063990265e-01,
+  -2.34374999999996419531e-01, 7.91065110850298847112e-01,
+  -2.49999999999996280753e-01, 7.78800783071407765057e-01,
+  -2.65624999999999888978e-01, 7.66726596070820165529e-01,
+  -2.81249999999989397370e-01, 7.54839601989015340777e-01,
+  -2.96874999999996114219e-01, 7.43136898668761203268e-01,
+  -3.12499999999999555911e-01, 7.31615628946642115871e-01,
+  -3.28124999999993782751e-01, 7.20272979955444259126e-01,
+  -3.43749999999997946087e-01, 7.09106182437399867879e-01,
+  -3.59374999999994337863e-01, 6.98112510068129799023e-01,
+  -3.74999999999994615418e-01, 6.87289278790975899369e-01,
+  -3.90624999999999000799e-01, 6.76633846161729612945e-01,
+  -4.06249999999947264406e-01, 6.66143610703522903727e-01,
+  -4.21874999999988453681e-01, 6.55816011271509125002e-01,
+  -4.37499999999999111822e-01, 6.45648526427892610613e-01,
+  -4.53124999999999278355e-01, 6.35638673826052436056e-01,
+  -4.68749999999999278355e-01, 6.25784009604591573428e-01,
+  -4.84374999999992894573e-01, 6.16082127790682609891e-01,
+  -4.99999999999998168132e-01, 6.06530659712634534486e-01,
+  -5.15625000000000000000e-01, 5.97127273421627413619e-01,
+  -5.31249999999989785948e-01, 5.87869673122352498496e-01,
+  -5.46874999999972688514e-01, 5.78755598612500032907e-01,
+  -5.62500000000000000000e-01, 5.69782824730923009859e-01,
+  -5.78124999999992339461e-01, 5.60949160814475100700e-01,
+  -5.93749999999948707696e-01, 5.52252450163048691500e-01,
+  -6.09374999999552580121e-01, 5.43690569513243682209e-01,
+  -6.24999999999984789945e-01, 5.35261428518998383375e-01,
+  -6.40624999999983457677e-01, 5.26962969243379708573e-01,
+  -6.56249999999998334665e-01, 5.18793165653890220312e-01,
+  -6.71874999999943378626e-01, 5.10750023129039609771e-01,
+  -6.87499999999997002398e-01, 5.02831577970942467104e-01,
+  -7.03124999999991118216e-01, 4.95035896926202978463e-01,
+  -7.18749999999991340260e-01, 4.87361076713623331269e-01,
+  -7.34374999999985678123e-01, 4.79805243559684402310e-01,
+  -7.49999999999997335465e-01, 4.72366552741015965911e-01,
+  -7.65624999999993782751e-01, 4.65043188134059204408e-01,
+  -7.81249999999863220523e-01, 4.57833361771676883301e-01,
+  -7.96874999999998112621e-01, 4.50735313406363247157e-01,
+  -8.12499999999990119015e-01, 4.43747310081084256339e-01,
+  -8.28124999999996003197e-01, 4.36867645705559026759e-01,
+  -8.43749999999988120614e-01, 4.30094640640067360504e-01,
+  -8.59374999999994115818e-01, 4.23426641285265303871e-01,
+  -8.74999999999977129406e-01, 4.16862019678517936594e-01,
+  -8.90624999999983346655e-01, 4.10399173096376801428e-01,
+  -9.06249999999991784350e-01, 4.04036523663345414903e-01,
+  -9.21874999999994004796e-01, 3.97772517966614058693e-01,
+  -9.37499999999994337863e-01, 3.91605626676801210628e-01,
+  -9.53124999999999444888e-01, 3.85534344174578935682e-01,
+  -9.68749999999986677324e-01, 3.79557188183094640355e-01,
+  -9.84374999999992339461e-01, 3.73672699406045860648e-01,
+  -9.99999999999995892175e-01, 3.67879441171443832825e-01,
+  -1.01562499999994315658e+00, 3.62175999080846300338e-01,
+  -1.03124999999991096011e+00, 3.56560980663978732697e-01,
+  -1.04687499999999067413e+00, 3.51033015038813400732e-01,
+};
+
+static const double zC[] = {
+  0.5,
+  4.61662413084468283841e+01,	/* 0x40471547, 0x652b82fe */
+  2.16608493865351192653e-02,	/* 0x3f962e42, 0xfee00000 */
+  5.96317165397058656257e-12,	/* 0x3d9a39ef, 0x35793c76 */
+  1.6666666666526086527e-1,	/* 3fc5555555548f7c */
+  4.1666666666226079285e-2,	/* 3fa5555555545d4e */
+  8.3333679843421958056e-3,	/* 3f811115b7aa905e */
+  1.3888949086377719040e-3,	/* 3f56c1728d739765 */
+  1.0,
+  0.0,
+  7.09782712893383973096e+02,	/* 0x40862E42, 0xFEFA39EF */
+  7.45133219101941108420e+02,	/* 0x40874910, 0xD52D3051 */
+  5.55111512312578270212e-17,	/* 0x3c900000, 0x00000000 */
+  1.0e-100,
+  1.0e-300,
+  1.0e300
+};
+
+#ifndef DBL_MIN
+#define DBL_MIN 2.2250738585072014E-308
+#endif
+
+#define	half		zC[0]
+#define	invln2_32	zC[1]
+#define	ln2_32hi	zC[2]
+#define	ln2_32lo	zC[3]
+#define	t2		zC[4]
+#define	t3		zC[5]
+#define	t4		zC[6]
+#define	t5		zC[7]
+#define	one		zC[8]
+#define	zero		zC[9]
+#define	threshold1	zC[10]
+#define	threshold2	zC[11]
+#define	twom54		zC[12]
+#define	small		zC[13]
+#define	tiny		zC[14]
+#define	hhuge		zC[15]
+
+
+void
+__set_err_exp (double x, double result)
+{
+  if (_LIB_VERSION != _IEEE_)
+    {
+      if (_LIB_VERSION == _POSIX_)
+	{
+	  __set_errno (ERANGE);
+	}
+      else
+	{
+	  struct exception exc;
+	  exc.arg1 = x;
+	  exc.type = DOMAIN;
+	  exc.name = (char *) "exp";
+	  exc.retval = result;
+	  if (!matherr (&exc))
+	    {
+	      __set_errno (ERANGE);
+	    }
+	}
+    }
+}
+
+
+double
+__exp_niagara4 (double x_arg)
+{
+  double z, t;
+  double retval;
+  int hx, ix, k, j, m;
+  union
+  {
+    int i_part[2];
+    double x;
+  } xx;
+  union
+  {
+    int y_part[2];
+    double y;
+  } yy;
+  xx.x = x_arg;
+
+  ix = xx.i_part[HIWORD];
+  hx = ix & ~0x80000000;
+
+  if (hx < 0x3ff0a2b2)
+    {				/* |x| < 3/2 ln 2 */
+      if (hx < 0x3f862e42)
+	{			/* |x| < 1/64 ln 2 */
+	  if (hx < 0x3ed00000)
+	    {			/* |x| < 2^-18 */
+	      /* raise inexact if x != 0 */
+	      if (hx < 0x3e300000)
+		{
+		  retval = one + xx.x;
+		  return (retval);
+		}
+	      retval = one + xx.x * (one + half * xx.x);
+	      return (retval);
+	    }
+	  t = xx.x * xx.x;
+	  yy.y = xx.x + (t * (half + xx.x * t2) +
+			 (t * t) * (t3 + xx.x * t4 + t * t5));
+	  retval = one + yy.y;
+	  return (retval);
+	}
+
+      /* find the multiple of 2^-6 nearest x */
+      k = hx >> 20;
+      j = (0x00100000 | (hx & 0x000fffff)) >> (0x40c - k);
+      j = (j - 1) & ~1;
+      if (ix < 0)
+	j += 134;
+      z = xx.x - TBL2[j];
+      t = z * z;
+      /* the "small" term below guarantees inexact will be raised */
+      yy.y = z + (t * (half + (z * t2 + small)) +
+		  (t * t) * (t3 + z * t4 + t * t5));
+      retval = TBL2[j + 1] + TBL2[j + 1] * yy.y;
+      return (retval);
+    }
+
+  if (hx >= 0x40862e42)
+    {				/* x is large, infinite, or nan */
+      if (hx >= 0x7ff00000)
+	{
+	  if (ix == 0xfff00000 && xx.i_part[LOWORD] == 0)
+	    return (zero);	/* exp(-inf) = 0 */
+	  return (xx.x * xx.x);	/* exp(nan/inf) is nan or inf */
+	}
+      if (xx.x > threshold1)
+	{			/* set overflow error condition */
+	  retval = hhuge * hhuge;
+	  __set_err_exp (xx.x, retval);
+	  return retval;
+	}
+      if (-xx.x > threshold2)
+	{			/* set underflow error condition */
+	  double force_underflow = tiny * tiny;
+	  math_force_eval (force_underflow);
+	  retval = zero;
+	  __set_err_exp (xx.x, retval);
+	  return retval;
+	}
+    }
+
+  t = invln2_32 * xx.x;
+  if (ix < 0)
+    t -= half;
+  else
+    t += half;
+  k = (int) t;
+  j = (k & 0x1f) << 1;
+  m = k >> 5;
+  z = (xx.x - k * ln2_32hi) - k * ln2_32lo;
+
+  /* z is now in primary range */
+  t = z * z;
+  yy.y = z + (t * (half + z * t2) + (t * t) * (t3 + z * t4 + t * t5));
+  yy.y = TBL[j] + (TBL[j + 1] + TBL[j] * yy.y);
+  if (m < -1021)
+    {
+      yy.y_part[HIWORD] += (m + 54) << 20;
+      retval = twom54 * yy.y;
+      if (retval < DBL_MIN)
+	{
+	  double force_underflow = tiny * tiny;
+	  math_force_eval (force_underflow);
+	  __set_err_exp (xx.x, retval);
+	}
+      return (retval);
+    }
+  yy.y_part[HIWORD] += m << 20;
+  return (yy.y);
+}
diff --git a/sysdeps/sparc/sparc64/fpu/multiarch/e_expf-niagara4.c b/sysdeps/sparc/sparc64/fpu/multiarch/e_expf-niagara4.c
new file mode 100644
index 0000000..ba0b1e7
--- /dev/null
+++ b/sysdeps/sparc/sparc64/fpu/multiarch/e_expf-niagara4.c
@@ -0,0 +1,445 @@ 
+/* EXPF function
+   Contributed by Oracle Inc.
+   Copyright (C) 2017 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+
+/* INDENT OFF */
+/*
+ * float expf(float x);
+ * Code by K.C. Ng for SUN 5.0 libmopt
+ * 11/5/99
+ * Method :
+ *	1. For |x| >= 2^7, either underflow/overflow.
+ *	   More precisely:
+ *		x > 88.722839355...(0x42B17218) => overflow;
+ *		x < -103.97207642..(0xc2CFF1B4) => underflow.
+ *	2. For |x| <  2^-6, use polynomail
+ *		exp(x) = 1 + x + p1*x^2 + p2*x^3
+ *      3. Otherwise, write |x|=(1+r)*2^n, where 0<=r<1.
+ *	   Let t = 2^n * (1+r) .... x > 0;
+ *	       t = 2^n * (1-r) .... x < 0. (x= -2**(n+1)+t)
+ *	   Since -6 <= n <= 6, we may break t into
+ *	   six 6-bits chunks:
+ *                    -5     -11     -17     -23     -29
+ *         t=j *2+j *2  +j *2   +j *2   +j *2   +j *2
+ *            1    2      3       4       5       6
+ *
+ *	   where 0 <= j  < 64 for i = 1,...,6.
+ *		       i
+ *	   Note that since t has only 24 significant bits,
+ *	   either j  or j  must be 0.
+ *		   1     6
+ *					       7-6i
+ *	   One may define j  by   (int) ( t * 2     ) mod 64
+ *			   i
+ *	   mathematically. In actual implementation, they can
+ *	   be obtained by manipulating the exponent and
+ *	   mantissa bits as follow:
+ *		Let ix = (HEX(x)&0x007fffff)|0x00800000.
+ *		If n>=0, let ix=ix<<n, then j =0 and
+ *					     6
+ *		    j  = ix>>(30-6i)) mod 64  ...i=1,...,5
+ *		     i
+ *		Otherwise, let ix=ix<<(j+6), then j = 0 and
+ *						   1
+ *		    j  = ix>>(36-6i)) mod 64  ...i=2,...,6
+ *		     i
+ *
+ *	4. Compute exp(t) by table look-up method.
+ *	   Precompute ET[k] = exp(j*2^(7-6i)), k=j+64*(6-i).
+ *	   Then
+ *	   exp(t) = ET[j +320]*ET[j +256]*ET[j +192]*
+ *		        1          2          3
+ *
+ *		    ET[j +128]*ET[j +64]*ET[j ]
+ *			4          5         6
+ *
+ *				  n+1
+ *	5. If x < 0, return exp(-2   )* exp(t). Note that
+ *	   -6 <= n <= 6. Let k = n - 6, then we can
+ *	   precompute
+ *	                 k-5          n+1
+ *         EN[k] = exp(-2   ) = exp(-2   ) for k=0,1,...,12.
+ *
+ *
+ * Special cases:
+ *	exp(INF) is INF, exp(NaN) is NaN;
+ *	exp(-INF) = 0;
+ *	for finite argument, only exp(0) = 1 is exact.
+ *
+ * Accuracy:
+ *      All calculations are done in double precision except for
+ *      the case |x| < 2^-6.  When |x| < 2^-6, the error is less
+ *      than 0.55 ulp.  When |x| >= 2^-6, the error is less than
+ *      0.51 ulp.
+ */
+/* INDENT ON */
+
+/* #pragma weak expf = __expf_niagara4 */
+#include <sparc-ifunc.h>
+#include "libm_endian.h"
+#include <math.h>
+#include <errno.h>
+
+extern float expf (float);
+extern float __expf_niagara4 (float);
+extern float __ieee754_expf (float);
+
+sparc_libm_ifunc (expf, hwcap & HWCAP_SPARC_CRYPTO ?
+		  __expf_niagara4 : __ieee754_expf);
+
+
+void
+__set_err_expf (float x, float result)
+{
+  if (_LIB_VERSION != _IEEE_)
+    {
+      if (_LIB_VERSION == _POSIX_)
+	{
+	  __set_errno (ERANGE);
+	}
+      else
+	{
+	  struct exception exc;
+	  exc.arg1 = x;
+	  exc.type = DOMAIN;
+	  exc.name = (char *) "exp";
+	  exc.retval = result;
+	  if (!matherr (&exc))
+	    {
+	      __set_errno (ERANGE);
+	    }
+	}
+    }
+}
+
+
+/*
+ * ET[k] = exp(j*2^(7-6i)) , where j = k mod 64, i = k/64
+ */
+static const double ET[] = {
+  1.00000000000000000000e+00, 1.00000000186264514923e+00,
+  1.00000000372529029846e+00, 1.00000000558793544769e+00,
+  1.00000000745058059692e+00, 1.00000000931322574615e+00,
+  1.00000001117587089539e+00, 1.00000001303851604462e+00,
+  1.00000001490116119385e+00, 1.00000001676380656512e+00,
+  1.00000001862645171435e+00, 1.00000002048909686359e+00,
+  1.00000002235174201282e+00, 1.00000002421438716205e+00,
+  1.00000002607703253332e+00, 1.00000002793967768255e+00,
+  1.00000002980232283178e+00, 1.00000003166496798102e+00,
+  1.00000003352761335229e+00, 1.00000003539025850152e+00,
+  1.00000003725290365075e+00, 1.00000003911554879998e+00,
+  1.00000004097819417126e+00, 1.00000004284083932049e+00,
+  1.00000004470348446972e+00, 1.00000004656612984100e+00,
+  1.00000004842877499023e+00, 1.00000005029142036150e+00,
+  1.00000005215406551073e+00, 1.00000005401671088201e+00,
+  1.00000005587935603124e+00, 1.00000005774200140252e+00,
+  1.00000005960464655175e+00, 1.00000006146729192302e+00,
+  1.00000006332993707225e+00, 1.00000006519258244353e+00,
+  1.00000006705522759276e+00, 1.00000006891787296404e+00,
+  1.00000007078051811327e+00, 1.00000007264316348454e+00,
+  1.00000007450580863377e+00, 1.00000007636845400505e+00,
+  1.00000007823109937632e+00, 1.00000008009374452556e+00,
+  1.00000008195638989683e+00, 1.00000008381903526811e+00,
+  1.00000008568168063938e+00, 1.00000008754432578861e+00,
+  1.00000008940697115989e+00, 1.00000009126961653116e+00,
+  1.00000009313226190244e+00, 1.00000009499490705167e+00,
+  1.00000009685755242295e+00, 1.00000009872019779422e+00,
+  1.00000010058284316550e+00, 1.00000010244548853677e+00,
+  1.00000010430813368600e+00, 1.00000010617077905728e+00,
+  1.00000010803342442856e+00, 1.00000010989606979983e+00,
+  1.00000011175871517111e+00, 1.00000011362136054238e+00,
+  1.00000011548400591366e+00, 1.00000011734665128493e+00,
+  1.00000000000000000000e+00, 1.00000011920929665621e+00,
+  1.00000023841860752327e+00, 1.00000035762793260119e+00,
+  1.00000047683727188996e+00, 1.00000059604662538959e+00,
+  1.00000071525599310007e+00, 1.00000083446537502141e+00,
+  1.00000095367477115360e+00, 1.00000107288418149665e+00,
+  1.00000119209360605055e+00, 1.00000131130304481530e+00,
+  1.00000143051249779091e+00, 1.00000154972196497738e+00,
+  1.00000166893144637470e+00, 1.00000178814094198287e+00,
+  1.00000190735045180190e+00, 1.00000202655997583179e+00,
+  1.00000214576951407253e+00, 1.00000226497906652412e+00,
+  1.00000238418863318657e+00, 1.00000250339821405987e+00,
+  1.00000262260780914403e+00, 1.00000274181741843904e+00,
+  1.00000286102704194491e+00, 1.00000298023667966163e+00,
+  1.00000309944633158921e+00, 1.00000321865599772764e+00,
+  1.00000333786567807692e+00, 1.00000345707537263706e+00,
+  1.00000357628508140806e+00, 1.00000369549480438991e+00,
+  1.00000381470454158261e+00, 1.00000393391429298617e+00,
+  1.00000405312405860059e+00, 1.00000417233383842586e+00,
+  1.00000429154363246198e+00, 1.00000441075344070896e+00,
+  1.00000452996326316679e+00, 1.00000464917309983548e+00,
+  1.00000476838295071502e+00, 1.00000488759281580542e+00,
+  1.00000500680269510667e+00, 1.00000512601258861878e+00,
+  1.00000524522249634174e+00, 1.00000536443241827556e+00,
+  1.00000548364235442023e+00, 1.00000560285230477575e+00,
+  1.00000572206226934213e+00, 1.00000584127224811937e+00,
+  1.00000596048224110746e+00, 1.00000607969224830640e+00,
+  1.00000619890226971620e+00, 1.00000631811230533685e+00,
+  1.00000643732235516836e+00, 1.00000655653241921073e+00,
+  1.00000667574249746394e+00, 1.00000679495258992802e+00,
+  1.00000691416269660294e+00, 1.00000703337281748873e+00,
+  1.00000715258295258536e+00, 1.00000727179310189285e+00,
+  1.00000739100326541120e+00, 1.00000751021344314040e+00,
+  1.00000000000000000000e+00, 1.00000762942363508046e+00,
+  1.00001525890547848796e+00, 1.00002288844553022251e+00,
+  1.00003051804379095024e+00, 1.00003814770026133729e+00,
+  1.00004577741494138365e+00, 1.00005340718783175546e+00,
+  1.00006103701893311886e+00, 1.00006866690824547383e+00,
+  1.00007629685576948653e+00, 1.00008392686150582307e+00,
+  1.00009155692545448346e+00, 1.00009918704761613384e+00,
+  1.00010681722799144033e+00, 1.00011444746658040295e+00,
+  1.00012207776338368781e+00, 1.00012970811840196106e+00,
+  1.00013733853163522269e+00, 1.00014496900308413885e+00,
+  1.00015259953274937565e+00, 1.00016023012063093311e+00,
+  1.00016786076672947736e+00, 1.00017549147104567453e+00,
+  1.00018312223357952462e+00, 1.00019075305433191581e+00,
+  1.00019838393330284809e+00, 1.00020601487049298761e+00,
+  1.00021364586590300050e+00, 1.00022127691953288675e+00,
+  1.00022890803138353455e+00, 1.00023653920145494389e+00,
+  1.00024417042974778091e+00, 1.00025180171626271175e+00,
+  1.00025943306099973640e+00, 1.00026706446395974304e+00,
+  1.00027469592514273167e+00, 1.00028232744454959047e+00,
+  1.00028995902218031944e+00, 1.00029759065803558471e+00,
+  1.00030522235211605242e+00, 1.00031285410442172257e+00,
+  1.00032048591495348333e+00, 1.00032811778371155675e+00,
+  1.00033574971069616488e+00, 1.00034338169590819589e+00,
+  1.00035101373934764979e+00, 1.00035864584101541475e+00,
+  1.00036627800091149076e+00, 1.00037391021903676602e+00,
+  1.00038154249539146257e+00, 1.00038917482997580244e+00,
+  1.00039680722279067382e+00, 1.00040443967383629875e+00,
+  1.00041207218311289928e+00, 1.00041970475062136359e+00,
+  1.00042733737636191371e+00, 1.00043497006033499375e+00,
+  1.00044260280254104778e+00, 1.00045023560298029786e+00,
+  1.00045786846165363215e+00, 1.00046550137856127272e+00,
+  1.00047313435370366363e+00, 1.00048076738708124900e+00,
+  1.00000000000000000000e+00, 1.00048840047869447289e+00,
+  1.00097703949241645383e+00, 1.00146591715766675179e+00,
+  1.00195503359100279717e+00, 1.00244438890903908579e+00,
+  1.00293398322844673487e+00, 1.00342381666595459322e+00,
+  1.00391388933834746489e+00, 1.00440420136246855165e+00,
+  1.00489475285521656645e+00, 1.00538554393354861993e+00,
+  1.00587657471447822211e+00, 1.00636784531507639251e+00,
+  1.00685935585247099411e+00, 1.00735110644384739942e+00,
+  1.00784309720644804642e+00, 1.00833532825757243856e+00,
+  1.00882779971457803292e+00, 1.00932051169487890796e+00,
+  1.00981346431594687374e+00, 1.01030665769531102782e+00,
+  1.01080009195055753324e+00, 1.01129376719933050666e+00,
+  1.01178768355933157430e+00, 1.01228184114831898377e+00,
+  1.01277624008410960244e+00, 1.01327088048457714109e+00,
+  1.01376576246765282008e+00, 1.01426088615132625748e+00,
+  1.01475625165364347069e+00, 1.01525185909270931894e+00,
+  1.01574770858668572693e+00, 1.01624380025379235093e+00,
+  1.01674013421230657883e+00, 1.01723671058056375216e+00,
+  1.01773352947695694404e+00, 1.01823059101993673714e+00,
+  1.01872789532801233392e+00, 1.01922544251975000229e+00,
+  1.01972323271377418585e+00, 1.02022126602876750390e+00,
+  1.02071954258347008526e+00, 1.02121806249668067856e+00,
+  1.02171682588725554197e+00, 1.02221583287410910934e+00,
+  1.02271508357621376817e+00, 1.02321457811260052573e+00,
+  1.02371431660235789884e+00, 1.02421429916463280207e+00,
+  1.02471452591863054771e+00, 1.02521499698361440167e+00,
+  1.02571571247890602763e+00, 1.02621667252388526492e+00,
+  1.02671787723799012859e+00, 1.02721932674071725344e+00,
+  1.02772102115162167202e+00, 1.02822296059031659254e+00,
+  1.02872514517647339893e+00, 1.02922757502982276101e+00,
+  1.02973025027015285815e+00, 1.03023317101731093359e+00,
+  1.03073633739120262831e+00, 1.03123974951179242510e+00,
+  1.00000000000000000000e+00, 1.03174340749910276038e+00,
+  1.06449445891785954288e+00, 1.09828514030782575794e+00,
+  1.13314845306682632220e+00, 1.16911844616950433284e+00,
+  1.20623024942098067136e+00, 1.24452010776609522935e+00,
+  1.28402541668774139438e+00, 1.32478475872886569675e+00,
+  1.36683794117379631139e+00, 1.41022603492571074746e+00,
+  1.45499141461820125087e+00, 1.50117780000012279729e+00,
+  1.54883029863413312910e+00, 1.59799544995063325104e+00,
+  1.64872127070012819416e+00, 1.70105730184840076014e+00,
+  1.75505465696029849809e+00, 1.81076607211938722664e+00,
+  1.86824595743222232613e+00, 1.92755045016754467113e+00,
+  1.98873746958229191684e+00, 2.05186677348797674725e+00,
+  2.11700001661267478426e+00, 2.18420081081561789915e+00,
+  2.25353478721320854561e+00, 2.32506966027712103084e+00,
+  2.39887529396709808793e+00, 2.47502376996302508871e+00,
+  2.55358945806292680913e+00, 2.63464908881563086851e+00,
+  2.71828182845904553488e+00, 2.80456935623722669604e+00,
+  2.89359594417176113623e+00, 2.98544853936535581340e+00,
+  3.08021684891803104733e+00, 3.17799342753883840018e+00,
+  3.27887376793867346692e+00, 3.38295639409246895468e+00,
+  3.49034295746184142217e+00, 3.60113833627217561073e+00,
+  3.71545073794110392029e+00, 3.83339180475841034834e+00,
+  3.95507672292057721464e+00, 4.08062433502646015882e+00,
+  4.21015725614395996956e+00, 4.34380199356104235164e+00,
+  4.48168907033806451778e+00, 4.62395315278208052234e+00,
+  4.77073318196760265408e+00, 4.92217250943229078786e+00,
+  5.07841903718008147450e+00, 5.23962536212848917216e+00,
+  5.40594892514116676097e+00, 5.57755216479125959239e+00,
+  5.75460267600573072144e+00, 5.93727337374560715233e+00,
+  6.12574266188198635064e+00, 6.32019460743274397174e+00,
+  6.52081912033011246166e+00, 6.72781213889469142941e+00,
+  6.94137582119703555605e+00, 7.16171874249371143151e+00,
+  1.00000000000000000000e+00, 7.38905609893065040694e+00,
+  5.45981500331442362040e+01, 4.03428793492735110249e+02,
+  2.98095798704172830185e+03, 2.20264657948067178950e+04,
+  1.62754791419003915507e+05, 1.20260428416477679275e+06,
+  8.88611052050787210464e+06, 6.56599691373305097222e+07,
+  4.85165195409790277481e+08, 3.58491284613159179688e+09,
+  2.64891221298434715271e+10, 1.95729609428838775635e+11,
+  1.44625706429147509766e+12, 1.06864745815244628906e+13,
+  7.89629601826806875000e+13, 5.83461742527454875000e+14,
+  4.31123154711519500000e+15, 3.18559317571137560000e+16,
+  2.35385266837020000000e+17, 1.73927494152050099200e+18,
+  1.28516001143593082880e+19, 9.49611942060244828160e+19,
+  7.01673591209763143680e+20, 5.18470552858707204506e+21,
+  3.83100800071657691546e+22, 2.83075330327469394756e+23,
+  2.09165949601299610311e+24, 1.54553893559010391826e+25,
+  1.14200738981568423454e+26, 8.43835666874145383188e+26,
+  6.23514908081161674391e+27, 4.60718663433129178064e+28,
+  3.40427604993174075827e+29, 2.51543867091916687979e+30,
+  1.85867174528412788702e+31, 1.37338297954017610775e+32,
+  1.01480038811388874615e+33, 7.49841699699012090701e+33,
+  5.54062238439350983445e+34, 4.09399696212745451138e+35,
+  3.02507732220114256223e+36, 2.23524660373471497416e+37,
+  1.65163625499400180987e+38, 1.22040329431784083418e+39,
+  9.01762840503429851945e+39, 6.66317621641089618500e+40,
+  4.92345828601205826106e+41, 3.63797094760880474988e+42,
+  2.68811714181613560943e+43, 1.98626483613765434356e+44,
+  1.46766223015544238535e+45, 1.08446385529002313207e+46,
+  8.01316426400059069850e+46, 5.92097202766466993617e+47,
+  4.37503944726134096988e+48, 3.23274119108485947460e+49,
+  2.38869060142499127023e+50, 1.76501688569176554670e+51,
+  1.30418087839363225614e+52, 9.63666567360320166416e+52,
+  7.12058632688933793173e+53, 5.26144118266638596909e+54,
+};
+
+/*
+ * EN[k] = exp(-2^(k-5))
+ */
+static const double EN[] = {
+  9.69233234476344129860e-01, 9.39413062813475807644e-01,
+  8.82496902584595455110e-01, 7.78800783071404878477e-01,
+  6.06530659712633424263e-01, 3.67879441171442334024e-01,
+  1.35335283236612702318e-01, 1.83156388887341786686e-02,
+  3.35462627902511853224e-04, 1.12535174719259116458e-07,
+  1.26641655490941755372e-14, 1.60381089054863792659e-28,
+  2.57220937264241481170e-56,
+};
+
+static const float zF[] = {
+  0.0f,
+  1.0f,
+  5.0000000951292138e-01F,
+  1.6666518897347284e-01F,
+  3.4028234663852885981170E+38F,
+  1.1754943508222875079688E-38F,
+  88.72283935546875,
+  -103.972076416
+};
+
+#define	zero	zF[0]
+#define	one	zF[1]
+#define	p1	zF[2]
+#define	p2	zF[3]
+#define	big	zF[4]
+#define	tiny	zF[5]
+#define	thresh1	zF[6]
+#define	thresh2	zF[7]
+
+
+float
+__expf_niagara4 (float xf)
+{
+  double w, p, q;
+  int hx, ix, n;
+  float retval;
+  union
+  {
+    int hx_part;
+    float my_xf;
+  } hxx;
+  hxx.my_xf = xf;
+  hx = hxx.hx_part;
+  ix = hx & ~0x80000000;
+
+  if (ix < 0x3c800000)
+    {				/* |x| < 2**-6 */
+      if (ix < 0x38800000)	/* |x| < 2**-14 */
+	return (one + xf);
+      return (one + (xf + (xf * xf) * (p1 + xf * p2)));
+    }
+
+  n = ix >> 23;			/* biased exponent */
+
+  if (n >= 0x85)
+    {				/* |x| >= 2^6 */
+      if (n >= 0x86)
+	{			/* |x| >= 2^7 */
+	  if (n >= 0xff)
+	    {			/* x is nan of +-inf */
+	      if (hx == 0xff800000)
+		return (zero);	/* exp(-inf)=0 */
+	      return (xf * xf);	/* exp(nan/inf) is nan or inf */
+	    }
+	  if (hx > 0)
+	    {
+	      retval = big * big;	/* overflow */
+	      __set_err_expf (hxx.my_xf, retval);
+	      return (retval);
+	    }
+	  else
+	    {
+	      retval = tiny * tiny;	/* underflow */
+	      __set_err_expf (hxx.my_xf, retval);
+	      return (retval);
+	    }
+	}
+      if (hxx.my_xf >= thresh1)
+	{
+	  retval = big * big;	/* overflow */
+	  __set_err_expf (hxx.my_xf, retval);
+	  return (retval);
+	}
+      if (hxx.my_xf <= thresh2)
+	{
+	  retval = tiny * tiny;	/* underflow */
+	  __set_err_expf (hxx.my_xf, retval);
+	  return (retval);
+	}
+    }
+  ix -= n << 23;
+  if (hx > 0)
+    ix += 0x800000;
+  else
+    ix = 0x800000 - ix;
+  if (n >= 0x7f)
+    {				/* n >= 0 */
+      ix <<= n - 0x7f;
+      w = ET[(ix & 0x3f) + 64] * ET[((ix >> 6) & 0x3f) + 128];
+      p = ET[((ix >> 12) & 0x3f) + 192] * ET[((ix >> 18) & 0x3f) + 256];
+      q = ET[((ix >> 24) & 0x3f) + 320];
+    }
+  else
+    {
+      ix <<= n - 0x79;
+      w = ET[ix & 0x3f] * ET[((ix >> 6) & 0x3f) + 64];
+      p = ET[((ix >> 12) & 0x3f) + 128] * ET[((ix >> 18) & 0x3f) + 192];
+      q = ET[((ix >> 24) & 0x3f) + 256];
+    }
+  xf = (float) ((w * p) * (hx < 0 ? q * EN[n - 0x79] : q));
+  return (xf);
+}
-- 
1.7.1