without wrapper on aarch64:
powf reciprocal-throughput: 4.2x faster
powf latency: 2.6x faster
old worst-case error: 1.11 ulp
new worst-case error: 0.82 ulp
aarch64 .text size: -780 bytes
aarch64 .rodata size: +144 bytes
powf(x,y) is implemented as exp2(y*log2(x)) with the same algorithms
that are used in exp2f and log2f, except that the log2f polynomial is
larger for extra precision and its output (and exp2f input) may be
scaled by a power of 2 (POWF_SCALE) to simplify the argument reduction
step of exp2 (possible when efficient round and convert toint operation
is available).
The special case handling tries to minimize the checks in the hot path.
When the input of exp2_inline is checked, int arithmetics is used as
that was faster on the tested aarch64 cores.
* math/Makefile (type-float-routines): Add e_powf_log2_data.
* sysdeps/ieee754/flt-32/e_powf.c: New implementation.
* sysdeps/ieee754/flt-32/e_powf_log2_data.c: New file.
* sysdeps/ieee754/flt-32/math_config.h (__powf_log2_data): Define.
(issignalingf_inline): Likewise.
(POWF_LOG2_TABLE_BITS): Likewise.
(POWF_LOG2_POLY_ORDER): Likewise.
(POWF_SCALE_BITS): Likewise.
(POWF_SCALE): Likewise.
* sysdeps/i386/fpu/e_powf_log2_data.c: New file.
* sysdeps/ia64/fpu/e_powf_log2_data.c: New file.
* sysdeps/m68k/m680x0/fpu/e_powf_log2_data.c: New file.
Most significant changes are code simplification and use of doubles for
intermediate values. Also, some rearrangement to move early
non-dependent code later, out of the faster paths.
* sysdeps/ieee754/flt-32/e_powf.c: Optimized implementation utilizing
rearranged code and doubles float types.
Bug 21112 reports a case where powf is substantially inaccurate. This
results from a multiplication where cp_h*p_h is required to be exact,
and p_h is masked to have only 12 leading nonzero bits in its
mantissa, but the value of cp_h has the 13th bit nonzero, leading to
inexact multiplication results in some cases that can result in large
errors in the final result of powf. This patch fixes this by using a
value of cp_h correctly rounded to nearest to 12 bits, with a
corresponding updated value of cp_l.
Tested for x86_64 and x86.
[BZ #21112]
* sysdeps/ieee754/flt-32/e_powf.c (cp_h): Use value with trailing
12 bits zero.
(cp_l): Update for new value of cp_h.
* math/auto-libm-test-in: Add another test of pow.
* math/auto-libm-test-out-pow: Regenerated.
Various pow function implementations mishandle sNaN arguments in
various ways. This includes returning sNaN instead of qNaN for sNaN
arguments. For arguments (1, sNaN) and (sNaN, 0), TS 18661-1
semantics are also that the result should be qNaN, whereas with a qNaN
argument there the result should be 1, but for the dbl-64
implementation of pow there are issues with sNaN arguments beyond not
implementing the TS 18661-1 semantics in those special cases.
This patch makes the implementations in sysdeps/ieee754 follow the TS
18661-1 semantics consistently. Because x86 / x86_64 implementations
still need fixing, testcases are not included with this patch; they
will be included with the fix for the x86 / x86_64 versions.
Tested for x86_64, x86, mips64 and powerpc (with such testcases, which
pass in the mips64 and powerpc cases).
[BZ #20916]
* sysdeps/ieee754/dbl-64/e_pow.c (__ieee754_pow): Do not return 1
for arguments (sNaN, 0) or (1, sNaN). Do arithmetic on NaN
arguments to compute result.
* sysdeps/ieee754/flt-32/e_powf.c (__ieee754_powf): Do not return
1 for arguments (sNaN, 0) or (1, sNaN).
* sysdeps/ieee754/ldbl-128/e_powl.c (__ieee754_powl): Likewise.
* sysdeps/ieee754/ldbl-128ibm/e_powl.c (__ieee754_powl): Likewise.
The flt-32 version of powf can be inaccurate because of bugs in the
extra-precision calculation of (x-1)/(x+1) or (x-1.5)/(x+1.5) as part
of calculating log(x) with extra precision: a constant used (as part
of adding 1 or 1.5 through integer arithmetic) is incorrect, and then
the code fails to mask a computed high part before using it in
arithmetic that relies on s_h*t_h being exactly representable. This
patch fixes these bugs.
Tested for x86_64 and x86. x86_64 ulps for powf removed and
regenerated to reflect reduced ulps from the increased accuracy for
existing tests.
[BZ #18956]
* sysdeps/ieee754/flt-32/e_powf.c (__ieee754_powf): Add 0x00400000
not 0x0040000 for high bit of mantissa. Mask with 0xfffff000 when
extracting high part.
* math/auto-libm-test-in: Add another test of pow.
* math/auto-libm-test-out: Regenerated.
* sysdeps/x86_64/fpu/libm-test-ulps: Update.
Similar to various other bugs in this area, pow functions can fail to
raise the underflow exception when the result is tiny and inexact but
one or more low bits of the intermediate result that is scaled down
(or, in the i386 case, converted from a wider evaluation format) are
zero. This patch forces the exception in a similar way to previous
fixes, thereby concluding the fixes for known bugs with missing
underflow exceptions currently filed in Bugzilla.
Tested for x86_64, x86, mips64 and powerpc.
[BZ #18825]
* sysdeps/i386/fpu/i386-math-asm.h (FLT_NARROW_EVAL_UFLOW_NONNAN):
New macro.
(DBL_NARROW_EVAL_UFLOW_NONNAN): Likewise.
(LDBL_CHECK_FORCE_UFLOW_NONNAN): Likewise.
* sysdeps/i386/fpu/e_pow.S: Use DEFINE_DBL_MIN.
(__ieee754_pow): Use DBL_NARROW_EVAL_UFLOW_NONNAN instead of
DBL_NARROW_EVAL, reloading the PIC register as needed.
* sysdeps/i386/fpu/e_powf.S: Use DEFINE_FLT_MIN.
(__ieee754_powf): Use FLT_NARROW_EVAL_UFLOW_NONNAN instead of
FLT_NARROW_EVAL. Use separate return path for case when first
argument is NaN.
* sysdeps/i386/fpu/e_powl.S: Include <i386-math-asm.h>. Use
DEFINE_LDBL_MIN.
(__ieee754_powl): Use LDBL_CHECK_FORCE_UFLOW_NONNAN, reloading the
PIC register.
* sysdeps/ieee754/dbl-64/e_pow.c (__ieee754_pow): Use
math_check_force_underflow_nonneg.
* sysdeps/ieee754/flt-32/e_powf.c (__ieee754_powf): Force
underflow for subnormal result.
* sysdeps/ieee754/ldbl-128/e_powl.c (__ieee754_powl): Likewise.
* sysdeps/ieee754/ldbl-128ibm/e_powl.c (__ieee754_powl): Use
math_check_force_underflow_nonneg.
* sysdeps/x86/fpu/powl_helper.c (__powl_helper): Use
math_check_force_underflow.
* sysdeps/x86_64/fpu/x86_64-math-asm.h
(LDBL_CHECK_FORCE_UFLOW_NONNAN): New macro.
* sysdeps/x86_64/fpu/e_powl.S: Include <x86_64-math-asm.h>. Use
DEFINE_LDBL_MIN.
(__ieee754_powl): Use LDBL_CHECK_FORCE_UFLOW_NONNAN.
* math/auto-libm-test-in: Add more tests of pow.
* math/auto-libm-test-out: Regenerated.
The flt-32 implementation of powf wrongly uses x-1 instead of |x|-1
when computing log (x) for the case where |x| is close to 1 and y is
large. This patch fixes the logic accordingly. Relevant tests
existed for x close to 1, and corresponding tests are added for x
close to -1, as well as for some new variant cases.
Tested for x86_64 and x86.
[BZ #18647]
* sysdeps/ieee754/flt-32/e_powf.c (__ieee754_powf): For large y
and |x| close to 1, use absolute value of x when computing log.
* math/auto-libm-test-in: Add more tests of pow.
* math/auto-libm-test-out: Regenerated.
