alpha: cpow math testsuite failures with GCC 9

[Aurelien Jarno <aurelien at aurel32 dot net>]

Hi,

When building the GNU libc with GCC 9 on alpha, there are a few new math
testsuite failures. Among them, I have been investigating the following
ones:

| FAIL: math/test-double-cpow
| FAIL: math/test-double-finite-cpow
| FAIL: math/test-float-cpow
| FAIL: math/test-float-finite-cpow
| FAIL: math/test-float32-cpow
| FAIL: math/test-float32-finite-cpow
| FAIL: math/test-float32x-cpow
| FAIL: math/test-float32x-finite-cpow
| FAIL: math/test-float64-cpow
| FAIL: math/test-float64-finite-cpow
| FAIL: math/test-idouble-cpow
| FAIL: math/test-ifloat-cpow
| FAIL: math/test-ifloat32-cpow
| FAIL: math/test-ifloat32x-cpow
| FAIL: math/test-ifloat64-cpow

Here is the detailed failure for math/test-double-cpow:

| FAIL: math/test-double-cpow
| original exit status 1
| testing double (without inline functions)
| Failure: Test: Real part of: cpow_downward (0x2p+0 + +0 i, 0xap+0 + +0 i)
| Result:
|  is:          1.0239999999999989e+03   0x1.ffffffffffff7p+9
|  should be:   1.0240000000000000e+03   0x1.0000000000000p+10
|  difference:  1.0231815394945442e-12   0x1.2000000000000p-40
|  ulp       :  4.5000
|  max.ulp   :  2.0000
| Maximal error of real part of: cpow_downward
|  is      : 5 ulp
|  accepted: 2 ulp
| Maximal error of imaginary part of: cpow_downward
|  is      : 1 ulp
|  accepted: 1 ulp
| Failure: Test: Real part of: cpow_towardzero (0x2p+0 + +0 i, 0xap+0 + +0 i)
| Result:
|  is:          1.0239999999999989e+03   0x1.ffffffffffff7p+9
|  should be:   1.0240000000000000e+03   0x1.0000000000000p+10
|  difference:  1.0231815394945442e-12   0x1.2000000000000p-40
|  ulp       :  4.5000
|  max.ulp   :  2.0000
| Maximal error of real part of: cpow_towardzero
|  is      : 5 ulp
|  accepted: 2 ulp
| Maximal error of imaginary part of: cpow_towardzero
|  is      : 1 ulp
|  accepted: 1 ulp
| 
| Test suite completed:
|   50 test cases plus 46 tests for exception flags and
|     46 tests for errno executed.
|   4 errors occurred.

cpow(x, c) is basically cexp(c * clog(x)). The change is in the way the
rounding in done in the complex muliplication.

There have been 2 changes in that regard in GCC 9:
- r268430: config/alpha/t-linux: Add -mfp-rounding-mode=d
  to HOST_LIBGCC2_CFLAGS.
- r259889: [tree-complex.c] PR tree-optimization/70291: Inline
  floating-point complex multiplication more aggressively.

The first one changes the muldc3 routine in libgcc to use /sud floating
point instructions (dynamic rounding) instead of /su (nearest rounding).
The second one inlines the multiplication for normal numbers.
Interestingly the inline code uses /su by default, however as s_cpow.c 
is compiled with -mfp-rounding-mode=d it ends up with the /sud version.

I don't really know how this should be fixed on the glibc side.

Thanks,
Aurelien

-- 
Aurelien Jarno                          GPG: 4096R/1DDD8C9B
aurelien@aurel32.net                 http://www.aurel32.net