diff options
author | Joseph Myers <joseph@codesourcery.com> | 2015-02-12 23:05:37 +0000 |
---|---|---|
committer | Joseph Myers <joseph@codesourcery.com> | 2015-02-12 23:05:37 +0000 |
commit | e8bd5286c68bc35be3b41e94c15c4387dcb3bec9 (patch) | |
tree | c62c20ff732d5a7955bf3eaeb73454c4a61bb92d /sysdeps/powerpc | |
parent | 96a157490c8db4c3be4495c4f9c0ca9ecaf5ae7c (diff) | |
download | glibc-e8bd5286c68bc35be3b41e94c15c4387dcb3bec9.tar.gz |
Fix powerpc software sqrt (bug 17964).
As Adhemerval noted in
<https://sourceware.org/ml/libc-alpha/2015-01/msg00451.html>, the
powerpc sqrt implementation for when _ARCH_PPCSQ is not defined is
inaccurate in some cases.
The problem is that this code relies on fused multiply-add, and relies
on the compiler contracting a * b + c to get a fused operation. But
sysdeps/ieee754/dbl-64/Makefile disables contraction for e_sqrt.c,
because the implementation in that directory relies on *not* having
contracted operations.
While it would be possible to arrange makefiles so that an earlier
sysdeps directory can disable the setting in
sysdeps/ieee754/dbl-64/Makefile, it seems a lot cleaner to make the
dependence on fused operations explicit in the .c file. GCC 4.6
introduced support for __builtin_fma on powerpc and other
architectures with such instructions, so we can rely on that; this
patch duly makes the code use __builtin_fma for all such fused
operations.
Tested for powerpc32 (hard float).
2015-02-12 Joseph Myers <joseph@codesourcery.com>
[BZ #17964]
* sysdeps/powerpc/fpu/e_sqrt.c (__slow_ieee754_sqrt): Use
__builtin_fma instead of relying on contraction of a * b + c.
Diffstat (limited to 'sysdeps/powerpc')
-rw-r--r-- | sysdeps/powerpc/fpu/e_sqrt.c | 33 |
1 files changed, 18 insertions, 15 deletions
diff --git a/sysdeps/powerpc/fpu/e_sqrt.c b/sysdeps/powerpc/fpu/e_sqrt.c index 0934faa5fe..9b55ef8390 100644 --- a/sysdeps/powerpc/fpu/e_sqrt.c +++ b/sysdeps/powerpc/fpu/e_sqrt.c @@ -99,38 +99,41 @@ __slow_ieee754_sqrt (double x) /* Here we have three Newton-Raphson iterations each of a division and a square root and the remainder of the argument reduction, all interleaved. */ - sd = -(sg * sg - sx); + sd = -__builtin_fma (sg, sg, -sx); fsgi = (xi0 + 0x40000000) >> 1 & 0x7ff00000; sy2 = sy + sy; - sg = sy * sd + sg; /* 16-bit approximation to sqrt(sx). */ + sg = __builtin_fma (sy, sd, sg); /* 16-bit approximation to + sqrt(sx). */ /* schedule the INSERT_WORDS (fsg, fsgi, 0) to get separation between the store and the load. */ INSERT_WORDS (fsg, fsgi, 0); iw_u.parts.msw = fsgi; iw_u.parts.lsw = (0); - e = -(sy * sg - almost_half); - sd = -(sg * sg - sx); + e = -__builtin_fma (sy, sg, -almost_half); + sd = -__builtin_fma (sg, sg, -sx); if ((xi0 & 0x7ff00000) == 0) goto denorm; - sy = sy + e * sy2; - sg = sg + sy * sd; /* 32-bit approximation to sqrt(sx). */ + sy = __builtin_fma (e, sy2, sy); + sg = __builtin_fma (sy, sd, sg); /* 32-bit approximation to + sqrt(sx). */ sy2 = sy + sy; /* complete the INSERT_WORDS (fsg, fsgi, 0) operation. */ fsg = iw_u.value; - e = -(sy * sg - almost_half); - sd = -(sg * sg - sx); - sy = sy + e * sy2; + e = -__builtin_fma (sy, sg, -almost_half); + sd = -__builtin_fma (sg, sg, -sx); + sy = __builtin_fma (e, sy2, sy); shx = sx * fsg; - sg = sg + sy * sd; /* 64-bit approximation to sqrt(sx), - but perhaps rounded incorrectly. */ + sg = __builtin_fma (sy, sd, sg); /* 64-bit approximation to + sqrt(sx), but perhaps + rounded incorrectly. */ sy2 = sy + sy; g = sg * fsg; - e = -(sy * sg - almost_half); - d = -(g * sg - shx); - sy = sy + e * sy2; + e = -__builtin_fma (sy, sg, -almost_half); + d = -__builtin_fma (g, sg, -shx); + sy = __builtin_fma (e, sy2, sy); fesetenv_register (fe); - return g + sy * d; + return __builtin_fma (sy, d, g); denorm: /* For denormalised numbers, we normalise, calculate the square root, and return an adjusted result. */ |