| Message ID | 20190329133529.22523-15-adhemerval.zanella@linaro.org |
|---|---|
| State | New |
| Headers | show |
| Series | powerpc floating-point optimization refactor | expand |
On Fri, Mar 29 2019, Adhemerval Zanella wrote: > > - The resulting binary difference on 32 bits architecture is minimum > for the non hotspot symbol. > > - It helps wordsize-64 architectures that use ldbl-opt. Same comments I made for patch 12/28 [1], apply here. [1] https://sourceware.org/ml/libc-alpha/2019-06/msg00085.html
On 29/03/2019 13:35, Adhemerval Zanella wrote: > - math.h will use compiler builtin for gcc 4.4+ when built without > -fsignaling-nans and the builtin is expanded inline for all > supported architectures. As an example, there is no intra isnan > call on libm for the architecture I checked (x86, arm, aarch64, > and powerpc). > > - The resulting binary difference on 32 bits architecture is minimum > for the non hotspot symbol. > > - It helps wordsize-64 architectures that use ldbl-opt. > > - It add some code simplification with reduction of duplicated > implementations. > > Checked on powerpc-linux-gnu (built without --with-cpu, with > --with-cpu=power4 and with --with-cpu=power5+ and --disable-multi-arch), > powerpc64-linux-gnu (built without --with-cp and with --with-cpu=power5+ > and --disable-multi-arch). > > * sysdeps/ieee754/dbl-64/wordsize-64/s_isinf.c: Move to ... > * sysdeps/ieee754/dbl-64/s_isinf.c: ... here and format code. on aarch64 buildbot i see make[2]: *** No rule to make target '../sysdeps/ieee754/dbl-64/wordsize-64/s_isinf.c', needed by '/B/math/s_isinf.o'. make[2]: *** No rule to make target '../sysdeps/ieee754/dbl-64/wordsize-64/s_isnan.c', needed by '/B/math/s_isnan.o'. make[2]: *** No rule to make target '../sysdeps/ieee754/dbl-64/wordsize-64/s_finite.c', needed by '/B/math/s_finite.o'. make[2]: Target 'subdir_lib' not remade because of errors. make[1]: *** [Makefile:259: math/subdir_lib] Error 2
On 13/06/2019 09:57, Szabolcs Nagy wrote: > on aarch64 buildbot i see > > make[2]: *** No rule to make target '../sysdeps/ieee754/dbl-64/wordsize-64/s_isinf.c', needed by '/B/math/s_isinf.o'. > make[2]: *** No rule to make target '../sysdeps/ieee754/dbl-64/wordsize-64/s_isnan.c', needed by '/B/math/s_isnan.o'. > make[2]: *** No rule to make target '../sysdeps/ieee754/dbl-64/wordsize-64/s_finite.c', needed by '/B/math/s_finite.o'. > make[2]: Target 'subdir_lib' not remade because of errors. > make[1]: *** [Makefile:259: math/subdir_lib] Error 2 sorry this is just a buildbot thing. (it first runs the build in an unclean dir, which of course had the deps for removed files, when it rerun in a clean dir there was no failure.)
diff --git a/sysdeps/ieee754/dbl-64/s_isinf.c b/sysdeps/ieee754/dbl-64/s_isinf.c index 93eb65c147..e5300fd2a5 100644 --- a/sysdeps/ieee754/dbl-64/s_isinf.c +++ b/sysdeps/ieee754/dbl-64/s_isinf.c @@ -4,10 +4,6 @@ * Public domain. */ -#if defined(LIBM_SCCS) && !defined(lint) -static char rcsid[] = "$NetBSD: s_isinf.c,v 1.3 1995/05/11 23:20:14 jtc Exp $"; -#endif - /* * isinf(x) returns 1 is x is inf, -1 if x is -inf, else 0; * no branching! @@ -21,11 +17,12 @@ static char rcsid[] = "$NetBSD: s_isinf.c,v 1.3 1995/05/11 23:20:14 jtc Exp $"; int __isinf (double x) { - int32_t hx, lx; - EXTRACT_WORDS (hx, lx, x); - lx |= (hx & 0x7fffffff) ^ 0x7ff00000; - lx |= -lx; - return ~(lx >> 31) & (hx >> 30); + int64_t ix; + EXTRACT_WORDS64 (ix,x); + int64_t t = ix & UINT64_C (0x7fffffffffffffff); + t ^= UINT64_C (0x7ff0000000000000); + t |= -t; + return ~(t >> 63) & (ix >> 62); } hidden_def (__isinf) weak_alias (__isinf, isinf) diff --git a/sysdeps/ieee754/dbl-64/wordsize-64/s_isinf.c b/sysdeps/ieee754/dbl-64/wordsize-64/s_isinf.c deleted file mode 100644 index 2b427a8b4c..0000000000 --- a/sysdeps/ieee754/dbl-64/wordsize-64/s_isinf.c +++ /dev/null @@ -1,34 +0,0 @@ -/* - * Written by J.T. Conklin <jtc@netbsd.org>. - * Changed to return -1 for -Inf by Ulrich Drepper <drepper@cygnus.com>. - * Public domain. - */ - -/* - * isinf(x) returns 1 is x is inf, -1 if x is -inf, else 0; - * no branching! - */ - -#include <math.h> -#include <math_private.h> -#include <ldbl-classify-compat.h> -#include <shlib-compat.h> - -int -__isinf (double x) -{ - int64_t ix; - EXTRACT_WORDS64(ix,x); - int64_t t = ix & UINT64_C(0x7fffffffffffffff); - t ^= UINT64_C(0x7ff0000000000000); - t |= -t; - return ~(t >> 63) & (ix >> 62); -} -hidden_def (__isinf) -weak_alias (__isinf, isinf) -#ifdef NO_LONG_DOUBLE -# if LDBL_CLASSIFY_COMPAT && SHLIB_COMPAT (libc, GLIBC_2_0, GLIBC_2_23) -compat_symbol (libc, __isinf, __isinfl, GLIBC_2_0); -# endif -weak_alias (__isinf, isinfl) -#endif
- math.h will use compiler builtin for gcc 4.4+ when built without -fsignaling-nans and the builtin is expanded inline for all supported architectures. As an example, there is no intra isnan call on libm for the architecture I checked (x86, arm, aarch64, and powerpc). - The resulting binary difference on 32 bits architecture is minimum for the non hotspot symbol. - It helps wordsize-64 architectures that use ldbl-opt. - It add some code simplification with reduction of duplicated implementations. Checked on powerpc-linux-gnu (built without --with-cpu, with --with-cpu=power4 and with --with-cpu=power5+ and --disable-multi-arch), powerpc64-linux-gnu (built without --with-cp and with --with-cpu=power5+ and --disable-multi-arch). * sysdeps/ieee754/dbl-64/wordsize-64/s_isinf.c: Move to ... * sysdeps/ieee754/dbl-64/s_isinf.c: ... here and format code. --- sysdeps/ieee754/dbl-64/s_isinf.c | 15 ++++----- sysdeps/ieee754/dbl-64/wordsize-64/s_isinf.c | 34 -------------------- 2 files changed, 6 insertions(+), 43 deletions(-) delete mode 100644 sysdeps/ieee754/dbl-64/wordsize-64/s_isinf.c -- 2.17.1