Message ID | DB6PR0801MB20534C59695D4E214A1133CA830D0@DB6PR0801MB2053.eurprd08.prod.outlook.com |
---|---|

State | New, archived |

Headers | show |

On Thu, 21 Dec 2017, Wilco Dijkstra wrote: > > (Direct calling of sqrtf128 with similar redirects to those for sqrt / > > sqrtf / sqrtl is also appropriate: GCC 8 has it as a built-in function so > > can inline calls with -fno-math-errno on powerpc64le for POWER9. So the > > patch should include a sqrtf128 redirect if __HAVE_DISTINCT_FLOAT128.) > > I can't get this to work, some targets don't like __float128, others > don't like _Float128 despite __HAVE_DISTINCT_FLOAT128 being defined. I It's not being defined, it's being 1 rather than 0 (#if, not #ifdef). You should use _Float128, not __float128; bits/floatn.h handles defining _Float128 as a typedef if necessary for older compilers. > don't see any calls to sqrtf128, so I think it's best to leave this for > now. It's used (more precisely, __ieee754_sqrtf128 is used, which ought to change to calling sqrtf128 to allow inlining on POWER9 with GCC 8, just as with the changes to other __ieee754_sqrt* calls) (a) via the #defines in sysdeps/ieee754/float128/float128_private.h before ldbl-128 files are included #define __ieee754_sqrtl __ieee754_sqrtf128 #define __sqrtl __sqrtf128 #define sqrtl sqrtf128 (which I expect won't need changing in the __ieee754_sqrt* -> sqrt* changes, since they already handle all variants of the function name) and (b) via sysdeps/generic/math-type-macros.h defining #define M_SQRT M_SUF (__ieee754_sqrt) which is then used in various type-generic templates, which thus end up using __ieee754_sqrtf128 when built for float128 (and changing that M_SQRT definition will obviously result in all types, including float128, changing at the same time to use sqrt* instead of __ieee754_sqrt*). > There are many calls to simple functions like floor, round etc that > aren't currently inlined at all, so inlining those seems like a more > useful thing to do right now. I've renamed the define to > NO_MATH_REDIRECT to make this simple. All those functions will of course need to handle float128 versions exactly the same. (The GCC patch to support floorf128 etc. as built-in functions is still pending review, but doing the glibc changes does not in any way depend on such GCC changes, it simply means that such GCC changes will result in better optimization in glibc.)

Joseph Myers wrote: > It's not being defined, it's being 1 rather than 0 (#if, not #ifdef). > You should use _Float128, not __float128; bits/floatn.h handles defining > _Float128 as a typedef if necessary for older compilers. OK I can get that to work, but I still get odd issues with sqrtf128 being called from the testsuite, while libm.so only defines __sqrtf128, so it gives undefined symbol errors on x64. So I wonder whether in external headers we should never do this: #define sqrtl sqrtf128 Unless there is an easy fix for these issues, I can't see any point in including it in my current patches - improving float/double inlining in this release gives much more benefit to all targets. Wilco

On Thu, 21 Dec 2017, Wilco Dijkstra wrote: > Joseph Myers wrote: > > > It's not being defined, it's being 1 rather than 0 (#if, not #ifdef). > > You should use _Float128, not __float128; bits/floatn.h handles defining > > _Float128 as a typedef if necessary for older compilers. > > OK I can get that to work, but I still get odd issues with sqrtf128 being > called from the testsuite, while libm.so only defines __sqrtf128, so it > gives undefined symbol errors on x64. So I wonder whether in external > headers we should never do this: > > #define sqrtl sqrtf128 float128_private.h isn't an external header. It includes math.h and other headers, then redefines symbols from them, and is included before the main implementation of a float128 function. Naturally your problem needs debugging, since of course libm should be exporting sqrtf128. (And float128_private.h shouldn't be included at all when the type-generic wrappers are, so shouldn't have any opportunity to cause trouble with the automatically generated w_sqrtf128 which is what ought to be defining sqrtf128 as an alias to __sqrtf128.) > Unless there is an easy fix for these issues, I can't see any point in > including it in my current patches - improving float/double inlining in this > release gives much more benefit to all targets. Using sqrt consistently for float / double means changing the M_SQRT definition, which means sqrtf128 gets used as well, which means you need to do something for sqrtf128 simply to avoid localplt failures on non-POWER9 systems with distinct float128, which clearly indicates handling all types consistently.

diff --git a/Makeconfig b/Makeconfig index 34bed9790fac1fd0be9e16b7fe2fa6f5c479b32b..b9da1c137bc495a2f63064765225dc8878cde306 100644 --- a/Makeconfig +++ b/Makeconfig @@ -831,6 +831,9 @@ endif # disable any optimization that assume default rounding mode. +math-flags = -frounding-math +# Build libc/libm using -fno-math-errno, but run testsuite with -fmath-errno. ++extra-math-flags = $(if $(filter nonlib testsuite,$(in-module)),-fmath-errno,-fno-math-errno) + # We might want to compile with some stack-protection flag. ifneq ($(stack-protector),) +stack-protector=$(stack-protector) @@ -966,6 +969,7 @@ endif override CFLAGS = -std=gnu11 -fgnu89-inline $(config-extra-cflags) \ $(filter-out %frame-pointer,$(+cflags)) $(+gccwarn-c) \ + $(+extra-math-flags) \ $(sysdep-CFLAGS) $(CFLAGS-$(suffix $@)) $(CFLAGS-$(<F)) \ $(CFLAGS-$(@F)) $(tls-model) \ $(foreach lib,$(libof-$(basename $(@F))) \ diff --git a/include/math.h b/include/math.h index 7ee291fc972088a1409949326f7024e8e3fe5415..adac78f497b0787757c20c4b9ba52fc9e43a7629 100644 --- a/include/math.h +++ b/include/math.h @@ -54,5 +54,17 @@ libm_hidden_proto (__expf128) libm_hidden_proto (__expm1f128) # endif +# if !(defined __FINITE_MATH_ONLY__ && __FINITE_MATH_ONLY__ > 0) +# ifndef NO_MATH_REDIRECT +/* Declare sqrt for use within GLIBC. Sqrt will typically inline into a + single instruction. Use an asm to avoid use of PLTs if it doesn't. */ +float (sqrtf) (float) asm ("__ieee754_sqrtf"); +double (sqrt) (double) asm ("__ieee754_sqrt"); +# ifndef __NO_LONG_DOUBLE_MATH +long double (sqrtl) (long double) asm ("__ieee754_sqrtl"); +# endif +# endif +# endif + #endif #endif diff --git a/math/w_sqrt_compat.c b/math/w_sqrt_compat.c index 3280d2fbb86af2aeaf6e686fd38579b209c901aa..1a8db2fdef87e57df2f7dda117ef3ee48581ca1d 100644 --- a/math/w_sqrt_compat.c +++ b/math/w_sqrt_compat.c @@ -16,6 +16,7 @@ License along with the GNU C Library; if not, see <http://www.gnu.org/licenses/>. */ +#define NO_MATH_REDIRECT #include <math.h> #include <math_private.h> #include <math-svid-compat.h> diff --git a/math/w_sqrt_template.c b/math/w_sqrt_template.c index 5fae302382d10e9b05df2665c9cb05126cd62f02..f03f04751a9334fc7cdb8c954fbbcd845ac8670f 100644 --- a/math/w_sqrt_template.c +++ b/math/w_sqrt_template.c @@ -21,6 +21,7 @@ for each floating-point type. */ #if __USE_WRAPPER_TEMPLATE +# define NO_MATH_REDIRECT # include <errno.h> # include <fenv.h> # include <math.h> diff --git a/math/w_sqrtf_compat.c b/math/w_sqrtf_compat.c index 6c8c7e3857c7dacf52de6b575a2386095688b5dc..fe7418dbdcdbdb571cb1b9743cf9b17eaa5c8cb2 100644 --- a/math/w_sqrtf_compat.c +++ b/math/w_sqrtf_compat.c @@ -16,6 +16,7 @@ License along with the GNU C Library; if not, see <http://www.gnu.org/licenses/>. */ +#define NO_MATH_REDIRECT #include <math.h> #include <math_private.h> #include <math-svid-compat.h> diff --git a/math/w_sqrtl_compat.c b/math/w_sqrtl_compat.c index 0590f6d155fee27d41bfd42c1dbdf19c381ba1c8..3fda9cf991e430f259b9d26f15219181b3e6dcbb 100644 --- a/math/w_sqrtl_compat.c +++ b/math/w_sqrtl_compat.c @@ -16,6 +16,7 @@ License along with the GNU C Library; if not, see <http://www.gnu.org/licenses/>. */ +#define NO_MATH_REDIRECT #include <math.h> #include <math_private.h> #include <math-svid-compat.h> diff --git a/sysdeps/i386/fpu/w_sqrt.c b/sysdeps/i386/fpu/w_sqrt.c index d37a5d55bf439cebdff05f05dd66f910d6d48c18..8bef04e68a7be3d21d89e6eb41dcfa2561f3f6a5 100644 --- a/sysdeps/i386/fpu/w_sqrt.c +++ b/sysdeps/i386/fpu/w_sqrt.c @@ -1,5 +1,6 @@ /* The inline __ieee754_sqrt is not correctly rounding; it's OK for most internal uses in glibc, but not for sqrt itself. */ +#define NO_MATH_REDIRECT #define __ieee754_sqrt __avoid_ieee754_sqrt #include <math.h> #include <math_private.h> diff --git a/sysdeps/i386/fpu/w_sqrt_compat.c b/sysdeps/i386/fpu/w_sqrt_compat.c index ddd36d0964c2945579bc59b02f127af88384acca..dd485f4b88c7f5348abaab81a1308aeadc9594ec 100644 --- a/sysdeps/i386/fpu/w_sqrt_compat.c +++ b/sysdeps/i386/fpu/w_sqrt_compat.c @@ -1,5 +1,6 @@ /* The inline __ieee754_sqrt is not correctly rounding; it's OK for most internal uses in glibc, but not for sqrt itself. */ +#define NO_MATH_REDIRECT #define __ieee754_sqrt __avoid_ieee754_sqrt #include <math.h> #include <math_private.h>