diff options
author | joseph <joseph@7b3dc134-2b1b-0410-93df-9e9f96275f8d> | 2010-10-22 00:19:23 +0000 |
---|---|---|
committer | joseph <joseph@7b3dc134-2b1b-0410-93df-9e9f96275f8d> | 2010-10-22 00:19:23 +0000 |
commit | 362c8265449ccbe244cf34c4a55d9bf2e6d36d64 (patch) | |
tree | c1856abcbb0288fd55e5b66a2628dac9f839460a | |
parent | 9a13037b9f1354b7405124e83087b72c4c44d388 (diff) |
Merge changes between r11760 and r11842 from /fsf/trunk.
git-svn-id: svn://svn.eglibc.org/trunk@11843 7b3dc134-2b1b-0410-93df-9e9f96275f8d
36 files changed, 1163 insertions, 59 deletions
diff --git a/libc/ChangeLog b/libc/ChangeLog index e2a4a7051..e162666af 100644 --- a/libc/ChangeLog +++ b/libc/ChangeLog @@ -1,3 +1,91 @@ +2010-10-19 Ulrich Drepper <drepper@gmail.com> + + * sysdeps/i386/bits/mathdef.h (FP_FAST_FMA): If the GCC 4.6 port has + fast fma builtins, define the macros in the C99 standard. + (FP_FAST_FMAF): Likewise. + (FP_FAST_FMAL): Likewise. + * sysdeps/x86_64/bits/mathdef.h: Likewise. + + * bits/mathdef.h: Update copyright year. + * sysdeps/powerpc/bits/mathdef.h: Likewise. + +2010-10-19 Michael Meissner <meissner@linux.vnet.ibm.com> + + * bits/mathdef.h (FP_FAST_FMA): If the GCC 4.6 port has fast fma + builtins, define the macros in the C99 standard. + (FP_FAST_FMAF): Likewise. + (FP_FAST_FMAL): Likewise. + * sysdeps/powerpc/bits/mathdef.h (FP_FAST_FMA): Define, ppc as + multiply/add. + (FP_FAST_FMAF): Likewise. + +2010-10-15 Jakub Jelinek <jakub@redhat.com> + + [BZ #3268] + * math/libm-test.inc (fma_test): Some new testcases. + * sysdeps/ieee754/ldbl-128/s_fmal.c: New file. + * sysdeps/ieee754/ldbl-96/s_fma.c (__fma): Fix fma with finite x and + y and infinite z. Do multiplication by C already in long double. + * sysdeps/ieee754/ldbl-96/s_fmal.c: New file. + * sysdeps/ieee754/dbl-64/s_fma.c (__fma): Fix fma with finite x and + y and infinite z. Do bitwise or of inexact bit into u.d. + * sysdeps/ieee754/ldbl-64-128/s_fmal.c: New file. + * sysdeps/i386/fpu/s_fmaf.S: Removed. + * sysdeps/i386/fpu/s_fma.S: Removed. + * sysdeps/i386/fpu/s_fmal.S: Removed. + +2010-10-16 Jakub Jelinek <jakub@redhat.com> + + [BZ #3268] + * math/libm-test.inc (fma_test): Add IEEE quad long double fmal tests. + * sysdeps/ieee754/ldbl-128/s_fmal.c (__fmal): Ensure a1 + u.d + computation is not scheduled after fetestexcept. Fix value + of minimum denormal long double. + +2010-10-14 Jakub Jelinek <jakub@redhat.com> + + [BZ #3268] + * math/libm-test.inc (fma_test): Add some more tests. + * sysdeps/ieee754/dbl-64/s_fma.c (__fma): Handle underflows + correctly. + +2010-10-15 Andreas Schwab <schwab@redhat.com> + + * scripts/data/localplt-s390-linux-gnu.data: New file. + * scripts/data/localplt-s390x-linux-gnu.data: New file. + +2010-10-13 Jakub Jelinek <jakub@redhat.com> + + [BZ #3268] + * math/libm-test.inc (fma_test): Some more fmaf and fma tests. + * sysdeps/i386/i686/multiarch/s_fma.c: Include ldbl-96 version + instead of dbl-64. + * sysdeps/i386/fpu/bits/mathinline.h (fma, fmaf, fmal): Remove + inlines. + * sysdeps/ieee754/ldbl-96/s_fma.c: New file. + * sysdeps/ieee754/dbl-64/s_fma.c (__fma): Fix exponent adjustment + if one of x and y is very large and the other is subnormal. + * sysdeps/s390/fpu/s_fmaf.c: New file. + * sysdeps/s390/fpu/s_fma.c: New file. + * sysdeps/powerpc/fpu/s_fmaf.S: New file. + * sysdeps/powerpc/fpu/s_fma.S: New file. + * sysdeps/powerpc/powerpc32/fpu/s_fma.S: New file. + * sysdeps/powerpc/powerpc64/fpu/s_fma.S: New file. + * sysdeps/unix/sysv/linux/s390/fpu/s_fma.c: New file. + +2010-10-12 Jakub Jelinek <jakub@redhat.com> + + [BZ #3268] + * math/libm-test.inc (fma_test): Add some more fmaf tests, add + fma tests. + * sysdeps/ieee754/dbl-64/s_fmaf.c (__fmaf): Fix Inf/Nan check. + * sysdeps/ieee754/dbl-64/s_fma.c: New file. + * sysdeps/i386/i686/multiarch/s_fma.c: Include + sysdeps/ieee754/dbl-64/s_fma.c instead of math/s_fma.c. + * sysdeps/x86_64/multiarch/s_fma.c: Likewise. + * sysdeps/ieee754/ldbl-opt/s_fma.c: Likewise. + * sysdeps/ieee754/ldbl-128/s_fma.c: New file. + 2010-10-12 Ulrich Drepper <drepper@redhat.com> [BZ #12078] @@ -1,4 +1,4 @@ -GNU C Library NEWS -- history of user-visible changes. 2010-10-11 +GNU C Library NEWS -- history of user-visible changes. 2010-10-13 Copyright (C) 1992-2009, 2010 Free Software Foundation, Inc. See the end for copying conditions. @@ -10,7 +10,8 @@ Version 2.13 * The following bugs are resolved with this release: 3268, 7066, 10851, 11611, 11640, 11701, 11840, 11856, 11883, 11903, 11904, - 11968, 11979, 12005, 12037, 12067, 12077, 12078, 12092, 12093, 12107, 12108 + 11968, 11979, 12005, 12037, 12067, 12077, 12078, 12092, 12093, 12107, 12108, + 12113 * New Linux interfaces: prlimit, prlimit64, fanotify_init, fanotify_mark diff --git a/libc/bits/mathdef.h b/libc/bits/mathdef.h index 00c67241a..3cea5847c 100644 --- a/libc/bits/mathdef.h +++ b/libc/bits/mathdef.h @@ -1,4 +1,4 @@ -/* Copyright (C) 1997, 1998, 1999, 2000, 2004 Free Software Foundation, Inc. +/* Copyright (C) 1997,1998,1999,2000,2004,2010 Free Software Foundation, Inc. This file is part of the GNU C Library. The GNU C Library is free software; you can redistribute it and/or @@ -34,6 +34,20 @@ typedef double double_t; /* `double' expressions are evaluated as # define FP_ILOGB0 (-2147483647) # define FP_ILOGBNAN 2147483647 +/* The GCC 4.6 compiler will define __FP_FAST_FMA{,F,L} if the fma{,f,l} + builtins are supported. */ +#if __FP_FAST_FMA +# define FP_FAST_FMA 1 +#endif + +#if __FP_FAST_FMAF +# define FP_FAST_FMAF 1 +#endif + +#if __FP_FAST_FMAL +# define FP_FAST_FMAL 1 +#endif + #endif /* ISO C99 */ #ifndef __NO_LONG_DOUBLE_MATH diff --git a/libc/math/libm-test.inc b/libc/math/libm-test.inc index 784cdb6e4..453088528 100644 --- a/libc/math/libm-test.inc +++ b/libc/math/libm-test.inc @@ -2800,11 +2800,77 @@ fma_test (void) TEST_fff_f (fma, minus_infty, plus_infty, plus_infty, nan_value, INVALID_EXCEPTION); TEST_fff_f (fma, plus_infty, minus_infty, plus_infty, nan_value, INVALID_EXCEPTION); TEST_fff_f (fma, minus_infty, minus_infty, minus_infty, nan_value, INVALID_EXCEPTION); + TEST_fff_f (fma, plus_infty, 3.5L, minus_infty, nan_value, INVALID_EXCEPTION); + TEST_fff_f (fma, minus_infty, -7.5L, minus_infty, nan_value, INVALID_EXCEPTION); + TEST_fff_f (fma, -13.5L, plus_infty, plus_infty, nan_value, INVALID_EXCEPTION); + TEST_fff_f (fma, minus_infty, 7.5L, plus_infty, nan_value, INVALID_EXCEPTION); TEST_fff_f (fma, 1.25L, 0.75L, 0.0625L, 1.0L); -#ifdef TEST_FLOAT + + FLOAT fltmax = CHOOSE (LDBL_MAX, DBL_MAX, FLT_MAX, + LDBL_MAX, DBL_MAX, FLT_MAX); + TEST_fff_f (fma, -fltmax, -fltmax, minus_infty, minus_infty); + TEST_fff_f (fma, fltmax / 2, fltmax / 2, minus_infty, minus_infty); + TEST_fff_f (fma, -fltmax, fltmax, plus_infty, plus_infty); + TEST_fff_f (fma, fltmax / 2, -fltmax / 4, plus_infty, plus_infty); + TEST_fff_f (fma, plus_infty, 4, plus_infty, plus_infty); + TEST_fff_f (fma, 2, minus_infty, minus_infty, minus_infty); + TEST_fff_f (fma, minus_infty, minus_infty, plus_infty, plus_infty); + TEST_fff_f (fma, plus_infty, minus_infty, minus_infty, minus_infty); + +#if defined (TEST_FLOAT) && FLT_MANT_DIG == 24 TEST_fff_f (fma, 0x1.7ff8p+13, 0x1.000002p+0, 0x1.ffffp-24, 0x1.7ff802p+13); TEST_fff_f (fma, 0x1.fffp+0, 0x1.00001p+0, -0x1.fffp+0, 0x1.fffp-20); + TEST_fff_f (fma, 0x1.9abcdep+127, 0x0.9abcdep-126, -0x1.f08948p+0, 0x1.bb421p-25); + TEST_fff_f (fma, 0x1.9abcdep+100, 0x0.9abcdep-126, -0x1.f08948p-27, 0x1.bb421p-52); + TEST_fff_f (fma, 0x1.fffffep+127, 0x1.001p+0, -0x1.fffffep+127, 0x1.fffffep+115); + TEST_fff_f (fma, -0x1.fffffep+127, 0x1.fffffep+0, 0x1.fffffep+127, -0x1.fffffap+127); + TEST_fff_f (fma, 0x1.fffffep+127, 2.0, -0x1.fffffep+127, 0x1.fffffep+127); +#endif +#if defined (TEST_DOUBLE) && DBL_MANT_DIG == 53 + TEST_fff_f (fma, 0x1.7fp+13, 0x1.0000000000001p+0, 0x1.ffep-48, 0x1.7f00000000001p+13); + TEST_fff_f (fma, 0x1.fffp+0, 0x1.0000000000001p+0, -0x1.fffp+0, 0x1.fffp-52); + TEST_fff_f (fma, 0x1.0000002p+0, 0x1.ffffffcp-1, 0x1p-300, 1.0); + TEST_fff_f (fma, 0x1.0000002p+0, 0x1.ffffffcp-1, -0x1p-300, 0x1.fffffffffffffp-1); + TEST_fff_f (fma, 0x1.deadbeef2feedp+1023, 0x0.deadbeef2feedp-1022, -0x1.a05f8c01a4bfbp+1, 0x1.0989687bc9da4p-53); + TEST_fff_f (fma, 0x1.deadbeef2feedp+900, 0x0.deadbeef2feedp-1022, -0x1.a05f8c01a4bfbp-122, 0x1.0989687bc9da4p-176); + TEST_fff_f (fma, 0x1.fffffffffffffp+1023, 0x1.001p+0, -0x1.fffffffffffffp+1023, 0x1.fffffffffffffp+1011); + TEST_fff_f (fma, -0x1.fffffffffffffp+1023, 0x1.fffffffffffffp+0, 0x1.fffffffffffffp+1023, -0x1.ffffffffffffdp+1023); + TEST_fff_f (fma, 0x1.fffffffffffffp+1023, 2.0, -0x1.fffffffffffffp+1023, 0x1.fffffffffffffp+1023); + TEST_fff_f (fma, 0x1.6a09e667f3bccp-538, 0x1.6a09e667f3bccp-538, 0.0, 0.0); + TEST_fff_f (fma, 0x1.deadbeef2feedp-495, 0x1.deadbeef2feedp-495, -0x1.bf86a5786a574p-989, 0x0.0000042625a1fp-1022); + TEST_fff_f (fma, 0x1.deadbeef2feedp-503, 0x1.deadbeef2feedp-503, -0x1.bf86a5786a574p-1005, 0x0.0000000004262p-1022); + TEST_fff_f (fma, 0x1p-537, 0x1p-538, 0x1p-1074, 0x0.0000000000002p-1022); + TEST_fff_f (fma, 0x1.7fffff8p-968, 0x1p-106, 0x0.000001p-1022, 0x0.0000010000001p-1022); + TEST_fff_f (fma, 0x1.4000004p-967, 0x1p-106, 0x0.000001p-1022, 0x0.0000010000003p-1022); + TEST_fff_f (fma, 0x1.4p-967, -0x1p-106, -0x0.000001p-1022, -0x0.0000010000002p-1022); + TEST_fff_f (fma, -0x1.19cab66d73e17p-959, 0x1.c7108a8c5ff51p-107, -0x0.80b0ad65d9b64p-1022, -0x0.80b0ad65d9d59p-1022); + TEST_fff_f (fma, -0x1.d2eaed6e8e9d3p-979, -0x1.4e066c62ac9ddp-63, -0x0.9245e6b003454p-1022, -0x0.9245c09c5fb5dp-1022); + TEST_fff_f (fma, 0x1.153d650bb9f06p-907, 0x1.2d01230d48407p-125, -0x0.b278d5acfc3cp-1022, -0x0.b22757123bbe9p-1022); + TEST_fff_f (fma, -0x1.fffffffffffffp-711, 0x1.fffffffffffffp-275, 0x1.fffffe00007ffp-983, 0x1.7ffffe00007ffp-983); +#endif +#if defined (TEST_LDOUBLE) && LDBL_MANT_DIG == 64 + TEST_fff_f (fma, -0x8.03fcp+3696L, 0xf.fffffffffffffffp-6140L, 0x8.3ffffffffffffffp-2450L, -0x8.01ecp-2440L); + TEST_fff_f (fma, 0x9.fcp+2033L, -0x8.000e1f000ff800fp-3613L, -0xf.fffffffffffc0ffp-1579L, -0xd.fc119fb093ed092p-1577L); + TEST_fff_f (fma, 0xc.7fc000003ffffffp-1194L, 0x8.1e0003fffffffffp+15327L, -0x8.fffep+14072L, 0xc.ae9f164020effffp+14136L); + TEST_fff_f (fma, -0x8.0001fc000000003p+1798L, 0xcp-2230L, 0x8.f7e000000000007p-468L, -0xc.0002f9ffee10404p-429L); + TEST_fff_f (fma, 0xc.0000000000007ffp+10130L, -0x8.000000000000001p+4430L, 0xc.07000000001ffffp+14513L, -0xb.fffffffffffd7e4p+14563L); + TEST_fff_f (fma, 0xb.ffffp-4777L, 0x8.000000fffffffffp-11612L, -0x0.3800fff8p-16385L, 0x5.c7fe80c7ffeffffp-16385L); +#endif +#if defined (TEST_LDOUBLE) && LDBL_MANT_DIG == 113 + TEST_fff_f (fma, 0x1.bb2de33e02ccbbfa6e245a7c1f71p-2584L, -0x1.6b500daf0580d987f1bc0cadfcddp-13777L, 0x1.613cd91d9fed34b33820e5ab9d8dp-16378L, -0x1.3a79fb50eb9ce887cffa0f09bd9fp-16360L); + TEST_fff_f (fma, -0x1.f949b880cacb0f0c61540105321dp-5954L, -0x1.3876cec84b4140f3bd6198731b7ep-10525L, -0x0.a5dc1c6cfbc498c54fb0b504bf19p-16382L, -0x0.a5dc1c6cfbc498c54fb0b5038abbp-16382L); + TEST_fff_f (fma, -0x1.0000fffffffffp-16221L, 0x1.0000001fffff8007fep-239L, 0x0.ff87ffffffffffffe000003fffffp-16382L, 0x0.ff87ffffffffffffdffc003bff7fp-16382L); + TEST_fff_f (fma, -0x1.ac79c9376ef447f3827c9e9de008p-2228L, -0x1.5ba830022b6139e21fbe7270cad8p-6314L, 0x1.e8282b6a26bb6a9daf5c8e73e9f9p-8616L, 0x1.22f14a0253878a730cd1aee373adp-8541L); + TEST_fff_f (fma, -0x1.c69749ec574caaa2ab8e97ddb9f3p+2652L, 0x1.f34235ff9d095449c29b4831b62dp+3311L, 0x1.fbe4302df23354dbd0c4d3cfe606p+5879L, -0x1.bb473bfdfb7a6e18886ce6e57eafp+5964L); + TEST_fff_f (fma, -0x1.ca8835fc6ecfb5398625fc891be5p-1686L, 0x1.621e1972bbe2180e5be9dd7d8df5p-7671L, -0x1.7d2d21b73b52cf20dec2a83902a4p-9395L, -0x1.3d2322191c9c88bc68a62ab8042cp-9356L); + TEST_fff_f (fma, -0x1.55cff679ec49c2541fab41fc843ep-11819L, 0x1.e60e9f464f9e8df0509647c7c971p+12325L, 0x1.eaa2a7649d765c2f564f7a5beca7p+454L, -0x1.447e29fa7e406a285f4e350fcf86p+507L); + TEST_fff_f (fma, 0x1.f0e7b1454908576f2537d863cf9bp+11432L, 0x1.cdce52f09d4ca76e68706f34b5d5p-1417L, -0x1.2e986187c70f146235ea2066e486p+9979L, 0x1.c030dad3cc5643f3dd0f5619f661p+10016L); + TEST_fff_f (fma, 0x1.f102f7da4a57a3a4aab620e29452p-3098L, -0x1.cc06a4ff40248f9e2dcc4b6afd84p-11727L, 0x1.d512a11126b5ac8ed8973b8580c8p-14849L, -0x1.be8f1cf737ab4d1c31c54f5ec23bp-14824L); + TEST_fff_f (fma, -0x1.fc47ac7434b993cd8dcb2b431f25p-3816L, 0x1.fbc9750da8468852d84558e1db6dp-5773L, -0x1.00a98abf783f75c40fe5b7a37d86p-9607L, -0x1.f81917b166f45e763cfcc057e2adp-9588L); + TEST_fff_f (fma, 0x1.00000000000007ffffffffffffffp-9045L, -0x1.ffffffffffff80000001ffffffffp+4773L, -0x1.f8p-4316L, -0x1.00000000000f88000000fffffdffp-4271L); + TEST_fff_f (fma, 0x1.4e922764c90701d4a2f21d01893dp-8683L, -0x1.955a12e2d7c9447c27fa022fc865p+212L, -0x1.e9634462eaef96528b90b6944578p-8521L, -0x1.08e1783184a371943d3598e10865p-8470L); + TEST_fff_f (fma, 0x1.801181509c03bdbef10d6165588cp-15131L, 0x1.ad86f8e57d3d40bfa8007780af63p-368L, -0x1.6e9df0dab1c9f1d7a6043c390741p-15507L, 0x1.417c9b2b15e2ad57dc9e0e920844p-15498L); #endif END (fma); diff --git a/libc/nptl/ChangeLog b/libc/nptl/ChangeLog index dcb94b220..e3bfd0000 100644 --- a/libc/nptl/ChangeLog +++ b/libc/nptl/ChangeLog @@ -1,3 +1,10 @@ +2010-10-13 H.J. Lu <hongjiu.lu@intel.com> + + [BZ #12113] + * sysdeps/x86_64/pthreaddef.h (TCB_ALIGNMENT): Changed to 32. + * sysdeps/x86_64/tls.h (TLS_TCB_ALIGN): Defined with alignment + of "struct pthread". + 2010-09-21 Andreas Schwab <schwab@redhat.com> * sysdeps/pthread/pthread.h (pthread_cleanup_push) diff --git a/libc/nptl/sysdeps/x86_64/pthreaddef.h b/libc/nptl/sysdeps/x86_64/pthreaddef.h index be63a8c74..8ec135c04 100644 --- a/libc/nptl/sysdeps/x86_64/pthreaddef.h +++ b/libc/nptl/sysdeps/x86_64/pthreaddef.h @@ -27,8 +27,9 @@ /* Minimal stack size after allocating thread descriptor and guard size. */ #define MINIMAL_REST_STACK 2048 -/* Alignment requirement for TCB. */ -#define TCB_ALIGNMENT 16 +/* Alignment requirement for TCB. Need to store post-AVX vector registers + in the TCB and we want the storage to be aligned at 32-byte. */ +#define TCB_ALIGNMENT 32 /* Location of current stack frame. The frame pointer is not usable. */ diff --git a/libc/nptl/sysdeps/x86_64/tls.h b/libc/nptl/sysdeps/x86_64/tls.h index e39eb5f69..41b5e6daa 100644 --- a/libc/nptl/sysdeps/x86_64/tls.h +++ b/libc/nptl/sysdeps/x86_64/tls.h @@ -117,12 +117,7 @@ typedef struct # define TLS_TCB_SIZE sizeof (struct pthread) /* Alignment requirements for the TCB. */ -//# define TLS_TCB_ALIGN __alignof__ (struct pthread) -// Normally the above would be correct But we have to store post-AVX -// vector registers in the TCB and we want the storage to be aligned. -// unfortunately there isn't yet a type for these values and hence no -// 32-byte alignment requirement. Make this explicit, for now. -# define TLS_TCB_ALIGN 32 +# define TLS_TCB_ALIGN __alignof__ (struct pthread) /* The TCB can have any size and the memory following the address the thread pointer points to is unspecified. Allocate the TCB there. */ diff --git a/libc/scripts/data/localplt-s390-linux-gnu.data b/libc/scripts/data/localplt-s390-linux-gnu.data new file mode 100644 index 000000000..8fb56b608 --- /dev/null +++ b/libc/scripts/data/localplt-s390-linux-gnu.data @@ -0,0 +1,7 @@ +libc.so: _Unwind_Find_FDE +libc.so: calloc +libc.so: free +libc.so: malloc +libc.so: memalign +libc.so: realloc +libm.so: matherr diff --git a/libc/scripts/data/localplt-s390x-linux-gnu.data b/libc/scripts/data/localplt-s390x-linux-gnu.data new file mode 100644 index 000000000..8fb56b608 --- /dev/null +++ b/libc/scripts/data/localplt-s390x-linux-gnu.data @@ -0,0 +1,7 @@ +libc.so: _Unwind_Find_FDE +libc.so: calloc +libc.so: free +libc.so: malloc +libc.so: memalign +libc.so: realloc +libm.so: matherr diff --git a/libc/sysdeps/i386/bits/mathdef.h b/libc/sysdeps/i386/bits/mathdef.h index ec42ed5df..479eb96f9 100644 --- a/libc/sysdeps/i386/bits/mathdef.h +++ b/libc/sysdeps/i386/bits/mathdef.h @@ -1,4 +1,4 @@ -/* Copyright (C) 1997, 1998, 1999, 2000, 2004 Free Software Foundation, Inc. +/* Copyright (C) 1997, 1998, 1999, 2000, 2004, 2010 Free Software Foundation, Inc. This file is part of the GNU C Library. The GNU C Library is free software; you can redistribute it and/or @@ -43,4 +43,18 @@ typedef long double double_t; /* `double' expressions are evaluated as # define FP_ILOGB0 (-2147483647 - 1) # define FP_ILOGBNAN (-2147483647 - 1) +/* The GCC 4.6 compiler will define __FP_FAST_FMA{,F,L} if the fma{,f,l} + builtins are supported. */ +# if __FP_FAST_FMA +# define FP_FAST_FMA 1 +# endif + +# if __FP_FAST_FMAF +# define FP_FAST_FMAF 1 +# endif + +# if __FP_FAST_FMAL +# define FP_FAST_FMAL 1 +# endif + #endif /* ISO C99 */ diff --git a/libc/sysdeps/i386/fpu/bits/mathinline.h b/libc/sysdeps/i386/fpu/bits/mathinline.h index 800599752..666ec2bc2 100644 --- a/libc/sysdeps/i386/fpu/bits/mathinline.h +++ b/libc/sysdeps/i386/fpu/bits/mathinline.h @@ -1,6 +1,6 @@ /* Inline math functions for i387. - Copyright (C) 1995,1996,1997,1998,1999,2000,2001,2003,2004,2006,2007,2009 - Free Software Foundation, Inc. + Copyright (C) 1995,1996,1997,1998,1999,2000,2001,2003,2004,2006,2007,2009, + 2010 Free Software Foundation, Inc. This file is part of the GNU C Library. Contributed by John C. Bowman <bowman@math.ualberta.ca>, 1995. @@ -657,8 +657,6 @@ __NTH (ldexpl (long double __x, int __y)) __ldexp_code; } -__inline_mathcodeNP3 (fma, __x, __y, __z, return (__x * __y) + __z) - __inline_mathopNP (rint, "frndint") # endif /* __FAST_MATH__ */ diff --git a/libc/sysdeps/i386/i686/multiarch/s_fma.c b/libc/sysdeps/i386/i686/multiarch/s_fma.c index d9291b0be..2373caabc 100644 --- a/libc/sysdeps/i386/i686/multiarch/s_fma.c +++ b/libc/sysdeps/i386/i686/multiarch/s_fma.c @@ -33,4 +33,4 @@ weak_alias (__fma, fma) # define __fma __fma_ia32 #endif -#include <math/s_fma.c> +#include <sysdeps/ieee754/ldbl-96/s_fma.c> diff --git a/libc/sysdeps/ieee754/dbl-64/s_fma.c b/libc/sysdeps/ieee754/dbl-64/s_fma.c new file mode 100644 index 000000000..3b0bfd5ce --- /dev/null +++ b/libc/sysdeps/ieee754/dbl-64/s_fma.c @@ -0,0 +1,227 @@ +/* Compute x * y + z as ternary operation. + Copyright (C) 2010 Free Software Foundation, Inc. + This file is part of the GNU C Library. + Contributed by Jakub Jelinek <jakub@redhat.com>, 2010. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, write to the Free + Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA + 02111-1307 USA. */ + +#include <float.h> +#include <math.h> +#include <fenv.h> +#include <ieee754.h> + +/* This implementation uses rounding to odd to avoid problems with + double rounding. See a paper by Boldo and Melquiond: + http://www.lri.fr/~melquion/doc/08-tc.pdf */ + +double +__fma (double x, double y, double z) +{ + union ieee754_double u, v, w; + int adjust = 0; + u.d = x; + v.d = y; + w.d = z; + if (__builtin_expect (u.ieee.exponent + v.ieee.exponent + >= 0x7ff + IEEE754_DOUBLE_BIAS - DBL_MANT_DIG, 0) + || __builtin_expect (u.ieee.exponent >= 0x7ff - DBL_MANT_DIG, 0) + || __builtin_expect (v.ieee.exponent >= 0x7ff - DBL_MANT_DIG, 0) + || __builtin_expect (w.ieee.exponent >= 0x7ff - DBL_MANT_DIG, 0) + || __builtin_expect (u.ieee.exponent + v.ieee.exponent + <= IEEE754_DOUBLE_BIAS + DBL_MANT_DIG, 0)) + { + /* If z is Inf, but x and y are finite, the result should be + z rather than NaN. */ + if (w.ieee.exponent == 0x7ff + && u.ieee.exponent != 0x7ff + && v.ieee.exponent != 0x7ff) + return (z + x) + y; + /* If x or y or z is Inf/NaN, or if fma will certainly overflow, + or if x * y is less than half of DBL_DENORM_MIN, + compute as x * y + z. */ + if (u.ieee.exponent == 0x7ff + || v.ieee.exponent == 0x7ff + || w.ieee.exponent == 0x7ff + || u.ieee.exponent + v.ieee.exponent + > 0x7ff + IEEE754_DOUBLE_BIAS + || u.ieee.exponent + v.ieee.exponent + < IEEE754_DOUBLE_BIAS - DBL_MANT_DIG - 2) + return x * y + z; + if (u.ieee.exponent + v.ieee.exponent + >= 0x7ff + IEEE754_DOUBLE_BIAS - DBL_MANT_DIG) + { + /* Compute 1p-53 times smaller result and multiply + at the end. */ + if (u.ieee.exponent > v.ieee.exponent) + u.ieee.exponent -= DBL_MANT_DIG; + else + v.ieee.exponent -= DBL_MANT_DIG; + /* If x + y exponent is very large and z exponent is very small, + it doesn't matter if we don't adjust it. */ + if (w.ieee.exponent > DBL_MANT_DIG) + w.ieee.exponent -= DBL_MANT_DIG; + adjust = 1; + } + else if (w.ieee.exponent >= 0x7ff - DBL_MANT_DIG) + { + /* Similarly. + If z exponent is very large and x and y exponents are + very small, it doesn't matter if we don't adjust it. */ + if (u.ieee.exponent > v.ieee.exponent) + { + if (u.ieee.exponent > DBL_MANT_DIG) + u.ieee.exponent -= DBL_MANT_DIG; + } + else if (v.ieee.exponent > DBL_MANT_DIG) + v.ieee.exponent -= DBL_MANT_DIG; + w.ieee.exponent -= DBL_MANT_DIG; + adjust = 1; + } + else if (u.ieee.exponent >= 0x7ff - DBL_MANT_DIG) + { + u.ieee.exponent -= DBL_MANT_DIG; + if (v.ieee.exponent) + v.ieee.exponent += DBL_MANT_DIG; + else + v.d *= 0x1p53; + } + else if (v.ieee.exponent >= 0x7ff - DBL_MANT_DIG) + { + v.ieee.exponent -= DBL_MANT_DIG; + if (u.ieee.exponent) + u.ieee.exponent += DBL_MANT_DIG; + else + u.d *= 0x1p53; + } + else /* if (u.ieee.exponent + v.ieee.exponent + <= IEEE754_DOUBLE_BIAS + DBL_MANT_DIG) */ + { + if (u.ieee.exponent > v.ieee.exponent) + u.ieee.exponent += 2 * DBL_MANT_DIG; + else + v.ieee.exponent += 2 * DBL_MANT_DIG; + if (w.ieee.exponent <= 4 * DBL_MANT_DIG + 4) + { + if (w.ieee.exponent) + w.ieee.exponent += 2 * DBL_MANT_DIG; + else + w.d *= 0x1p106; + adjust = -1; + } + /* Otherwise x * y should just affect inexact + and nothing else. */ + } + x = u.d; + y = v.d; + z = w.d; + } + /* Multiplication m1 + m2 = x * y using Dekker's algorithm. */ +#define C ((1 << (DBL_MANT_DIG + 1) / 2) + 1) + double x1 = x * C; + double y1 = y * C; + double m1 = x * y; + x1 = (x - x1) + x1; + y1 = (y - y1) + y1; + double x2 = x - x1; + double y2 = y - y1; + double m2 = (((x1 * y1 - m1) + x1 * y2) + x2 * y1) + x2 * y2; + + /* Addition a1 + a2 = z + m1 using Knuth's algorithm. */ + double a1 = z + m1; + double t1 = a1 - z; + double t2 = a1 - t1; + t1 = m1 - t1; + t2 = z - t2; + double a2 = t1 + t2; + + fenv_t env; + feholdexcept (&env); + fesetround (FE_TOWARDZERO); + /* Perform m2 + a2 addition with round to odd. */ + u.d = a2 + m2; + + if (__builtin_expect (adjust == 0, 1)) + { + if ((u.ieee.mantissa1 & 1) == 0 && u.ieee.exponent != 0x7ff) + u.ieee.mantissa1 |= fetestexcept (FE_INEXACT) != 0; + feupdateenv (&env); + /* Result is a1 + u.d. */ + return a1 + u.d; + } + else if (__builtin_expect (adjust > 0, 1)) + { + if ((u.ieee.mantissa1 & 1) == 0 && u.ieee.exponent != 0x7ff) + u.ieee.mantissa1 |= fetestexcept (FE_INEXACT) != 0; + feupdateenv (&env); + /* Result is a1 + u.d, scaled up. */ + return (a1 + u.d) * 0x1p53; + } + else + { + if ((u.ieee.mantissa1 & 1) == 0) + u.ieee.mantissa1 |= fetestexcept (FE_INEXACT) != 0; + v.d = a1 + u.d; + int j = fetestexcept (FE_INEXACT) != 0; + feupdateenv (&env); + /* Ensure the following computations are performed in default rounding + mode instead of just reusing the round to zero computation. */ + asm volatile ("" : "=m" (u) : "m" (u)); + /* If a1 + u.d is exact, the only rounding happens during + scaling down. */ + if (j == 0) + return v.d * 0x1p-106; + /* If result rounded to zero is not subnormal, no double + rounding will occur. */ + if (v.ieee.exponent > 106) + return (a1 + u.d) * 0x1p-106; + /* If v.d * 0x1p-106 with round to zero is a subnormal above + or equal to DBL_MIN / 2, then v.d * 0x1p-106 shifts mantissa + down just by 1 bit, which means v.ieee.mantissa1 |= j would + change the round bit, not sticky or guard bit. + v.d * 0x1p-106 never normalizes by shifting up, + so round bit plus sticky bit should be already enough + for proper rounding. */ + if (v.ieee.exponent == 106) + { + /* v.ieee.mantissa1 & 2 is LSB bit of the result before rounding, + v.ieee.mantissa1 & 1 is the round bit and j is our sticky + bit. In round-to-nearest 001 rounds down like 00, + 011 rounds up, even though 01 rounds down (thus we need + to adjust), 101 rounds down like 10 and 111 rounds up + like 11. */ + if ((v.ieee.mantissa1 & 3) == 1) + { + v.d *= 0x1p-106; + if (v.ieee.negative) + return v.d - 0x1p-1074 /* __DBL_DENORM_MIN__ */; + else + return v.d + 0x1p-1074 /* __DBL_DENORM_MIN__ */; + } + else + return v.d * 0x1p-106; + } + v.ieee.mantissa1 |= j; + return v.d * 0x1p-106; + } +} +#ifndef __fma +weak_alias (__fma, fma) +#endif + +#ifdef NO_LONG_DOUBLE +strong_alias (__fma, __fmal) +weak_alias (__fmal, fmal) +#endif diff --git a/libc/sysdeps/ieee754/dbl-64/s_fmaf.c b/libc/sysdeps/ieee754/dbl-64/s_fmaf.c index 972187660..cd16cd1dc 100644 --- a/libc/sysdeps/ieee754/dbl-64/s_fmaf.c +++ b/libc/sysdeps/ieee754/dbl-64/s_fmaf.c @@ -39,7 +39,7 @@ __fmaf (float x, float y, float z) fesetround (FE_TOWARDZERO); /* Perform addition with round to odd. */ u.d = temp + (double) z; - if ((u.ieee.mantissa1 & 1) == 0 && u.ieee.exponent != 0xff) + if ((u.ieee.mantissa1 & 1) == 0 && u.ieee.exponent != 0x7ff) u.ieee.mantissa1 |= fetestexcept (FE_INEXACT) != 0; feupdateenv (&env); /* And finally truncation with round to nearest. */ diff --git a/libc/sysdeps/ieee754/ldbl-128/s_fma.c b/libc/sysdeps/ieee754/ldbl-128/s_fma.c new file mode 100644 index 000000000..78055e5c6 --- /dev/null +++ b/libc/sysdeps/ieee754/ldbl-128/s_fma.c @@ -0,0 +1,50 @@ +/* Compute x * y + z as ternary operation. + Copyright (C) 2010 Free Software Foundation, Inc. + This file is part of the GNU C Library. + Contributed by Jakub Jelinek <jakub@redhat.com>, 2010. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, write to the Free + Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA + 02111-1307 USA. */ + +#include <math.h> +#include <fenv.h> +#include <ieee754.h> + +/* This implementation relies on long double being more than twice as + precise as double and uses rounding to odd in order to avoid problems + with double rounding. + See a paper by Boldo and Melquiond: + http://www.lri.fr/~melquion/doc/08-tc.pdf */ + +double +__fma (double x, double y, double z) +{ + fenv_t env; + /* Multiplication is always exact. */ + long double temp = (long double) x * (long double) y; + union ieee854_long_double u; + feholdexcept (&env); + fesetround (FE_TOWARDZERO); + /* Perform addition with round to odd. */ + u.d = temp + (long double) z; + if ((u.ieee.mantissa3 & 1) == 0 && u.ieee.exponent != 0x7fff) + u.ieee.mantissa3 |= fetestexcept (FE_INEXACT) != 0; + feupdateenv (&env); + /* And finally truncation with round to nearest. */ + return (double) u.d; +} +#ifndef __fma +weak_alias (__fma, fma) +#endif diff --git a/libc/sysdeps/ieee754/ldbl-128/s_fmal.c b/libc/sysdeps/ieee754/ldbl-128/s_fmal.c new file mode 100644 index 000000000..b750538b8 --- /dev/null +++ b/libc/sysdeps/ieee754/ldbl-128/s_fmal.c @@ -0,0 +1,223 @@ +/* Compute x * y + z as ternary operation. + Copyright (C) 2010 Free Software Foundation, Inc. + This file is part of the GNU C Library. + Contributed by Jakub Jelinek <jakub@redhat.com>, 2010. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, write to the Free + Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA + 02111-1307 USA. */ + +#include <float.h> +#include <math.h> +#include <fenv.h> +#include <ieee754.h> + +/* This implementation uses rounding to odd to avoid problems with + double rounding. See a paper by Boldo and Melquiond: + http://www.lri.fr/~melquion/doc/08-tc.pdf */ + +long double +__fmal (long double x, long double y, long double z) +{ + union ieee854_long_double u, v, w; + int adjust = 0; + u.d = x; + v.d = y; + w.d = z; + if (__builtin_expect (u.ieee.exponent + v.ieee.exponent + >= 0x7fff + IEEE854_LONG_DOUBLE_BIAS + - LDBL_MANT_DIG, 0) + || __builtin_expect (u.ieee.exponent >= 0x7fff - LDBL_MANT_DIG, 0) + || __builtin_expect (v.ieee.exponent >= 0x7fff - LDBL_MANT_DIG, 0) + || __builtin_expect (w.ieee.exponent >= 0x7fff - LDBL_MANT_DIG, 0) + || __builtin_expect (u.ieee.exponent + v.ieee.exponent + <= IEEE854_LONG_DOUBLE_BIAS + LDBL_MANT_DIG, 0)) + { + /* If z is Inf, but x and y are finite, the result should be + z rather than NaN. */ + if (w.ieee.exponent == 0x7fff + && u.ieee.exponent != 0x7fff + && v.ieee.exponent != 0x7fff) + return (z + x) + y; + /* If x or y or z is Inf/NaN, or if fma will certainly overflow, + or if x * y is less than half of LDBL_DENORM_MIN, + compute as x * y + z. */ + if (u.ieee.exponent == 0x7fff + || v.ieee.exponent == 0x7fff + || w.ieee.exponent == 0x7fff + || u.ieee.exponent + v.ieee.exponent + > 0x7fff + IEEE854_LONG_DOUBLE_BIAS + || u.ieee.exponent + v.ieee.exponent + < IEEE854_LONG_DOUBLE_BIAS - LDBL_MANT_DIG - 2) + return x * y + z; + if (u.ieee.exponent + v.ieee.exponent + >= 0x7fff + IEEE854_LONG_DOUBLE_BIAS - LDBL_MANT_DIG) + { + /* Compute 1p-113 times smaller result and multiply + at the end. */ + if (u.ieee.exponent > v.ieee.exponent) + u.ieee.exponent -= LDBL_MANT_DIG; + else + v.ieee.exponent -= LDBL_MANT_DIG; + /* If x + y exponent is very large and z exponent is very small, + it doesn't matter if we don't adjust it. */ + if (w.ieee.exponent > LDBL_MANT_DIG) + w.ieee.exponent -= LDBL_MANT_DIG; + adjust = 1; + } + else if (w.ieee.exponent >= 0x7fff - LDBL_MANT_DIG) + { + /* Similarly. + If z exponent is very large and x and y exponents are + very small, it doesn't matter if we don't adjust it. */ + if (u.ieee.exponent > v.ieee.exponent) + { + if (u.ieee.exponent > LDBL_MANT_DIG) + u.ieee.exponent -= LDBL_MANT_DIG; + } + else if (v.ieee.exponent > LDBL_MANT_DIG) + v.ieee.exponent -= LDBL_MANT_DIG; + w.ieee.exponent -= LDBL_MANT_DIG; + adjust = 1; + } + else if (u.ieee.exponent >= 0x7fff - LDBL_MANT_DIG) + { + u.ieee.exponent -= LDBL_MANT_DIG; + if (v.ieee.exponent) + v.ieee.exponent += LDBL_MANT_DIG; + else + v.d *= 0x1p113L; + } + else if (v.ieee.exponent >= 0x7fff - LDBL_MANT_DIG) + { + v.ieee.exponent -= LDBL_MANT_DIG; + if (u.ieee.exponent) + u.ieee.exponent += LDBL_MANT_DIG; + else + u.d *= 0x1p113L; + } + else /* if (u.ieee.exponent + v.ieee.exponent + <= IEEE854_LONG_DOUBLE_BIAS + LDBL_MANT_DIG) */ + { + if (u.ieee.exponent > v.ieee.exponent) + u.ieee.exponent += 2 * LDBL_MANT_DIG; + else + v.ieee.exponent += 2 * LDBL_MANT_DIG; + if (w.ieee.exponent <= 4 * LDBL_MANT_DIG + 4) + { + if (w.ieee.exponent) + w.ieee.exponent += 2 * LDBL_MANT_DIG; + else + w.d *= 0x1p226L; + adjust = -1; + } + /* Otherwise x * y should just affect inexact + and nothing else. */ + } + x = u.d; + y = v.d; + z = w.d; + } + /* Multiplication m1 + m2 = x * y using Dekker's algorithm. */ +#define C ((1LL << (LDBL_MANT_DIG + 1) / 2) + 1) + long double x1 = x * C; + long double y1 = y * C; + long double m1 = x * y; + x1 = (x - x1) + x1; + y1 = (y - y1) + y1; + long double x2 = x - x1; + long double y2 = y - y1; + long double m2 = (((x1 * y1 - m1) + x1 * y2) + x2 * y1) + x2 * y2; + + /* Addition a1 + a2 = z + m1 using Knuth's algorithm. */ + long double a1 = z + m1; + long double t1 = a1 - z; + long double t2 = a1 - t1; + t1 = m1 - t1; + t2 = z - t2; + long double a2 = t1 + t2; + + fenv_t env; + feholdexcept (&env); + fesetround (FE_TOWARDZERO); + /* Perform m2 + a2 addition with round to odd. */ + u.d = a2 + m2; + + if (__builtin_expect (adjust == 0, 1)) + { + if ((u.ieee.mantissa3 & 1) == 0 && u.ieee.exponent != 0x7fff) + u.ieee.mantissa3 |= fetestexcept (FE_INEXACT) != 0; + feupdateenv (&env); + /* Result is a1 + u.d. */ + return a1 + u.d; + } + else if (__builtin_expect (adjust > 0, 1)) + { + if ((u.ieee.mantissa3 & 1) == 0 && u.ieee.exponent != 0x7fff) + u.ieee.mantissa3 |= fetestexcept (FE_INEXACT) != 0; + feupdateenv (&env); + /* Result is a1 + u.d, scaled up. */ + return (a1 + u.d) * 0x1p113L; + } + else + { + if ((u.ieee.mantissa3 & 1) == 0) + u.ieee.mantissa3 |= fetestexcept (FE_INEXACT) != 0; + v.d = a1 + u.d; + /* Ensure the addition is not scheduled after fetestexcept call. */ + asm volatile ("" : "m" (v)); + int j = fetestexcept (FE_INEXACT) != 0; + feupdateenv (&env); + /* Ensure the following computations are performed in default rounding + mode instead of just reusing the round to zero computation. */ + asm volatile ("" : "=m" (u) : "m" (u)); + /* If a1 + u.d is exact, the only rounding happens during + scaling down. */ + if (j == 0) + return v.d * 0x1p-226L; + /* If result rounded to zero is not subnormal, no double + rounding will occur. */ + if (v.ieee.exponent > 226) + return (a1 + u.d) * 0x1p-226L; + /* If v.d * 0x1p-226L with round to zero is a subnormal above + or equal to LDBL_MIN / 2, then v.d * 0x1p-226L shifts mantissa + down just by 1 bit, which means v.ieee.mantissa3 |= j would + change the round bit, not sticky or guard bit. + v.d * 0x1p-226L never normalizes by shifting up, + so round bit plus sticky bit should be already enough + for proper rounding. */ + if (v.ieee.exponent == 226) + { + /* v.ieee.mantissa3 & 2 is LSB bit of the result before rounding, + v.ieee.mantissa3 & 1 is the round bit and j is our sticky + bit. In round-to-nearest 001 rounds down like 00, + 011 rounds up, even though 01 rounds down (thus we need + to adjust), 101 rounds down like 10 and 111 rounds up + like 11. */ + if ((v.ieee.mantissa3 & 3) == 1) + { + v.d *= 0x1p-226L; + if (v.ieee.negative) + return v.d - 0x1p-16494L /* __LDBL_DENORM_MIN__ */; + else + return v.d + 0x1p-16494L /* __LDBL_DENORM_MIN__ */; + } + else + return v.d * 0x1p-226L; + } + v.ieee.mantissa3 |= j; + return v.d * 0x1p-226L; + } +} +weak_alias (__fmal, fmal) diff --git a/libc/sysdeps/ieee754/ldbl-64-128/s_fmal.c b/libc/sysdeps/ieee754/ldbl-64-128/s_fmal.c new file mode 100644 index 000000000..218aa52b3 --- /dev/null +++ b/libc/sysdeps/ieee754/ldbl-64-128/s_fmal.c @@ -0,0 +1,5 @@ +#include <math_ldbl_opt.h> +#undef weak_alias +#define weak_alias(n,a) +#include <sysdeps/ieee754/ldbl-128/s_fmal.c> +long_double_symbol (libm, __fmal, fmal); diff --git a/libc/sysdeps/ieee754/ldbl-96/s_fma.c b/libc/sysdeps/ieee754/ldbl-96/s_fma.c new file mode 100644 index 000000000..6c7e9d0d3 --- /dev/null +++ b/libc/sysdeps/ieee754/ldbl-96/s_fma.c @@ -0,0 +1,79 @@ +/* Compute x * y + z as ternary operation. + Copyright (C) 2010 Free Software Foundation, Inc. + This file is part of the GNU C Library. + Contributed by Jakub Jelinek <jakub@redhat.com>, 2010. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, write to the Free + Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA + 02111-1307 USA. */ + +#include <float.h> +#include <math.h> +#include <fenv.h> +#include <ieee754.h> + +/* This implementation uses rounding to odd to avoid problems with + double rounding. See a paper by Boldo and Melquiond: + http://www.lri.fr/~melquion/doc/08-tc.pdf */ + +double +__fma (double x, double y, double z) +{ + if (__builtin_expect (isinf (z), 0)) + { + /* If z is Inf, but x and y are finite, the result should be + z rather than NaN. */ + if (finite (x) && finite (y)) + return (z + x) + y; + return (x * y) + z; + } + + /* Multiplication m1 + m2 = x * y using Dekker's algorithm. */ +#define C ((1ULL << (LDBL_MANT_DIG + 1) / 2) + 1) + long double x1 = (long double) x * C; + long double y1 = (long double) y * C; + long double m1 = (long double) x * y; + x1 = (x - x1) + x1; + y1 = (y - y1) + y1; + long double x2 = x - x1; + long double y2 = y - y1; + long double m2 = (((x1 * y1 - m1) + x1 * y2) + x2 * y1) + x2 * y2; + + /* Addition a1 + a2 = z + m1 using Knuth's algorithm. */ + long double a1 = z + m1; + long double t1 = a1 - z; + long double t2 = a1 - t1; + t1 = m1 - t1; + t2 = z - t2; + long double a2 = t1 + t2; + + fenv_t env; + feholdexcept (&env); + fesetround (FE_TOWARDZERO); + /* Perform m2 + a2 addition with round to odd. */ + a2 = a2 + m2; + + /* Add that to a1 again using rounding to odd. */ + union ieee854_long_double u; + u.d = a1 + a2; + if ((u.ieee.mantissa1 & 1) == 0 && u.ieee.exponent != 0x7fff) + u.ieee.mantissa1 |= fetestexcept (FE_INEXACT) != 0; + feupdateenv (&env); + + /* Add finally round to double precision. */ + return u.d; +} +#ifndef __fma +weak_alias (__fma, fma) +#endif diff --git a/libc/sysdeps/ieee754/ldbl-96/s_fmal.c b/libc/sysdeps/ieee754/ldbl-96/s_fmal.c new file mode 100644 index 000000000..340f02536 --- /dev/null +++ b/libc/sysdeps/ieee754/ldbl-96/s_fmal.c @@ -0,0 +1,221 @@ +/* Compute x * y + z as ternary operation. + Copyright (C) 2010 Free Software Foundation, Inc. + This file is part of the GNU C Library. + Contributed by Jakub Jelinek <jakub@redhat.com>, 2010. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, write to the Free + Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA + 02111-1307 USA. */ + +#include <float.h> +#include <math.h> +#include <fenv.h> +#include <ieee754.h> + +/* This implementation uses rounding to odd to avoid problems with + double rounding. See a paper by Boldo and Melquiond: + http://www.lri.fr/~melquion/doc/08-tc.pdf */ + +long double +__fmal (long double x, long double y, long double z) +{ + union ieee854_long_double u, v, w; + int adjust = 0; + u.d = x; + v.d = y; + w.d = z; + if (__builtin_expect (u.ieee.exponent + v.ieee.exponent + >= 0x7fff + IEEE854_LONG_DOUBLE_BIAS + - LDBL_MANT_DIG, 0) + || __builtin_expect (u.ieee.exponent >= 0x7fff - LDBL_MANT_DIG, 0) + || __builtin_expect (v.ieee.exponent >= 0x7fff - LDBL_MANT_DIG, 0) + || __builtin_expect (w.ieee.exponent >= 0x7fff - LDBL_MANT_DIG, 0) + || __builtin_expect (u.ieee.exponent + v.ieee.exponent + <= IEEE854_LONG_DOUBLE_BIAS + LDBL_MANT_DIG, 0)) + { + /* If z is Inf, but x and y are finite, the result should be + z rather than NaN. */ + if (w.ieee.exponent == 0x7fff + && u.ieee.exponent != 0x7fff + && v.ieee.exponent != 0x7fff) + return (z + x) + y; + /* If x or y or z is Inf/NaN, or if fma will certainly overflow, + or if x * y is less than half of LDBL_DENORM_MIN, + compute as x * y + z. */ + if (u.ieee.exponent == 0x7fff + || v.ieee.exponent == 0x7fff + || w.ieee.exponent == 0x7fff + || u.ieee.exponent + v.ieee.exponent + > 0x7fff + IEEE854_LONG_DOUBLE_BIAS + || u.ieee.exponent + v.ieee.exponent + < IEEE854_LONG_DOUBLE_BIAS - LDBL_MANT_DIG - 2) + return x * y + z; + if (u.ieee.exponent + v.ieee.exponent + >= 0x7fff + IEEE854_LONG_DOUBLE_BIAS - LDBL_MANT_DIG) + { + /* Compute 1p-64 times smaller result and multiply + at the end. */ + if (u.ieee.exponent > v.ieee.exponent) + u.ieee.exponent -= LDBL_MANT_DIG; + else + v.ieee.exponent -= LDBL_MANT_DIG; + /* If x + y exponent is very large and z exponent is very small, + it doesn't matter if we don't adjust it. */ + if (w.ieee.exponent > LDBL_MANT_DIG) + w.ieee.exponent -= LDBL_MANT_DIG; + adjust = 1; + } + else if (w.ieee.exponent >= 0x7fff - LDBL_MANT_DIG) + { + /* Similarly. + If z exponent is very large and x and y exponents are + very small, it doesn't matter if we don't adjust it. */ + if (u.ieee.exponent > v.ieee.exponent) + { + if (u.ieee.exponent > LDBL_MANT_DIG) + u.ieee.exponent -= LDBL_MANT_DIG; + } + else if (v.ieee.exponent > LDBL_MANT_DIG) + v.ieee.exponent -= LDBL_MANT_DIG; + w.ieee.exponent -= LDBL_MANT_DIG; + adjust = 1; + } + else if (u.ieee.exponent >= 0x7fff - LDBL_MANT_DIG) + { + u.ieee.exponent -= LDBL_MANT_DIG; + if (v.ieee.exponent) + v.ieee.exponent += LDBL_MANT_DIG; + else + v.d *= 0x1p64L; + } + else if (v.ieee.exponent >= 0x7fff - LDBL_MANT_DIG) + { + v.ieee.exponent -= LDBL_MANT_DIG; + if (u.ieee.exponent) + u.ieee.exponent += LDBL_MANT_DIG; + else + u.d *= 0x1p64L; + } + else /* if (u.ieee.exponent + v.ieee.exponent + <= IEEE854_LONG_DOUBLE_BIAS + LDBL_MANT_DIG) */ + { + if (u.ieee.exponent > v.ieee.exponent) + u.ieee.exponent += 2 * LDBL_MANT_DIG; + else + v.ieee.exponent += 2 * LDBL_MANT_DIG; + if (w.ieee.exponent <= 4 * LDBL_MANT_DIG + 4) + { + if (w.ieee.exponent) + w.ieee.exponent += 2 * LDBL_MANT_DIG; + else + w.d *= 0x1p128L; + adjust = -1; + } + /* Otherwise x * y should just affect inexact + and nothing else. */ + } + x = u.d; + y = v.d; + z = w.d; + } + /* Multiplication m1 + m2 = x * y using Dekker's algorithm. */ +#define C ((1LL << (LDBL_MANT_DIG + 1) / 2) + 1) + long double x1 = x * C; + long double y1 = y * C; + long double m1 = x * y; + x1 = (x - x1) + x1; + y1 = (y - y1) + y1; + long double x2 = x - x1; + long double y2 = y - y1; + long double m2 = (((x1 * y1 - m1) + x1 * y2) + x2 * y1) + x2 * y2; + + /* Addition a1 + a2 = z + m1 using Knuth's algorithm. */ + long double a1 = z + m1; + long double t1 = a1 - z; + long double t2 = a1 - t1; + t1 = m1 - t1; + t2 = z - t2; + long double a2 = t1 + t2; + + fenv_t env; + feholdexcept (&env); + fesetround (FE_TOWARDZERO); + /* Perform m2 + a2 addition with round to odd. */ + u.d = a2 + m2; + + if (__builtin_expect (adjust == 0, 1)) + { + if ((u.ieee.mantissa1 & 1) == 0 && u.ieee.exponent != 0x7fff) + u.ieee.mantissa1 |= fetestexcept (FE_INEXACT) != 0; + feupdateenv (&env); + /* Result is a1 + u.d. */ + return a1 + u.d; + } + else if (__builtin_expect (adjust > 0, 1)) + { + if ((u.ieee.mantissa1 & 1) == 0 && u.ieee.exponent != 0x7fff) + u.ieee.mantissa1 |= fetestexcept (FE_INEXACT) != 0; + feupdateenv (&env); + /* Result is a1 + u.d, scaled up. */ + return (a1 + u.d) * 0x1p64L; + } + else + { + if ((u.ieee.mantissa1 & 1) == 0) + u.ieee.mantissa1 |= fetestexcept (FE_INEXACT) != 0; + v.d = a1 + u.d; + int j = fetestexcept (FE_INEXACT) != 0; + feupdateenv (&env); + /* Ensure the following computations are performed in default rounding + mode instead of just reusing the round to zero computation. */ + asm volatile ("" : "=m" (u) : "m" (u)); + /* If a1 + u.d is exact, the only rounding happens during + scaling down. */ + if (j == 0) + return v.d * 0x1p-128L; + /* If result rounded to zero is not subnormal, no double + rounding will occur. */ + if (v.ieee.exponent > 128) + return (a1 + u.d) * 0x1p-128L; + /* If v.d * 0x1p-128L with round to zero is a subnormal above + or equal to LDBL_MIN / 2, then v.d * 0x1p-128L shifts mantissa + down just by 1 bit, which means v.ieee.mantissa1 |= j would + change the round bit, not sticky or guard bit. + v.d * 0x1p-128L never normalizes by shifting up, + so round bit plus sticky bit should be already enough + for proper rounding. */ + if (v.ieee.exponent == 128) + { + /* v.ieee.mantissa1 & 2 is LSB bit of the result before rounding, + v.ieee.mantissa1 & 1 is the round bit and j is our sticky + bit. In round-to-nearest 001 rounds down like 00, + 011 rounds up, even though 01 rounds down (thus we need + to adjust), 101 rounds down like 10 and 111 rounds up + like 11. */ + if ((v.ieee.mantissa1 & 3) == 1) + { + v.d *= 0x1p-128L; + if (v.ieee.negative) + return v.d - 0x1p-16445L /* __LDBL_DENORM_MIN__ */; + else + return v.d + 0x1p-16445L /* __LDBL_DENORM_MIN__ */; + } + else + return v.d * 0x1p-128L; + } + v.ieee.mantissa1 |= j; + return v.d * 0x1p-128L; + } +} +weak_alias (__fmal, fmal) diff --git a/libc/sysdeps/ieee754/ldbl-opt/s_fma.c b/libc/sysdeps/ieee754/ldbl-opt/s_fma.c index 2ee721413..1723c5c30 100644 --- a/libc/sysdeps/ieee754/ldbl-opt/s_fma.c +++ b/libc/sysdeps/ieee754/ldbl-opt/s_fma.c @@ -1,5 +1,5 @@ #include <math_ldbl_opt.h> -#include <math/s_fma.c> +#include <sysdeps/ieee754/dbl-64/s_fma.c> #if LONG_DOUBLE_COMPAT(libm, GLIBC_2_1) compat_symbol (libm, __fma, fmal, GLIBC_2_1); #endif diff --git a/libc/sysdeps/powerpc/bits/mathdef.h b/libc/sysdeps/powerpc/bits/mathdef.h index e3d53922c..5207062ee 100644 --- a/libc/sysdeps/powerpc/bits/mathdef.h +++ b/libc/sysdeps/powerpc/bits/mathdef.h @@ -1,4 +1,4 @@ -/* Copyright (C) 1997,1998,1999,2000,2003,2004,2006 +/* Copyright (C) 1997,1998,1999,2000,2003,2004,2006,2010 Free Software Foundation, Inc. This file is part of the GNU C Library. @@ -39,4 +39,8 @@ typedef double double_t; # define FP_ILOGB0 (-2147483647) # define FP_ILOGBNAN (2147483647) +/* The powerpc has a combined multiply/add instruction. */ +# define FP_FAST_FMA 1 +# define FP_FAST_FMAF 1 + #endif /* ISO C99 */ diff --git a/libc/sysdeps/i386/fpu/s_fma.S b/libc/sysdeps/powerpc/fpu/s_fma.S index db4959ccf..9c650eab0 100644 --- a/libc/sysdeps/i386/fpu/s_fma.S +++ b/libc/sysdeps/powerpc/fpu/s_fma.S @@ -1,7 +1,6 @@ -/* Compute (X * Y) + Z as ternary operation. - Copyright (C) 1997, 1998 Free Software Foundation, Inc. +/* Compute x * y + z as ternary operation. PowerPC version. + Copyright (C) 2010 Free Software Foundation, Inc. This file is part of the GNU C Library. - Contributed by Ulrich Drepper <drepper@cygnus.com>, 1997. The GNU C Library is free software; you can redistribute it and/or modify it under the terms of the GNU Lesser General Public @@ -20,12 +19,15 @@ #include <sysdep.h> - .text ENTRY(__fma) - fldl 4(%esp) // x - fmull 12(%esp) // x * y - fldl 20(%esp) // z : x * y - faddp // (x * y) + z - ret +/* double [f1] fma (double [f1] x, double [f2] y, double [f3] z); */ + fmadd fp1,fp1,fp2,fp3 + blr END(__fma) -weak_alias (__fma, fma) + +weak_alias (__fma,fma) + +#ifdef NO_LONG_DOUBLE +weak_alias (__fma,__fmal) +weak_alias (__fma,fmal) +#endif diff --git a/libc/sysdeps/i386/fpu/s_fmaf.S b/libc/sysdeps/powerpc/fpu/s_fmaf.S index 5f875532a..1a755fe2d 100644 --- a/libc/sysdeps/i386/fpu/s_fmaf.S +++ b/libc/sysdeps/powerpc/fpu/s_fmaf.S @@ -1,7 +1,6 @@ -/* Compute (X * Y) + Z as ternary operation. - Copyright (C) 1997 Free Software Foundation, Inc. +/* Compute x * y + z as ternary operation. PowerPC version. + Copyright (C) 2010 Free Software Foundation, Inc. This file is part of the GNU C Library. - Contributed by Ulrich Drepper <drepper@cygnus.com>, 1997. The GNU C Library is free software; you can redistribute it and/or modify it under the terms of the GNU Lesser General Public @@ -20,12 +19,10 @@ #include <sysdep.h> - .text ENTRY(__fmaf) - flds 4(%esp) // x - fmuls 8(%esp) // x * y - flds 12(%esp) // z : x * y - faddp // (x * y) + z - ret +/* float [f1] fmaf (float [f1] x, float [f2] y, float [f3] z); */ + fmadds fp1,fp1,fp2,fp3 + blr END(__fmaf) -weak_alias (__fmaf, fmaf) + +weak_alias (__fmaf,fmaf) diff --git a/libc/sysdeps/powerpc/powerpc32/fpu/s_fma.S b/libc/sysdeps/powerpc/powerpc32/fpu/s_fma.S new file mode 100644 index 000000000..d40695c63 --- /dev/null +++ b/libc/sysdeps/powerpc/powerpc32/fpu/s_fma.S @@ -0,0 +1,5 @@ +#include <math_ldbl_opt.h> +#include <sysdeps/powerpc/fpu/s_fma.S> +#if LONG_DOUBLE_COMPAT(libm, GLIBC_2_1) +compat_symbol (libm, __fma, fmal, GLIBC_2_1) +#endif diff --git a/libc/sysdeps/powerpc/powerpc64/fpu/s_fma.S b/libc/sysdeps/powerpc/powerpc64/fpu/s_fma.S new file mode 100644 index 000000000..d40695c63 --- /dev/null +++ b/libc/sysdeps/powerpc/powerpc64/fpu/s_fma.S @@ -0,0 +1,5 @@ +#include <math_ldbl_opt.h> +#include <sysdeps/powerpc/fpu/s_fma.S> +#if LONG_DOUBLE_COMPAT(libm, GLIBC_2_1) +compat_symbol (libm, __fma, fmal, GLIBC_2_1) +#endif diff --git a/libc/sysdeps/s390/fpu/s_fma.c b/libc/sysdeps/s390/fpu/s_fma.c new file mode 100644 index 000000000..f68236231 --- /dev/null +++ b/libc/sysdeps/s390/fpu/s_fma.c @@ -0,0 +1,37 @@ +/* Compute x * y + z as ternary operation. S/390 version. + Copyright (C) 2010 Free Software Foundation, Inc. + This file is part of the GNU C Library. + Contributed by Jakub Jelinek <jakub@redhat.com>, 2010. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, write to the Free + Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA + 02111-1307 USA. */ + +#include <math.h> + +double +__fma (double x, double y, double z) +{ + double r; + asm ("madbr %0,%1,%2" : "=f" (r) : "%f" (x), "fR" (y), "0" (z)); + return r; +} +#ifndef __fma +weak_alias (__fma, fma) +#endif + +#ifdef NO_LONG_DOUBLE +strong_alias (__fma, __fmal) +weak_alias (__fmal, fmal) +#endif diff --git a/libc/sysdeps/i386/fpu/s_fmal.S b/libc/sysdeps/s390/fpu/s_fmaf.c index 1837f848c..030d7b229 100644 --- a/libc/sysdeps/i386/fpu/s_fmal.S +++ b/libc/sysdeps/s390/fpu/s_fmaf.c @@ -1,7 +1,7 @@ -/* Compute (X * Y) + Z as ternary operation. - Copyright (C) 1997 Free Software Foundation, Inc. +/* Compute x * y + z as ternary operation. S/390 version. + Copyright (C) 2010 Free Software Foundation, Inc. This file is part of the GNU C Library. - Contributed by Ulrich Drepper <drepper@cygnus.com>, 1997. + Contributed by Jakub Jelinek <jakub@redhat.com>, 2010. The GNU C Library is free software; you can redistribute it and/or modify it under the terms of the GNU Lesser General Public @@ -18,15 +18,15 @@ Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA. */ -#include <sysdep.h> +#include <math.h> - .text -ENTRY(__fmal) - fldt 4(%esp) // x - fldt 16(%esp) // x : y - fmulp // x * y - fldt 28(%esp) // z : x * y - faddp // (x * y) + z - ret -END(__fmal) -weak_alias (__fmal, fmal) +float +__fmaf (float x, float y, float z) +{ + float r; + asm ("maebr %0,%1,%2" : "=f" (r) : "%f" (x), "fR" (y), "0" (z)); + return r; +} +#ifndef __fmaf +weak_alias (__fmaf, fmaf) +#endif diff --git a/libc/sysdeps/unix/sysv/linux/s390/fpu/s_fma.c b/libc/sysdeps/unix/sysv/linux/s390/fpu/s_fma.c new file mode 100644 index 000000000..2b2d50503 --- /dev/null +++ b/libc/sysdeps/unix/sysv/linux/s390/fpu/s_fma.c @@ -0,0 +1,5 @@ +#include <math_ldbl_opt.h> +#include <sysdeps/s390/fpu/s_fma.c> +#if LONG_DOUBLE_COMPAT(libm, GLIBC_2_1) +compat_symbol (libm, __fma, fmal, GLIBC_2_1); +#endif diff --git a/libc/sysdeps/x86_64/bits/mathdef.h b/libc/sysdeps/x86_64/bits/mathdef.h index 7b1618959..91463921b 100644 --- a/libc/sysdeps/x86_64/bits/mathdef.h +++ b/libc/sysdeps/x86_64/bits/mathdef.h @@ -1,4 +1,4 @@ -/* Copyright (C) 2001, 2004 Free Software Foundation, Inc. +/* Copyright (C) 2001, 2004, 2010 Free Software Foundation, Inc. This file is part of the GNU C Library. The GNU C Library is free software; you can redistribute it and/or @@ -45,4 +45,18 @@ typedef long double double_t; /* `double' expressions are evaluated as # define FP_ILOGB0 (-2147483647 - 1) # define FP_ILOGBNAN (-2147483647 - 1) +/* The GCC 4.6 compiler will define __FP_FAST_FMA{,F,L} if the fma{,f,l} + builtins are supported. */ +# if __FP_FAST_FMA +# define FP_FAST_FMA 1 +# endif + +# if __FP_FAST_FMAF +# define FP_FAST_FMAF 1 +# endif + +# if __FP_FAST_FMAL +# define FP_FAST_FMAL 1 +# endif + #endif /* ISO C99 */ diff --git a/libc/sysdeps/x86_64/multiarch/s_fma.c b/libc/sysdeps/x86_64/multiarch/s_fma.c index cfecf9b41..9a680c68f 100644 --- a/libc/sysdeps/x86_64/multiarch/s_fma.c +++ b/libc/sysdeps/x86_64/multiarch/s_fma.c @@ -1,5 +1,5 @@ /* FMA version of fma. - Copyright (C) 2009 Free Software Foundation, Inc. + Copyright (C) 2009, 2010 Free Software Foundation, Inc. Contributed by Intel Corporation. This file is part of the GNU C Library. @@ -40,4 +40,4 @@ weak_alias (__fma, fma) # define __fma __fma_sse2 #endif -#include <math/s_fma.c> +#include <sysdeps/ieee754/dbl-64/s_fma.c> diff --git a/ports/ChangeLog.arm b/ports/ChangeLog.arm index 02affa8ce..69a28c6ba 100644 --- a/ports/ChangeLog.arm +++ b/ports/ChangeLog.arm @@ -1,3 +1,8 @@ +2010-10-19 Joseph Myers <joseph@codesourcery.com> + + * sysdeps/arm/bits/mathdef.h (FP_FAST_FMA, FP_FAST_FMAF, + FP_FAST_FMAL): Define if associated GCC macros are defined. + 2010-10-11 Joseph Myers <joseph@codesourcery.com> * sysdeps/arm/eabi/feupdateenv.c (feupdateenv): Add diff --git a/ports/ChangeLog.mips b/ports/ChangeLog.mips index 8daa71c15..8ea85cce9 100644 --- a/ports/ChangeLog.mips +++ b/ports/ChangeLog.mips @@ -1,3 +1,10 @@ +2010-10-21 Joseph Myers <joseph@codesourcery.com> + + * sysdeps/unix/sysv/linux/mips/mips32/syscalls.list, + sysdeps/unix/sysv/linux/mips/mips64/n32/syscalls.list, + sysdeps/unix/sysv/linux/mips/mips64/n64/syscalls.list: Add entries + for fanotify_mark. + 2010-10-11 Joseph Myers <joseph@codesourcery.com> * sysdeps/mips/fpu/feupdateenv.c (feupdateenv): Add diff --git a/ports/sysdeps/arm/bits/mathdef.h b/ports/sysdeps/arm/bits/mathdef.h index daca1102f..97f5d4d22 100644 --- a/ports/sysdeps/arm/bits/mathdef.h +++ b/ports/sysdeps/arm/bits/mathdef.h @@ -1,4 +1,4 @@ -/* Copyright (C) 1999, 2000, 2004, 2006, 2007 +/* Copyright (C) 1999, 2000, 2004, 2006, 2007, 2010 Free Software Foundation, Inc. This file is part of the GNU C Library. @@ -34,6 +34,20 @@ typedef double double_t; /* `double' expressions are evaluated as # define FP_ILOGB0 (-2147483647) # define FP_ILOGBNAN (2147483647) +/* The GCC 4.6 compiler will define __FP_FAST_FMA{,F,L} if the fma{,f,l} + builtins are supported. */ +# if __FP_FAST_FMA +# define FP_FAST_FMA 1 +# endif + +# if __FP_FAST_FMAF +# define FP_FAST_FMAF 1 +# endif + +# if __FP_FAST_FMAL +# define FP_FAST_FMAL 1 +# endif + #endif /* ISO C99 */ #ifndef __NO_LONG_DOUBLE_MATH diff --git a/ports/sysdeps/unix/sysv/linux/mips/mips32/syscalls.list b/ports/sysdeps/unix/sysv/linux/mips/mips32/syscalls.list index dbeb184d7..f357b5c91 100644 --- a/ports/sysdeps/unix/sysv/linux/mips/mips32/syscalls.list +++ b/ports/sysdeps/unix/sysv/linux/mips/mips32/syscalls.list @@ -1,3 +1,5 @@ # File name Caller Syscall name # args Strong name Weak names prlimit64 EXTRA prlimit64 i:iipp prlimit64 + +fanotify_mark EXTRA fanotify_mark i:iiiiis fanotify_mark diff --git a/ports/sysdeps/unix/sysv/linux/mips/mips64/n32/syscalls.list b/ports/sysdeps/unix/sysv/linux/mips/mips64/n32/syscalls.list index 7a6a77779..87c194990 100644 --- a/ports/sysdeps/unix/sysv/linux/mips/mips64/n32/syscalls.list +++ b/ports/sysdeps/unix/sysv/linux/mips/mips64/n32/syscalls.list @@ -6,3 +6,5 @@ ftruncate - ftruncate i:ii __ftruncate ftruncate ftruncate64 __ftruncate64 truncate - truncate i:si truncate truncate64 prlimit64 EXTRA prlimit64 i:iipp prlimit64 + +fanotify_mark EXTRA fanotify_mark i:iiiis fanotify_mark diff --git a/ports/sysdeps/unix/sysv/linux/mips/mips64/n64/syscalls.list b/ports/sysdeps/unix/sysv/linux/mips/mips64/n64/syscalls.list index a7bccdd37..2c8c4d484 100644 --- a/ports/sysdeps/unix/sysv/linux/mips/mips64/n64/syscalls.list +++ b/ports/sysdeps/unix/sysv/linux/mips/mips64/n64/syscalls.list @@ -1,3 +1,5 @@ # File name Caller Syscall name # args Strong name Weak names prlimit EXTRA prlimit64 i:iipp prlimit prlimit64 + +fanotify_mark EXTRA fanotify_mark i:iiiis fanotify_mark |