This consolidates the guts of the three endian.h files (endian.h,
sys/endian.h, and machine/endian.h) into just sys/endian.h, and turns
the others into headers that forward to it.
This also consolidates all the architecture specific machine/endian.h
files into a single generic header.
Finally, this is also the update version of sys/endian.h, which just
forwards to compiler builtins rather than implementing it ourselves.
Change-Id: Ifce64dc684ce3d3231f3f43a94c083bbd8661840
This is all macros, so there's no reason to have one per API level.
This is still quite a ways from matching what we have in bionic, but
that's a problem for another day.
Change-Id: I804edfd6299d9ba15a3ba59e1091f301eace2142
char16_t and char32_t should not be defined when building as C. For
GCC they are, but not for clang. Pull in the updated bionic header
which has the fix for this.
Change-Id: I41b292ece36af79bbe835706bee4441b298215ff
From the comment about __NDK_FPABI__ vs __NDK_FPABI_MATH__:
NOTE: Disable for clang for now unless _NDK_MATH_NO_SOFTFP=1, because clang before 3.4 doesn't
allow change of calling convension for builtin and produces error message reads:
a.i:564:6: error: function declared 'aapcs' here was previously declared without calling convention
int sin(double d) __attribute__((pcs("aapcs")));
^
a.i:564:6: note: previous declaration is here
This applies to current clang as well, not just pre-3.4. Mark these
math functions as math functions.
Change-Id: I001b31df540aa3f61eec049f89346cbf2f11f128
This relies on a change for the NDK build system to recognize and
accept the neon tag in these builds as well.
The intrinsics work just fine in these build configurations.
Change-Id: Ib652ab69e822b56edaa9387400bad326fab10133
GCC and Clang will not define __GNUC_STDC_INLINE__ when compiling C++, so this
macro currently implies regular C++ inline semantics in C++. This may cause
the compiler to emit an external definition of the function, which will cause
link errors when compiling with -fvisibility=hidden; see e.g. crbug.com/481902.
To fix this, also test the __cplusplus macro when deciding whether to assume
non-GNU inline semantics.
Change-Id: Icbd1f42279c2f65610e62f21d4a0a09d3b0e091c
On LP64, this could have caused the top bits of the pointer returned by
malloc to be lost because sizeof(int) != sizeof(void*).
Change-Id: I866a5071c1d359478d020d7fa4d466b53d52edab
Cherry pick bionic 86d2feef9f4df517dd5755b39b993bec0c3e1ea5 which
fixes struct stat64. This may look like breaking mips64 ABI, but mips64
support is stil work in progress post-L
Also see prebuilts/ndk e6639a4c0ba9296a19fd477154aace3585221d9d
Change-Id: Ia8859762191220b9d3127b37970f05594375ca40
This CL unconditionally removes the existance of __builtin_isinf and
__builtin_isnan from math.h for clang, because (unlike other math functions)
clang doesn't like those to be declared.
See LLVM bugzilla entry: http://llvm.org/bugs/show_bug.cgi?id=20958
The reason for decorating all math functions with __attribute__((pcs("aapcs")))
is to offer developers the ability to compile their code with hardfp and still
link Android's libm.so which follows softfp calling convention.
Before issue 20958 is fixed, those who want clang to compile their code with
hardfp in NDK has to go all the way hard, ie.
1) LOCAL_CFLAGS += -mhard-float -D_NDK_MATH_NO_SOFTFP=1
Note that _NDK_MATH_NO_SOFTFP=1 remove __attribute__((pcs("aapcs"))) for all math functions
2) LOCAL_LDLIBS += -lm_hard # link libm_hard.a recompiled with -mhard-float instead of bionic libm.so
Change-Id: Ie2f95e73e58e1574e8cadbcab92a5209d94448e1
