Commit Graph

6 Commits

Author SHA1 Message Date
Andrew Hsieh
3dbef8fe11 Bump up clang vesion accepting __builtin_isnan __attribute__((pcs("aapcs")))
Upstream clang3.6 still doesn't allow change of calling convension of
__builtin_isnan from the default to "softfp" (via __attribute__((pcs("aapcs"))))
as required by bionic libm.so for 32-bit ARM.  Let's hope clang3.7
fix that.  See upstream bug http://llvm.org/bugs/show_bug.cgi?id=20958

Change-Id: Ibdcc1783ffa7cf2f36e733983bd6d912ad38df70
2014-09-26 16:20:56 +08:00
Andrew Hsieh
b280e72a97 Bump __clang_minor__ check to 6 for __builtin_isnan
It appears that upstream 3.5 still can't accept __builtin_isnan redeclared
with calling convention, eg.

cat > isnan.c <<EOF
int __builtin_isnan(double) __attribute__((pcs("aapcs")));
EOF

  clang  -target armv5te-none-linux-androideabi -c isnan.c
  isnan.c:1:5: error: function declared 'aapcs' here was previously declared without calling convention
  int __builtin_isnan(double) __attribute__((pcs("aapcs")));
      ^
  isnan.c:1:5: note: previous declaration is here

Change-Id: Iff466fd913c3deec27aa3ff355a208543028d368
2014-04-25 09:55:31 -07:00
Logan Chien
ad40d952a1 Workaround the __builtin_isnan() type error for clang 3.4.
Similar to clang 3.3, it seems that clang 3.4 declares an incompatible
function prototypes for __builtin_isnan().  To workaround this issue,
we should not declare the function prototype for __builtin_isnan().

Note: Though it may be good to fix this in clang type checker, like
__builtin_nexttoward(), however it seems non-trivial since __builtin_isnsn()
might be overloaded.  We can revert this after we have fixed the
clang builtin definition.

Change-Id: I329b56225f3d229b0c141bb8414c4369f154efda
2013-12-25 01:59:29 +00:00
Andrew Hsieh
2a571fda2b Fix -mhard-float for calling __builtin_* math functions
Annotate __builtin math functions with __NDK_FPABI_MATH__ too.
1. __attribute__((pcs("aapcs"))) has no effect on codegen when __buitlin
   function is emitted inlined.
2. When __builtin call into regular function in Android's libm, it has
   proper attribute which observes soft float-abi

Change-Id: I11e8b74fc22f3072b9ce849193e4a56e98c92d64
2013-12-04 11:45:04 +08:00
Andrew Hsieh
5881b5cbe7 Add __NDK_FPABI__ to function taking/returning float/double
Allow user code to be compiled with "-mhard-float" for armeabi-v7a
(which implies -mfloat-abi=hard), and either link with
1. Android native APIs which use softfp, or
2. A customized libm_hard.a which is compiled with -mhard-float. Need
   -D_NDK_MATH_NO_SOFTFP=1

See tests/device/hard-float/jni/Android.mk for details and restriction
on Clang

Change-Id: I773a842c73368e08b9a6cda0441e95a96fa303b2
2013-10-17 19:15:22 +08:00
Andrew Hsieh
0df3e70b99 Add missing math functions
1. Add sincos, sincosf, and sincosl to API >= 9 (for ARM only,
   because X86 and MIPS have it already)
2. Add nan, nanf, nanl, and tgammaf to API >= 13 (11 actually,
   but we don't want to add another API which is not released)
3. Add log2, log2f, log2l, logbl, nexttoward, and nexttowardl to
   API >= 18

See b.android.com/38423

Change-Id: I0a756aeeddafc3862f3c0ae38218cb3301608231
2013-10-15 15:55:12 +08:00