94 lines
3.6 KiB
Makefile
94 lines
3.6 KiB
Makefile
LOCAL_PATH := $(call my-dir)
|
|
|
|
# This is the way used to be: For ARM, ndk-build forces "softfp"
|
|
# regardless armeabi (default to use -msoft-float) or armeabi-v7a
|
|
# (default to use -mfpu=vfpv3-d16)
|
|
include $(CLEAR_VARS)
|
|
LOCAL_MODULE := float-softfp-abi
|
|
LOCAL_SRC_FILES := test-float.c
|
|
include $(BUILD_EXECUTABLE)
|
|
|
|
|
|
ifeq ($(filter %armeabi-v7a,$(TARGET_ARCH_ABI)),)
|
|
|
|
#
|
|
# The following two examples are relevant only to armeabi-v7a
|
|
#
|
|
|
|
$(warning Examples hard-float-softfp-abi and hard-float-hard-abi not relevant to ABI $(TARGET_ARCH_ABI))
|
|
|
|
else
|
|
|
|
ifeq (,$(filter clang%,$(NDK_TOOLCHAIN_VERSION)))
|
|
|
|
# This is to compile all user code in -mhard-float (which implies
|
|
# -mfloat-abi=hard and overrides -mfloat-abi=softfp previously added
|
|
# by ndk-build), but still link system libm.so which use softfp.
|
|
# All functions in Android platform headers taking or returning float/double
|
|
# has __attribute__((pcs("aapcs"))), so even under the implied -mfloat-abi=hard
|
|
# softfp ABI is still observed by compiler for Android native APIs.
|
|
#
|
|
# You also need to ensure that
|
|
# 1. All code and library *must* be recompiled with consistent soft-abi.
|
|
# If your project use ndk-build and import other project, you may want to do
|
|
#
|
|
# $NDK/ndk-build -B APP_ABI=armeabi-v7a APP_CFLAGS=-mhard-float APP_LDFLAGS=-Wl,--no-warn-mismatch
|
|
#
|
|
# If your project use libraries compiled by others, it's likely those were built with
|
|
# -msoft-abi=softfp or evan -msoft-float (for armeabi). Recompile them with -mhard-float
|
|
#
|
|
# 2. Correct headers (eg. #include <math.h>) are always included, and there is no
|
|
# declaration like "extern double sin(double)" w/o proper __attribute__ in
|
|
# your code instead of including math.h, etc. See the end of sys/cdefs.h
|
|
# the conditions upon which __NDK_FPABI__ and __NDK_FPABI_MATH__ are defined
|
|
#
|
|
# 3. All JNI functions should have JNICALL which is defined to __NDK_FPABI__ in jni.h.
|
|
#
|
|
# 4. If you use undocumented APIs which takes/returns float/double, be careful
|
|
# to add __attribute__((pcs("aapcs"))) for arm
|
|
#
|
|
# Note that "--no-warn-mismatch" is needed for linker (except mclinker which check correctly)
|
|
# to suppress linker error about not all functions use VFP register to pass argument, eg.
|
|
#
|
|
# .../arm-linux-androideabi/bin/ld: error: ..../test-float.o
|
|
# uses VFP register arguments, output does not
|
|
#
|
|
include $(CLEAR_VARS)
|
|
LOCAL_MODULE := hard-float-softfp-abi
|
|
LOCAL_CFLAGS += -mhard-float
|
|
LOCAL_SRC_FILES := test-float.c
|
|
LOCAL_LDFLAGS := -Wl,--no-warn-mismatch
|
|
include $(BUILD_EXECUTABLE)
|
|
|
|
else
|
|
|
|
# Clang before 3.4 doesn't allow change of calling convenstion 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")));544
|
|
# ^
|
|
# a.i:564:6: note: previous declaration is here
|
|
#
|
|
# As a result, "-mhard-float" doesn't work properly for now (3.3 and 3.2), unless
|
|
# libm_hard is also used. See example below.
|
|
#
|
|
|
|
$(warning Skip example hard-float-softfp-abi for clang for now)
|
|
|
|
endif # check clang
|
|
|
|
# This is to compile all user code in -mhard-float and link a customized
|
|
# libm_hard.a which follows -mfloat-abi=hard. _NDK_MATH_NO_SOFTFP=1
|
|
# is to turn off __attribute__((pcs("aapcs"))) in math.h for all math
|
|
# functions to accept float-abi (implicit "hard" in this case) as specified by user.
|
|
#
|
|
include $(CLEAR_VARS)
|
|
LOCAL_MODULE := hard-float-hard-abi
|
|
LOCAL_CFLAGS += -mhard-float -D_NDK_MATH_NO_SOFTFP=1
|
|
LOCAL_LDLIBS += -lm_hard
|
|
LOCAL_SRC_FILES := test-float.c
|
|
LOCAL_LDFLAGS := -Wl,--no-warn-mismatch
|
|
include $(BUILD_EXECUTABLE)
|
|
|
|
endif # check armeabi-v7a
|