hardfp calling support

This is commented out for android, but it is tested and works!

core/core.mk

@@ -52,7 +52,7 @@ else
 RZDCY_CXXFLAGS	:= \
 	$(CFLAGS) -c -g -O3 -I$(RZDCY_SRC_DIR) -I$(RZDCY_SRC_DIR)/deps \
 	-D_ANDROID -DRELEASE -DTARGET_BEAGLE\
-	-march=armv7-a -mtune=cortex-a9 -mfpu=vfpv3-d16 -mfloat-abi=softfp \
+	-march=armv7-a -mtune=cortex-a9 -mfpu=vfpv3-d16 \
 	-frename-registers -fsingle-precision-constant -ffast-math \
 	-ftree-vectorize -fomit-frame-pointer -fno-exceptions -fno-rtti -std=gnu++11
 endif

core/rec-ARM/arm_dyna.cpp

@@ -565,6 +565,22 @@ void ngen_CC_Param(shil_opcode* op,shil_param* par,CanonicalParamType tp)
 	switch(tp)
 	{
 		case CPT_f32rv:
+		#ifdef ARM_HARDFP
+			{
+				if (reg.IsAllocg(*par))
+				{
+					//printf("MOV(reg.map(*par),r0); %d\n",reg.map(*par));
+					VMOV(reg.mapg(*par),f0);
+				}
+				else if (reg.IsAllocf(*par))
+				{
+					//VMOV(reg.mapf(*par),0,r0); %d\n",reg.map(*par));
+					VMOV(reg.mapfs(*par),f0);
+				}
+			}
+			break;
+		#endif
+
 		case CPT_u32rv:
 		case CPT_u64rvL:
 			{
@@ -606,6 +622,8 @@ void ngen_CC_Param(shil_opcode* op,shil_param* par,CanonicalParamType tp)
 void ngen_CC_Call(shil_opcode* op,void* function) 
 {
 	u32 rd=r0;
+	u32 fd=f0;
+
 	for (int i=CC_pars.size();i-->0;)
 	{
 		if (CC_pars[i].type==CPT_ptr)
@@ -616,6 +634,25 @@ void ngen_CC_Call(shil_opcode* op,void* function)
 		{
 			if (CC_pars[i].par->is_reg())
 			{
+				#ifdef ARM_HARDFP
+				if (CC_pars[i].type == CPT_f32) 
+				{
+					if (reg.IsAllocg(*CC_pars[i].par))
+					{
+						//printf("MOV((eReg)rd,reg.map(*CC_pars[i].par)); %d %d\n",rd,reg.map(*CC_pars[i].par));
+						VMOV((eFSReg)fd,reg.mapg(*CC_pars[i].par));
+					}
+					else if (reg.IsAllocf(*CC_pars[i].par))
+					{
+						//printf("LoadSh4Reg_mem((eReg)rd, *CC_pars[i].par); %d\n",rd);
+						VMOV((eFSReg)fd,reg.mapfs(*CC_pars[i].par));
+					}
+					else
+						die("Must not happen!\n");
+					continue;
+				}
+				#endif
+
 				if (reg.IsAllocg(*CC_pars[i].par))
 				{
 					//printf("MOV((eReg)rd,reg.map(*CC_pars[i].par)); %d %d\n",rd,reg.map(*CC_pars[i].par));
@@ -631,6 +668,7 @@ void ngen_CC_Call(shil_opcode* op,void* function)
 			}
 			else
 			{
+				verify(CC_pars[i].type != CPT_f32);
 				//printf("MOV32((eReg)rd, CC_pars[i].par->_imm); %d\n",rd);
 				MOV32((eReg)rd, CC_pars[i].par->_imm);
 			}

shell/android/jni/Android.mk

@@ -32,10 +32,28 @@ LOCAL_MODULE	:= dc
 LOCAL_CFLAGS	:= $(LOCAL_CXXFLAGS) -DHAS_VMU
 LOCAL_DISABLE_FORMAT_STRING_CHECKS=true
 LOCAL_ASFLAGS := -fvisibility=hidden
-LOCAL_LDLIBS	:= -llog -lGLESv2 -lEGL -lz 
+LOCAL_LDLIBS	:= -llog -lGLESv2 -lEGL -lz
 #-Wl,-Map,./res/raw/syms.mp3
 LOCAL_ARM_MODE	:= arm
 
+#
+# android has poor support for hardfp calling.
+# r9b+ is required, and it only works for internal calls
+# the opengl drivers would really benefit from this, but they are still using softfp
+# the header files tell gcc to automatically use aapcs for calling system/etc
+# so there is no real perfomance difference
+#
+# The way this is implemented is a huge hack on the android/linux side
+# (but then again, which part of android isn't a huge hack?)
+
+#ifneq ($(filter %armeabi-v7a,$(TARGET_ARCH_ABI)),)
+#LOCAL_CFLAGS += -mhard-float -D_NDK_MATH_NO_SOFTFP=1 -DARM_HARDFP
+#LOCAL_LDLIBS += -lm_hard
+#ifeq (,$(filter -fuse-ld=mcld,$(APP_LDFLAGS) $(LOCAL_LDFLAGS)))
+#LOCAL_LDFLAGS += -Wl,--no-warn-mismatch
+#endif
+#endif
+
 include $(BUILD_SHARED_LIBRARY)
 
 $(call import-module,android/native_app_glue)