On 03/27/12 02:24, Jia Liu wrote:
> +ifeq ($(TARGET_BASE_ARCH), mips)
> +libobj-y += dsp_helper.o
> +endif
...
> +#include "dyngen-exec.h"
> +
> +/*** MIPS DSP internal functions begin ***/
> +static inline void set_DSPControl_overflow_flag(uint32_t flag, int position)
> +{
> + env->active_tc.DSPControl |= (target_ulong)flag << position;
> +}
In order to use dyngen-exec.h, and the global ENV variable,
you must also do
dsp_helper.o: QEMU_CFLAGS += $(HELPER_CFLAGS)
in the Makefile.target. If you had tested i386 host instead
of x86_64, you should have seen this failure.
Better, from the point of view of one working toward eliminating
dyngen-exec.h and these special calling conventions, is to actually
pass down env in the helper. E.g.
+/** DSP Arithmetic Sub-class insns **/
+uint32_t helper_addq_ph(CPUMIPSState *env, uint32_t rs, uint32_t rt)
+DEF_HELPER_FLAGS_3(addq_ph, 0, i32, env, i32, i32)
where mipsdsp_add_i16 and thence set_DSPControl_overflow_flag also
grow ENV parameters.
In addition, we can't use TCG_CALL_CONST | TCG_CALL_PURE on any
function that sets DSPControl, because that's the cpu_dspctrl tcg
variable.
You might get better performance out of TCG if you do *not* make
DSPControl a TCG variable, but instead perform explicit loads from
that slot in the few places that need it. That way you can at
least set TCG_CALL_CONST. (Though not TCG_CALL_PURE, because the
functions do have side effects.)
> + if (len == 2)
> + env->active_tc.DSPControl &= 0xFCFFFFFF;
> + else if (len == 4)
> + env->active_tc.DSPControl &= 0xF0FFFFFF;
Run all your patches through ./scripts/checkpatch.pl and fix the
errors that will report.
r~
