Re: [Qemu-devel] Re: Compile files only once: some planning

[Richard Henderson]

On 03/24/2010 07:45 AM, Paolo Bonzini wrote:
> On 03/24/2010 12:19 PM, Richard Henderson wrote:
>> On 03/24/2010 02:47 AM, Paolo Bonzini wrote:
>>> 1) make CPUState define only common fields. Include CPUState at the
>>> beginning of each per-target CPUXYZState.
>>
>> Irritatingly, the common fields contain quite big TLBs. And the
>> offsets from the start of env affect the compactness of the code
>> generated from TCG. We really really want the general registers
>> to come first to make sure that those offsets fit the host's
>> reg+offset addressing mode.
> 
> What about adding a 512-bytes (or more) block or something like that at
> the beginning of CPUState with a union, so you can put the per-target
> stuff there?

I think that would be confusing.

What might be just as good (although possibly just as confusing)
is to move the big members into a different structure.  E.g.

struct CPUSmallCommonState
{
    // most of the stuff from CPU_COMMON.
    // sorted for some thought of padding elimination.  ;-)
};

struct CPULargeCommonState
{
    CPUTLBEntry tlb_table[NB_MMU_MODES][CPU_TLB_SIZE];
    target_phys_addr_t iotlb[NB_MMU_MODES][CPU_TLB_SIZE];
    struct TranslationBlock *tb_jmp_cache[TB_JMP_CACHE_SIZE];
    jmp_buf jmp_env;
};

struct CPUXYZSmallState
{
    CPUSmallCommonState common_s;
    // the rest of the cpu-specific stuff.
};

struct CPUXYZLargeState
{
    CPUXYZSmallState s;
    CPUBigCommonState common_l;
};

extern int cpu_large_state_offset = offsetof(CPUXYZLargeState, common_l);

Now.  If you're compiling a file for which cpu-specific code is ok:

register CPUXYZLargeState *env __asm__(AREG0);
#define ENV_SMALL_COMMON_STATE    (&env->s.common_s)
#define ENV_LARGE_COMMON_STATE    (&env->common_l)

If you're compiling a file which is supposed to be independant of cpu:

register CPUSmallCommonState *env __asm__(AREG0);
#define ENV_SMALL_COMMON_STATE    (env)
#define ENV_LARGE_COMMON_STATE    ((CPULargeCommonState *)((char *)env + 
cpu_large_state_offset))

For the gcc-compiled code, the addition of the cpu_large_state_offset
is probably more or less on par in efficiency with indirection.  But
for TCG generated code, the variable read happens at code generation
time, which means we *still* have a constant in the generated code.


r~