Re: [Qemu-devel] [PATCH] Porting TCG to alpha platform

[Richard Henderson <rth@twiddle.net>]

On 01/28/2010 05:55 PM, identifier scorpio wrote:
> 1. adding constraints for operation mov_i32/movi_i32/mov_i64/movi_i64
> 2. replacing "goto do_arith" with "break" for INDEX_op_not_i32/i64 case
> handling in tcg_out_op

Oops.

> 3. removing the redundant declaration for "alpha_tb_set_jmp_target()" at
> tail of tcg-target.c, it has been declared in exec-all.h

Ah, yes, a line I had to add temporarily for the x86 compiler; I meant 
to take it out again.

> I just find that a suspect bug in tcg_out_brcond() and have not tried yet.
>
> if ( opc == 0) {
> opc = cmp_opc[cond];
> if ( const_arg2) { /* should be if (const_arg2 && arg2 & ~0xff == 0) ?
> tcg_out_fmt_opi(s, opc, ...)

     { INDEX_op_brcond_i32,      { "rJ", "rI" } },
     { INDEX_op_brcond_i64,      { "rJ", "rI" } },

The "I" constraint is supposed to have already limited the constant to 0 
.. 0xff.  You could put an assert there if you like, to make sure 
everything is going as it should.

> Your code is more skilled and I learn a lot from it, but I haven't yet
> recognized the key difference between your version and mine. In last
> mail, you've said that "Mainly, tcg_target_long was used places that
> could be holding HOST values, which means that the host values would
> have been truncated to 32 bits.", and I can't fully understand it, may
> you put it more clearly.

It was a mis-statement made while tired.  Of course tcg_target_long is 
defined to int64_t when TCG_TARGET_REG_BITS == 64, so there was no 
truncation to 32 bits.

Have you tested of any of the other embedded targets?  You may be able 
to find the error with one of those easier than finding it emulating 
full-scale Windows.  See the images at the bottom of the QEMU download page.


r~