Re: [PATCH] target/openrisc: Do not reset delay slot flag on early tb exit

[Stafford Horne]

On Wed, May 11, 2022 at 07:32:58AM -0700, Richard Henderson wrote:
> On 5/11/22 05:05, Stafford Horne wrote:
> > +static void openrisc_cpu_synchronize_from_tb(CPUState *cs,
> > +                                             const TranslationBlock *tb)
> > +{
> > +    OpenRISCCPU *cpu = OPENRISC_CPU(cs);
> > +
> > +    cpu->env.pc = tb->pc;
> > +}
> 
> If mips is a guide, you'd want to set dflag based on
> 
>   tb->flags & TB_FLAGS_DFLAG
> 
> as well.   But I think openrisc is more careful to keep dflag up-to-date.

I was thinking that too so I left it out.

For example:

    0xc01e3ffc:  l.bf      3
    0xc01e4000:   l.ori     r12, r0, 1

    ---

     ---- c01e3ffc 00000000                
     movcond_i32 jmp_pc,sr_f,$0x0,$0xc01e4008,$0xc01e4004,ne  sync: 0  dead: 0 
1 2 3 4  pref=0xffff
     mov_i32 dflag,$0x1                       sync: 0  dead: 0 1  pref=0xffff
     mov_i32 ppc,$0xc01e3ffc                  sync: 0  dead: 0 1  pref=0xffff
     mov_i32 pc,$0xc01e4000                   sync: 0  dead: 0 1  pref=0xffff
     call lookup_tb_ptr,$0x6,$1,tmp7,env      dead: 1  pref=none
     goto_ptr tmp7                            dead: 0
     set_label $L0                          
     exit_tb $0x7f7b047f3b43                

    ---

     ld_i32 tmp0,env,$0xfffffffffffffff0      pref=0xffff
     brcond_i32 tmp0,$0x0,lt,$L0              dead: 0

     ---- c01e4000 00000001                
     mov_i32 r12,$0x1                         sync: 0  dead: 0 1  pref=0xffff
     mov_i32 dflag,$0x0                       sync: 0  dead: 0 1  pref=0xffff
     mov_i32 ppc,$0xc01e4000                  sync: 0  dead: 0 1  pref=0xffff
     mov_i32 pc,jmp_pc                        sync: 0  dead: 0 1  pref=0xffff
     discard jmp_pc                           pref=none
     call lookup_tb_ptr,$0x6,$1,tmp4,env      dead: 1  pref=none
     goto_ptr tmp4                            dead: 0
     set_label $L0                          
     exit_tb $0x7f7b047f3c83    


This is an example of a branch followed by a branch delay slot.  If we exit the
branch delay slot via `exit_tb $0x7f7b047f3c83`.  The `mov_i32 dflag,$0x1` 
instruction
would have run from `c01e3ffc` having env already updated.

At this point how would tb->flags have the right value?  Would it always be set
correctly by `cpu_get_tb_cpu_state` in the `lookup_tb_ptr`call?

-Stafford