Re: [Qemu-devel] SH: Implement MOVCO.L and MOVLI.L

[Vladimir Prus]

On Friday 13 February 2009 23:11:05 Aurelien Jarno wrote:
> On Thu, Feb 12, 2009 at 12:05:15AM +0300, Vladimir Prus wrote:
> > On Saturday 07 February 2009 17:24:59 Aurelien Jarno wrote:
> > > On Thu, Dec 11, 2008 at 09:48:00PM +0300, Vladimir Prus wrote:
> > > > On Thursday 11 December 2008 21:36:13 Paul Brook wrote:
> > > > > On Thursday 11 December 2008, Vladimir Prus wrote:
> > > > > > This patch implements a couple more instructions
> > > > > > specific to SH4A. 
> > > > > 
> > > > > > +          For now, do unconditional move. FIXME.
> > > > > 
> > > > > I can't make sense of this FIXME.
> > > > 
> > > > At it happens, it's stale. The entire line you quoted can be removed.
> > > > I attach the revised patch.
> > > > 
> > > > - Volodya
> > > > 
> > > > 
> > > 
> > > > commit 9ba7a6294a809dc11a0348ccca501513dd90ae6e
> > > > Author: Vladimir Prus <address@hidden>
> > > > Date:   Sun Oct 12 19:36:22 2008 +0400
> > > > 
> > > >     SH: Implement MOVCO.L and MOVLI.L
> > > >     
> > > >         * target-sh4/cpu.h (struct CPUSH4State): New field ldst.
> > > >         * target-sh4/helper.c (do_interrupt): Initialize ldst.
> > > >         * target-sh4/translate.c (cpu_ldst): New.
> > > >         (sh4_translate_init): Initialize cpu_ldst.
> > > >         (_decode_opc): Support MOVCO.L and MOVLI.L.
> > > > 
> > > > diff --git a/target-sh4/cpu.h b/target-sh4/cpu.h
> > > > index a0108b4..ae434d1 100644
> > > > --- a/target-sh4/cpu.h
> > > > +++ b/target-sh4/cpu.h
> > > > @@ -135,6 +135,8 @@ typedef struct CPUSH4State {
> > > >      uint32_t prr;              /* Processor Revision Register */
> > > >      uint32_t cvr;              /* Cache Version Register */
> > > >  
> > > > +    uint32_t ldst;
> > > > +
> > > >       CPU_COMMON tlb_t utlb[UTLB_SIZE]; /* unified translation table */
> > > >      tlb_t itlb[ITLB_SIZE];     /* instruction translation table */
> > > >      void *intc_handle;
> > > > diff --git a/target-sh4/helper.c b/target-sh4/helper.c
> > > > index 882bc9c..61b668b 100644
> > > > --- a/target-sh4/helper.c
> > > > +++ b/target-sh4/helper.c
> > > > @@ -160,6 +160,7 @@ void do_interrupt(CPUState * env)
> > > >      env->spc = env->pc;
> > > >      env->sgr = env->gregs[15];
> > > >      env->sr |= SR_BL | SR_MD | SR_RB;
> > > > +    env->ldst = 0;
> > > >  
> > > >      if (env->flags & (DELAY_SLOT | DELAY_SLOT_CONDITIONAL)) {
> > > >          /* Branch instruction should be executed again before delay 
> > > > slot. */
> > > > diff --git a/target-sh4/translate.c b/target-sh4/translate.c
> > > > index d38e95f..e4a1b28 100644
> > > > --- a/target-sh4/translate.c
> > > > +++ b/target-sh4/translate.c
> > > > @@ -72,7 +72,7 @@ static TCGv_ptr cpu_env;
> > > >  static TCGv cpu_gregs[24];
> > > >  static TCGv cpu_pc, cpu_sr, cpu_ssr, cpu_spc, cpu_gbr;
> > > >  static TCGv cpu_vbr, cpu_sgr, cpu_dbr, cpu_mach, cpu_macl;
> > > > -static TCGv cpu_pr, cpu_fpscr, cpu_fpul;
> > > > +static TCGv cpu_pr, cpu_fpscr, cpu_fpul, cpu_ldst;
> > > >  static TCGv cpu_fregs[32];
> > > >  
> > > >  /* internal register indexes */
> > > > @@ -144,6 +144,8 @@ static void sh4_translate_init(void)
> > > >      cpu_delayed_pc = tcg_global_mem_new_i32(TCG_AREG0,
> > > >                                             offsetof(CPUState, 
> > > > delayed_pc),
> > > >                                             "_delayed_pc_");
> > > > +    cpu_ldst = tcg_global_mem_new_i32(TCG_AREG0,
> > > > +                                     offsetof(CPUState, ldst), 
> > > > "_ldst_");
> > > >  
> > > >      for (i = 0; i < 32; i++)
> > > >          cpu_fregs[i] = tcg_global_mem_new_i32(TCG_AREG0,
> > > > @@ -1543,6 +1545,36 @@ static void _decode_opc(DisasContext * ctx)
> > > >      case 0x0029:               /* movt Rn */
> > > >         tcg_gen_andi_i32(REG(B11_8), cpu_sr, SR_T);
> > > >         return;
> > > > +    case 0x0073:
> > > > +        /* MOVCO.L 
> > > > +              LDST -> T
> > > > +               If (T == 1) R0 -> (Rn)
> > > > +               0 -> LDST
> > > > +        */        
> > > > +        if (ctx->features & SH_FEATURE_SH4A) {
> > > > +           int label = gen_new_label();
> > > > +           gen_clr_t();
> > > > +           tcg_gen_or_i32(cpu_sr, cpu_sr, cpu_ldst);
> > > > +           tcg_gen_brcondi_i32(TCG_COND_EQ, cpu_ldst, 0, label);
> > > > +           tcg_gen_qemu_st32(REG(0), REG(B11_8), ctx->memidx);
> > > > +           gen_set_label(label);
> > > > +           tcg_gen_movi_i32(cpu_ldst, 0);
> > > > +           return;
> > > > +       } else
> > > > +           break;
> > > > +    case 0x0063:
> > > > +        /* MOVLI.L @Rm,R0 
> > > > +               1 -> LDST
> > > > +               (Rm) -> R0
> > > > +               When interrupt/exception
> > > > +               occurred 0 -> LDST
> > > > +        */     
> > > > +       if (ctx->features & SH_FEATURE_SH4A) {
> > > > +           tcg_gen_movi_i32(cpu_ldst, 1);
> > > > +           tcg_gen_qemu_ld32s(REG(0), REG(B11_8), ctx->memidx);
> > > > +           return;
> > > 
> > > This TCG code doesn't seems to match the comments. With the current
> > > code, if a load exception occurs, LDST end with a value of 1 instead of
> > > 0.
> > 
> > I might be missing something -- it was my intention that if any exception
> > or interrupt happens, do_interrupt will set ldst to zero. And if the 
> > exception
> > happens during load, there's nothing else that will set ldst back to 1. Did
> > I miss something?
> > 
> 
> In system mode, I am not sure that doing that in do_interrupt() is the
> best way, especially given it looks like we can exit this subroutine
> before actually setting the value to 0, though I haven't really analyzed
> those cases. In user mode, it's surely wrong, given that do_interrupt()
> does nothing.
> 
> In my opinion you should use the fact that the TCG code will stop it's
> execution at the point were the exception occurs. Doing something like
> this should work:
> 
> tcg_gen_movi_i32(cpu_ldst, 0);
> tcg_gen_qemu_ld32s(REG(0), REG(B11_8), ctx->memidx);
> tcg_gen_movi_i32(cpu_ldst, 1);
> 
> This way if everything works fine, cpu_ldst ends with 1. If an exception
> occurs, it stays to zero.

Here's a patch implementing this idea -- it appears to still work
fine with 7785.

- Volodya

movcol.diff
Description: Text Data