On 9/25/20 1:51 AM, Frank Chang wrote:
> trans_vsetvli() uses gen_goto_tb() to save the computation in the link to the
> next TB.
> I know there was a discussion about this back in RVV v0.7.1:
> https://patchew.org/QEMU/20200103033347.20909-1-zhiwei_liu@c-sky.com/20200103033347.20909-5-zhiwei_liu@c-sky.com/
>
> However, we had encountered an issue that looked like it was caused by the
> linked TB.
> The code snippet which cause the issue is:
>
> 00000000000104a8 <loop>: 104a8: 0122ffd7 vsetvli t6,t0,e32,m4,tu,mu,d1 104ac:
> 02036407 vle32.v v8,(t1) 104b0: 028a0a57 vadd.vv v20,v8,v20 104b4: 41f282b3 sub
> t0,t0,t6 104b8: 002f9893 slli a7,t6,0x2 104bc: 9346 add t1,t1,a7 104be:
> fe0295e3 bnez t0,104a8 <loop> 104c2: 012f7057 vsetvli zero,t5,e32,m4,tu,mu,d1
> .....
>
> If $t0 is given with the value, e.g. 68.
> <loop> is expected to process 32 elements in each iteration.
> That's it, the env->vl after vsetvli at 0x104a8 in each iteration would be:
> 1st iteration: 32 (remaining elements to be processed: 68 - 32 = 36)
> 2nd iteration: 32 (remaining elements to be processed: 36 - 32 = 4)
> 3rd iteration: 4 (remaining elements to be processed: 4 - 4 = 0, will leave
> <loop> after 0x104be)
>
> vadd.vv at 0x104b0 is implemented with gvec for acceleration:
>
> if (a->vm && s->vl_eq_vlmax) {
> gvec_fn(s->sew, vreg_ofs(s, a->rd),
> vreg_ofs(s, a->rs2), vreg_ofs(s, a->rs1),
> MAXSZ(s), MAXSZ(s));
> } else {
> uint32_t data = "">
>
> data = "" VDATA, VM, a->vm);
> data = "" VDATA, LMUL, s->lmul);
> tcg_gen_gvec_4_ptr(vreg_ofs(s, a->rd), vreg_ofs(s, 0),
> vreg_ofs(s, a->rs1), vreg_ofs(s, a->rs2),
> cpu_env, 0, s->vlen / 8, data, fn);
> }
>
> gvec function is used when a->vm and s->vl_eq_vlmax are both true.
> However, s->vl_eq_vlmax, for the above case, is only true in 1st and 2nd
> iterations.
> In third iteration, env->vl is 4 which is not equal to vlmax = 32.
> But as the TB where vadd.vv resides are already linked with vsetvli's TB,
> it won't be retranslated and still use the same gvec function in the third
> iteration.
> The total elemented being proceeded would be: 32 + 32 + 32 = 96, instead of 68.
>
> I'm wondering under such conditions, is it still correct to use gen_goto_tb() here?
> Or we should use lookup_and_goto_ptr() as in trans_vsetvl() to not link the TBs.
You're correct -- because of vl_eq_vlmax we can't use goto_tb when using a
variable input.
It would be possible when using xN,x0 for VLMAX, or x0,x0 for reuse of the
current vl, but I doubt it's worth special-casing that.
I wonder if the goto_tb conversation happened before we introduced vl_eq_vlmax
and forgot to re-evaluate, or if I just missed that in the first place.
Anyway, thanks for finding this.
r~
Thanks Richard, I'll include the fix in my next version patchset.
Frank Chang