[PATCH PR95696] regrename creates overlapping register allocations for vliw

[Richard Sandiford]

Hi,

Zhongyunde <zhongyunde@huawei.com> writes:
> Hi，
>
> In most target, it is limited to issue two insns with change the same register. So a register is not realy unused if there is another insn, which set the register in the save VLIW.
>
> For example, The insn 73 start with insn:TI, so it will be issued together with others insns until a new insn start with insn:TI, here is insn 243.
>
> The regrename pass known the mode V2VF in insn 73 need two successive registers, i.e. v2 and v3, here is dump snippet before the regrename:
>
>
>
> ==========================================================================================
>
> (insn:TI 73 76 71 4 (set (reg/v:V2VF 37 v2 [orig:180 _62 ] [180])
>
>         (unspec:V2VF [
>
>                 (reg/v:VHF 43 v8 [orig:210 Dest_value ] [210])
>
>                 (reg/v:VHF 43 v8 [orig:210 Dest_value ] [210])
>
>             ] UNSPEC_HFSQMAG_32X32)) "../test_modify.c":57 710 {hfsqmag_v2vf}
>
>      (expr_list:REG_DEAD (reg/v:VHF 43 v8 [orig:210 Dest_value ] [210])
>
>         (expr_list:REG_UNUSED (reg:VHF 38 v3)
>
>             (expr_list:REG_STAGE (const_int 2 [0x2])
>
>                 (expr_list:REG_CYCLE (const_int 2 [0x2])
>
>                     (expr_list:REG_UNITS (const_int 256 [0x100])
>
>                         (nil)))))))
>
>
>
> (insn 71 73 243 4 (set (reg:VHF 43 v8 [orig:265 MEM[(const vfloat32x16 *)Src_base_134] ] [265])
>
>         (mem:VHF (reg/v/f:DI 13 a13 [orig:207 Src_base ] [207]) [1 MEM[(const vfloat32x16 *)Src_base_134]+0 S64 A512])) "../test_modify.c":56 450 {movvhf_internal}
>
>      (expr_list:REG_STAGE (const_int 1 [0x1])
>
>         (expr_list:REG_CYCLE (const_int 2 [0x2])
>
>             (nil))))
>
>
>
> (insn:TI 243 …
>
>
>
> Then, in the regrename, the insn 71 will be transformed as following code with register v3, and there is an conflict between insn 73 and insn 71 （as both of them set the v3 register）.
>
>
>
> Register v2 (2): 73 [SVEC_REGS]
>
> Register v8 (1): 71 [VEC_ALL_REGS]
>
>
>
> ....
>
>
>
> (insn 71 73 243 4 (set (reg:VHF 38 v3 [orig:265 MEM[(const vfloat32x16 *)Src_base_134] ] [265])
>
>         (mem:VHF (reg/v/f:DI 13 a13 [orig:207 Src_base ] [207]) [1 MEM[(const vfloat32x16 *)Src_base_134]+0 S64 A512])) "../test_modify.c":56 450 {movvhf_internal}
>
>      (expr_list:REG_STAGE (const_int 1 [0x1])
>
>         (expr_list:REG_CYCLE (const_int 2 [0x2])

Do you need this for correctness, or is it “just” an optimisation?

The reason for asking is that regrename runs quite a long time after
the first scheduling pass, with things like register allocation
happening in between.  The :TI markers on the instructions are
therefore going to be very stale by the time that regrename runs.

If it's just a heuristic then it might be OK to use them anyway,
but it would be worth having a comment to say what's going on.

One alternative would be to run the DFA over the instructions
to recompute the VLIW bundles, but that isn't simple.  It also isn't
necessarily a better heuristic, since the second scheduling pass is
likely to change things anyway.

There's also the problem that :TI markers are used on non-VLIW
targets too, whereas the problem is really specific to VLIW targets.

Unfortunately, I can't think of any good counter-suggestions
at the moment…

Thanks,
Richard

>
> diff --git a/gcc/regrename.c b/gcc/regrename.c
> index c38173a77..e54794413 100644
> --- a/gcc/regrename.c
> +++ b/gcc/regrename.c
> @@ -1614,12 +1614,26 @@ record_out_operands (rtx_insn *insn, bool earlyclobber, insn_rr_info *insn_info)
>    cur_operand = NULL;
>  }
>  
> +/* Get the first real insn of next vliw in current BB.  */
> +static rtx_insn *
> +get_next_vliw_first_insn (rtx_insn *cur_insn, basic_block bb)
> +{
> +  rtx_insn *insn = next_real_insn (cur_insn);
> +
> +  for (; insn && insn != BB_END (bb); insn = next_real_insn (insn))
> +    if (GET_MODE (insn) == TImode)
> +      return insn;
> +
> +  return cur_insn;
> +}
> +
>  /* Build def/use chain.  */
>  
>  static bool
>  build_def_use (basic_block bb)
>  {
>    rtx_insn *insn;
> +  rtx_insn *vliw_start_insn = NULL;
>    unsigned HOST_WIDE_INT untracked_operands;
>  
>    fail_current_block = false;
> @@ -1663,6 +1677,9 @@ build_def_use (basic_block bb)
>  	     to be marked unrenamable or even cause us to abort the entire
>  	     basic block.  */
>  
> +	  if (GET_MODE (insn) == TImode)
> +	    vliw_start_insn = insn;
> +
>  	  extract_constrain_insn (insn);
>  	  preprocess_constraints (insn);
>  	  const operand_alternative *op_alt = which_op_alt ();
> @@ -1858,17 +1875,26 @@ build_def_use (basic_block bb)
>  	      scan_rtx (insn, &XEXP (note, 0), ALL_REGS, mark_access,
>  			OP_INOUT);
>  
> -	  /* Step 7: Close chains for registers that were never
> -	     really used here.  */
> -	  for (note = REG_NOTES (insn); note; note = XEXP (note, 1))
> -	    if (REG_NOTE_KIND (note) == REG_UNUSED)
> -	      {
> -		remove_from_hard_reg_set (&live_hard_regs,
> -					  GET_MODE (XEXP (note, 0)),
> -					  REGNO (XEXP (note, 0)));
> -		scan_rtx (insn, &XEXP (note, 0), NO_REGS, terminate_dead,
> -			  OP_IN);
> -	      }
> +	  /* Step 7: Close chains for registers that were never
> +	     really used delayed at the end of vliw.  */
> +	  if (vliw_start_insn
> +	      && next_real_insn (insn) == get_next_vliw_first_insn (insn, bb))
> +	    {
> +	      rtx_insn *member;
> +
> +	      for (member = vliw_start_insn;
> +	           member != next_active_insn (insn);
> +	           member = next_active_insn (member))
> +	        for (note = REG_NOTES (member); note; note = XEXP (note, 1))
> +	          if (REG_NOTE_KIND (note) == REG_UNUSED)
> +	            {
> +	              remove_from_hard_reg_set (&live_hard_regs,
> +	                                        GET_MODE (XEXP (note, 0)),
> +	                                        REGNO (XEXP (note, 0)));
> +	              scan_rtx (member, &XEXP (note, 0), NO_REGS, terminate_dead,
> +	                        OP_IN);
> +	            }
> +	    }
>  
>  	  /* Step 8: Kill the chains involving register restores.  Those
>  	     should restore _that_ register.  */