Re: Avoid recomputing data references in BB SLP

[Richard Biener <richard dot guenther at gmail dot com>]

On October 19, 2019 5:04:51 PM GMT+02:00, Richard Sandiford <richard.sandiford@arm.com> wrote:
>If the first attempt at applying BB SLP to a region fails, the main
>loop
>in vect_slp_bb recomputes the region's bounds and datarefs for the next
>vector size.  AFAICT this isn't needed any more; we should be able
>to reuse the datarefs from the first attempt instead.
>
>Tested on aarch64-linux-gnu and x86_64-linux-gnu.  OK to install?


OK. 

Richard. 
>Richard
>
>
>2019-10-19  Richard Sandiford  <richard.sandiford@arm.com>
>
>gcc/
>	* tree-vect-slp.c (vect_slp_analyze_bb_1): Call save_datarefs
>	when processing the given datarefs for the first time and
>	check_datarefs subsequently.
>	(vect_slp_bb_region): New function, split out of...
>	(vect_slp_bb): ...here.  Don't recompute the region bounds and
>	dataref sets when retrying with a different vector size.
>
>Index: gcc/tree-vect-slp.c
>===================================================================
>*** gcc/tree-vect-slp.c	2019-10-11 15:43:55.439484181 +0100
>--- gcc/tree-vect-slp.c	2019-10-19 16:03:30.000000000 +0100
>*************** vect_slp_analyze_bb_1 (gimple_stmt_itera
>*** 2852,2857 ****
>--- 2852,2858 ----
>    slp_instance instance;
>    int i;
>    poly_uint64 min_vf = 2;
>+   bool first_time_p = shared->datarefs.is_empty ();
>  
>    /* The first group of checks is independent of the vector size.  */
>    fatal = true;
>*************** vect_slp_analyze_bb_1 (gimple_stmt_itera
>*** 2871,2877 ****
>      return NULL;
>  
>    BB_VINFO_DATAREFS (bb_vinfo) = datarefs;
>!   bb_vinfo->shared->save_datarefs ();
>  
>    /* Analyze the data references.  */
>  
>--- 2872,2881 ----
>      return NULL;
>  
>    BB_VINFO_DATAREFS (bb_vinfo) = datarefs;
>!   if (first_time_p)
>!     bb_vinfo->shared->save_datarefs ();
>!   else
>!     bb_vinfo->shared->check_datarefs ();
>  
>    /* Analyze the data references.  */
>  
>*************** vect_slp_analyze_bb_1 (gimple_stmt_itera
>*** 3007,3022 ****
>    return bb_vinfo;
>  }
>  
>! 
>! /* Main entry for the BB vectorizer.  Analyze and transform BB,
>returns
>!    true if anything in the basic-block was vectorized.  */
>! 
>! bool
>! vect_slp_bb (basic_block bb)
>  {
>    bb_vec_info bb_vinfo;
>-   gimple_stmt_iterator gsi;
>-   bool any_vectorized = false;
>    auto_vector_sizes vector_sizes;
>  
>    /* Autodetect first vector size we try.  */
>--- 3011,3028 ----
>    return bb_vinfo;
>  }
>  
>! /* Subroutine of vect_slp_bb.  Try to vectorize the statements
>between
>!    REGION_BEGIN (inclusive) and REGION_END (exclusive), returning
>true
>!    on success.  The region has N_STMTS statements and has the
>datarefs
>!    given by DATAREFS.  */
>! 
>! static bool
>! vect_slp_bb_region (gimple_stmt_iterator region_begin,
>! 		    gimple_stmt_iterator region_end,
>! 		    vec<data_reference_p> datarefs,
>! 		    unsigned int n_stmts)
>  {
>    bb_vec_info bb_vinfo;
>    auto_vector_sizes vector_sizes;
>  
>    /* Autodetect first vector size we try.  */
>*************** vect_slp_bb (basic_block bb)
>*** 3024,3069 ****
>   targetm.vectorize.autovectorize_vector_sizes (&vector_sizes, false);
>    unsigned int next_size = 0;
>  
>!   gsi = gsi_start_bb (bb);
>  
>    poly_uint64 autodetected_vector_size = 0;
>    while (1)
>      {
>-       if (gsi_end_p (gsi))
>- 	break;
>- 
>-       gimple_stmt_iterator region_begin = gsi;
>-       vec<data_reference_p> datarefs = vNULL;
>-       int insns = 0;
>- 
>-       for (; !gsi_end_p (gsi); gsi_next (&gsi))
>- 	{
>- 	  gimple *stmt = gsi_stmt (gsi);
>- 	  if (is_gimple_debug (stmt))
>- 	    continue;
>- 	  insns++;
>- 
>- 	  if (gimple_location (stmt) != UNKNOWN_LOCATION)
>- 	    vect_location = stmt;
>- 
>- 	  if (!vect_find_stmt_data_reference (NULL, stmt, &datarefs))
>- 	    break;
>- 	}
>- 
>-       /* Skip leading unhandled stmts.  */
>-       if (gsi_stmt (region_begin) == gsi_stmt (gsi))
>- 	{
>- 	  gsi_next (&gsi);
>- 	  continue;
>- 	}
>- 
>-       gimple_stmt_iterator region_end = gsi;
>- 
>        bool vectorized = false;
>        bool fatal = false;
>-       vec_info_shared shared;
>        bb_vinfo = vect_slp_analyze_bb_1 (region_begin, region_end,
>! 					datarefs, insns, fatal, &shared);
>        if (bb_vinfo
>  	  && dbg_cnt (vect_slp))
>  	{
>--- 3030,3044 ----
>   targetm.vectorize.autovectorize_vector_sizes (&vector_sizes, false);
>    unsigned int next_size = 0;
>  
>!   vec_info_shared shared;
>  
>    poly_uint64 autodetected_vector_size = 0;
>    while (1)
>      {
>        bool vectorized = false;
>        bool fatal = false;
>        bb_vinfo = vect_slp_analyze_bb_1 (region_begin, region_end,
>! 					datarefs, n_stmts, fatal, &shared);
>        if (bb_vinfo
>  	  && dbg_cnt (vect_slp))
>  	{
>*************** vect_slp_bb (basic_block bb)
>*** 3090,3097 ****
>  	}
>        delete bb_vinfo;
>  
>-       any_vectorized |= vectorized;
>- 
>        if (next_size == 0)
>  	autodetected_vector_size = current_vector_size;
>  
>--- 3065,3070 ----
>*************** vect_slp_bb (basic_block bb)
>*** 3105,3137 ****
>  	  /* If vect_slp_analyze_bb_1 signaled that analysis for all
>  	     vector sizes will fail do not bother iterating.  */
>  	  || fatal)
>  	{
>! 	  if (gsi_end_p (region_end))
>! 	    break;
>  
>! 	  /* Skip the unhandled stmt.  */
>! 	  gsi_next (&gsi);
>  
>! 	  /* And reset vector sizes.  */
>! 	  current_vector_size = 0;
>! 	  next_size = 0;
>! 	}
>!       else
>  	{
>! 	  /* Try the next biggest vector size.  */
>! 	  current_vector_size = vector_sizes[next_size++];
>! 	  if (dump_enabled_p ())
>! 	    {
>! 	      dump_printf_loc (MSG_NOTE, vect_location,
>! 			       "***** Re-trying analysis with "
>! 			       "vector size ");
>! 	      dump_dec (MSG_NOTE, current_vector_size);
>! 	      dump_printf (MSG_NOTE, "\n");
>! 	    }
>  
>! 	  /* Start over.  */
>! 	  gsi = region_begin;
>  	}
>      }
>  
>    return any_vectorized;
>--- 3078,3145 ----
>  	  /* If vect_slp_analyze_bb_1 signaled that analysis for all
>  	     vector sizes will fail do not bother iterating.  */
>  	  || fatal)
>+ 	return vectorized;
>+ 
>+       /* Try the next biggest vector size.  */
>+       current_vector_size = vector_sizes[next_size++];
>+       if (dump_enabled_p ())
>  	{
>! 	  dump_printf_loc (MSG_NOTE, vect_location,
>! 			   "***** Re-trying analysis with "
>! 			   "vector size ");
>! 	  dump_dec (MSG_NOTE, current_vector_size);
>! 	  dump_printf (MSG_NOTE, "\n");
>! 	}
>!     }
>! }
>  
>! /* Main entry for the BB vectorizer.  Analyze and transform BB,
>returns
>!    true if anything in the basic-block was vectorized.  */
>  
>! bool
>! vect_slp_bb (basic_block bb)
>! {
>!   gimple_stmt_iterator gsi;
>!   bool any_vectorized = false;
>! 
>!   gsi = gsi_start_bb (bb);
>!   while (!gsi_end_p (gsi))
>!     {
>!       gimple_stmt_iterator region_begin = gsi;
>!       vec<data_reference_p> datarefs = vNULL;
>!       int insns = 0;
>! 
>!       for (; !gsi_end_p (gsi); gsi_next (&gsi))
>  	{
>! 	  gimple *stmt = gsi_stmt (gsi);
>! 	  if (is_gimple_debug (stmt))
>! 	    continue;
>! 	  insns++;
>! 
>! 	  if (gimple_location (stmt) != UNKNOWN_LOCATION)
>! 	    vect_location = stmt;
>  
>! 	  if (!vect_find_stmt_data_reference (NULL, stmt, &datarefs))
>! 	    break;
>  	}
>+ 
>+       /* Skip leading unhandled stmts.  */
>+       if (gsi_stmt (region_begin) == gsi_stmt (gsi))
>+ 	{
>+ 	  gsi_next (&gsi);
>+ 	  continue;
>+ 	}
>+ 
>+       gimple_stmt_iterator region_end = gsi;
>+ 
>+       if (vect_slp_bb_region (region_begin, region_end, datarefs,
>insns))
>+ 	any_vectorized = true;
>+ 
>+       if (gsi_end_p (region_end))
>+ 	break;
>+ 
>+       /* Skip the unhandled stmt.  */
>+       gsi_next (&gsi);
>      }
>  
>    return any_vectorized;