diff mbox series

tree-optimization/106365 - DSE of LEN_STORE and MASK_STORE

Message ID 20220721090553.6DCFD13A1B@imap2.suse-dmz.suse.de
State New
Headers show
Series tree-optimization/106365 - DSE of LEN_STORE and MASK_STORE | expand

Commit Message

Richard Biener July 21, 2022, 9:05 a.m. UTC 
The following enhances DSE to handle LEN_STORE (optimally) and
MASK_STORE (conservatively).

Bootstrapped on x86_64-unknown-linux-gnu, testing in progress.
Kewen is testing on powerpc.  Handling MASK_STORE_LANES in
a similar way to MASK_STORE is probably possible but I couldn't
figure a way to generate one for testing.  STORE_LANES is
probably handled already since it's ECF_CONST.

	PR tree-optimization/106365
	* tree-ssa-dse.cc (initialize_ao_ref_for_dse): Handle
	LEN_STORE, add mode to initialize a may-def and handle
	MASK_STORE that way.
	(dse_optimize_stmt): Query may-defs.  Handle internal
	functions LEN_STORE and MASK_STORE similar to how
	we handle memory builtins but without byte tracking.
---
 gcc/tree-ssa-dse.cc | 55 +++++++++++++++++++++++++++++++++++++++++----
 1 file changed, 51 insertions(+), 4 deletions(-)

Comments

Jeff Law July 31, 2022, 2:45 a.m. UTC | #1 
On 7/21/2022 3:05 AM, Richard Biener via Gcc-patches wrote:
> The following enhances DSE to handle LEN_STORE (optimally) and
> MASK_STORE (conservatively).
>
> Bootstrapped on x86_64-unknown-linux-gnu, testing in progress.
> Kewen is testing on powerpc.  Handling MASK_STORE_LANES in
> a similar way to MASK_STORE is probably possible but I couldn't
> figure a way to generate one for testing.  STORE_LANES is
> probably handled already since it's ECF_CONST.
>
> 	PR tree-optimization/106365
> 	* tree-ssa-dse.cc (initialize_ao_ref_for_dse): Handle
> 	LEN_STORE, add mode to initialize a may-def and handle
> 	MASK_STORE that way.
> 	(dse_optimize_stmt): Query may-defs.  Handle internal
> 	functions LEN_STORE and MASK_STORE similar to how
> 	we handle memory builtins but without byte tracking.
LGTM.  Obviously if you can massage the PR into a testcase, it's even 
better :-)

jeff
diff mbox series

Patch

diff --git a/gcc/tree-ssa-dse.cc b/gcc/tree-ssa-dse.cc
index 8d1739a4510..34cfd1a8802 100644
--- a/gcc/tree-ssa-dse.cc
+++ b/gcc/tree-ssa-dse.cc
@@ -93,7 +93,9 @@  static bitmap need_eh_cleanup;
 static bitmap need_ab_cleanup;
 
 /* STMT is a statement that may write into memory.  Analyze it and
-   initialize WRITE to describe how STMT affects memory.
+   initialize WRITE to describe how STMT affects memory.  When
+   MAY_DEF_OK is true then the function initializes WRITE to what
+   the stmt may define.
 
    Return TRUE if the statement was analyzed, FALSE otherwise.
 
@@ -101,7 +103,7 @@  static bitmap need_ab_cleanup;
    can be achieved by analyzing more statements.  */
 
 static bool
-initialize_ao_ref_for_dse (gimple *stmt, ao_ref *write)
+initialize_ao_ref_for_dse (gimple *stmt, ao_ref *write, bool may_def_ok = false)
 {
   /* It's advantageous to handle certain mem* functions.  */
   if (gimple_call_builtin_p (stmt, BUILT_IN_NORMAL))
@@ -146,6 +148,32 @@  initialize_ao_ref_for_dse (gimple *stmt, ao_ref *write)
 	  break;
 	}
     }
+  else if (is_gimple_call (stmt)
+	   && gimple_call_internal_p (stmt))
+    {
+      switch (gimple_call_internal_fn (stmt))
+	{
+	case IFN_LEN_STORE:
+	  ao_ref_init_from_ptr_and_size
+	      (write, gimple_call_arg (stmt, 0),
+	       int_const_binop (MINUS_EXPR,
+				gimple_call_arg (stmt, 2),
+				gimple_call_arg (stmt, 4)));
+	  return true;
+	case IFN_MASK_STORE:
+	  /* We cannot initialize a must-def ao_ref (in all cases) but we
+	     can provide a may-def variant.  */
+	  if (may_def_ok)
+	    {
+	      ao_ref_init_from_ptr_and_size
+		  (write, gimple_call_arg (stmt, 0),
+		   TYPE_SIZE_UNIT (TREE_TYPE (gimple_call_arg (stmt, 2))));
+	      return true;
+	    }
+	  break;
+	default:;
+	}
+    }
   else if (tree lhs = gimple_get_lhs (stmt))
     {
       if (TREE_CODE (lhs) != SSA_NAME)
@@ -1328,8 +1356,10 @@  dse_optimize_stmt (function *fun, gimple_stmt_iterator *gsi, sbitmap live_bytes)
 
   ao_ref ref;
   /* If this is not a store we can still remove dead call using
-     modref summary.  */
-  if (!initialize_ao_ref_for_dse (stmt, &ref))
+     modref summary.  Note we specifically allow ref to be initialized
+     to a conservative may-def since we are looking for followup stores
+     to kill all of it.  */
+  if (!initialize_ao_ref_for_dse (stmt, &ref, true))
     {
       dse_optimize_call (gsi, live_bytes);
       return;
@@ -1398,6 +1428,23 @@  dse_optimize_stmt (function *fun, gimple_stmt_iterator *gsi, sbitmap live_bytes)
 	  return;
 	}
     }
+  else if (is_gimple_call (stmt)
+	   && gimple_call_internal_p (stmt))
+    {
+      switch (gimple_call_internal_fn (stmt))
+	{
+	case IFN_LEN_STORE:
+	case IFN_MASK_STORE:
+	  {
+	    enum dse_store_status store_status;
+	    store_status = dse_classify_store (&ref, stmt, false, live_bytes);
+	    if (store_status == DSE_STORE_DEAD)
+	      delete_dead_or_redundant_call (gsi, "dead");
+	    return;
+	  }
+	default:;
+	}
+    }
 
   bool by_clobber_p = false;