From: Richard Biener <rguenther@suse.de>
Date: Tue, 3 Dec 2019 11:59:13 +0000 (+0000)
Subject: re PR tree-optimization/92645 (Hand written vector code is 450 times slower when... 
X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=5105b576dfdcae0c6cc30a89a68b4ef3cbf56a1c;p=gcc.git

re PR tree-optimization/92645 (Hand written vector code is 450 times slower when compiled with GCC compared to Clang)

2019-12-03  Richard Biener  <rguenther@suse.de>

	PR tree-optimization/92645
	* gimple-fold.c (gimple_fold_builtin_memory_op): Fold memcpy
	from or to a properly aligned register variable.

	* gcc.target/i386/pr92645-5.c: New testcase.

From-SVN: r278934
---

diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index 9917ec2eb09..e9119f99745 100644
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1,3 +1,9 @@
+2019-12-03  Richard Biener  <rguenther@suse.de>
+
+	PR tree-optimization/92645
+	* gimple-fold.c (gimple_fold_builtin_memory_op): Fold memcpy
+	from or to a properly aligned register variable.
+
 2019-12-03  Matthias Klose  <doko@ubuntu.com>
 
 	* Makefile.in (SOURCES): Add doc/lto-dump.1.
diff --git a/gcc/gimple-fold.c b/gcc/gimple-fold.c
index 849bee2e069..24a478a8007 100644
--- a/gcc/gimple-fold.c
+++ b/gcc/gimple-fold.c
@@ -986,36 +986,33 @@ gimple_fold_builtin_memory_op (gimple_stmt_iterator *gsi,
 
       src_align = get_pointer_alignment (src);
       dest_align = get_pointer_alignment (dest);
-      if (dest_align < TYPE_ALIGN (desttype)
-	  || src_align < TYPE_ALIGN (srctype))
-	return false;
 
+      /* Choose between src and destination type for the access based
+         on alignment, whether the access constitutes a register access
+	 and whether it may actually expose a declaration for SSA rewrite
+	 or SRA decomposition.  */
       destvar = NULL_TREE;
+      srcvar = NULL_TREE;
       if (TREE_CODE (dest) == ADDR_EXPR
 	  && var_decl_component_p (TREE_OPERAND (dest, 0))
-	  && tree_int_cst_equal (TYPE_SIZE_UNIT (desttype), len))
+	  && tree_int_cst_equal (TYPE_SIZE_UNIT (desttype), len)
+	  && dest_align >= TYPE_ALIGN (desttype)
+	  && (is_gimple_reg_type (desttype)
+	      || src_align >= TYPE_ALIGN (desttype)))
 	destvar = fold_build2 (MEM_REF, desttype, dest, off0);
-
-      srcvar = NULL_TREE;
-      if (TREE_CODE (src) == ADDR_EXPR
-	  && var_decl_component_p (TREE_OPERAND (src, 0))
-	  && tree_int_cst_equal (TYPE_SIZE_UNIT (srctype), len))
-	{
-	  if (!destvar
-	      || src_align >= TYPE_ALIGN (desttype))
-	    srcvar = fold_build2 (MEM_REF, destvar ? desttype : srctype,
-				  src, off0);
-	  else if (!STRICT_ALIGNMENT)
-	    {
-	      srctype = build_aligned_type (TYPE_MAIN_VARIANT (desttype),
-					    src_align);
-	      srcvar = fold_build2 (MEM_REF, srctype, src, off0);
-	    }
-	}
-
+      else if (TREE_CODE (src) == ADDR_EXPR
+	       && var_decl_component_p (TREE_OPERAND (src, 0))
+	       && tree_int_cst_equal (TYPE_SIZE_UNIT (srctype), len)
+	       && src_align >= TYPE_ALIGN (srctype)
+	       && (is_gimple_reg_type (srctype)
+		   || dest_align >= TYPE_ALIGN (srctype)))
+	srcvar = fold_build2 (MEM_REF, srctype, src, off0);
       if (srcvar == NULL_TREE && destvar == NULL_TREE)
 	return false;
 
+      /* Now that we chose an access type express the other side in
+         terms of it if the target allows that with respect to alignment
+	 constraints.  */
       if (srcvar == NULL_TREE)
 	{
 	  if (src_align >= TYPE_ALIGN (desttype))
diff --git a/gcc/testsuite/ChangeLog b/gcc/testsuite/ChangeLog
index e7e07882a43..158e09073b1 100644
--- a/gcc/testsuite/ChangeLog
+++ b/gcc/testsuite/ChangeLog
@@ -1,3 +1,8 @@
+2019-12-03  Richard Biener  <rguenther@suse.de>
+
+	PR tree-optimization/92645
+	* gcc.target/i386/pr92645-5.c: New testcase.
+
 2019-12-03  Richard Biener  <rguenther@suse.de>
 
 	PR tree-optimization/92751
diff --git a/gcc/testsuite/gcc.target/i386/pr92645-5.c b/gcc/testsuite/gcc.target/i386/pr92645-5.c
new file mode 100644
index 00000000000..af3a4e23dca
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/pr92645-5.c
@@ -0,0 +1,21 @@
+/* { dg-do compile } */
+/* { dg-options "-O -fdump-tree-cddce1 -mavx2 -Wno-psabi" } */
+typedef long long v4di __attribute__((vector_size(32)));
+struct Vec
+{
+  unsigned int v[8];
+};
+
+v4di pun (struct Vec *s)
+{
+  v4di tem;
+  __builtin_memcpy (&tem, s, 32);
+  return tem;
+}
+
+/* We're expecting exactly two stmts, in particular no BIT_INSERT_EXPR
+   and no memcpy call.
+    _3 = MEM <vector(4) long long int> [(char * {ref-all})s_2(D)];
+    return _3;  */
+/* { dg-final { scan-tree-dump-times " = MEM" 1 "cddce1" } } */
+/* { dg-final { scan-tree-dump-not "memcpy" "cddce1" } } */