diff mbox

Patch: Add #pragma ivdep support to the ME and C FE

Message ID 526927B2.3080000@net-b.de
State New
Headers show

Commit Message

Tobias Burnus Oct. 24, 2013, 1:59 p.m. UTC
Richard Biener wrote:
> Please instead amend tree-cfg.c:execute_build_cfg and do after loop_optimizer_init sth like
> [...]
> possibly doing after it
>    FOR_EACH_BB (bb)
> [...]
> +DEFTREECODE (ANNOTATE_EXPR, "annotate_expr", tcc_expression, 1)
>
> FIXME?  Btw, not using a tree operand here is somewhat of a premature
> optimization I think ...

I have now moved the code to a new function in tree-cfg.c. As it was not 
completely clear to me after the IRC discussion whether it is sufficient 
to assume that loop->latch == NULL won't occur (and rely on the assert 
in internal_fn), I created a version with FOR_EACH_BB and without. (see 
#if 0 code in tree-cfg.c).

For  both versions, I did a true all-language bootstrap+ regtesting on 
x86-64-gnu-linux.

Additionally, the ID is now stored in an INTEGER_CST and I also 
documented ANNOTATE_EXPR in generic.texi.

OK? Or are there more comments?

Tobias

PS: Before committal, I will remove the #if 0 code - or enable it, as 
preferred.

Comments

Richard Biener Oct. 24, 2013, 2:14 p.m. UTC | #1
On Thu, 24 Oct 2013, Tobias Burnus wrote:

> Richard Biener wrote:
> > Please instead amend tree-cfg.c:execute_build_cfg and do after
> > loop_optimizer_init sth like
> > [...]
> > possibly doing after it
> >    FOR_EACH_BB (bb)
> > [...]
> > +DEFTREECODE (ANNOTATE_EXPR, "annotate_expr", tcc_expression, 1)
> > 
> > FIXME?  Btw, not using a tree operand here is somewhat of a premature
> > optimization I think ...
> 
> I have now moved the code to a new function in tree-cfg.c. As it was not
> completely clear to me after the IRC discussion whether it is sufficient to
> assume that loop->latch == NULL won't occur (and rely on the assert in
> internal_fn), I created a version with FOR_EACH_BB and without. (see #if 0
> code in tree-cfg.c).
> 
> For  both versions, I did a true all-language bootstrap+ regtesting on
> x86-64-gnu-linux.
> 
> Additionally, the ID is now stored in an INTEGER_CST and I also documented
> ANNOTATE_EXPR in generic.texi.
> 
> OK? Or are there more comments?

Looks good to me now ...

> Tobias
> 
> PS: Before committal, I will remove the #if 0 code - or enable it, as
> preferred.

... I think it's "easy" to trigger.

entry:
#pragma GCC ivdep
  for (i = 0; i < n; ++i) {
    a[i] = b[i] + c[i];
    if (i == 5)
      goto entry;
    a[i] = a[i] + 1;
  }

or by using setjmp/longjmp and thus having some abnormal edges
into a function call inside a loop.

The whole point is that execute_build_cfg passes AVOID_CFG_MODIFICATIONS
to loop_optimizer_init and thus it won't disambiguate loops with
multiple latches.  We could eventually change that and pass
LOOPS_HAVE_SIMPLE_LATCHES, but I'm not sure if that works well enough
alone (it's not really tested besides of in combination LOOPS_NORMAL).

Richard.
Jakub Jelinek Oct. 24, 2013, 2:20 p.m. UTC | #2
On Thu, Oct 24, 2013 at 03:59:14PM +0200, Tobias Burnus wrote:
> +  FOR_EACH_LOOP (li, loop, 0)
> +    if (loop->latch)
> +      FOR_EACH_EDGE (e, ei, loop->latch->succs)

Isn't the only successor of loop->latch loop->header?
Thus, can't you just remove the above two lines and use loop->header
instead of e->dest?

	Jakub
Richard Biener Oct. 24, 2013, 2:31 p.m. UTC | #3
On Thu, 24 Oct 2013, Jakub Jelinek wrote:

> On Thu, Oct 24, 2013 at 03:59:14PM +0200, Tobias Burnus wrote:
> > +  FOR_EACH_LOOP (li, loop, 0)
> > +    if (loop->latch)
> > +      FOR_EACH_EDGE (e, ei, loop->latch->succs)
> 
> Isn't the only successor of loop->latch loop->header?
> Thus, can't you just remove the above two lines and use loop->header
> instead of e->dest?

Yes, of course.

Richard.
diff mbox

Patch

2013-08-24  Tobias Burnus  <burnus@net-b.de>

	PR other/33426
        * c-pragma.c (init_pragma) Add #pragma ivdep handling.
        * c-pragma.h (pragma_kind): Add PRAGMA_IVDEP.

	PR other/33426
        * c-parser.c (c_parser_pragma, c_parser_for_statement):
        Handle PRAGMA_IVDEP.
        (c_parser_statement_after_labels): Update call.

	PR other/33426
        * tree-cfg.c (replace_loop_annotate): New function.
        (execute_build_cfg): Call it.
        * gimplify.c (gimple_boolify, gimplify_expr): Handle ANNOTATE_EXPR.
        * internal-fn.c (expand_ANNOTATE): New function.
        * internal-fn.def (ANNOTATE): Define as new internal function.
        * tree-core.h (tree_node_kind): Add annot_expr_ivdep_kind.
        * tree-pretty-print.c (dump_generic_node): Handle ANNOTATE_EXPR.
        * tree.def (ANNOTATE_EXPR): New DEFTREECODE.
	* doc/extend.texi (Pragmas): Document #pragma ivdep.
	* doc/generic.texi (Expressions): Document ANNOTATE_EXPR.

	PR other/33426
	* testsuite/gcc.dg/ivdep.c: New.
	* testsuite/gcc.dg/vect/vect-ivdep-1.c: New.

diff --git a/gcc/c-family/c-pragma.c b/gcc/c-family/c-pragma.c
index be5748b..1656000 100644
--- a/gcc/c-family/c-pragma.c
+++ b/gcc/c-family/c-pragma.c
@@ -1362,6 +1362,8 @@  init_pragma (void)
     cpp_register_deferred_pragma (parse_in, "GCC", "pch_preprocess",
 				  PRAGMA_GCC_PCH_PREPROCESS, false, false);
 
+  cpp_register_deferred_pragma (parse_in, "GCC", "ivdep", PRAGMA_IVDEP, false,
+				false);
 #ifdef HANDLE_PRAGMA_PACK_WITH_EXPANSION
   c_register_pragma_with_expansion (0, "pack", handle_pragma_pack);
 #else
diff --git a/gcc/c-family/c-pragma.h b/gcc/c-family/c-pragma.h
index c421284..705bcb4 100644
--- a/gcc/c-family/c-pragma.h
+++ b/gcc/c-family/c-pragma.h
@@ -53,6 +53,7 @@  typedef enum pragma_kind {
   PRAGMA_OMP_TEAMS,
 
   PRAGMA_GCC_PCH_PREPROCESS,
+  PRAGMA_IVDEP,
 
   PRAGMA_FIRST_EXTERNAL
 } pragma_kind;
diff --git a/gcc/c/c-parser.c b/gcc/c/c-parser.c
index 4d6c930..4f25078 100644
--- a/gcc/c/c-parser.c
+++ b/gcc/c/c-parser.c
@@ -1159,7 +1159,7 @@  static void c_parser_if_statement (c_parser *);
 static void c_parser_switch_statement (c_parser *);
 static void c_parser_while_statement (c_parser *);
 static void c_parser_do_statement (c_parser *);
-static void c_parser_for_statement (c_parser *);
+static void c_parser_for_statement (c_parser *, bool);
 static tree c_parser_asm_statement (c_parser *);
 static tree c_parser_asm_operands (c_parser *);
 static tree c_parser_asm_goto_operands (c_parser *);
@@ -4585,7 +4585,7 @@  c_parser_statement_after_labels (c_parser *parser)
 	  c_parser_do_statement (parser);
 	  break;
 	case RID_FOR:
-	  c_parser_for_statement (parser);
+	  c_parser_for_statement (parser, false);
 	  break;
 	case RID_GOTO:
 	  c_parser_consume_token (parser);
@@ -5038,7 +5038,7 @@  c_parser_do_statement (c_parser *parser)
 */
 
 static void
-c_parser_for_statement (c_parser *parser)
+c_parser_for_statement (c_parser *parser, bool ivdep)
 {
   tree block, cond, incr, save_break, save_cont, body;
   /* The following are only used when parsing an ObjC foreach statement.  */
@@ -5144,8 +5144,17 @@  c_parser_for_statement (c_parser *parser)
 	{
 	  if (c_parser_next_token_is (parser, CPP_SEMICOLON))
 	    {
-	      c_parser_consume_token (parser);
-	      cond = NULL_TREE;
+	      if (ivdep)
+		{
+		  c_parser_error (parser, "missing loop condition in loop with "
+				  "%<GCC ivdep%> pragma");
+		  cond = error_mark_node;
+		}
+	      else
+		{
+		  c_parser_consume_token (parser);
+		  cond = NULL_TREE;
+		}
 	    }
 	  else
 	    {
@@ -5159,6 +5168,10 @@  c_parser_for_statement (c_parser *parser)
 	      c_parser_skip_until_found (parser, CPP_SEMICOLON,
 					 "expected %<;%>");
 	    }
+	  if (ivdep && cond != error_mark_node)
+	    cond = build2 (ANNOTATE_EXPR, TREE_TYPE (cond), cond,
+			   build_int_cst (integer_type_node,
+			   annot_expr_ivdep_kind));
 	}
       /* Parse the increment expression (the third expression in a
 	 for-statement).  In the case of a foreach-statement, this is
@@ -9086,6 +9099,16 @@  c_parser_pragma (c_parser *parser, enum pragma_context context)
     case PRAGMA_OMP_DECLARE_REDUCTION:
       c_parser_omp_declare (parser, context);
       return false;
+    case PRAGMA_IVDEP:
+      c_parser_consume_pragma (parser);
+      c_parser_skip_to_pragma_eol (parser);
+      if (!c_parser_next_token_is_keyword (parser, RID_FOR))
+	{
+	  c_parser_error (parser, "for statement expected");
+	  return false;
+	}
+      c_parser_for_statement (parser, true);
+      return false;
 
     case PRAGMA_GCC_PCH_PREPROCESS:
       c_parser_error (parser, "%<#pragma GCC pch_preprocess%> must be first");
diff --git a/gcc/doc/extend.texi b/gcc/doc/extend.texi
index 562b676..7623a5e 100644
--- a/gcc/doc/extend.texi
+++ b/gcc/doc/extend.texi
@@ -15473,6 +15473,7 @@  for further explanation.
 * Visibility Pragmas::
 * Push/Pop Macro Pragmas::
 * Function Specific Option Pragmas::
+* Loop-Specific Pragmas::
 @end menu
 
 @node ARM Pragmas
@@ -15995,6 +15996,48 @@  The @samp{#pragma GCC reset_options} pragma is not implemented in GCC
 versions earlier than 4.4.
 @end table
 
+@node Loop-Specific Pragmas
+@subsection Loop-Specific Pragmas
+
+@table @code
+@item #pragma GCC ivdep
+@cindex pragma GCC ivdep
+@end table
+
+With this pragma, the programmer asserts that there are no loop-carried
+dependencies which would prevent that consecutive iterations of
+the following loop can be executed concurrently with SIMD
+(single instruction multiple data) instructions.
+
+For example, the compiler can only unconditionally vectorize the following
+loop with the pragma:
+
+@smallexample
+void foo (int n, int *a, int *b, int *c)
+@{
+  int i, j;
+#pragma GCC ivdep
+  for (i = 0; i < n; ++i)
+    a[i] = b[i] + c[i];
+@}
+@end smallexample
+
+@noindent
+In this example, using the @code{restrict} qualifier had the same
+effect. In the following example, that would not be possible. Assume
+@math{k < -m} or @math{k >= m}. Only with the pragma, the compiler knows
+that it can unconditionally vectorize the following loop:
+
+@smallexample
+void ignore_vec_dep (int *a, int k, int c, int m)
+@{
+#pragma GCC ivdep
+  for (int i = 0; i < m; i++)
+    a[i] = a[i + k] * c;
+@}
+@end smallexample
+
+
 @node Unnamed Fields
 @section Unnamed struct/union fields within structs/unions
 @cindex @code{struct}
diff --git a/gcc/doc/generic.texi b/gcc/doc/generic.texi
index ccecd6e..75c4745 100644
--- a/gcc/doc/generic.texi
+++ b/gcc/doc/generic.texi
@@ -1279,6 +1279,7 @@  the byte offset of the field, but should not be used directly; call
 @tindex SAVE_EXPR
 @tindex TARGET_EXPR
 @tindex VA_ARG_EXPR
+@tindex ANNOTATE_EXPR
 
 @table @code
 @item NEGATE_EXPR
@@ -1692,8 +1693,13 @@  mechanism.  It represents expressions like @code{va_arg (ap, type)}.
 Its @code{TREE_TYPE} yields the tree representation for @code{type} and
 its sole argument yields the representation for @code{ap}.
 
+@item ANNOTATE_EXPR
+This node is used to attach markers to an expression. The first operand
+is the annotated expression, the second is an @code{INTEGER_CST} with
+a value from @code{enum annot_expr_kind}.
 @end table
 
+
 @node Vectors
 @subsection Vectors
 @tindex VEC_LSHIFT_EXPR
diff --git a/gcc/gimplify.c b/gcc/gimplify.c
index fa1b155..8d78440 100644
--- a/gcc/gimplify.c
+++ b/gcc/gimplify.c
@@ -3076,6 +3076,17 @@  gimple_boolify (tree expr)
 	TREE_TYPE (expr) = boolean_type_node;
       return expr;
 
+    case ANNOTATE_EXPR:
+      if ((enum annot_expr_kind) TREE_INT_CST_LOW (TREE_OPERAND (expr, 1))
+	  == annot_expr_ivdep_kind)
+	{
+	  TREE_OPERAND (expr, 0) = gimple_boolify (TREE_OPERAND (expr, 0));
+	  if (TREE_CODE (type) != BOOLEAN_TYPE)
+	    TREE_TYPE (expr) = boolean_type_node;
+	  return expr;
+	}
+      /* FALLTHRU */
+
     default:
       if (COMPARISON_CLASS_P (expr))
 	{
@@ -7731,6 +7742,21 @@  gimplify_expr (tree *expr_p, gimple_seq *pre_p, gimple_seq *post_p,
 	  ret = gimplify_addr_expr (expr_p, pre_p, post_p);
 	  break;
 
+	case ANNOTATE_EXPR:
+	  {
+	    tree cond = TREE_OPERAND (*expr_p, 0);
+	    tree id = TREE_OPERAND (*expr_p, 1);
+	    tree tmp = create_tmp_var_raw (TREE_TYPE(cond), NULL);
+	    gimplify_arg (&cond, pre_p, EXPR_LOCATION (*expr_p));
+	    gimple call = gimple_build_call_internal (IFN_ANNOTATE, 2,
+						      cond, id);
+	    gimple_call_set_lhs (call, tmp);
+	    gimplify_seq_add_stmt (pre_p, call);
+	    *expr_p = tmp;
+	    ret = GS_ALL_DONE;
+	    break;
+	  }
+
 	case VA_ARG_EXPR:
 	  ret = gimplify_va_arg_expr (expr_p, pre_p, post_p);
 	  break;
diff --git a/gcc/internal-fn.c b/gcc/internal-fn.c
index 983efeb..a22f222 100644
--- a/gcc/internal-fn.c
+++ b/gcc/internal-fn.c
@@ -109,6 +109,12 @@  expand_STORE_LANES (gimple stmt)
   expand_insn (get_multi_vector_move (type, vec_store_lanes_optab), 2, ops);
 }
 
+static void
+expand_ANNOTATE (gimple stmt ATTRIBUTE_UNUSED)
+{
+  gcc_unreachable ();
+}
+
 /* This should get expanded in adjust_simduid_builtins.  */
 
 static void
diff --git a/gcc/internal-fn.def b/gcc/internal-fn.def
index 5427664..0f5cc3c 100644
--- a/gcc/internal-fn.def
+++ b/gcc/internal-fn.def
@@ -43,3 +43,4 @@  DEF_INTERNAL_FN (STORE_LANES, ECF_CONST | ECF_LEAF)
 DEF_INTERNAL_FN (GOMP_SIMD_LANE, ECF_NOVOPS | ECF_LEAF | ECF_NOTHROW)
 DEF_INTERNAL_FN (GOMP_SIMD_VF, ECF_CONST | ECF_LEAF | ECF_NOTHROW)
 DEF_INTERNAL_FN (GOMP_SIMD_LAST_LANE, ECF_CONST | ECF_LEAF | ECF_NOTHROW)
+DEF_INTERNAL_FN (ANNOTATE,  ECF_CONST | ECF_LEAF | ECF_NOTHROW)
diff --git a/gcc/testsuite/gcc.dg/ivdep.c b/gcc/testsuite/gcc.dg/ivdep.c
new file mode 100644
index 0000000..23d51de
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/ivdep.c
@@ -0,0 +1,11 @@ 
+/* { dg-do compile } */
+
+/* PR other/33426 */
+
+void foo(int n, int *a, int *b, int *c, int *d, int *e) {
+  int i, j;
+#pragma GCC ivdep
+  for (i = 0; ; ++i) { /* { dg-error "missing loop condition in loop with 'GCC ivdep' pragma before ';' token" } */
+    a[i] = b[i] + c[i];
+  }
+}
diff --git a/gcc/testsuite/gcc.dg/vect/vect-ivdep-1.c b/gcc/testsuite/gcc.dg/vect/vect-ivdep-1.c
new file mode 100644
index 0000000..8767bb8
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/vect/vect-ivdep-1.c
@@ -0,0 +1,19 @@ 
+/* { dg-do compile } */
+/* { dg-require-effective-target vect_float } */
+/* { dg-options "-O3 -fopt-info-vec-optimized" } */
+
+/* PR other/33426 */
+/* Testing whether #pragma ivdep is working.  */
+
+void foo(int n, int *a, int *b, int *c, int *d, int *e) {
+  int i, j;
+#pragma GCC ivdep
+  for (i = 0; i < n; ++i) {
+    a[i] = b[i] + c[i];
+  }
+}
+
+/* { dg-message "loop vectorized" "" { target *-*-* } 0 } */
+/* { dg-bogus "version" "" { target *-*-* } 0 } */
+/* { dg-bogus "alias" "" { target *-*-* } 0 } */
+/* { dg-final { cleanup-tree-dump "vect" } } */
diff --git a/gcc/tree-cfg.c b/gcc/tree-cfg.c
index 2021d94..5348c1f 100644
--- a/gcc/tree-cfg.c
+++ b/gcc/tree-cfg.c
@@ -250,6 +250,77 @@  build_gimple_cfg (gimple_seq seq)
   discriminator_per_locus.dispose ();
 }
 
+
+/* Search for ANNOTATE call with annot_expr_ivdep_kind; if found, remove
+   it and set loop->safelen to INT_MAX.  We assume that the annotation
+   comes immediately before the condition.  */
+
+static void
+replace_loop_annotate ()
+{
+  struct loop *loop;
+  loop_iterator li;
+/*  basic_block bb; */
+  edge e;
+  edge_iterator ei;
+  gimple_stmt_iterator gsi;
+  gimple stmt;
+
+  FOR_EACH_LOOP (li, loop, 0)
+    if (loop->latch)
+      FOR_EACH_EDGE (e, ei, loop->latch->succs)
+	{
+	  gsi = gsi_last_bb (e->dest);
+	  stmt = gsi_stmt (gsi);
+	  if (stmt && gimple_code (stmt) == GIMPLE_COND)
+	    {
+	      gsi_prev_nondebug (&gsi);
+	      if (gsi_end_p (gsi))
+		continue;
+	      stmt = gsi_stmt (gsi);
+	      if (gimple_code (stmt) != GIMPLE_CALL)
+		continue;
+	      if (!gimple_call_internal_p (stmt)
+		  || gimple_call_internal_fn (stmt) != IFN_ANNOTATE)
+		continue;
+	      if ((annot_expr_kind) tree_low_cst (gimple_call_arg (stmt, 1), 0)
+		  != annot_expr_ivdep_kind)
+		continue;
+	      stmt = gimple_build_assign (gimple_call_lhs (stmt),
+					  gimple_call_arg (stmt, 0));
+	      gsi_replace (&gsi, stmt, true);
+	      loop->safelen = INT_MAX;
+	    }
+	}
+
+#if 0
+  /* Remove IFN_ANNOTATE. Safeguard for the case loop->latch == NULL.  */
+  FOR_EACH_BB (bb)
+    {
+      gsi = gsi_last_bb (bb);
+      stmt = gsi_stmt (gsi);
+      if (stmt && gimple_code (stmt) == GIMPLE_COND)
+	gsi_prev_nondebug (&gsi);
+      if (gsi_end_p (gsi))
+	continue;
+      stmt = gsi_stmt (gsi);
+      if (gimple_code (stmt) != GIMPLE_CALL)
+	continue;
+      if (!gimple_call_internal_p (stmt)
+	  || gimple_call_internal_fn (stmt) != IFN_ANNOTATE)
+	continue;
+      if ((annot_expr_kind) tree_low_cst (gimple_call_arg (stmt, 1), 0)
+	  != annot_expr_ivdep_kind)
+	continue;
+      warning (0, "ignoring %<GCC ivdep%> annotation");
+      stmt = gimple_build_assign (gimple_call_lhs (stmt),
+				  gimple_call_arg (stmt, 0));
+      gsi_replace (&gsi, stmt, true);
+    }
+#endif
+}
+
+
 static unsigned int
 execute_build_cfg (void)
 {
@@ -264,6 +335,7 @@  execute_build_cfg (void)
     }
   cleanup_tree_cfg ();
   loop_optimizer_init (AVOID_CFG_MODIFICATIONS);
+  replace_loop_annotate ();
   return 0;
 }
 
diff --git a/gcc/tree-core.h b/gcc/tree-core.h
index c7fc84f..ea110bc 100644
--- a/gcc/tree-core.h
+++ b/gcc/tree-core.h
@@ -647,6 +647,10 @@  enum tree_node_kind {
   all_kinds
 };
 
+enum annot_expr_kind {
+  annot_expr_ivdep_kind
+};
+
 
 /*---------------------------------------------------------------------------
                                 Type definitions
diff --git a/gcc/tree-pretty-print.c b/gcc/tree-pretty-print.c
index b3bd8fa..150333b 100644
--- a/gcc/tree-pretty-print.c
+++ b/gcc/tree-pretty-print.c
@@ -2095,6 +2095,18 @@  dump_generic_node (pretty_printer *buffer, tree node, int spc, int flags,
       pp_string (buffer, " predictor.");
       break;
 
+    case ANNOTATE_EXPR:
+      pp_string (buffer, "ANNOTATE_EXPR <");
+      switch ((enum annot_expr_kind) TREE_INT_CST_LOW (TREE_OPERAND (node, 1)))
+	{
+	case annot_expr_ivdep_kind:
+	  pp_string (buffer, "ivdep, ");
+	  break;
+	}
+      dump_generic_node (buffer, TREE_OPERAND (node, 0), spc, flags, false);
+      pp_greater (buffer);
+      break;
+
     case RETURN_EXPR:
       pp_string (buffer, "return");
       op0 = TREE_OPERAND (node, 0);
diff --git a/gcc/tree.def b/gcc/tree.def
index 88c850a..9bd1b3d 100644
--- a/gcc/tree.def
+++ b/gcc/tree.def
@@ -1261,6 +1261,12 @@  DEFTREECODE (OPTIMIZATION_NODE, "optimization_node", tcc_exceptional, 0)
 /* TARGET_OPTION_NODE.  Node to store the target specific options.  */
 DEFTREECODE (TARGET_OPTION_NODE, "target_option_node", tcc_exceptional, 0)
 
+/* ANNOTATE_EXPR.
+   Operand 0 is the expression to be annotated.
+   Operand 1 is the annotation id. */
+DEFTREECODE (ANNOTATE_EXPR, "annotate_expr", tcc_expression, 2)
+
+
 /*
 Local variables:
 mode:c