diff mbox

[fortran] PR 37131, inline matmul

Message ID 55486270.8010909@netcologne.de
State New
Headers show

Commit Message

Thomas Koenig May 5, 2015, 6:25 a.m. UTC
Hello world,

this is an update of the matmul inline patch.  The only difference to
the last version is that it has the ubound simplification taken out.

Any further comments?  OK for trunk?

	Thomas

2015-05-05  Thomas Koenig  <tkoenig@gcc.gnu.org>

	PR fortran/37131
	* gfortran.h (gfc_isym_id):  Add GFC_ISYM_FE_RUNTIME_ERROR.
	(gfc_array_spec):  Add resolved flag.
	(gfc_intrinsic_sym):  Add vararg.
	* intrinsic.h (gfc_check_fe_runtime_error):  Add prototype.
	(gfc_resolve_re_runtime_error):  Likewise.
	Add prototype for gfc_is_reallocatable_lhs.
	* array.c (gfc_resolve_array_spec):  Do not resolve if it has
	already been resolved.
	* trans-array.h (gfc_is_reallocatable_lhs):  Remove prototype.
	* check.c (gfc_check_fe_runtime_error):  New function.
	* intrinsic.c (add_sym_1p):  New function.
	(make_vararg):  New function.
	(add_subroutines):  Add fe_runtime_error.
	(gfc_intrinsic_sub_interface): Skip sorting for variable number
	of arguments.
	* iresolve.c (gfc_resolve_fe_runtime_error):  New function.
	* lang.opt (inline-matmul-limit):  New option.
	(gfc_post_options): If no inline matmul limit has been set and
	BLAS is called externally, use the BLAS limit.
	* frontend-passes.c:  Include intrinsic.h.
	(var_num):  New global counter for naming temporary variablbles.
	(matrix_case):  Enum for differentiating the different matmul
	cases.
	(realloc_string_callback):  Add "trim" to the variable name.
	(create_var): Add optional argument vname as part of the name.
	Use var_num. Set dimension of result correctly. Split off block
	creation into
	(insert_block): New function.
	(cfe_expr_0): Use "fcn" as part of temporary variable name.
	(optimize_namesapce): Also set gfc_current_ns. Call
	inline_matmul_assign.
	(combine_array_constructor):  Use "constr" as part of
	temporary name.
	(get_array_inq_function):  New function.
	(build_logical_expr):  New function.
	(get_operand):  new function.
	(inline_limit_check):  New function.
	(runtime_error_ne):  New function.
	(matmul_lhs_realloc):  New function.
	(is_functino_or_op):  New function.
	(has_function_or_op):  New function.
	(freeze_expr):  New function.
	(freeze_references):  New function.
	(convert_to_index_kind):  New function.
	(create_do_loop):  New function.
	(get_size_m1):  New function.
	(scalarized_expr):  New function.
	(inline_matmul_assign):  New function.
	* simplify.c (simplify_bound):  Simplify the case of the
	lower bound of an assumed-shape argument.

2015-05-05  Thomas Koenig  <tkoenig@gcc.gnu.org>

	PR fortran/37131
	* gfortran.dg/dependency_26.f90: Add option to suppress inlining
	matmul.
	* gfortran.dg/function_optimize_1.f90:  Likewise.
	* gfortran.dg/function_optimize_2.f90:  Likewise.
	* gfortran.dg/function_optimize_5.f90:  Likewise.
	* gfortran.dg/function_optimize_7.f90:  Likewise.
	* gfortran.dg/inline_matmul_1.f90:  New test.
	* gfortran.dg/inline_matmul_2.f90:  New test.
	* gfortran.dg/inline_matmul_3.f90:  New test.
	* gfortran.dg/inline_matmul_4.f90:  New test.
	* gfortran.dg/inline_matmul_5.f90:  New test.

Comments

Mikael Morin May 5, 2015, 12:58 p.m. UTC | #1
Le 05/05/2015 08:25, Thomas Koenig a écrit :
> Hello world,
> 
> this is an update of the matmul inline patch.  The only difference to
> the last version is that it has the ubound simplification taken out.

Sorry, I delayed this, but it wasn't (yet) forgotten.

Below a few answers to
https://gcc.gnu.org/ml/gcc-patches/2015-04/msg01247.html
and documentation fixes.


> 	* simplify.c (simplify_bound):  Simplify the case of the
> 	lower bound of an assumed-shape argument.
Entry to be removed. ;-)


>>> Index: fortran/array.c
>>> ===================================================================
>>> --- fortran/array.c	(Revision 222218)
>>> +++ fortran/array.c	(Arbeitskopie)
>>> @@ -338,6 +338,9 @@ gfc_resolve_array_spec (gfc_array_spec *as, int ch
>>>    if (as == NULL)
>>>      return true;
>>>  
>>> +  if (as->resolved)
>>> +    return true;
>>> +
>> Why this?
> 
> Because you get regressions otherwise.  Not resolving an array spec
> twice should do no harm, and resolving it twice does so - I hit the
> error message in check_restricted.  I'm not sure what is wrong, maybe
> PR 23466 was not fully fixed, but this works.
> 
Hum, it seems to work without it here.  Can you double check?

>>> @@ -524,29 +542,11 @@ constant_string_length (gfc_expr *e)
>>>  
>>>  }
>>>  
>>> -/* Returns a new expression (a variable) to be used in place of the old one,
>>> -   with an assignment statement before the current statement to set
>>> -   the value of the variable. Creates a new BLOCK for the statement if
>>> -   that hasn't already been done and puts the statement, plus the
>>> -   newly created variables, in that block.  Special cases:  If the
>>> -   expression is constant or a temporary which has already
>>> -   been created, just copy it.  */
>>> -
>>> -static gfc_expr*
>>> -create_var (gfc_expr * e)
>> Keep a comment here.
> 
> Still exists, further down.
> 
I don't mind the comment being moved around together with create_var. :-)
I was asking for a comment for the new function insert_block.


> 
>>> +	      gfc_simplify_expr (ar->start[i], 0);
>>> +	    }
>>> +	  else if (was_fullref)
>>> +	    {
>>> +	      ar->dimen_type[i] = DIMEN_RANGE;
>>> +	      ar->start[i] = NULL;
>>> +	      ar->end[i] = NULL;
>>> +	      ar->stride[i] = NULL;
>>> +	    }
>> Is this reachable ?
> 
> Not in the current incarnation, I wanted to keep it around for
> a full segment later.  I can also remove this.
> 
At least add an assert, a comment, something telling that it's not used.
I would rather remove it until it's actually used, but I can live with
it, if it becomes used shortly.

>>
>> [...]
>>
>>> Index: fortran/options.c
>>> ===================================================================
>>> --- fortran/options.c	(Revision 222218)
>>> +++ fortran/options.c	(Arbeitskopie)
[...]
>>> +
>>> +  if (flag_external_blas && flag_inline_matmul_limit < 0)
>>> +    flag_inline_matmul_limit = flag_blas_matmul_limit;
>> Hum, shouldn't we do something for flag_inline_matmul_limit > 0 as well?
> 
> 
> This is done automatically, by the options machinery.  That is cool 
> 
Huh? is it?
I was talking about this:
Using -fblas-matmul-limit=10, one gets inlining until 10.
Using -fblas-matmul-limit=10 -finline-matmul-limit=20 the inlining limit
is _increased_ to 20.

It is of course questionable for a user to specify a low limit for blas
being lower than the high limit for inlining, so that we have a
contradictory specification for the matrix sizes in between.
But I think it would make more sense to have blas take precedence, that
is have flag_inline_matmul_limit clamped to flag_blas_matmul_limit, even
for flag_inline_matmul_limit > 0.
Is this done by the options machinery?

Either way, I don't mind too much, this is a corner case.


> Index: invoke.texi
> ===================================================================
> --- invoke.texi	(Revision 222218)
> +++ invoke.texi	(Arbeitskopie)
> @@ -1537,6 +1538,20 @@ geometric mean of the dimensions of the argument a
>  
>  The default value for @var{n} is 30.
>  
> +@item -finline-matmul-limit=@var{n}
> +@opindex @code{finline-matmul-limit}
> +When front-end optimiztion is active,
optimization
> some calls to the @code{MATMUL}
> +intrinsic function will be inlined.
s/some calls ... inlined/calls ... inlined for small matrix sizes/
or something like that.

>  Setting
> +@code{-finline-matmul-limit=0} will disable inlining in all cases.

> +Setting this option it to a specified value will call the library
> +routines for matrices with size larger than @var{n}. 
s/it //
Setting ... value @var{n} will produce inline code for matrix sizes up
to @var{n}; the library routines will be called for bigger matrices.

Maybe add something about code bloat here.  Suggestion:

This may result in code size increase if the matrix size can't be
determined at compile time, as code for both cases is generated.

> If the matrices
> +involved are not square, the size comparison is performed using the
> +geometric mean of the dimensions of the argument and result matrices.
> +
> +The default value for @var{n} is the value specified for
> +@code{-fblas-matmul-limit} if this option is specified, or unlimitited
unlimitted
> +otherwise.
> +
>  @item -frecursive
>  @opindex @code{frecursive}
>  Allow indirect recursion by forcing all local arrays to be allocated



With the resolved thing removed, the comment before insert_block and the
doc fixes, the patch is OK.
The rest is up to you.  Thanks.

Mikael
H.J. Lu May 8, 2015, 4:53 p.m. UTC | #2
On Mon, May 4, 2015 at 11:25 PM, Thomas Koenig <tkoenig@netcologne.de> wrote:
> Hello world,
>
> this is an update of the matmul inline patch.  The only difference to
> the last version is that it has the ubound simplification taken out.
>
> Any further comments?  OK for trunk?
>
>         Thomas
>
> 2015-05-05  Thomas Koenig  <tkoenig@gcc.gnu.org>
>
>         PR fortran/37131
>         * gfortran.h (gfc_isym_id):  Add GFC_ISYM_FE_RUNTIME_ERROR.
>         (gfc_array_spec):  Add resolved flag.
>         (gfc_intrinsic_sym):  Add vararg.
>         * intrinsic.h (gfc_check_fe_runtime_error):  Add prototype.
>         (gfc_resolve_re_runtime_error):  Likewise.
>         Add prototype for gfc_is_reallocatable_lhs.
>         * array.c (gfc_resolve_array_spec):  Do not resolve if it has
>         already been resolved.
>         * trans-array.h (gfc_is_reallocatable_lhs):  Remove prototype.
>         * check.c (gfc_check_fe_runtime_error):  New function.
>         * intrinsic.c (add_sym_1p):  New function.
>         (make_vararg):  New function.
>         (add_subroutines):  Add fe_runtime_error.
>         (gfc_intrinsic_sub_interface): Skip sorting for variable number
>         of arguments.
>         * iresolve.c (gfc_resolve_fe_runtime_error):  New function.
>         * lang.opt (inline-matmul-limit):  New option.
>         (gfc_post_options): If no inline matmul limit has been set and
>         BLAS is called externally, use the BLAS limit.
>         * frontend-passes.c:  Include intrinsic.h.
>         (var_num):  New global counter for naming temporary variablbles.
>         (matrix_case):  Enum for differentiating the different matmul
>         cases.
>         (realloc_string_callback):  Add "trim" to the variable name.
>         (create_var): Add optional argument vname as part of the name.
>         Use var_num. Set dimension of result correctly. Split off block
>         creation into
>         (insert_block): New function.
>         (cfe_expr_0): Use "fcn" as part of temporary variable name.
>         (optimize_namesapce): Also set gfc_current_ns. Call
>         inline_matmul_assign.
>         (combine_array_constructor):  Use "constr" as part of
>         temporary name.
>         (get_array_inq_function):  New function.
>         (build_logical_expr):  New function.
>         (get_operand):  new function.
>         (inline_limit_check):  New function.
>         (runtime_error_ne):  New function.
>         (matmul_lhs_realloc):  New function.
>         (is_functino_or_op):  New function.
>         (has_function_or_op):  New function.
>         (freeze_expr):  New function.
>         (freeze_references):  New function.
>         (convert_to_index_kind):  New function.
>         (create_do_loop):  New function.
>         (get_size_m1):  New function.
>         (scalarized_expr):  New function.
>         (inline_matmul_assign):  New function.
>         * simplify.c (simplify_bound):  Simplify the case of the
>         lower bound of an assumed-shape argument.
>

This caused:

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66073
diff mbox

Patch

Index: fortran/array.c
===================================================================
--- fortran/array.c	(Revision 222771)
+++ fortran/array.c	(Arbeitskopie)
@@ -338,6 +338,9 @@  gfc_resolve_array_spec (gfc_array_spec *as, int ch
   if (as == NULL)
     return true;
 
+  if (as->resolved)
+    return true;
+
   for (i = 0; i < as->rank + as->corank; i++)
     {
       e = as->lower[i];
@@ -364,6 +367,8 @@  gfc_resolve_array_spec (gfc_array_spec *as, int ch
 	}
     }
 
+  as->resolved = true;
+
   return true;
 }
 
Index: fortran/check.c
===================================================================
--- fortran/check.c	(Revision 222771)
+++ fortran/check.c	(Arbeitskopie)
@@ -5527,7 +5527,37 @@  gfc_check_random_seed (gfc_expr *size, gfc_expr *p
   return true;
 }
 
+bool
+gfc_check_fe_runtime_error (gfc_actual_arglist *a)
+{
+  gfc_expr *e;
+  int len, i;
+  int num_percent, nargs;
 
+  e = a->expr;
+  if (e->expr_type != EXPR_CONSTANT)
+    return true;
+
+  len = e->value.character.length;
+  if (e->value.character.string[len-1] != '\0')
+    gfc_internal_error ("fe_runtime_error string must be null terminated");
+
+  num_percent = 0;
+  for (i=0; i<len-1; i++)
+    if (e->value.character.string[i] == '%')
+      num_percent ++;
+
+  nargs = 0;
+  for (; a; a = a->next)
+    nargs ++;
+
+  if (nargs -1 != num_percent)
+    gfc_internal_error ("fe_runtime_error: Wrong number of arguments (%d instead of %d)",
+			nargs, num_percent++);
+
+  return true;
+}
+
 bool
 gfc_check_second_sub (gfc_expr *time)
 {
Index: fortran/frontend-passes.c
===================================================================
--- fortran/frontend-passes.c	(Revision 222771)
+++ fortran/frontend-passes.c	(Arbeitskopie)
@@ -27,6 +27,7 @@  along with GCC; see the file COPYING3.  If not see
 #include "dependency.h"
 #include "constructor.h"
 #include "opts.h"
+#include "intrinsic.h"
 
 /* Forward declarations.  */
 
@@ -43,7 +44,11 @@  static void doloop_warn (gfc_namespace *);
 static void optimize_reduction (gfc_namespace *);
 static int callback_reduction (gfc_expr **, int *, void *);
 static void realloc_strings (gfc_namespace *);
-static gfc_expr *create_var (gfc_expr *);
+static gfc_expr *create_var (gfc_expr *, const char *vname=NULL);
+static int inline_matmul_assign (gfc_code **, int *, void *);
+static gfc_code * create_do_loop (gfc_expr *, gfc_expr *, gfc_expr *,
+				  locus *, gfc_namespace *, 
+				  char *vname=NULL);
 
 /* How deep we are inside an argument list.  */
 
@@ -93,6 +98,19 @@  struct my_struct *evec;
 
 static bool in_assoc_list;
 
+/* Counter for temporary variables.  */
+
+static int var_num = 1;
+
+/* What sort of matrix we are dealing with when inlining MATMUL.  */
+
+enum matrix_case { none=0, A2B2, A2B1, A1B2 };
+
+/* Keep track of the number of expressions we have inserted so far 
+   using create_var.  */
+
+int n_vars;
+
 /* Entry point - run all passes for a namespace.  */
 
 void
@@ -157,7 +175,7 @@  realloc_string_callback (gfc_code **c, int *walk_s
     return 0;
   
   current_code = c;
-  n = create_var (expr2);
+  n = create_var (expr2, "trim");
   co->expr2 = n;
   return 0;
 }
@@ -524,29 +542,11 @@  constant_string_length (gfc_expr *e)
 
 }
 
-/* Returns a new expression (a variable) to be used in place of the old one,
-   with an assignment statement before the current statement to set
-   the value of the variable. Creates a new BLOCK for the statement if
-   that hasn't already been done and puts the statement, plus the
-   newly created variables, in that block.  Special cases:  If the
-   expression is constant or a temporary which has already
-   been created, just copy it.  */
-
-static gfc_expr*
-create_var (gfc_expr * e)
+static gfc_namespace*
+insert_block ()
 {
-  char name[GFC_MAX_SYMBOL_LEN +1];
-  static int num = 1;
-  gfc_symtree *symtree;
-  gfc_symbol *symbol;
-  gfc_expr *result;
-  gfc_code *n;
   gfc_namespace *ns;
-  int i;
 
-  if (e->expr_type == EXPR_CONSTANT || is_fe_temp (e))
-    return gfc_copy_expr (e);
-
   /* If the block hasn't already been created, do so.  */
   if (inserted_block == NULL)
     {
@@ -578,7 +578,37 @@  constant_string_length (gfc_expr *e)
   else
     ns = inserted_block->ext.block.ns;
 
-  sprintf(name, "__var_%d",num++);
+  return ns;
+}
+
+/* Returns a new expression (a variable) to be used in place of the old one,
+   with an optional assignment statement before the current statement to set
+   the value of the variable. Creates a new BLOCK for the statement if that
+   hasn't already been done and puts the statement, plus the newly created
+   variables, in that block.  Special cases: If the expression is constant or
+   a temporary which has already been created, just copy it.  */
+
+static gfc_expr*
+create_var (gfc_expr * e, const char *vname)
+{
+  char name[GFC_MAX_SYMBOL_LEN +1];
+  gfc_symtree *symtree;
+  gfc_symbol *symbol;
+  gfc_expr *result;
+  gfc_code *n;
+  gfc_namespace *ns;
+  int i;
+
+  if (e->expr_type == EXPR_CONSTANT || is_fe_temp (e))
+    return gfc_copy_expr (e);
+
+  ns = insert_block ();
+
+  if (vname)
+    snprintf (name, GFC_MAX_SYMBOL_LEN, "__var_%d_%s", var_num++, vname);
+  else
+    snprintf (name, GFC_MAX_SYMBOL_LEN, "__var_%d", var_num++);
+
   if (gfc_get_sym_tree (name, ns, &symtree, false) != 0)
     gcc_unreachable ();
 
@@ -651,6 +681,7 @@  constant_string_length (gfc_expr *e)
       result->ref->type = REF_ARRAY;
       result->ref->u.ar.type = AR_FULL;
       result->ref->u.ar.where = e->where;
+      result->ref->u.ar.dimen = e->rank;
       result->ref->u.ar.as = symbol->ts.type == BT_CLASS
 			     ? CLASS_DATA (symbol)->as : symbol->as;
       if (warn_array_temporaries)
@@ -666,6 +697,7 @@  constant_string_length (gfc_expr *e)
   n->expr1 = gfc_copy_expr (result);
   n->expr2 = e;
   *changed_statement = n;
+  n_vars ++;
 
   return result;
 }
@@ -724,7 +756,7 @@  cfe_expr_0 (gfc_expr **e, int *walk_subtrees,
 	  if (gfc_dep_compare_functions (*ei, *ej, true) == 0)
 	    {
 	      if (newvar == NULL)
-		newvar = create_var (*ei);
+		newvar = create_var (*ei, "fcn");
 
 	      if (warn_function_elimination)
 		do_warn_function_elimination (*ej);
@@ -931,13 +963,15 @@  convert_elseif (gfc_code **c, int *walk_subtrees A
   /*  Don't walk subtrees.  */
   return 0;
 }
+
 /* Optimize a namespace, including all contained namespaces.  */
 
 static void
 optimize_namespace (gfc_namespace *ns)
 {
-
+  gfc_namespace *saved_ns = gfc_current_ns;
   current_ns = ns;
+  gfc_current_ns = ns;
   forall_level = 0;
   iterator_level = 0;
   in_assoc_list = false;
@@ -947,6 +981,9 @@  optimize_namespace (gfc_namespace *ns)
   gfc_code_walker (&ns->code, convert_elseif, dummy_expr_callback, NULL);
   gfc_code_walker (&ns->code, cfe_code, cfe_expr_0, NULL);
   gfc_code_walker (&ns->code, optimize_code, optimize_expr, NULL);
+  if (flag_inline_matmul_limit != 0)
+    gfc_code_walker (&ns->code, inline_matmul_assign, dummy_expr_callback,
+		     NULL);
 
   /* BLOCKs are handled in the expression walker below.  */
   for (ns = ns->contained; ns; ns = ns->sibling)
@@ -954,6 +991,7 @@  optimize_namespace (gfc_namespace *ns)
       if (ns->code == NULL || ns->code->op != EXEC_BLOCK)
 	optimize_namespace (ns);
     }
+  gfc_current_ns = saved_ns;
 }
 
 /* Handle dependencies for allocatable strings which potentially redefine
@@ -968,10 +1006,7 @@  realloc_strings (gfc_namespace *ns)
   for (ns = ns->contained; ns; ns = ns->sibling)
     {
       if (ns->code == NULL || ns->code->op != EXEC_BLOCK)
-	{
-	  // current_ns = ns;
-	  realloc_strings (ns);
-	}
+	realloc_strings (ns);
     }
 
 }
@@ -1222,7 +1257,7 @@  combine_array_constructor (gfc_expr *e)
   if (op2->ts.type == BT_CHARACTER)
     return false;
 
-  scalar = create_var (gfc_copy_expr (op2));
+  scalar = create_var (gfc_copy_expr (op2), "constr");
 
   oldbase = op1->value.constructor;
   newbase = NULL;
@@ -1939,7 +1974,1049 @@  doloop_warn (gfc_namespace *ns)
   gfc_code_walker (&ns->code, doloop_code, do_function, NULL);
 }
 
+/* This selction deals with inlining calls to MATMUL.  */
 
+/* Auxiliary function to build and simplify an array inquiry function.
+   dim is zero-based.  */
+
+static gfc_expr *
+get_array_inq_function (gfc_isym_id id, gfc_expr *e, int dim)
+{
+  gfc_expr *fcn;
+  gfc_expr *dim_arg, *kind;
+  const char *name;
+  gfc_expr *ec;
+
+  switch (id)
+    {
+    case GFC_ISYM_LBOUND:
+      name = "_gfortran_lbound";
+      break;
+
+    case GFC_ISYM_UBOUND:
+      name = "_gfortran_ubound";
+      break;
+
+    case GFC_ISYM_SIZE:
+      name = "_gfortran_size";
+      break;
+
+    default:
+      gcc_unreachable ();
+    }
+
+  dim_arg =  gfc_get_int_expr (gfc_default_integer_kind, &e->where, dim);
+  kind = gfc_get_int_expr (gfc_default_integer_kind, &e->where,
+			   gfc_index_integer_kind);
+
+  ec = gfc_copy_expr (e);
+  fcn = gfc_build_intrinsic_call (current_ns, id, name, e->where, 3,
+				  ec, dim_arg,  kind);
+  gfc_simplify_expr (fcn, 0);
+  return fcn;
+}
+
+/* Builds a logical expression.  */
+
+static gfc_expr*
+build_logical_expr (gfc_intrinsic_op op, gfc_expr *e1, gfc_expr *e2)
+{
+  gfc_typespec ts;
+  gfc_expr *res;
+
+  ts.type = BT_LOGICAL;
+  ts.kind = gfc_default_logical_kind;
+  res = gfc_get_expr ();
+  res->where = e1->where;
+  res->expr_type = EXPR_OP;
+  res->value.op.op = op;
+  res->value.op.op1 = e1;
+  res->value.op.op2 = e2;
+  res->ts = ts;
+
+  return res;
+}
+
+
+/* Return an operation of one two gfc_expr (one if e2 is NULL). This assumes
+   compatible typespecs.  */
+
+static gfc_expr *
+get_operand (gfc_intrinsic_op op, gfc_expr *e1, gfc_expr *e2)
+{
+  gfc_expr *res;
+
+  res = gfc_get_expr ();
+  res->ts = e1->ts;
+  res->where = e1->where;
+  res->expr_type = EXPR_OP;
+  res->value.op.op = op;
+  res->value.op.op1 = e1;
+  res->value.op.op2 = e2;
+  gfc_simplify_expr (res, 0);
+  return res;
+}
+
+/* Generate the IF statement for a runtime check if we want to do inlining or
+   not - putting in the code for both branches and putting it into the syntax
+   tree is the caller's responsibility.  For fixed array sizes, this should be
+   removed by DCE. Only called for rank-two matrices A and B.  */
+
+static gfc_code *
+inline_limit_check (gfc_expr *a, gfc_expr *b, enum matrix_case m_case)
+{
+  gfc_expr *inline_limit;
+  gfc_code *if_1, *if_2, *else_2;
+  gfc_expr *b2, *a2, *a1, *m1, *m2;
+  gfc_typespec ts;
+  gfc_expr *cond;
+
+  gcc_assert (m_case == A2B2);
+
+  /* Calculation is done in real to avoid integer overflow.  */
+
+  inline_limit = gfc_get_constant_expr (BT_REAL, gfc_default_real_kind,
+					&a->where);
+  mpfr_set_si (inline_limit->value.real, flag_inline_matmul_limit,
+	       GFC_RND_MODE);
+  mpfr_pow_ui (inline_limit->value.real, inline_limit->value.real, 3,
+	       GFC_RND_MODE);
+
+  a1 = get_array_inq_function (GFC_ISYM_SIZE, a, 1);
+  a2 = get_array_inq_function (GFC_ISYM_SIZE, a, 2);
+  b2 = get_array_inq_function (GFC_ISYM_SIZE, b, 2);
+
+  gfc_clear_ts (&ts);
+  ts.type = BT_REAL;
+  ts.kind = gfc_default_real_kind;
+  gfc_convert_type_warn (a1, &ts, 2, 0);
+  gfc_convert_type_warn (a2, &ts, 2, 0);
+  gfc_convert_type_warn (b2, &ts, 2, 0);
+
+  m1 = get_operand (INTRINSIC_TIMES, a1, a2);
+  m2 = get_operand (INTRINSIC_TIMES, m1, b2);
+
+  cond = build_logical_expr (INTRINSIC_LE, m2, inline_limit);
+  gfc_simplify_expr (cond, 0);
+
+  else_2 = XCNEW (gfc_code);
+  else_2->op = EXEC_IF;
+  else_2->loc = a->where;
+
+  if_2 = XCNEW (gfc_code);
+  if_2->op = EXEC_IF;
+  if_2->expr1 = cond;
+  if_2->loc = a->where;
+  if_2->block = else_2;
+
+  if_1 = XCNEW (gfc_code);
+  if_1->op = EXEC_IF;
+  if_1->block = if_2;
+  if_1->loc = a->where;
+
+  return if_1;
+}
+
+
+/* Insert code to issue a runtime error if the expressions are not equal.  */
+
+static gfc_code *
+runtime_error_ne (gfc_expr *e1, gfc_expr *e2, const char *msg)
+{
+  gfc_expr *cond;
+  gfc_code *if_1, *if_2;
+  gfc_code *c;
+  gfc_actual_arglist *a1, *a2, *a3;
+
+  gcc_assert (e1->where.lb);
+  /* Build the call to runtime_error.  */
+  c = XCNEW (gfc_code);
+  c->op = EXEC_CALL;
+  c->loc = e1->where;
+
+  /* Get a null-terminated message string.  */
+
+  a1 = gfc_get_actual_arglist ();
+  a1->expr = gfc_get_character_expr (gfc_default_character_kind, &e1->where,
+				     msg, strlen(msg)+1);
+  c->ext.actual = a1;
+
+  /* Pass the value of the first expression.  */
+  a2 = gfc_get_actual_arglist ();
+  a2->expr = gfc_copy_expr (e1);
+  a1->next = a2;
+
+  /* Pass the value of the second expression.  */
+  a3 = gfc_get_actual_arglist ();
+  a3->expr = gfc_copy_expr (e2);
+  a2->next = a3;
+
+  gfc_check_fe_runtime_error (c->ext.actual);
+  gfc_resolve_fe_runtime_error (c);
+
+  if_2 = XCNEW (gfc_code);
+  if_2->op = EXEC_IF;
+  if_2->loc = e1->where;
+  if_2->next = c;
+
+  if_1 = XCNEW (gfc_code);
+  if_1->op = EXEC_IF;
+  if_1->block = if_2;
+  if_1->loc = e1->where;
+
+  cond = build_logical_expr (INTRINSIC_NE, e1, e2);
+  gfc_simplify_expr (cond, 0);
+  if_2->expr1 = cond;
+
+  return if_1;
+}
+
+/* Handle matrix reallocation.  Caller is responsible to insert into
+   the code tree.
+
+   For the two-dimensional case, build 
+
+  if (allocated(c)) then
+     if (size(c,1) /= size(a,1) .or. size(c,2) /= size(b,2)) then
+        deallocate(c)
+        allocate (c(size(a,1), size(b,2)))
+     end if
+  else
+     allocate (c(size(a,1),size(b,2)))
+  end if
+
+  and for the other cases correspondingly.
+*/
+
+static gfc_code *
+matmul_lhs_realloc (gfc_expr *c, gfc_expr *a, gfc_expr *b,
+		    enum matrix_case m_case)
+{
+
+  gfc_expr *allocated, *alloc_expr;
+  gfc_code *if_alloc_1, *if_alloc_2, *if_size_1, *if_size_2;
+  gfc_code *else_alloc;
+  gfc_code *deallocate, *allocate1, *allocate_else;
+  gfc_array_ref *ar;
+  gfc_expr *cond, *ne1, *ne2;
+
+  if (warn_realloc_lhs)
+    gfc_warning (OPT_Wrealloc_lhs,
+		 "Code for reallocating the allocatable array at %L will "
+		 "be added", &c->where);
+
+  alloc_expr = gfc_copy_expr (c);
+
+  ar = gfc_find_array_ref (alloc_expr);
+  gcc_assert (ar && ar->type == AR_FULL);
+
+  /* c comes in as a full ref.  Change it into a copy and make it into an
+     element ref so it has the right form for for ALLOCATE.  In the same
+     switch statement, also generate the size comparison for the secod IF
+     statement.  */
+
+  ar->type = AR_ELEMENT;
+
+  switch (m_case)
+    {
+    case A2B2:
+      ar->start[0] = get_array_inq_function (GFC_ISYM_SIZE, a, 1);
+      ar->start[1] = get_array_inq_function (GFC_ISYM_SIZE, b, 2);
+      ne1 = build_logical_expr (INTRINSIC_NE,
+				get_array_inq_function (GFC_ISYM_SIZE, c, 1),
+				get_array_inq_function (GFC_ISYM_SIZE, a, 1));
+      ne2 = build_logical_expr (INTRINSIC_NE,
+				get_array_inq_function (GFC_ISYM_SIZE, c, 2),
+				get_array_inq_function (GFC_ISYM_SIZE, b, 2));
+      cond = build_logical_expr (INTRINSIC_OR, ne1, ne2);
+      break;
+
+    case A2B1:
+      ar->start[0] = get_array_inq_function (GFC_ISYM_SIZE, a, 1);
+      cond = build_logical_expr (INTRINSIC_NE,
+				 get_array_inq_function (GFC_ISYM_SIZE, c, 1),
+				 get_array_inq_function (GFC_ISYM_SIZE, a, 2));
+      break;
+
+    case A1B2:
+      ar->start[0] = get_array_inq_function (GFC_ISYM_SIZE, b, 1);
+      cond = build_logical_expr (INTRINSIC_NE,
+				 get_array_inq_function (GFC_ISYM_SIZE, c, 1),
+				 get_array_inq_function (GFC_ISYM_SIZE, b, 2));
+      break;
+
+    default:
+      gcc_unreachable();
+
+    }
+
+  gfc_simplify_expr (cond, 0);
+
+  /* We need two identical allocate statements in two
+     branches of the IF statement.  */
+  
+  allocate1 = XCNEW (gfc_code);
+  allocate1->op = EXEC_ALLOCATE;
+  allocate1->ext.alloc.list = gfc_get_alloc ();
+  allocate1->loc = c->where;
+  allocate1->ext.alloc.list->expr = gfc_copy_expr (alloc_expr);
+
+  allocate_else = XCNEW (gfc_code);
+  allocate_else->op = EXEC_ALLOCATE;
+  allocate_else->ext.alloc.list = gfc_get_alloc ();
+  allocate_else->loc = c->where;
+  allocate_else->ext.alloc.list->expr = alloc_expr;
+
+  allocated = gfc_build_intrinsic_call (current_ns, GFC_ISYM_ALLOCATED,
+					"_gfortran_allocated", c->where,
+					1, gfc_copy_expr (c));
+
+  deallocate = XCNEW (gfc_code);
+  deallocate->op = EXEC_DEALLOCATE;
+  deallocate->ext.alloc.list = gfc_get_alloc ();
+  deallocate->ext.alloc.list->expr = gfc_copy_expr (c);
+  deallocate->next = allocate1;
+  deallocate->loc = c->where;
+  
+  if_size_2 = XCNEW (gfc_code);
+  if_size_2->op = EXEC_IF;
+  if_size_2->expr1 = cond;
+  if_size_2->loc = c->where;
+  if_size_2->next = deallocate;
+
+  if_size_1 = XCNEW (gfc_code);
+  if_size_1->op = EXEC_IF;
+  if_size_1->block = if_size_2;
+  if_size_1->loc = c->where;
+
+  else_alloc = XCNEW (gfc_code);
+  else_alloc->op = EXEC_IF;
+  else_alloc->loc = c->where;
+  else_alloc->next = allocate_else;
+
+  if_alloc_2 = XCNEW (gfc_code);
+  if_alloc_2->op = EXEC_IF;
+  if_alloc_2->expr1 = allocated;
+  if_alloc_2->loc = c->where;
+  if_alloc_2->next = if_size_1;
+  if_alloc_2->block = else_alloc;
+
+  if_alloc_1 = XCNEW (gfc_code);
+  if_alloc_1->op = EXEC_IF;
+  if_alloc_1->block = if_alloc_2;
+  if_alloc_1->loc = c->where;
+
+  return if_alloc_1;
+}
+
+/* Callback function for has_function_or_op.  */
+
+static int
+is_function_or_op (gfc_expr **e, int *walk_subtrees ATTRIBUTE_UNUSED,
+	     void *data ATTRIBUTE_UNUSED)
+{
+  if ((*e) == 0)
+    return 0;
+  else
+    return (*e)->expr_type == EXPR_FUNCTION
+      || (*e)->expr_type == EXPR_OP;
+}
+
+/* Returns true if the expression contains a function.  */
+
+static bool
+has_function_or_op (gfc_expr **e)
+{
+  if (e == NULL)
+    return false;
+  else
+    return gfc_expr_walker (e, is_function_or_op, NULL);
+}
+
+/* Freeze (assign to a temporary variable) a single expression.  */
+
+static void
+freeze_expr (gfc_expr **ep)
+{
+  gfc_expr *ne;
+  if (has_function_or_op (ep))
+    {
+      ne = create_var (*ep, "freeze");
+      *ep = ne;
+    }
+}
+
+/* Go through an expression's references and assign them to temporary
+   variables if they contain functions.  This is usually done prior to
+   front-end scalarization to avoid multiple invocations of functions.  */
+
+static void
+freeze_references (gfc_expr *e)
+{
+  gfc_ref *r;
+  gfc_array_ref *ar;
+  int i;
+
+  for (r=e->ref; r; r=r->next)
+    {
+      if (r->type == REF_SUBSTRING)
+	{
+	  if (r->u.ss.start != NULL)
+	    freeze_expr (&r->u.ss.start);
+
+	  if (r->u.ss.end != NULL)
+	    freeze_expr (&r->u.ss.end);
+	}
+      else if (r->type == REF_ARRAY)
+	{
+	  ar = &r->u.ar;
+	  switch (ar->type)
+	    {
+	    case AR_FULL:
+	      break;
+
+	    case AR_SECTION:
+	      for (i=0; i<ar->dimen; i++)
+		{
+		  if (ar->dimen_type[i] == DIMEN_RANGE)
+		    {
+		      freeze_expr (&ar->start[i]);
+		      freeze_expr (&ar->end[i]);
+		      freeze_expr (&ar->stride[i]);
+		    }
+		  else if (ar->dimen_type[i] == DIMEN_ELEMENT)
+		    {
+		      freeze_expr (&ar->start[i]);
+		    }
+		}
+	      break;
+
+	    case AR_ELEMENT:
+	      for (i=0; i<ar->dimen; i++)
+		freeze_expr (&ar->start[i]);
+	      break;
+
+	    default:
+	      break;
+	    }
+	}
+    }
+}
+
+/* Convert to gfc_index_integer_kind if needed, just do a copy otherwise.  */
+
+static gfc_expr *
+convert_to_index_kind (gfc_expr *e)
+{
+  gfc_expr *res;
+
+  gcc_assert (e != NULL);
+
+  res = gfc_copy_expr (e);
+
+  gcc_assert (e->ts.type == BT_INTEGER);
+
+  if (res->ts.kind != gfc_index_integer_kind)
+    {
+      gfc_typespec ts;
+      gfc_clear_ts (&ts);
+      ts.type = BT_INTEGER;
+      ts.kind = gfc_index_integer_kind;
+
+      gfc_convert_type_warn (e, &ts, 2, 0);
+    }
+
+  return res;
+}
+
+/* Function to create a DO loop including creation of the
+   iteration variable.  gfc_expr are copied.*/
+
+static gfc_code *
+create_do_loop (gfc_expr *start, gfc_expr *end, gfc_expr *step, locus *where,
+		gfc_namespace *ns, char *vname)
+{
+
+  char name[GFC_MAX_SYMBOL_LEN +1];
+  gfc_symtree *symtree;
+  gfc_symbol *symbol;
+  gfc_expr *i;
+  gfc_code *n, *n2;
+
+  /* Create an expression for the iteration variable.  */
+  if (vname)
+    sprintf (name, "__var_%d_do_%s", var_num++, vname);
+  else
+    sprintf (name, "__var_%d_do", var_num++);
+
+
+  if (gfc_get_sym_tree (name, ns, &symtree, false) != 0)
+    gcc_unreachable ();
+
+  /* Create the loop variable.  */
+
+  symbol = symtree->n.sym;
+  symbol->ts.type = BT_INTEGER;
+  symbol->ts.kind = gfc_index_integer_kind;
+  symbol->attr.flavor = FL_VARIABLE;
+  symbol->attr.referenced = 1;
+  symbol->attr.dimension = 0;
+  symbol->attr.fe_temp = 1;
+  gfc_commit_symbol (symbol);
+
+  i = gfc_get_expr ();
+  i->expr_type = EXPR_VARIABLE;
+  i->ts = symbol->ts;
+  i->rank = 0;
+  i->where = *where;
+  i->symtree = symtree;
+
+  /* ... and the nested DO statements.  */
+  n = XCNEW (gfc_code);
+  n->op = EXEC_DO;
+  n->loc = *where;
+  n->ext.iterator = gfc_get_iterator ();
+  n->ext.iterator->var = i;
+  n->ext.iterator->start = convert_to_index_kind (start);
+  n->ext.iterator->end = convert_to_index_kind (end);
+  if (step)
+    n->ext.iterator->step = convert_to_index_kind (step);
+  else
+    n->ext.iterator->step = gfc_get_int_expr (gfc_index_integer_kind,
+					      where, 1);
+
+  n2 = XCNEW (gfc_code);
+  n2->op = EXEC_DO;
+  n2->loc = *where;
+  n2->next = NULL;
+  n->block = n2;
+  return n;
+}
+
+/* Get the upper bound of the DO loops for matmul along a dimension.  This
+ is one-based.  */
+
+static gfc_expr*
+get_size_m1 (gfc_expr *e, int dimen)
+{
+  mpz_t size;
+  gfc_expr *res;
+
+  if (gfc_array_dimen_size (e, dimen - 1, &size))
+    {
+      res = gfc_get_constant_expr (BT_INTEGER,
+				   gfc_index_integer_kind, &e->where);
+      mpz_sub_ui (res->value.integer, size, 1);
+      mpz_clear (size);
+    }
+  else
+    {
+      res = get_operand (INTRINSIC_MINUS,
+			 get_array_inq_function (GFC_ISYM_SIZE, e, dimen),
+			 gfc_get_int_expr (gfc_index_integer_kind,
+					   &e->where, 1));
+      gfc_simplify_expr (res, 0);
+    }
+
+  return res;
+}
+
+/* Function to return a scalarized expression. It is assumed that indices are
+ zero based to make generation of DO loops easier.  A zero as index will
+ access the first element along a dimension.  Single element references will
+ be skipped.  A NULL as an expression will be replaced by a full reference.
+ This assumes that the index loops have gfc_index_integer_kind, and that all
+ references have been frozen.  */
+
+static gfc_expr*
+scalarized_expr (gfc_expr *e_in, gfc_expr **index, int count_index)
+{
+  gfc_array_ref *ar;
+  int i;
+  int rank;
+  gfc_expr *e;
+  int i_index;
+  bool was_fullref;
+
+  e = gfc_copy_expr(e_in);
+
+  rank = e->rank;
+
+  ar = gfc_find_array_ref (e);
+
+  /* We scalarize count_index variables, reducing the rank by count_index.  */
+
+  e->rank = rank - count_index;
+
+  was_fullref = ar->type == AR_FULL;
+
+  if (e->rank == 0)
+    ar->type = AR_ELEMENT;
+  else
+    ar->type = AR_SECTION;
+
+  /* Loop over the indices.  For each index, create the expression
+     index * stride + lbound(e, dim).  */
+  
+  i_index = 0;
+  for (i=0; i < ar->dimen; i++)
+    {
+      if (was_fullref || ar->dimen_type[i] == DIMEN_RANGE)
+	{
+	  if (index[i_index] != NULL)
+	    {
+	      gfc_expr *lbound, *nindex;
+	      gfc_expr *loopvar;
+	      
+	      loopvar = gfc_copy_expr (index[i_index]); 
+	      
+	      if (ar->stride[i])
+		{
+		  gfc_expr *tmp;
+
+		  tmp = gfc_copy_expr(ar->stride[i]);
+		  if (tmp->ts.kind != gfc_index_integer_kind)
+		    {
+		      gfc_typespec ts;
+		      gfc_clear_ts (&ts);
+		      ts.type = BT_INTEGER;
+		      ts.kind = gfc_index_integer_kind;
+		      gfc_convert_type (tmp, &ts, 2);
+		    }
+		  nindex = get_operand (INTRINSIC_TIMES, loopvar, tmp);
+		}
+	      else
+		nindex = loopvar;
+	      
+	      /* Calculate the lower bound of the expression.  */
+	      if (ar->start[i])
+		{
+		  lbound = gfc_copy_expr (ar->start[i]);
+		  if (lbound->ts.kind != gfc_index_integer_kind)
+		    {
+		      gfc_typespec ts;
+		      gfc_clear_ts (&ts);
+		      ts.type = BT_INTEGER;
+		      ts.kind = gfc_index_integer_kind;
+		      gfc_convert_type (lbound, &ts, 2);
+
+		    }
+		}
+	      else
+		{
+		  if (!was_fullref)
+		    {
+		      /* Look at full individual sections, like a(:).  The first index
+			 is the lbound of a full ref.  */
+
+		      gfc_array_ref *ar;
+
+		      ar = gfc_find_array_ref (e_in);
+		      ar->type = AR_FULL;
+		    }
+		  lbound = get_array_inq_function (GFC_ISYM_LBOUND, e_in,
+						   i_index + 1);
+		}
+	      
+	      ar->dimen_type[i] = DIMEN_ELEMENT;
+
+	      gfc_free_expr (ar->start[i]);
+	      ar->start[i] = get_operand (INTRINSIC_PLUS, nindex, lbound);
+	      
+	      gfc_free_expr (ar->end[i]);
+	      ar->end[i] = NULL;
+	      gfc_free_expr (ar->stride[i]);
+	      ar->stride[i] = NULL;
+	      gfc_simplify_expr (ar->start[i], 0);
+	    }
+	  else if (was_fullref)
+	    {
+	      ar->dimen_type[i] = DIMEN_RANGE;
+	      ar->start[i] = NULL;
+	      ar->end[i] = NULL;
+	      ar->stride[i] = NULL;
+	    }
+	  i_index ++;
+	}
+    }
+  return e;
+}
+
+
+/* Inline assignments of the form c = matmul(a,b).
+   Handle only the cases currently where b and c are rank-two arrays.
+
+   This basically translates the code to
+
+   BLOCK
+     integer i,j,k
+     c = 0
+     do j=0, size(b,2)-1
+       do k=0, size(a, 2)-1
+         do i=0, size(a, 1)-1
+            c(i * stride(c,1) + lbound(c,1), j * stride(c,2) + lbound(c,2)) =
+	    c(i * stride(c,1) + lbound(c,1), j * stride(c,2) + lbound(c,2)) +
+            a(i * stride(a,1) + lbound(a,1), k * stride(a,2) + lbound(a,2)) *
+            b(k * stride(b,1) + lbound(b,1), j * stride(b,2) + lbound(b,2))
+         end do
+       end do
+     end do
+   END BLOCK
+   
+*/
+
+static int
+inline_matmul_assign (gfc_code **c, int *walk_subtrees,
+			  void *data ATTRIBUTE_UNUSED)
+{
+  gfc_code *co = *c;
+  gfc_expr *expr1, *expr2;
+  gfc_expr *matrix_a, *matrix_b;
+  gfc_actual_arglist *a, *b;
+  gfc_code *do_1, *do_2, *do_3, *assign_zero, *assign_matmul;
+  gfc_expr *zero_e;
+  gfc_expr *u1, *u2, *u3;
+  gfc_expr *list[2];
+  gfc_expr *ascalar, *bscalar, *cscalar;
+  gfc_expr *mult;
+  gfc_expr *var_1, *var_2, *var_3;
+  gfc_expr *zero;
+  gfc_namespace *ns;
+  gfc_intrinsic_op op_times, op_plus;
+  enum matrix_case m_case;
+  int i;
+  gfc_code *if_limit = NULL;
+  gfc_code **next_code_point;
+
+  if (co->op != EXEC_ASSIGN)
+    return 0;
+
+  expr1 = co->expr1;
+  expr2 = co->expr2;
+  if (expr2->expr_type != EXPR_FUNCTION
+      || expr2->value.function.isym == NULL
+      || expr2->value.function.isym->id != GFC_ISYM_MATMUL)
+    return 0;
+
+  current_code = c;
+  inserted_block = NULL;
+  changed_statement = NULL;
+
+  a = expr2->value.function.actual;
+  matrix_a = a->expr;
+  b = a->next;
+  matrix_b = b->expr;
+
+  /* Currently only handling direct variables.  Transpose etc. will come
+     later.  */
+
+  if (matrix_a->expr_type != EXPR_VARIABLE
+      || matrix_b->expr_type != EXPR_VARIABLE)
+    return 0;
+
+  if (matrix_a->rank == 2)
+    m_case = matrix_b->rank == 1 ? A2B1 : A2B2;
+  else
+    m_case = A1B2;
+
+  /* We do not handle data dependencies yet.  */
+  if (gfc_check_dependency (expr1, matrix_a, true)
+      || gfc_check_dependency (expr1, matrix_b, true))
+    return 0;
+
+  ns = insert_block ();
+
+  /* Assign the type of the zero expression for initializing the resulting
+     array, and the expression (+ and * for real, integer and complex;
+     .and. and .or for logical.  */
+
+  switch(expr1->ts.type)
+    {
+    case BT_INTEGER:
+      zero_e = gfc_get_int_expr (expr1->ts.kind, &expr1->where, 0);
+      op_times = INTRINSIC_TIMES;
+      op_plus = INTRINSIC_PLUS;
+      break;
+
+    case BT_LOGICAL:
+      op_times = INTRINSIC_AND;
+      op_plus = INTRINSIC_OR;
+      zero_e = gfc_get_logical_expr (expr1->ts.kind, &expr1->where,
+				     0);
+      break;
+    case BT_REAL:
+      zero_e = gfc_get_constant_expr (BT_REAL, expr1->ts.kind,
+				      &expr1->where);
+      mpfr_set_si (zero_e->value.real, 0, GFC_RND_MODE);
+      op_times = INTRINSIC_TIMES;
+      op_plus = INTRINSIC_PLUS;
+      break;
+
+    case BT_COMPLEX:
+      zero_e = gfc_get_constant_expr (BT_COMPLEX, expr1->ts.kind,
+				      &expr1->where);
+      mpc_set_si_si (zero_e->value.complex, 0, 0, GFC_RND_MODE);
+      op_times = INTRINSIC_TIMES;
+      op_plus = INTRINSIC_PLUS;
+
+      break;
+
+    default:
+      gcc_unreachable();
+    }
+
+  current_code = &ns->code;
+
+  /* Freeze the references, keeping track of how many temporary variables were
+     created.  */
+  n_vars = 0;
+  freeze_references (matrix_a);
+  freeze_references (matrix_b);
+  freeze_references (expr1);
+
+  if (n_vars == 0)
+    next_code_point = current_code;
+  else
+    {
+      next_code_point = &ns->code;
+      for (i=0; i<n_vars; i++)
+	next_code_point = &(*next_code_point)->next;
+    }
+
+  /* Take care of the inline flag.  If the limit check evaluates to a
+     constant, dead code elimination will eliminate the unneeded branch.  */
+
+  if (m_case == A2B2 && flag_inline_matmul_limit > 0)
+    {
+      if_limit = inline_limit_check (matrix_a, matrix_b, m_case);
+
+      /* Insert the original statement into the else branch.  */
+      if_limit->block->block->next = co;
+      co->next = NULL;
+
+      /* ... and the new ones go into the original one.  */
+      *next_code_point = if_limit;
+      next_code_point = &if_limit->block->next;
+    }
+
+  assign_zero = XCNEW (gfc_code);
+  assign_zero->op = EXEC_ASSIGN;
+  assign_zero->loc = co->loc;
+  assign_zero->expr1 = gfc_copy_expr (expr1);
+  assign_zero->expr2 = zero_e;
+
+  /* Handle the reallocation, if needed.  */
+  if (flag_realloc_lhs && gfc_is_reallocatable_lhs (expr1))
+    {
+      gfc_code *lhs_alloc;
+
+      /* Only need to check a single dimension for the A2B2 case for
+	 bounds checking, the rest will be allocated.  */
+
+      if (gfc_option.rtcheck & GFC_RTCHECK_BOUNDS && m_case == A2B2)
+	{
+	  gfc_code *test;
+	  gfc_expr *a2, *b1;
+
+	  a2 = get_array_inq_function (GFC_ISYM_SIZE, matrix_a, 2);
+	  b1 = get_array_inq_function (GFC_ISYM_SIZE, matrix_b, 1);
+	  test = runtime_error_ne (b1, a2, "Dimension of array B incorrect "
+				   "in MATMUL intrinsic: Is %ld, should be %ld");
+	  *next_code_point = test;
+	  next_code_point = &test->next;
+	}
+
+
+      lhs_alloc = matmul_lhs_realloc (expr1, matrix_a, matrix_b, m_case);
+
+      *next_code_point = lhs_alloc;
+      next_code_point = &lhs_alloc->next;
+
+    }
+  else if (gfc_option.rtcheck & GFC_RTCHECK_BOUNDS)
+    {
+      gfc_code *test;
+      gfc_expr *a2, *b1, *c1, *c2, *a1, *b2;
+
+      if (m_case == A2B2 || m_case == A2B1)
+	{
+	  a2 = get_array_inq_function (GFC_ISYM_SIZE, matrix_a, 2);
+	  b1 = get_array_inq_function (GFC_ISYM_SIZE, matrix_b, 1);
+	  test = runtime_error_ne (b1, a2, "Dimension of array B incorrect "
+				   "in MATMUL intrinsic: Is %ld, should be %ld");
+	  *next_code_point = test;
+	  next_code_point = &test->next;
+
+	  c1 = get_array_inq_function (GFC_ISYM_SIZE, expr1, 1);
+	  a1 = get_array_inq_function (GFC_ISYM_SIZE, matrix_a, 1);
+
+	  if (m_case == A2B2)
+	    test = runtime_error_ne (c1, a1, "Incorrect extent in return array in "
+				     "MATMUL intrinsic for dimension 1: "
+				     "is %ld, should be %ld");
+	  else if (m_case == A2B1)
+	    test = runtime_error_ne (c1, a1, "Incorrect extent in return array in "
+				     "MATMUL intrinsic: "
+				     "is %ld, should be %ld");
+
+
+	  *next_code_point = test;
+	  next_code_point = &test->next;
+	}
+      else if (m_case == A1B2)
+	{
+	  a1 = get_array_inq_function (GFC_ISYM_SIZE, matrix_a, 1);
+	  b1 = get_array_inq_function (GFC_ISYM_SIZE, matrix_b, 1);
+	  test = runtime_error_ne (b1, a1, "Dimension of array B incorrect "
+				   "in MATMUL intrinsic: Is %ld, should be %ld");
+	  *next_code_point = test;
+	  next_code_point = &test->next;
+
+	  c1 = get_array_inq_function (GFC_ISYM_SIZE, expr1, 1);
+	  b2 = get_array_inq_function (GFC_ISYM_SIZE, matrix_b, 2);
+
+	  test = runtime_error_ne (c1, b2, "Incorrect extent in return array in "
+				   "MATMUL intrinsic: "
+				   "is %ld, should be %ld");
+
+	  *next_code_point = test;
+	  next_code_point = &test->next;
+	}
+
+      if (m_case == A2B2)
+	{
+	  c2 = get_array_inq_function (GFC_ISYM_SIZE, expr1, 2);
+	  b2 = get_array_inq_function (GFC_ISYM_SIZE, matrix_b, 2);
+	  test = runtime_error_ne (c2, b2, "Incorrect extent in return array in "
+				   "MATMUL intrinsic for dimension 2: is %ld, should be %ld");
+
+	  *next_code_point = test;
+	  next_code_point = &test->next;
+	}
+    }
+
+  *next_code_point = assign_zero;
+
+  zero = gfc_get_int_expr (gfc_index_integer_kind, &co->loc, 0);
+
+  assign_matmul = XCNEW (gfc_code);
+  assign_matmul->op = EXEC_ASSIGN;
+  assign_matmul->loc = co->loc;
+
+  /* Get the bounds for the loops, create them and create the scalarized
+     expressions.  */
+
+  switch (m_case)
+    {
+    case A2B2:
+      inline_limit_check (matrix_a, matrix_b, m_case);
+
+      u1 = get_size_m1 (matrix_b, 2);
+      u2 = get_size_m1 (matrix_a, 2);
+      u3 = get_size_m1 (matrix_a, 1);
+
+      do_1 = create_do_loop (gfc_copy_expr (zero), u1, NULL, &co->loc, ns);
+      do_2 = create_do_loop (gfc_copy_expr (zero), u2, NULL, &co->loc, ns);
+      do_3 = create_do_loop (gfc_copy_expr (zero), u3, NULL, &co->loc, ns);
+
+      do_1->block->next = do_2;
+      do_2->block->next = do_3;
+      do_3->block->next = assign_matmul;
+
+      var_1 = do_1->ext.iterator->var;
+      var_2 = do_2->ext.iterator->var;
+      var_3 = do_3->ext.iterator->var;
+
+      list[0] = var_3;
+      list[1] = var_1;
+      cscalar = scalarized_expr (gfc_copy_expr (co->expr1), list, 2);
+
+      list[0] = var_3;
+      list[1] = var_2;
+      ascalar = scalarized_expr (gfc_copy_expr (matrix_a), list, 2);
+
+      list[0] = var_2;
+      list[1] = var_1;
+      bscalar = scalarized_expr (gfc_copy_expr (matrix_b), list, 2);
+
+      break;
+
+    case A2B1:
+      u1 = get_size_m1 (matrix_b, 1);
+      u2 = get_size_m1 (matrix_a, 1);
+
+      do_1 = create_do_loop (gfc_copy_expr (zero), u1, NULL, &co->loc, ns);
+      do_2 = create_do_loop (gfc_copy_expr (zero), u2, NULL, &co->loc, ns);
+
+      do_1->block->next = do_2;
+      do_2->block->next = assign_matmul;
+
+      var_1 = do_1->ext.iterator->var;
+      var_2 = do_2->ext.iterator->var;
+
+      list[0] = var_2;
+      cscalar = scalarized_expr (gfc_copy_expr (co->expr1), list, 1);
+
+      list[0] = var_2;
+      list[1] = var_1;
+      ascalar = scalarized_expr (gfc_copy_expr (matrix_a), list, 2);
+
+      list[0] = var_1;
+      bscalar = scalarized_expr (gfc_copy_expr (matrix_b), list, 1);
+
+      break;
+
+    case A1B2:
+      u1 = get_size_m1 (matrix_b, 2);
+      u2 = get_size_m1 (matrix_a, 1);
+
+      do_1 = create_do_loop (gfc_copy_expr (zero), u1, NULL, &co->loc, ns);
+      do_2 = create_do_loop (gfc_copy_expr (zero), u2, NULL, &co->loc, ns);
+
+      do_1->block->next = do_2;
+      do_2->block->next = assign_matmul;
+
+      var_1 = do_1->ext.iterator->var;
+      var_2 = do_2->ext.iterator->var;
+
+      list[0] = var_1;
+      cscalar = scalarized_expr (gfc_copy_expr (co->expr1), list, 1);
+
+      list[0] = var_2;
+      ascalar = scalarized_expr (gfc_copy_expr (matrix_a), list, 1);
+
+      list[0] = var_2;
+      list[1] = var_1;
+      bscalar = scalarized_expr (gfc_copy_expr (matrix_b), list, 2);
+
+      break;
+
+    default:
+      gcc_unreachable();
+    }
+
+  /* First loop comes after the zero assignment.  */
+  assign_zero->next = do_1;
+
+  /* Build the assignment expression in the loop.  */
+  assign_matmul->expr1 = gfc_copy_expr (cscalar);
+
+  mult = get_operand (op_times, ascalar, bscalar);
+  assign_matmul->expr2 = get_operand (op_plus, cscalar, mult);
+
+  /* If we don't want to keep the original statement around in
+     the else branch, we can free it.  */
+
+  if (if_limit == NULL)
+    gfc_free_statements(co);
+  else
+    co->next = NULL;
+
+  gfc_free_expr (zero);
+  *walk_subtrees = 0;
+  return 0;
+}
+
 #define WALK_SUBEXPR(NODE) \
   do							\
     {							\
Index: fortran/gfortran.h
===================================================================
--- fortran/gfortran.h	(Revision 222771)
+++ fortran/gfortran.h	(Arbeitskopie)
@@ -419,6 +419,7 @@  enum gfc_isym_id
   GFC_ISYM_EXPONENT,
   GFC_ISYM_EXTENDS_TYPE_OF,
   GFC_ISYM_FDATE,
+  GFC_ISYM_FE_RUNTIME_ERROR,
   GFC_ISYM_FGET,
   GFC_ISYM_FGETC,
   GFC_ISYM_FLOOR,
@@ -1001,6 +1002,7 @@  typedef struct
   bool cp_was_assumed; /* AS_ASSUMED_SIZE cp arrays are converted to
 			AS_EXPLICIT, but we want to remember that we
 			did this.  */
+  bool resolved;
 
 }
 gfc_array_spec;
@@ -1907,7 +1909,7 @@  typedef struct gfc_intrinsic_sym
   gfc_typespec ts;
   unsigned elemental:1, inquiry:1, transformational:1, pure:1,
     generic:1, specific:1, actual_ok:1, noreturn:1, conversion:1,
-    from_module:1;
+    from_module:1, vararg:1;
 
   int standard;
 
@@ -3231,4 +3233,8 @@  int gfc_code_walker (gfc_code **, walk_code_fn_t,
 
 void gfc_convert_mpz_to_signed (mpz_t, int);
 
+/* trans-array.c  */
+
+bool gfc_is_reallocatable_lhs (gfc_expr *);
+
 #endif /* GCC_GFORTRAN_H  */
Index: fortran/intrinsic.c
===================================================================
--- fortran/intrinsic.c	(Revision 222771)
+++ fortran/intrinsic.c	(Arbeitskopie)
@@ -520,7 +520,30 @@  add_sym_1s (const char *name, gfc_isym_id id, enum
 	   (void *) 0);
 }
 
+/* Add a symbol to the subroutine ilst where the subroutine takes one
+   printf-style character argument and a variable number of arguments
+   to follow.  */
 
+static void
+add_sym_1p (const char *name, gfc_isym_id id, enum klass cl, bt type, int kind,
+	    int standard, bool (*check) (gfc_actual_arglist *),
+	    gfc_expr *(*simplify) (gfc_expr*), void (*resolve) (gfc_code *),
+	    const char *a1, bt type1, int kind1, int optional1, sym_intent intent1)
+{
+  gfc_check_f cf;
+  gfc_simplify_f sf;
+  gfc_resolve_f rf;
+
+  cf.f1m = check;
+  sf.f1 = simplify;
+  rf.s1 = resolve;
+
+  add_sym (name, id, cl, ACTUAL_NO, type, kind, standard, cf, sf, rf,
+	   a1, type1, kind1, optional1, intent1,
+	   (void *) 0);
+}
+
+
 /* Add a symbol from the MAX/MIN family of intrinsic functions to the
    function.  MAX et al take 2 or more arguments.  */
 
@@ -1159,6 +1182,17 @@  make_from_module (void)
     next_sym[-1].from_module = 1;
 }
 
+
+/* Mark the current subroutine as having a variable number of
+   arguments.  */
+
+static void
+make_vararg (void)
+{
+  if (sizing == SZ_NOTHING)
+    next_sym[-1].vararg = 1;
+}
+
 /* Set the attr.value of the current procedure.  */
 
 static void
@@ -3292,6 +3326,17 @@  add_subroutines (void)
 	      "fptr", BT_UNKNOWN, 0, REQUIRED, INTENT_OUT);
   make_from_module();
 
+  /* Internal subroutine for emitting a runtime error.  */
+
+  add_sym_1p ("fe_runtime_error", GFC_ISYM_FE_RUNTIME_ERROR, CLASS_IMPURE,
+	      BT_UNKNOWN, 0, GFC_STD_GNU,
+	      gfc_check_fe_runtime_error, NULL, gfc_resolve_fe_runtime_error,
+	      "msg", BT_CHARACTER, dc, REQUIRED, INTENT_IN);
+
+  make_noreturn ();
+  make_vararg ();
+  make_from_module ();
+
   /* Coarray collectives.  */
   add_sym_4s ("co_broadcast", GFC_ISYM_CO_BROADCAST, CLASS_IMPURE,
 	      BT_UNKNOWN, 0, GFC_STD_F2008_TS,
@@ -4501,7 +4546,7 @@  gfc_intrinsic_sub_interface (gfc_code *c, int erro
 
   init_arglist (isym);
 
-  if (!sort_actual (name, &c->ext.actual, isym->formal, &c->loc))
+  if (!isym->vararg && !sort_actual (name, &c->ext.actual, isym->formal, &c->loc))
     goto fail;
 
   if (!do_ts29113_check (isym, c->ext.actual))
Index: fortran/intrinsic.h
===================================================================
--- fortran/intrinsic.h	(Revision 222771)
+++ fortran/intrinsic.h	(Arbeitskopie)
@@ -190,6 +190,7 @@  bool gfc_check_system_clock (gfc_expr *, gfc_expr
 bool gfc_check_date_and_time (gfc_expr *, gfc_expr *, gfc_expr *, gfc_expr *);
 bool gfc_check_exit (gfc_expr *);
 bool gfc_check_fdate_sub (gfc_expr *);
+bool gfc_check_fe_runtime_error (gfc_actual_arglist *);
 bool gfc_check_flush (gfc_expr *);
 bool gfc_check_free (gfc_expr *);
 bool gfc_check_fstat_sub (gfc_expr *, gfc_expr *, gfc_expr *);
@@ -602,6 +603,7 @@  void gfc_resolve_ctime_sub (gfc_code *);
 void gfc_resolve_execute_command_line (gfc_code *);
 void gfc_resolve_exit (gfc_code *);
 void gfc_resolve_fdate_sub (gfc_code *);
+void gfc_resolve_fe_runtime_error (gfc_code *);
 void gfc_resolve_flush (gfc_code *);
 void gfc_resolve_free (gfc_code *);
 void gfc_resolve_fseek_sub (gfc_code *);
Index: fortran/invoke.texi
===================================================================
--- fortran/invoke.texi	(Revision 222771)
+++ fortran/invoke.texi	(Arbeitskopie)
@@ -178,6 +178,7 @@  and warnings}.
 -finit-character=@var{n} -finit-integer=@var{n} -finit-local-zero @gol
 -finit-logical=@var{<true|false>}
 -finit-real=@var{<zero|inf|-inf|nan|snan>} @gol
+-finline-matmul-limit=@var{n} @gol
 -fmax-array-constructor=@var{n} -fmax-stack-var-size=@var{n}
 -fno-align-commons @gol
 -fno-automatic -fno-protect-parens -fno-underscoring @gol
@@ -1537,6 +1538,20 @@  geometric mean of the dimensions of the argument a
 
 The default value for @var{n} is 30.
 
+@item -finline-matmul-limit=@var{n}
+@opindex @code{finline-matmul-limit}
+When front-end optimiztion is active, some calls to the @code{MATMUL}
+intrinsic function will be inlined.  Setting
+@code{-finline-matmul-limit=0} will disable inlining in all cases.
+Setting this option it to a specified value will call the library
+routines for matrices with size larger than @var{n}. If the matrices
+involved are not square, the size comparison is performed using the
+geometric mean of the dimensions of the argument and result matrices.
+
+The default value for @var{n} is the value specified for
+@code{-fblas-matmul-limit} if this option is specified, or unlimitited
+otherwise.
+
 @item -frecursive
 @opindex @code{frecursive}
 Allow indirect recursion by forcing all local arrays to be allocated
@@ -1632,11 +1647,12 @@  if @option{-ffrontend-optimize} is in effect.
 @cindex Front-end optimization
 This option performs front-end optimization, based on manipulating
 parts the Fortran parse tree.  Enabled by default by any @option{-O}
-option.  Optimizations enabled by this option include elimination of
-identical function calls within expressions, removing unnecessary
-calls to @code{TRIM} in comparisons and assignments and replacing
-@code{TRIM(a)} with @code{a(1:LEN_TRIM(a))}. 
-It can be deselected by specifying @option{-fno-frontend-optimize}.
+option.  Optimizations enabled by this option include inlining calls
+to @code{MATMUL}, elimination of identical function calls within
+expressions, removing unnecessary calls to @code{TRIM} in comparisons
+and assignments and replacing @code{TRIM(a)} with
+@code{a(1:LEN_TRIM(a))}.  It can be deselected by specifying
+@option{-fno-frontend-optimize}.
 @end table
 
 @xref{Code Gen Options,,Options for Code Generation Conventions,
Index: fortran/iresolve.c
===================================================================
--- fortran/iresolve.c	(Revision 222771)
+++ fortran/iresolve.c	(Arbeitskopie)
@@ -2197,7 +2197,20 @@  gfc_resolve_rrspacing (gfc_expr *f, gfc_expr *x)
   f->value.function.name = gfc_get_string ("__rrspacing_%d", x->ts.kind);
 }
 
+void
+gfc_resolve_fe_runtime_error (gfc_code *c)
+{
+  const char *name;
+  gfc_actual_arglist *a;
 
+  name = gfc_get_string (PREFIX ("runtime_error"));
+
+  for (a = c->ext.actual->next; a; a = a->next)
+    a->name = "%VAL";
+
+  c->resolved_sym = gfc_get_intrinsic_sub_symbol (name);
+}
+
 void
 gfc_resolve_scale (gfc_expr *f, gfc_expr *x, gfc_expr *i ATTRIBUTE_UNUSED)
 {
Index: fortran/lang.opt
===================================================================
--- fortran/lang.opt	(Revision 222771)
+++ fortran/lang.opt	(Arbeitskopie)
@@ -542,6 +542,10 @@  Enum(gfc_init_local_real) String(inf) Value(GFC_IN
 EnumValue
 Enum(gfc_init_local_real) String(-inf) Value(GFC_INIT_REAL_NEG_INF)
 
+finline-matmul-limit=
+Fortran RejectNegative Joined UInteger Var(flag_inline_matmul_limit) Init(-1)
+-finline-matmul-limit=<n>	Specify the size of the largest matrix for which matmul will be inlined
+
 fmax-array-constructor=
 Fortran RejectNegative Joined UInteger Var(flag_max_array_constructor) Init(65535)
 -fmax-array-constructor=<n>	Maximum number of objects in an array constructor
Index: fortran/options.c
===================================================================
--- fortran/options.c	(Revision 222771)
+++ fortran/options.c	(Arbeitskopie)
@@ -378,6 +378,11 @@  gfc_post_options (const char **pfilename)
   if (!flag_automatic)
     flag_max_stack_var_size = 0;
   
+  /* If we call BLAS directly, only inline up to the BLAS limit.  */
+
+  if (flag_external_blas && flag_inline_matmul_limit < 0)
+    flag_inline_matmul_limit = flag_blas_matmul_limit;
+
   /* Optimization implies front end optimization, unless the user
      specified it directly.  */
 
Index: fortran/trans-array.h
===================================================================
--- fortran/trans-array.h	(Revision 222771)
+++ fortran/trans-array.h	(Arbeitskopie)
@@ -64,8 +64,6 @@  tree gfc_copy_only_alloc_comp (gfc_symbol *, tree,
 
 tree gfc_alloc_allocatable_for_assignment (gfc_loopinfo*, gfc_expr*, gfc_expr*);
 
-bool gfc_is_reallocatable_lhs (gfc_expr *);
-
 /* Add initialization for deferred arrays.  */
 void gfc_trans_deferred_array (gfc_symbol *, gfc_wrapped_block *);
 /* Generate an initializer for a static pointer or allocatable array.  */
Index: testsuite/gfortran.dg/dependency_26.f90
===================================================================
--- testsuite/gfortran.dg/dependency_26.f90	(Revision 222771)
+++ testsuite/gfortran.dg/dependency_26.f90	(Arbeitskopie)
@@ -1,5 +1,5 @@ 
 ! { dg-do compile }
-! { dg-options "-fdump-tree-original" }
+! { dg-options "-finline-matmul-limit=0 -fdump-tree-original" }
 !
 ! Test the fix for PR36932 and PR36933, in which unnecessary
 ! temporaries were being generated.  The module m2 tests the
Index: testsuite/gfortran.dg/function_optimize_1.f90
===================================================================
--- testsuite/gfortran.dg/function_optimize_1.f90	(Revision 222771)
+++ testsuite/gfortran.dg/function_optimize_1.f90	(Arbeitskopie)
@@ -1,5 +1,5 @@ 
 ! { dg-do compile }
-! { dg-options "-O -fdump-tree-original -Warray-temporaries" }
+! { dg-options "-O -fdump-tree-original -finline-matmul-limit=0 -Warray-temporaries" }
 program main
   implicit none
   real, dimension(2,2) :: a, b, c, d
Index: testsuite/gfortran.dg/function_optimize_2.f90
===================================================================
--- testsuite/gfortran.dg/function_optimize_2.f90	(Revision 222771)
+++ testsuite/gfortran.dg/function_optimize_2.f90	(Arbeitskopie)
@@ -1,5 +1,5 @@ 
 ! { dg-do compile }
-! { dg-options "-O -faggressive-function-elimination -fdump-tree-original" }
+! { dg-options "-O -finline-matmul-limit=0 -faggressive-function-elimination -fdump-tree-original" }
 program main
   implicit none
   real, dimension(2,2) :: a, b, c, d
Index: testsuite/gfortran.dg/function_optimize_5.f90
===================================================================
--- testsuite/gfortran.dg/function_optimize_5.f90	(Revision 222771)
+++ testsuite/gfortran.dg/function_optimize_5.f90	(Arbeitskopie)
@@ -1,5 +1,5 @@ 
 ! { dg-do compile }
-! { dg-options "-ffrontend-optimize -Wfunction-elimination" }
+! { dg-options "-ffrontend-optimize -finline-matmul-limit=0 -Wfunction-elimination" }
 ! Check the -ffrontend-optimize (in the absence of -O) and
 ! -Wfunction-elimination options.
 program main
Index: testsuite/gfortran.dg/function_optimize_7.f90
===================================================================
--- testsuite/gfortran.dg/function_optimize_7.f90	(Revision 222771)
+++ testsuite/gfortran.dg/function_optimize_7.f90	(Arbeitskopie)
@@ -1,5 +1,5 @@ 
 ! { dg-do compile }
-! { dg-options "-O -fdump-tree-original -Warray-temporaries" }
+! { dg-options "-O -fdump-tree-original -Warray-temporaries -finline-matmul-limit=0" }
 subroutine xx(n, m, a, b, c, d, x, z, i, s_in, s_out)
   implicit none
   integer, intent(in) :: n, m