PR fortran/26025

	* lang.opt: Add -fexternal-blas and -fblas-matmul-limit options.
	* options.c (gfc_init_options): Initialize new flags.
	(gfc_handle_option): Handle new flags.
	* gfortran.h (gfc_option): Add flag_external_blas and
	blas_matmul_limit flags.
	* trans-expr.c (gfc_conv_function_call): Use new argument
	append_args, appending it at the end of the argument list
	built for a function call.
	* trans-stmt.c (gfc_trans_call): Use NULL_TREE for the new
	append_args argument to gfc_trans_call.
	* trans.h (gfc_conv_function_call): Update prototype.
	* trans-decl.c (gfc_build_intrinsic_function_decls): Add
	prototypes for BLAS ?gemm routines.
	* trans-intrinsic.c (gfc_conv_intrinsic_funcall): Generate the
	extra arguments given to the library matmul function, and give
	them to gfc_conv_function_call.
	* invoke.texi: Add documentation for -fexternal-blas and
	-fblas-matmul-limit.

	* m4/matmul.m4: Add possible call to gemm routine.
	* generated/matmul_r8.c: Regenerate.
	* generated/matmul_r16.c: Regenerate.
	* generated/matmul_c8.c: Regenerate.
	* generated/matmul_i8.c: Regenerate.
	* generated/matmul_c16.c: Regenerate.
	* generated/matmul_r10.c: Regenerate.
	* generated/matmul_r4.c: Regenerate.
	* generated/matmul_c10.c: Regenerate.
	* generated/matmul_c4.c: Regenerate.
	* generated/matmul_i4.c: Regenerate.
	* generated/matmul_i16.c: Regenerate.


git-svn-id: svn+ssh://gcc.gnu.org/svn/gcc/trunk@117948 138bc75d-0d04-0410-961f-82ee72b054a4

13 files changed, 544 insertions, 36 deletions

diff --git a/libgfortran/ChangeLog b/libgfortran/ChangeLog
index a9e70825f80..063c62519d3 100644
--- a/libgfortran/ChangeLog
+++ b/libgfortran/ChangeLog
@@ -1,3 +1,19 @@
+2006-10-22  Francois-Xavier Coudert  <coudert@clipper.ens.fr>
+
+	PR fortran/26025
+	* m4/matmul.m4: Add possible call to gemm routine.
+	* generated/matmul_r8.c: Regenerate.
+	* generated/matmul_r16.c: Regenerate.
+	* generated/matmul_c8.c: Regenerate.
+	* generated/matmul_i8.c: Regenerate.
+	* generated/matmul_c16.c: Regenerate.
+	* generated/matmul_r10.c: Regenerate.
+	* generated/matmul_r4.c: Regenerate.
+	* generated/matmul_c10.c: Regenerate.
+	* generated/matmul_c4.c: Regenerate.
+	* generated/matmul_i4.c: Regenerate.
+	* generated/matmul_i16.c: Regenerate.
+
 2006-10-21  Steven G. Kargl  <kargl@gcc.gnu.org>
 
 	* runtime/error.c: Add errno.h
diff --git a/libgfortran/generated/matmul_c10.c b/libgfortran/generated/matmul_c10.c
index df2cd93c15f..5e3b281245c 100644
--- a/libgfortran/generated/matmul_c10.c
+++ b/libgfortran/generated/matmul_c10.c
@@ -36,6 +36,16 @@ Boston, MA 02110-1301, USA.  */
 
 #if defined (HAVE_GFC_COMPLEX_10)
 
+/* Prototype for the BLAS ?gemm subroutine, a pointer to which can be
+   passed to us by the front-end, in which case we'll call it for large
+   matrices.  */
+
+typedef void (*blas_call)(const char *, const char *, const int *, const int *,
+                          const int *, const GFC_COMPLEX_10 *, const GFC_COMPLEX_10 *,
+                          const int *, const GFC_COMPLEX_10 *, const int *,
+                          const GFC_COMPLEX_10 *, GFC_COMPLEX_10 *, const int *,
+                          int, int);
+
 /* The order of loops is different in the case of plain matrix
    multiplication C=MATMUL(A,B), and in the frequent special case where
    the argument A is the temporary result of a TRANSPOSE intrinsic:
@@ -56,18 +66,24 @@ Boston, MA 02110-1301, USA.  */
        DO I=1,M
          S = 0
          DO K=1,COUNT
-           S = S+A(I,K)+B(K,J)
+           S = S+A(I,K)*B(K,J)
          C(I,J) = S
    ENDIF
 */
 
+/* If try_blas is set to a nonzero value, then the matmul function will
+   see if there is a way to perform the matrix multiplication by a call
+   to the BLAS gemm function.  */
+
 extern void matmul_c10 (gfc_array_c10 * const restrict retarray, 
-	gfc_array_c10 * const restrict a, gfc_array_c10 * const restrict b);
+	gfc_array_c10 * const restrict a, gfc_array_c10 * const restrict b, int try_blas,
+	int blas_limit, blas_call gemm);
 export_proto(matmul_c10);
 
 void
 matmul_c10 (gfc_array_c10 * const restrict retarray, 
-	gfc_array_c10 * const restrict a, gfc_array_c10 * const restrict b)
+	gfc_array_c10 * const restrict a, gfc_array_c10 * const restrict b, int try_blas,
+	int blas_limit, blas_call gemm)
 {
   const GFC_COMPLEX_10 * restrict abase;
   const GFC_COMPLEX_10 * restrict bbase;
@@ -177,6 +193,31 @@ matmul_c10 (gfc_array_c10 * const restrict retarray,
   bbase = b->data;
   dest = retarray->data;
 
+
+  /* Now that everything is set up, we're performing the multiplication
+     itself.  */
+
+#define POW3(x) (((float) (x)) * ((float) (x)) * ((float) (x)))
+
+  if (try_blas && rxstride == 1 && (axstride == 1 || aystride == 1)
+      && (bxstride == 1 || bystride == 1)
+      && (((float) xcount) * ((float) ycount) * ((float) count)
+          > POW3(blas_limit)))
+  {
+    const int m = xcount, n = ycount, k = count, ldc = rystride;
+    const GFC_COMPLEX_10 one = 1, zero = 0;
+    const int lda = (axstride == 1) ? aystride : axstride,
+              ldb = (bxstride == 1) ? bystride : bxstride;
+
+    if (lda > 0 && ldb > 0 && ldc > 0 && m > 1 && n > 1 && k > 1)
+      {
+        assert (gemm != NULL);
+        gemm (axstride == 1 ? "N" : "T", bxstride == 1 ? "N" : "T", &m, &n, &k,
+              &one, abase, &lda, bbase, &ldb, &zero, dest, &ldc, 1, 1);
+        return;
+      }
+  }
+
   if (rxstride == 1 && axstride == 1 && bxstride == 1)
     {
       const GFC_COMPLEX_10 * restrict bbase_y;
diff --git a/libgfortran/generated/matmul_c16.c b/libgfortran/generated/matmul_c16.c
index 6425eb8d49d..f7301114b37 100644
--- a/libgfortran/generated/matmul_c16.c
+++ b/libgfortran/generated/matmul_c16.c
@@ -36,6 +36,16 @@ Boston, MA 02110-1301, USA.  */
 
 #if defined (HAVE_GFC_COMPLEX_16)
 
+/* Prototype for the BLAS ?gemm subroutine, a pointer to which can be
+   passed to us by the front-end, in which case we'll call it for large
+   matrices.  */
+
+typedef void (*blas_call)(const char *, const char *, const int *, const int *,
+                          const int *, const GFC_COMPLEX_16 *, const GFC_COMPLEX_16 *,
+                          const int *, const GFC_COMPLEX_16 *, const int *,
+                          const GFC_COMPLEX_16 *, GFC_COMPLEX_16 *, const int *,
+                          int, int);
+
 /* The order of loops is different in the case of plain matrix
    multiplication C=MATMUL(A,B), and in the frequent special case where
    the argument A is the temporary result of a TRANSPOSE intrinsic:
@@ -56,18 +66,24 @@ Boston, MA 02110-1301, USA.  */
        DO I=1,M
          S = 0
          DO K=1,COUNT
-           S = S+A(I,K)+B(K,J)
+           S = S+A(I,K)*B(K,J)
          C(I,J) = S
    ENDIF
 */
 
+/* If try_blas is set to a nonzero value, then the matmul function will
+   see if there is a way to perform the matrix multiplication by a call
+   to the BLAS gemm function.  */
+
 extern void matmul_c16 (gfc_array_c16 * const restrict retarray, 
-	gfc_array_c16 * const restrict a, gfc_array_c16 * const restrict b);
+	gfc_array_c16 * const restrict a, gfc_array_c16 * const restrict b, int try_blas,
+	int blas_limit, blas_call gemm);
 export_proto(matmul_c16);
 
 void
 matmul_c16 (gfc_array_c16 * const restrict retarray, 
-	gfc_array_c16 * const restrict a, gfc_array_c16 * const restrict b)
+	gfc_array_c16 * const restrict a, gfc_array_c16 * const restrict b, int try_blas,
+	int blas_limit, blas_call gemm)
 {
   const GFC_COMPLEX_16 * restrict abase;
   const GFC_COMPLEX_16 * restrict bbase;
@@ -177,6 +193,31 @@ matmul_c16 (gfc_array_c16 * const restrict retarray,
   bbase = b->data;
   dest = retarray->data;
 
+
+  /* Now that everything is set up, we're performing the multiplication
+     itself.  */
+
+#define POW3(x) (((float) (x)) * ((float) (x)) * ((float) (x)))
+
+  if (try_blas && rxstride == 1 && (axstride == 1 || aystride == 1)
+      && (bxstride == 1 || bystride == 1)
+      && (((float) xcount) * ((float) ycount) * ((float) count)
+          > POW3(blas_limit)))
+  {
+    const int m = xcount, n = ycount, k = count, ldc = rystride;
+    const GFC_COMPLEX_16 one = 1, zero = 0;
+    const int lda = (axstride == 1) ? aystride : axstride,
+              ldb = (bxstride == 1) ? bystride : bxstride;
+
+    if (lda > 0 && ldb > 0 && ldc > 0 && m > 1 && n > 1 && k > 1)
+      {
+        assert (gemm != NULL);
+        gemm (axstride == 1 ? "N" : "T", bxstride == 1 ? "N" : "T", &m, &n, &k,
+              &one, abase, &lda, bbase, &ldb, &zero, dest, &ldc, 1, 1);
+        return;
+      }
+  }
+
   if (rxstride == 1 && axstride == 1 && bxstride == 1)
     {
       const GFC_COMPLEX_16 * restrict bbase_y;
diff --git a/libgfortran/generated/matmul_c4.c b/libgfortran/generated/matmul_c4.c
index 2d47a134972..f2984ab48ab 100644
--- a/libgfortran/generated/matmul_c4.c
+++ b/libgfortran/generated/matmul_c4.c
@@ -36,6 +36,16 @@ Boston, MA 02110-1301, USA.  */
 
 #if defined (HAVE_GFC_COMPLEX_4)
 
+/* Prototype for the BLAS ?gemm subroutine, a pointer to which can be
+   passed to us by the front-end, in which case we'll call it for large
+   matrices.  */
+
+typedef void (*blas_call)(const char *, const char *, const int *, const int *,
+                          const int *, const GFC_COMPLEX_4 *, const GFC_COMPLEX_4 *,
+                          const int *, const GFC_COMPLEX_4 *, const int *,
+                          const GFC_COMPLEX_4 *, GFC_COMPLEX_4 *, const int *,
+                          int, int);
+
 /* The order of loops is different in the case of plain matrix
    multiplication C=MATMUL(A,B), and in the frequent special case where
    the argument A is the temporary result of a TRANSPOSE intrinsic:
@@ -56,18 +66,24 @@ Boston, MA 02110-1301, USA.  */
        DO I=1,M
          S = 0
          DO K=1,COUNT
-           S = S+A(I,K)+B(K,J)
+           S = S+A(I,K)*B(K,J)
          C(I,J) = S
    ENDIF
 */
 
+/* If try_blas is set to a nonzero value, then the matmul function will
+   see if there is a way to perform the matrix multiplication by a call
+   to the BLAS gemm function.  */
+
 extern void matmul_c4 (gfc_array_c4 * const restrict retarray, 
-	gfc_array_c4 * const restrict a, gfc_array_c4 * const restrict b);
+	gfc_array_c4 * const restrict a, gfc_array_c4 * const restrict b, int try_blas,
+	int blas_limit, blas_call gemm);
 export_proto(matmul_c4);
 
 void
 matmul_c4 (gfc_array_c4 * const restrict retarray, 
-	gfc_array_c4 * const restrict a, gfc_array_c4 * const restrict b)
+	gfc_array_c4 * const restrict a, gfc_array_c4 * const restrict b, int try_blas,
+	int blas_limit, blas_call gemm)
 {
   const GFC_COMPLEX_4 * restrict abase;
   const GFC_COMPLEX_4 * restrict bbase;
@@ -177,6 +193,31 @@ matmul_c4 (gfc_array_c4 * const restrict retarray,
   bbase = b->data;
   dest = retarray->data;
 
+
+  /* Now that everything is set up, we're performing the multiplication
+     itself.  */
+
+#define POW3(x) (((float) (x)) * ((float) (x)) * ((float) (x)))
+
+  if (try_blas && rxstride == 1 && (axstride == 1 || aystride == 1)
+      && (bxstride == 1 || bystride == 1)
+      && (((float) xcount) * ((float) ycount) * ((float) count)
+          > POW3(blas_limit)))
+  {
+    const int m = xcount, n = ycount, k = count, ldc = rystride;
+    const GFC_COMPLEX_4 one = 1, zero = 0;
+    const int lda = (axstride == 1) ? aystride : axstride,
+              ldb = (bxstride == 1) ? bystride : bxstride;
+
+    if (lda > 0 && ldb > 0 && ldc > 0 && m > 1 && n > 1 && k > 1)
+      {
+        assert (gemm != NULL);
+        gemm (axstride == 1 ? "N" : "T", bxstride == 1 ? "N" : "T", &m, &n, &k,
+              &one, abase, &lda, bbase, &ldb, &zero, dest, &ldc, 1, 1);
+        return;
+      }
+  }
+
   if (rxstride == 1 && axstride == 1 && bxstride == 1)
     {
       const GFC_COMPLEX_4 * restrict bbase_y;
diff --git a/libgfortran/generated/matmul_c8.c b/libgfortran/generated/matmul_c8.c
index f22719df505..65cc0a52c4b 100644
--- a/libgfortran/generated/matmul_c8.c
+++ b/libgfortran/generated/matmul_c8.c
@@ -36,6 +36,16 @@ Boston, MA 02110-1301, USA.  */
 
 #if defined (HAVE_GFC_COMPLEX_8)
 
+/* Prototype for the BLAS ?gemm subroutine, a pointer to which can be
+   passed to us by the front-end, in which case we'll call it for large
+   matrices.  */
+
+typedef void (*blas_call)(const char *, const char *, const int *, const int *,
+                          const int *, const GFC_COMPLEX_8 *, const GFC_COMPLEX_8 *,
+                          const int *, const GFC_COMPLEX_8 *, const int *,
+                          const GFC_COMPLEX_8 *, GFC_COMPLEX_8 *, const int *,
+                          int, int);
+
 /* The order of loops is different in the case of plain matrix
    multiplication C=MATMUL(A,B), and in the frequent special case where
    the argument A is the temporary result of a TRANSPOSE intrinsic:
@@ -56,18 +66,24 @@ Boston, MA 02110-1301, USA.  */
        DO I=1,M
          S = 0
          DO K=1,COUNT
-           S = S+A(I,K)+B(K,J)
+           S = S+A(I,K)*B(K,J)
          C(I,J) = S
    ENDIF
 */
 
+/* If try_blas is set to a nonzero value, then the matmul function will
+   see if there is a way to perform the matrix multiplication by a call
+   to the BLAS gemm function.  */
+
 extern void matmul_c8 (gfc_array_c8 * const restrict retarray, 
-	gfc_array_c8 * const restrict a, gfc_array_c8 * const restrict b);
+	gfc_array_c8 * const restrict a, gfc_array_c8 * const restrict b, int try_blas,
+	int blas_limit, blas_call gemm);
 export_proto(matmul_c8);
 
 void
 matmul_c8 (gfc_array_c8 * const restrict retarray, 
-	gfc_array_c8 * const restrict a, gfc_array_c8 * const restrict b)
+	gfc_array_c8 * const restrict a, gfc_array_c8 * const restrict b, int try_blas,
+	int blas_limit, blas_call gemm)
 {
   const GFC_COMPLEX_8 * restrict abase;
   const GFC_COMPLEX_8 * restrict bbase;
@@ -177,6 +193,31 @@ matmul_c8 (gfc_array_c8 * const restrict retarray,
   bbase = b->data;
   dest = retarray->data;
 
+
+  /* Now that everything is set up, we're performing the multiplication
+     itself.  */
+
+#define POW3(x) (((float) (x)) * ((float) (x)) * ((float) (x)))
+
+  if (try_blas && rxstride == 1 && (axstride == 1 || aystride == 1)
+      && (bxstride == 1 || bystride == 1)
+      && (((float) xcount) * ((float) ycount) * ((float) count)
+          > POW3(blas_limit)))
+  {
+    const int m = xcount, n = ycount, k = count, ldc = rystride;
+    const GFC_COMPLEX_8 one = 1, zero = 0;
+    const int lda = (axstride == 1) ? aystride : axstride,
+              ldb = (bxstride == 1) ? bystride : bxstride;
+
+    if (lda > 0 && ldb > 0 && ldc > 0 && m > 1 && n > 1 && k > 1)
+      {
+        assert (gemm != NULL);
+        gemm (axstride == 1 ? "N" : "T", bxstride == 1 ? "N" : "T", &m, &n, &k,
+              &one, abase, &lda, bbase, &ldb, &zero, dest, &ldc, 1, 1);
+        return;
+      }
+  }
+
   if (rxstride == 1 && axstride == 1 && bxstride == 1)
     {
       const GFC_COMPLEX_8 * restrict bbase_y;
diff --git a/libgfortran/generated/matmul_i16.c b/libgfortran/generated/matmul_i16.c
index 73c3fbc108d..a193669d108 100644
--- a/libgfortran/generated/matmul_i16.c
+++ b/libgfortran/generated/matmul_i16.c
@@ -36,6 +36,16 @@ Boston, MA 02110-1301, USA.  */
 
 #if defined (HAVE_GFC_INTEGER_16)
 
+/* Prototype for the BLAS ?gemm subroutine, a pointer to which can be
+   passed to us by the front-end, in which case we'll call it for large
+   matrices.  */
+
+typedef void (*blas_call)(const char *, const char *, const int *, const int *,
+                          const int *, const GFC_INTEGER_16 *, const GFC_INTEGER_16 *,
+                          const int *, const GFC_INTEGER_16 *, const int *,
+                          const GFC_INTEGER_16 *, GFC_INTEGER_16 *, const int *,
+                          int, int);
+
 /* The order of loops is different in the case of plain matrix
    multiplication C=MATMUL(A,B), and in the frequent special case where
    the argument A is the temporary result of a TRANSPOSE intrinsic:
@@ -56,18 +66,24 @@ Boston, MA 02110-1301, USA.  */
        DO I=1,M
          S = 0
          DO K=1,COUNT
-           S = S+A(I,K)+B(K,J)
+           S = S+A(I,K)*B(K,J)
          C(I,J) = S
    ENDIF
 */
 
+/* If try_blas is set to a nonzero value, then the matmul function will
+   see if there is a way to perform the matrix multiplication by a call
+   to the BLAS gemm function.  */
+
 extern void matmul_i16 (gfc_array_i16 * const restrict retarray, 
-	gfc_array_i16 * const restrict a, gfc_array_i16 * const restrict b);
+	gfc_array_i16 * const restrict a, gfc_array_i16 * const restrict b, int try_blas,
+	int blas_limit, blas_call gemm);
 export_proto(matmul_i16);
 
 void
 matmul_i16 (gfc_array_i16 * const restrict retarray, 
-	gfc_array_i16 * const restrict a, gfc_array_i16 * const restrict b)
+	gfc_array_i16 * const restrict a, gfc_array_i16 * const restrict b, int try_blas,
+	int blas_limit, blas_call gemm)
 {
   const GFC_INTEGER_16 * restrict abase;
   const GFC_INTEGER_16 * restrict bbase;
@@ -177,6 +193,31 @@ matmul_i16 (gfc_array_i16 * const restrict retarray,
   bbase = b->data;
   dest = retarray->data;
 
+
+  /* Now that everything is set up, we're performing the multiplication
+     itself.  */
+
+#define POW3(x) (((float) (x)) * ((float) (x)) * ((float) (x)))
+
+  if (try_blas && rxstride == 1 && (axstride == 1 || aystride == 1)
+      && (bxstride == 1 || bystride == 1)
+      && (((float) xcount) * ((float) ycount) * ((float) count)
+          > POW3(blas_limit)))
+  {
+    const int m = xcount, n = ycount, k = count, ldc = rystride;
+    const GFC_INTEGER_16 one = 1, zero = 0;
+    const int lda = (axstride == 1) ? aystride : axstride,
+              ldb = (bxstride == 1) ? bystride : bxstride;
+
+    if (lda > 0 && ldb > 0 && ldc > 0 && m > 1 && n > 1 && k > 1)
+      {
+        assert (gemm != NULL);
+        gemm (axstride == 1 ? "N" : "T", bxstride == 1 ? "N" : "T", &m, &n, &k,
+              &one, abase, &lda, bbase, &ldb, &zero, dest, &ldc, 1, 1);
+        return;
+      }
+  }
+
   if (rxstride == 1 && axstride == 1 && bxstride == 1)
     {
       const GFC_INTEGER_16 * restrict bbase_y;
diff --git a/libgfortran/generated/matmul_i4.c b/libgfortran/generated/matmul_i4.c
index 63bca0152cd..69b9b487a81 100644
--- a/libgfortran/generated/matmul_i4.c
+++ b/libgfortran/generated/matmul_i4.c
@@ -36,6 +36,16 @@ Boston, MA 02110-1301, USA.  */
 
 #if defined (HAVE_GFC_INTEGER_4)
 
+/* Prototype for the BLAS ?gemm subroutine, a pointer to which can be
+   passed to us by the front-end, in which case we'll call it for large
+   matrices.  */
+
+typedef void (*blas_call)(const char *, const char *, const int *, const int *,
+                          const int *, const GFC_INTEGER_4 *, const GFC_INTEGER_4 *,
+                          const int *, const GFC_INTEGER_4 *, const int *,
+                          const GFC_INTEGER_4 *, GFC_INTEGER_4 *, const int *,
+                          int, int);
+
 /* The order of loops is different in the case of plain matrix
    multiplication C=MATMUL(A,B), and in the frequent special case where
    the argument A is the temporary result of a TRANSPOSE intrinsic:
@@ -56,18 +66,24 @@ Boston, MA 02110-1301, USA.  */
        DO I=1,M
          S = 0
          DO K=1,COUNT
-           S = S+A(I,K)+B(K,J)
+           S = S+A(I,K)*B(K,J)
          C(I,J) = S
    ENDIF
 */
 
+/* If try_blas is set to a nonzero value, then the matmul function will
+   see if there is a way to perform the matrix multiplication by a call
+   to the BLAS gemm function.  */
+
 extern void matmul_i4 (gfc_array_i4 * const restrict retarray, 
-	gfc_array_i4 * const restrict a, gfc_array_i4 * const restrict b);
+	gfc_array_i4 * const restrict a, gfc_array_i4 * const restrict b, int try_blas,
+	int blas_limit, blas_call gemm);
 export_proto(matmul_i4);
 
 void
 matmul_i4 (gfc_array_i4 * const restrict retarray, 
-	gfc_array_i4 * const restrict a, gfc_array_i4 * const restrict b)
+	gfc_array_i4 * const restrict a, gfc_array_i4 * const restrict b, int try_blas,
+	int blas_limit, blas_call gemm)
 {
   const GFC_INTEGER_4 * restrict abase;
   const GFC_INTEGER_4 * restrict bbase;
@@ -177,6 +193,31 @@ matmul_i4 (gfc_array_i4 * const restrict retarray,
   bbase = b->data;
   dest = retarray->data;
 
+
+  /* Now that everything is set up, we're performing the multiplication
+     itself.  */
+
+#define POW3(x) (((float) (x)) * ((float) (x)) * ((float) (x)))
+
+  if (try_blas && rxstride == 1 && (axstride == 1 || aystride == 1)
+      && (bxstride == 1 || bystride == 1)
+      && (((float) xcount) * ((float) ycount) * ((float) count)
+          > POW3(blas_limit)))
+  {
+    const int m = xcount, n = ycount, k = count, ldc = rystride;
+    const GFC_INTEGER_4 one = 1, zero = 0;
+    const int lda = (axstride == 1) ? aystride : axstride,
+              ldb = (bxstride == 1) ? bystride : bxstride;
+
+    if (lda > 0 && ldb > 0 && ldc > 0 && m > 1 && n > 1 && k > 1)
+      {
+        assert (gemm != NULL);
+        gemm (axstride == 1 ? "N" : "T", bxstride == 1 ? "N" : "T", &m, &n, &k,
+              &one, abase, &lda, bbase, &ldb, &zero, dest, &ldc, 1, 1);
+        return;
+      }
+  }
+
   if (rxstride == 1 && axstride == 1 && bxstride == 1)
     {
       const GFC_INTEGER_4 * restrict bbase_y;
diff --git a/libgfortran/generated/matmul_i8.c b/libgfortran/generated/matmul_i8.c
index caaf9e8f976..23a87a904f7 100644
--- a/libgfortran/generated/matmul_i8.c
+++ b/libgfortran/generated/matmul_i8.c
@@ -36,6 +36,16 @@ Boston, MA 02110-1301, USA.  */
 
 #if defined (HAVE_GFC_INTEGER_8)
 
+/* Prototype for the BLAS ?gemm subroutine, a pointer to which can be
+   passed to us by the front-end, in which case we'll call it for large
+   matrices.  */
+
+typedef void (*blas_call)(const char *, const char *, const int *, const int *,
+                          const int *, const GFC_INTEGER_8 *, const GFC_INTEGER_8 *,
+                          const int *, const GFC_INTEGER_8 *, const int *,
+                          const GFC_INTEGER_8 *, GFC_INTEGER_8 *, const int *,
+                          int, int);
+
 /* The order of loops is different in the case of plain matrix
    multiplication C=MATMUL(A,B), and in the frequent special case where
    the argument A is the temporary result of a TRANSPOSE intrinsic:
@@ -56,18 +66,24 @@ Boston, MA 02110-1301, USA.  */
        DO I=1,M
          S = 0
          DO K=1,COUNT
-           S = S+A(I,K)+B(K,J)
+           S = S+A(I,K)*B(K,J)
          C(I,J) = S
    ENDIF
 */
 
+/* If try_blas is set to a nonzero value, then the matmul function will
+   see if there is a way to perform the matrix multiplication by a call
+   to the BLAS gemm function.  */
+
 extern void matmul_i8 (gfc_array_i8 * const restrict retarray, 
-	gfc_array_i8 * const restrict a, gfc_array_i8 * const restrict b);
+	gfc_array_i8 * const restrict a, gfc_array_i8 * const restrict b, int try_blas,
+	int blas_limit, blas_call gemm);
 export_proto(matmul_i8);
 
 void
 matmul_i8 (gfc_array_i8 * const restrict retarray, 
-	gfc_array_i8 * const restrict a, gfc_array_i8 * const restrict b)
+	gfc_array_i8 * const restrict a, gfc_array_i8 * const restrict b, int try_blas,
+	int blas_limit, blas_call gemm)
 {
   const GFC_INTEGER_8 * restrict abase;
   const GFC_INTEGER_8 * restrict bbase;
@@ -177,6 +193,31 @@ matmul_i8 (gfc_array_i8 * const restrict retarray,
   bbase = b->data;
   dest = retarray->data;
 
+
+  /* Now that everything is set up, we're performing the multiplication
+     itself.  */
+
+#define POW3(x) (((float) (x)) * ((float) (x)) * ((float) (x)))
+
+  if (try_blas && rxstride == 1 && (axstride == 1 || aystride == 1)
+      && (bxstride == 1 || bystride == 1)
+      && (((float) xcount) * ((float) ycount) * ((float) count)
+          > POW3(blas_limit)))
+  {
+    const int m = xcount, n = ycount, k = count, ldc = rystride;
+    const GFC_INTEGER_8 one = 1, zero = 0;
+    const int lda = (axstride == 1) ? aystride : axstride,
+              ldb = (bxstride == 1) ? bystride : bxstride;
+
+    if (lda > 0 && ldb > 0 && ldc > 0 && m > 1 && n > 1 && k > 1)
+      {
+        assert (gemm != NULL);
+        gemm (axstride == 1 ? "N" : "T", bxstride == 1 ? "N" : "T", &m, &n, &k,
+              &one, abase, &lda, bbase, &ldb, &zero, dest, &ldc, 1, 1);
+        return;
+      }
+  }
+
   if (rxstride == 1 && axstride == 1 && bxstride == 1)
     {
       const GFC_INTEGER_8 * restrict bbase_y;
diff --git a/libgfortran/generated/matmul_r10.c b/libgfortran/generated/matmul_r10.c
index 8fa1d6d9e49..e4dfd74ef03 100644
--- a/libgfortran/generated/matmul_r10.c
+++ b/libgfortran/generated/matmul_r10.c
@@ -36,6 +36,16 @@ Boston, MA 02110-1301, USA.  */
 
 #if defined (HAVE_GFC_REAL_10)
 
+/* Prototype for the BLAS ?gemm subroutine, a pointer to which can be
+   passed to us by the front-end, in which case we'll call it for large
+   matrices.  */
+
+typedef void (*blas_call)(const char *, const char *, const int *, const int *,
+                          const int *, const GFC_REAL_10 *, const GFC_REAL_10 *,
+                          const int *, const GFC_REAL_10 *, const int *,
+                          const GFC_REAL_10 *, GFC_REAL_10 *, const int *,
+                          int, int);
+
 /* The order of loops is different in the case of plain matrix
    multiplication C=MATMUL(A,B), and in the frequent special case where
    the argument A is the temporary result of a TRANSPOSE intrinsic:
@@ -56,18 +66,24 @@ Boston, MA 02110-1301, USA.  */
        DO I=1,M
          S = 0
          DO K=1,COUNT
-           S = S+A(I,K)+B(K,J)
+           S = S+A(I,K)*B(K,J)
          C(I,J) = S
    ENDIF
 */
 
+/* If try_blas is set to a nonzero value, then the matmul function will
+   see if there is a way to perform the matrix multiplication by a call
+   to the BLAS gemm function.  */
+
 extern void matmul_r10 (gfc_array_r10 * const restrict retarray, 
-	gfc_array_r10 * const restrict a, gfc_array_r10 * const restrict b);
+	gfc_array_r10 * const restrict a, gfc_array_r10 * const restrict b, int try_blas,
+	int blas_limit, blas_call gemm);
 export_proto(matmul_r10);
 
 void
 matmul_r10 (gfc_array_r10 * const restrict retarray, 
-	gfc_array_r10 * const restrict a, gfc_array_r10 * const restrict b)
+	gfc_array_r10 * const restrict a, gfc_array_r10 * const restrict b, int try_blas,
+	int blas_limit, blas_call gemm)
 {
   const GFC_REAL_10 * restrict abase;
   const GFC_REAL_10 * restrict bbase;
@@ -177,6 +193,31 @@ matmul_r10 (gfc_array_r10 * const restrict retarray,
   bbase = b->data;
   dest = retarray->data;
 
+
+  /* Now that everything is set up, we're performing the multiplication
+     itself.  */
+
+#define POW3(x) (((float) (x)) * ((float) (x)) * ((float) (x)))
+
+  if (try_blas && rxstride == 1 && (axstride == 1 || aystride == 1)
+      && (bxstride == 1 || bystride == 1)
+      && (((float) xcount) * ((float) ycount) * ((float) count)
+          > POW3(blas_limit)))
+  {
+    const int m = xcount, n = ycount, k = count, ldc = rystride;
+    const GFC_REAL_10 one = 1, zero = 0;
+    const int lda = (axstride == 1) ? aystride : axstride,
+              ldb = (bxstride == 1) ? bystride : bxstride;
+
+    if (lda > 0 && ldb > 0 && ldc > 0 && m > 1 && n > 1 && k > 1)
+      {
+        assert (gemm != NULL);
+        gemm (axstride == 1 ? "N" : "T", bxstride == 1 ? "N" : "T", &m, &n, &k,
+              &one, abase, &lda, bbase, &ldb, &zero, dest, &ldc, 1, 1);
+        return;
+      }
+  }
+
   if (rxstride == 1 && axstride == 1 && bxstride == 1)
     {
       const GFC_REAL_10 * restrict bbase_y;
diff --git a/libgfortran/generated/matmul_r16.c b/libgfortran/generated/matmul_r16.c
index 0f61b038168..ec760f2d3d8 100644
--- a/libgfortran/generated/matmul_r16.c
+++ b/libgfortran/generated/matmul_r16.c
@@ -36,6 +36,16 @@ Boston, MA 02110-1301, USA.  */
 
 #if defined (HAVE_GFC_REAL_16)
 
+/* Prototype for the BLAS ?gemm subroutine, a pointer to which can be
+   passed to us by the front-end, in which case we'll call it for large
+   matrices.  */
+
+typedef void (*blas_call)(const char *, const char *, const int *, const int *,
+                          const int *, const GFC_REAL_16 *, const GFC_REAL_16 *,
+                          const int *, const GFC_REAL_16 *, const int *,
+                          const GFC_REAL_16 *, GFC_REAL_16 *, const int *,
+                          int, int);
+
 /* The order of loops is different in the case of plain matrix
    multiplication C=MATMUL(A,B), and in the frequent special case where
    the argument A is the temporary result of a TRANSPOSE intrinsic:
@@ -56,18 +66,24 @@ Boston, MA 02110-1301, USA.  */
        DO I=1,M
          S = 0
          DO K=1,COUNT
-           S = S+A(I,K)+B(K,J)
+           S = S+A(I,K)*B(K,J)
          C(I,J) = S
    ENDIF
 */
 
+/* If try_blas is set to a nonzero value, then the matmul function will
+   see if there is a way to perform the matrix multiplication by a call
+   to the BLAS gemm function.  */
+
 extern void matmul_r16 (gfc_array_r16 * const restrict retarray, 
-	gfc_array_r16 * const restrict a, gfc_array_r16 * const restrict b);
+	gfc_array_r16 * const restrict a, gfc_array_r16 * const restrict b, int try_blas,
+	int blas_limit, blas_call gemm);
 export_proto(matmul_r16);
 
 void
 matmul_r16 (gfc_array_r16 * const restrict retarray, 
-	gfc_array_r16 * const restrict a, gfc_array_r16 * const restrict b)
+	gfc_array_r16 * const restrict a, gfc_array_r16 * const restrict b, int try_blas,
+	int blas_limit, blas_call gemm)
 {
   const GFC_REAL_16 * restrict abase;
   const GFC_REAL_16 * restrict bbase;
@@ -177,6 +193,31 @@ matmul_r16 (gfc_array_r16 * const restrict retarray,
   bbase = b->data;
   dest = retarray->data;
 
+
+  /* Now that everything is set up, we're performing the multiplication
+     itself.  */
+
+#define POW3(x) (((float) (x)) * ((float) (x)) * ((float) (x)))
+
+  if (try_blas && rxstride == 1 && (axstride == 1 || aystride == 1)
+      && (bxstride == 1 || bystride == 1)
+      && (((float) xcount) * ((float) ycount) * ((float) count)
+          > POW3(blas_limit)))
+  {
+    const int m = xcount, n = ycount, k = count, ldc = rystride;
+    const GFC_REAL_16 one = 1, zero = 0;
+    const int lda = (axstride == 1) ? aystride : axstride,
+              ldb = (bxstride == 1) ? bystride : bxstride;
+
+    if (lda > 0 && ldb > 0 && ldc > 0 && m > 1 && n > 1 && k > 1)
+      {
+        assert (gemm != NULL);
+        gemm (axstride == 1 ? "N" : "T", bxstride == 1 ? "N" : "T", &m, &n, &k,
+              &one, abase, &lda, bbase, &ldb, &zero, dest, &ldc, 1, 1);
+        return;
+      }
+  }
+
   if (rxstride == 1 && axstride == 1 && bxstride == 1)
     {
       const GFC_REAL_16 * restrict bbase_y;
diff --git a/libgfortran/generated/matmul_r4.c b/libgfortran/generated/matmul_r4.c
index d684dd2905c..cf2f45fb125 100644
--- a/libgfortran/generated/matmul_r4.c
+++ b/libgfortran/generated/matmul_r4.c
@@ -36,6 +36,16 @@ Boston, MA 02110-1301, USA.  */
 
 #if defined (HAVE_GFC_REAL_4)
 
+/* Prototype for the BLAS ?gemm subroutine, a pointer to which can be
+   passed to us by the front-end, in which case we'll call it for large
+   matrices.  */
+
+typedef void (*blas_call)(const char *, const char *, const int *, const int *,
+                          const int *, const GFC_REAL_4 *, const GFC_REAL_4 *,
+                          const int *, const GFC_REAL_4 *, const int *,
+                          const GFC_REAL_4 *, GFC_REAL_4 *, const int *,
+                          int, int);
+
 /* The order of loops is different in the case of plain matrix
    multiplication C=MATMUL(A,B), and in the frequent special case where
    the argument A is the temporary result of a TRANSPOSE intrinsic:
@@ -56,18 +66,24 @@ Boston, MA 02110-1301, USA.  */
        DO I=1,M
          S = 0
          DO K=1,COUNT
-           S = S+A(I,K)+B(K,J)
+           S = S+A(I,K)*B(K,J)
          C(I,J) = S
    ENDIF
 */
 
+/* If try_blas is set to a nonzero value, then the matmul function will
+   see if there is a way to perform the matrix multiplication by a call
+   to the BLAS gemm function.  */
+
 extern void matmul_r4 (gfc_array_r4 * const restrict retarray, 
-	gfc_array_r4 * const restrict a, gfc_array_r4 * const restrict b);
+	gfc_array_r4 * const restrict a, gfc_array_r4 * const restrict b, int try_blas,
+	int blas_limit, blas_call gemm);
 export_proto(matmul_r4);
 
 void
 matmul_r4 (gfc_array_r4 * const restrict retarray, 
-	gfc_array_r4 * const restrict a, gfc_array_r4 * const restrict b)
+	gfc_array_r4 * const restrict a, gfc_array_r4 * const restrict b, int try_blas,
+	int blas_limit, blas_call gemm)
 {
   const GFC_REAL_4 * restrict abase;
   const GFC_REAL_4 * restrict bbase;
@@ -177,6 +193,31 @@ matmul_r4 (gfc_array_r4 * const restrict retarray,
   bbase = b->data;
   dest = retarray->data;
 
+
+  /* Now that everything is set up, we're performing the multiplication
+     itself.  */
+
+#define POW3(x) (((float) (x)) * ((float) (x)) * ((float) (x)))
+
+  if (try_blas && rxstride == 1 && (axstride == 1 || aystride == 1)
+      && (bxstride == 1 || bystride == 1)
+      && (((float) xcount) * ((float) ycount) * ((float) count)
+          > POW3(blas_limit)))
+  {
+    const int m = xcount, n = ycount, k = count, ldc = rystride;
+    const GFC_REAL_4 one = 1, zero = 0;
+    const int lda = (axstride == 1) ? aystride : axstride,
+              ldb = (bxstride == 1) ? bystride : bxstride;
+
+    if (lda > 0 && ldb > 0 && ldc > 0 && m > 1 && n > 1 && k > 1)
+      {
+        assert (gemm != NULL);
+        gemm (axstride == 1 ? "N" : "T", bxstride == 1 ? "N" : "T", &m, &n, &k,
+              &one, abase, &lda, bbase, &ldb, &zero, dest, &ldc, 1, 1);
+        return;
+      }
+  }
+
   if (rxstride == 1 && axstride == 1 && bxstride == 1)
     {
       const GFC_REAL_4 * restrict bbase_y;
diff --git a/libgfortran/generated/matmul_r8.c b/libgfortran/generated/matmul_r8.c
index 41726bce2a5..c746f6c3519 100644
--- a/libgfortran/generated/matmul_r8.c
+++ b/libgfortran/generated/matmul_r8.c
@@ -36,6 +36,16 @@ Boston, MA 02110-1301, USA.  */
 
 #if defined (HAVE_GFC_REAL_8)
 
+/* Prototype for the BLAS ?gemm subroutine, a pointer to which can be
+   passed to us by the front-end, in which case we'll call it for large
+   matrices.  */
+
+typedef void (*blas_call)(const char *, const char *, const int *, const int *,
+                          const int *, const GFC_REAL_8 *, const GFC_REAL_8 *,
+                          const int *, const GFC_REAL_8 *, const int *,
+                          const GFC_REAL_8 *, GFC_REAL_8 *, const int *,
+                          int, int);
+
 /* The order of loops is different in the case of plain matrix
    multiplication C=MATMUL(A,B), and in the frequent special case where
    the argument A is the temporary result of a TRANSPOSE intrinsic:
@@ -56,18 +66,24 @@ Boston, MA 02110-1301, USA.  */
        DO I=1,M
          S = 0
          DO K=1,COUNT
-           S = S+A(I,K)+B(K,J)
+           S = S+A(I,K)*B(K,J)
          C(I,J) = S
    ENDIF
 */
 
+/* If try_blas is set to a nonzero value, then the matmul function will
+   see if there is a way to perform the matrix multiplication by a call
+   to the BLAS gemm function.  */
+
 extern void matmul_r8 (gfc_array_r8 * const restrict retarray, 
-	gfc_array_r8 * const restrict a, gfc_array_r8 * const restrict b);
+	gfc_array_r8 * const restrict a, gfc_array_r8 * const restrict b, int try_blas,
+	int blas_limit, blas_call gemm);
 export_proto(matmul_r8);
 
 void
 matmul_r8 (gfc_array_r8 * const restrict retarray, 
-	gfc_array_r8 * const restrict a, gfc_array_r8 * const restrict b)
+	gfc_array_r8 * const restrict a, gfc_array_r8 * const restrict b, int try_blas,
+	int blas_limit, blas_call gemm)
 {
   const GFC_REAL_8 * restrict abase;
   const GFC_REAL_8 * restrict bbase;
@@ -177,6 +193,31 @@ matmul_r8 (gfc_array_r8 * const restrict retarray,
   bbase = b->data;
   dest = retarray->data;
 
+
+  /* Now that everything is set up, we're performing the multiplication
+     itself.  */
+
+#define POW3(x) (((float) (x)) * ((float) (x)) * ((float) (x)))
+
+  if (try_blas && rxstride == 1 && (axstride == 1 || aystride == 1)
+      && (bxstride == 1 || bystride == 1)
+      && (((float) xcount) * ((float) ycount) * ((float) count)
+          > POW3(blas_limit)))
+  {
+    const int m = xcount, n = ycount, k = count, ldc = rystride;
+    const GFC_REAL_8 one = 1, zero = 0;
+    const int lda = (axstride == 1) ? aystride : axstride,
+              ldb = (bxstride == 1) ? bystride : bxstride;
+
+    if (lda > 0 && ldb > 0 && ldc > 0 && m > 1 && n > 1 && k > 1)
+      {
+        assert (gemm != NULL);
+        gemm (axstride == 1 ? "N" : "T", bxstride == 1 ? "N" : "T", &m, &n, &k,
+              &one, abase, &lda, bbase, &ldb, &zero, dest, &ldc, 1, 1);
+        return;
+      }
+  }
+
   if (rxstride == 1 && axstride == 1 && bxstride == 1)
     {
       const GFC_REAL_8 * restrict bbase_y;
diff --git a/libgfortran/m4/matmul.m4 b/libgfortran/m4/matmul.m4
index 3678c639f2a..ef2f0fb88dc 100644
--- a/libgfortran/m4/matmul.m4
+++ b/libgfortran/m4/matmul.m4
@@ -37,6 +37,16 @@ include(iparm.m4)dnl
 
 `#if defined (HAVE_'rtype_name`)'
 
+/* Prototype for the BLAS ?gemm subroutine, a pointer to which can be
+   passed to us by the front-end, in which case we'll call it for large
+   matrices.  */
+
+typedef void (*blas_call)(const char *, const char *, const int *, const int *,
+                          const int *, const rtype_name *, const rtype_name *,
+                          const int *, const rtype_name *, const int *,
+                          const rtype_name *, rtype_name *, const int *,
+                          int, int);
+
 /* The order of loops is different in the case of plain matrix
    multiplication C=MATMUL(A,B), and in the frequent special case where
    the argument A is the temporary result of a TRANSPOSE intrinsic:
@@ -57,18 +67,24 @@ include(iparm.m4)dnl
        DO I=1,M
          S = 0
          DO K=1,COUNT
-           S = S+A(I,K)+B(K,J)
+           S = S+A(I,K)*B(K,J)
          C(I,J) = S
    ENDIF
 */
 
+/* If try_blas is set to a nonzero value, then the matmul function will
+   see if there is a way to perform the matrix multiplication by a call
+   to the BLAS gemm function.  */
+
 extern void matmul_`'rtype_code (rtype * const restrict retarray, 
-	rtype * const restrict a, rtype * const restrict b);
+	rtype * const restrict a, rtype * const restrict b, int try_blas,
+	int blas_limit, blas_call gemm);
 export_proto(matmul_`'rtype_code);
 
 void
 matmul_`'rtype_code (rtype * const restrict retarray, 
-	rtype * const restrict a, rtype * const restrict b)
+	rtype * const restrict a, rtype * const restrict b, int try_blas,
+	int blas_limit, blas_call gemm)
 {
   const rtype_name * restrict abase;
   const rtype_name * restrict bbase;
@@ -179,6 +195,31 @@ sinclude(`matmul_asm_'rtype_code`.m4')dnl
   bbase = b->data;
   dest = retarray->data;
 
+
+  /* Now that everything is set up, we're performing the multiplication
+     itself.  */
+
+#define POW3(x) (((float) (x)) * ((float) (x)) * ((float) (x)))
+
+  if (try_blas && rxstride == 1 && (axstride == 1 || aystride == 1)
+      && (bxstride == 1 || bystride == 1)
+      && (((float) xcount) * ((float) ycount) * ((float) count)
+          > POW3(blas_limit)))
+  {
+    const int m = xcount, n = ycount, k = count, ldc = rystride;
+    const rtype_name one = 1, zero = 0;
+    const int lda = (axstride == 1) ? aystride : axstride,
+              ldb = (bxstride == 1) ? bystride : bxstride;
+
+    if (lda > 0 && ldb > 0 && ldc > 0 && m > 1 && n > 1 && k > 1)
+      {
+        assert (gemm != NULL);
+        gemm (axstride == 1 ? "N" : "T", bxstride == 1 ? "N" : "T", &m, &n, &k,
+              &one, abase, &lda, bbase, &ldb, &zero, dest, &ldc, 1, 1);
+        return;
+      }
+  }
+
   if (rxstride == 1 && axstride == 1 && bxstride == 1)
     {
       const rtype_name * restrict bbase_y;