===================================================================
@@ -1630,7 +1630,7 @@ square, the size comparison is performed using the
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
+@code{-fblas-matmul-limit} if this option is specified, or 30
otherwise.
@item -frecursive
===================================================================
@@ -388,10 +388,16 @@ 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 the user did not specify an inline matmul limit, inline up to the BLAS
+ limit or up to 30 if no external BLAS is specified. */
- if (flag_external_blas && flag_inline_matmul_limit < 0)
- flag_inline_matmul_limit = flag_blas_matmul_limit;
+ if (flag_inline_matmul_limit < 0)
+ {
+ if (flag_external_blas)
+ flag_inline_matmul_limit = flag_blas_matmul_limit;
+ else
+ flag_inline_matmul_limit = 30;
+ }
/* Optimization implies front end optimization, unless the user
specified it directly. */
===================================================================
@@ -1,5 +1,5 @@
! { dg-do run }
-! { dg-options "-ffrontend-optimize -fdump-tree-original -Wrealloc-lhs" }
+! { dg-options "-ffrontend-optimize -fdump-tree-optimized -Wrealloc-lhs" }
! PR 37131 - check basic functionality of inlined matmul, making
! sure that the library is not called, with and without reallocation.
@@ -149,4 +149,4 @@ program main
end program main
-! { dg-final { scan-tree-dump-times "_gfortran_matmul" 0 "original" } }
+! { dg-final { scan-tree-dump-times "_gfortran_matmul" 0 "optimized" } }
===================================================================
@@ -1,5 +1,5 @@
! { dg-do run }
-! { dg-additional-options "-ffrontend-optimize -fdump-tree-original" }
+! { dg-additional-options "-ffrontend-optimize -fdump-tree-optimized" }
! PR fortran/66176 - inline conjg for matml.
program main
complex, dimension(3,2) :: a
@@ -29,4 +29,4 @@ program main
c = matmul(conjg(a), b)
if (any(conjg(c) /= res2)) call abort
end program main
-! { dg-final { scan-tree-dump-times "_gfortran_matmul" 0 "original" } }
+! { dg-final { scan-tree-dump-times "_gfortran_matmul" 0 "optimized" } }
===================================================================
@@ -1,5 +1,5 @@
! { dg-do run }
-! { dg-options "-ffrontend-optimize -fdump-tree-original" }
+! { dg-options "-ffrontend-optimize -fdump-tree-optimized" }
! PR 66041 - this used to ICE with an incomplete fix for the PR.
program main
implicit none
@@ -21,4 +21,4 @@ program main
if (any (c2-reshape([248., -749.],shape(c2)) /= 0.)) call abort
end program main
-! { dg-final { scan-tree-dump-times "_gfortran_matmul" 0 "original" } }
+! { dg-final { scan-tree-dump-times "_gfortran_matmul" 0 "optimized" } }
Hello world, there still was one piece missing for the new matmul library version. To make sure that users (usually) benefit, we need to call the library by default up from a certain limit. The attached patch does that, with a limit of 30, which seems to be reasonable given a few benchmarks. Some test cases had to be changed to scan the optimized tree instead of the original because the version still had some if (0) statement in them. Regeression-tested. OK for trunk? Regards Thomas 2017-02-25 Thomas Koenig <tkoenig@gcc.gnu.org> PR fortran/51119 * options.c (gfc_post_options): Set default limit for matmul inlining to 30. * invoke.texi: Document change. 2017-02-25 Thomas Koenig <tkoenig@gcc.gnu.org> PR fortran/51119 * gfortran.dg/inline_matmul_1.f90: Scan optimized dump instead of original. * gfortran.dg/inline_matmul_11.f90: Likewise. * gfortran.dg/inline_matmul_9.f90: Likewise. * gfortran.dg/matmul_13.f90: New test. * gfortran.dg/matmul_14.f90: New test.