[PATCH] Fix PR90913

Richard Biener rguenther@suse.de
Fri Jun 21 11:09:00 GMT 2019


The following fixes the case where we're vectorizing the inner loop
of an if-converted scalar version of a outer loop.  In this case
we cannot re-use the if-conversion version because the not vectorized
path would fall into the code intended for outer loop vectorization.

Bootstrapped and tested on x86_64-unknown-linux-gnu, applied.

Richard.

2019-06-21  Richard Biener  <rguenther@suse.de>

	PR tree-optimization/90913
	* tree-vect-loop-manip.c (vect_loop_versioning): Do not re-use
	the scalar variant of if-conversion versioning.

	* gfortran.dg/vect/pr90913.f90: New testcase.

Index: gcc/tree-vect-loop-manip.c
===================================================================
--- gcc/tree-vect-loop-manip.c	(revision 272422)
+++ gcc/tree-vect-loop-manip.c	(working copy)
@@ -3095,9 +3095,12 @@ vect_loop_versioning (loop_vec_info loop
     }
 
   /* Apply versioning.  If there is already a scalar version created by
-     if-conversion re-use that.  */
+     if-conversion re-use that.  Note we cannot re-use the copy of
+     an if-converted outer-loop when vectorizing the inner loop only.  */
   gcond *cond;
-  if (gimple *call = vect_loop_vectorized_call (loop_to_version, &cond))
+  gimple *call;
+  if ((!loop_to_version->inner || loop == loop_to_version)
+      && (call = vect_loop_vectorized_call (loop_to_version, &cond)))
     {
       gcc_assert (scalar_loop);
       condition_bb = gimple_bb (cond);
Index: gcc/testsuite/gfortran.dg/vect/pr90913.f90
===================================================================
--- gcc/testsuite/gfortran.dg/vect/pr90913.f90	(nonexistent)
+++ gcc/testsuite/gfortran.dg/vect/pr90913.f90	(working copy)
@@ -0,0 +1,22 @@
+! { dg-do compile }
+! { dg-options "-O3 -ffast-math" }
+! { dg-additional-options "-mavx -mveclibabi=svml" { target i?86-*-* x86_64-*-* } }
+subroutine foo (a, b, c, d, e, f, g, h, k, l)
+  implicit none
+  integer :: d, e, f, g, i, j
+  real :: a, b(5,6), c(6), h(6,10,5), k(5,10,2), l(10,5), m, n, o
+  do i=1,5
+    do j=1,6
+      m=l(f,g)*log(c(j))
+      if (m<2) then
+        if (m<-2) then
+          h(j,f,g)=n
+        else
+          h(j,f,g)=o
+        endif
+      endif
+      b(i,j)=a+k(i,d,e)+k(i,1,e)**h(j,f,g)
+    enddo
+  enddo
+  write(*,'()') 
+end



More information about the Gcc-patches mailing list