It turns out solving this long-standing optimization regression
is now easy by exploiting implmenetation details in how we
canonicalize refs in LIM. This allows us to properly identifying
MEM[(integer(kind=4)[64] *)&a][0] and MEM[(c_char * {ref-all})&a]
the same, applying store-motion to an initialization (non-)loop
thereby eliminating it.
Bootstrap & regtest running on x86_64-unknown-linux-gnu.
I didn't go further trying to exploit alias subset relationship
instead but the alias-set zero case is obvious enough to be
correct.
Richard.
2019-04-15 Richard Biener <rguent...@suse.de>
PR tree-optimization/56049
* tree-ssa-loop-im.c (mem_ref_hasher::equal): Elide alias-set
equality check if alias-set zero will prevail.
* gfortran.dg/pr56049.f90: New testcase.
Index: gcc/tree-ssa-loop-im.c
===================================================================
--- gcc/tree-ssa-loop-im.c (revision 270366)
+++ gcc/tree-ssa-loop-im.c (working copy)
@@ -178,7 +178,17 @@ mem_ref_hasher::equal (const im_mem_ref
&& known_eq (mem1->mem.size, obj2->size)
&& known_eq (mem1->mem.max_size, obj2->max_size)
&& mem1->mem.volatile_p == obj2->volatile_p
- && mem1->mem.ref_alias_set == obj2->ref_alias_set
+ && (mem1->mem.ref_alias_set == obj2->ref_alias_set
+ /* We are not canonicalizing alias-sets but for the
+ special-case we didn't canonicalize yet and the
+ incoming ref is a alias-set zero MEM we pick
+ the correct one already. */
+ || (!mem1->ref_canonical
+ && (TREE_CODE (obj2->ref) == MEM_REF
+ || TREE_CODE (obj2->ref) == TARGET_MEM_REF)
+ && obj2->ref_alias_set == 0)
+ /* Likewise if there's a canonical ref with alias-set zero. */
+ || (mem1->ref_canonical && mem1->mem.ref_alias_set == 0))
&& types_compatible_p (TREE_TYPE (mem1->mem.ref),
TREE_TYPE (obj2->ref)));
else
Index: gcc/testsuite/gfortran.dg/pr56049.f90
===================================================================
--- gcc/testsuite/gfortran.dg/pr56049.f90 (nonexistent)
+++ gcc/testsuite/gfortran.dg/pr56049.f90 (working copy)
@@ -0,0 +1,29 @@
+! { dg-do compile }
+! { dg-options "-O3 -fdump-tree-optimized" }
+
+program inline
+
+ integer i
+ integer a(8,8), b(8,8)
+
+ a = 0
+ do i = 1, 10000000
+ call add(b, a, 1)
+ a = b
+ end do
+
+ print *, a
+
+contains
+
+ subroutine add(b, a, o)
+ integer, intent(inout) :: b(8,8)
+ integer, intent(in) :: a(8,8), o
+ b = a + o
+ end subroutine add
+
+end program inline
+
+! Check there's no loop left, just two bb 2 in two functions.
+! { dg-final { scan-tree-dump-times "<bb \[0-9\]*>" 2 "optimized" } }
+! { dg-final { scan-tree-dump-times "<bb 2>" 2 "optimized" } }
