[gcc(refs/users/rguenth/heads/slp-reorg)] store-merging: Fix up -fnon-call-exceptions handling [PR94224]

Richard Biener rguenth@gcc.gnu.org
Mon Mar 23 15:56:44 GMT 2020


https://gcc.gnu.org/g:4119cd693d27e9dd87c547de75283edd45bf6dce

commit 4119cd693d27e9dd87c547de75283edd45bf6dce
Author: Jakub Jelinek <jakub@redhat.com>
Date:   Fri Mar 20 09:33:38 2020 +0100

    store-merging: Fix up -fnon-call-exceptions handling [PR94224]
    
    When we are adding a single store into a store group, we are already
    checking that store->lp_nr matches, but we have also code to add further
    INTEGER_CST stores into the group right away if the ordering requires that
    either we put there all or none from a certain set of stores.  And in those
    cases we weren't doing these lp_nr checks, which means we could end up with
    stores with different lp_nr in the same group, which then ICEs during
    output_merged_store.
    
    2020-03-20  Jakub Jelinek  <jakub@redhat.com>
    
            PR tree-optimization/94224
            * gimple-ssa-store-merging.c
            (imm_store_chain_info::coalesce_immediate): Don't consider overlapping
            or adjacent INTEGER_CST rhs_code stores as mergeable if they have
            different lp_nr.
    
            * g++.dg/tree-ssa/pr94224.C: New test.

Diff:
---
 gcc/ChangeLog                           |  8 ++++++++
 gcc/gimple-ssa-store-merging.c          |  4 +++-
 gcc/testsuite/ChangeLog                 |  5 +++++
 gcc/testsuite/g++.dg/tree-ssa/pr94224.C | 34 +++++++++++++++++++++++++++++++++
 4 files changed, 50 insertions(+), 1 deletion(-)

diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index f7672cc275e..00dd10c7f1f 100644
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1,3 +1,11 @@
+2020-03-20  Jakub Jelinek  <jakub@redhat.com>
+
+	PR tree-optimization/94224
+	* gimple-ssa-store-merging.c
+	(imm_store_chain_info::coalesce_immediate): Don't consider overlapping
+	or adjacent INTEGER_CST rhs_code stores as mergeable if they have
+	different lp_nr.
+
 2020-03-20  Andre Vieira  <andre.simoesdiasvieira@arm.com>
 
 	* config/arm/arm.md (define_attr "conds"): Fix logic for neon and mve.
diff --git a/gcc/gimple-ssa-store-merging.c b/gcc/gimple-ssa-store-merging.c
index 4d4f5494788..83ae6c4c5b1 100644
--- a/gcc/gimple-ssa-store-merging.c
+++ b/gcc/gimple-ssa-store-merging.c
@@ -2773,7 +2773,8 @@ imm_store_chain_info::coalesce_immediate_stores ()
 			    break;
 			  if (info2->order < try_order)
 			    {
-			      if (info2->rhs_code != INTEGER_CST)
+			      if (info2->rhs_code != INTEGER_CST
+				  || info2->lp_nr != merged_store->lp_nr)
 				{
 				  /* Normally check_no_overlap makes sure this
 				     doesn't happen, but if end grows below,
@@ -2791,6 +2792,7 @@ imm_store_chain_info::coalesce_immediate_stores ()
 					      info2->bitpos + info2->bitsize);
 			    }
 			  else if (info2->rhs_code == INTEGER_CST
+				   && info2->lp_nr == merged_store->lp_nr
 				   && !last_iter)
 			    {
 			      max_order = MAX (max_order, info2->order + 1);
diff --git a/gcc/testsuite/ChangeLog b/gcc/testsuite/ChangeLog
index 212c6ffc06e..d11f3fe0171 100644
--- a/gcc/testsuite/ChangeLog
+++ b/gcc/testsuite/ChangeLog
@@ -1,3 +1,8 @@
+2020-03-20  Jakub Jelinek  <jakub@redhat.com>
+
+	PR tree-optimization/94224
+	* g++.dg/tree-ssa/pr94224.C: New test.
+
 2020-03-19  Jan Hubicka  <hubicka@ucw.cz>
 
 	PR ipa/94202
diff --git a/gcc/testsuite/g++.dg/tree-ssa/pr94224.C b/gcc/testsuite/g++.dg/tree-ssa/pr94224.C
new file mode 100644
index 00000000000..5251917bdaf
--- /dev/null
+++ b/gcc/testsuite/g++.dg/tree-ssa/pr94224.C
@@ -0,0 +1,34 @@
+// PR tree-optimization/94224
+// { dg-do compile }
+// { dg-options "-O2 -fnon-call-exceptions -Wno-return-type" }
+
+void foo (int, int, long);
+
+static inline int
+bar (int &x)
+{
+  x = 0;
+}
+
+struct U
+{
+  int n, p;
+  long q;
+  bool *z;
+  int a;
+  U () : n (), z (), a (1) {}
+  ~U () { if (n) foo (p, n, q); }
+  void baz () { bar (a); }
+};
+
+struct W
+{
+  U w[2];
+  W () { w[0].baz (); }
+};
+
+void
+qux ()
+{
+  new W;
+}


More information about the Gcc-cvs mailing list