On 10/12/2016 09:00 AM, Georg-Johann Lay wrote:
On 10.10.2016 23:06, Jeff Law wrote:
So if we have an equality conditional between A & B, we record into our
const/copy tables A = B and B = A.
This helps us discover some of the more obscure equivalences. But it also
creates problems with an expression like
A ^ B
Where we might cprop the first operand generating
B ^ B
Then the second generating
B ^ A
ANd we've lost the folding opportunity. At first I'd tried folding
after each
propagation step, but that turns into a bit of a nightmare because of
changes
in the underlying structure of the gimple statement and cycles that
may develop
if we re-build the operand cache after folding.
This approach is simpler and should catch all these cases for binary
operators. We just track the last copy propagated argument and refuse to
ping-pong propagations.
It fixes the tests from 71947 and 77647 without regressing
(obviously). I've
included an xfailed test for a more complex situation that we don't
currently
handle (would require backtracking from the equality comparison
through the
logicals that feed the equality comparison).
Bootstrapped and regression tested on x86_64. Applied to the trunk.
commit 6223e6e425b6de916f0330b9dbe5698765d4a73c
Author: law <law@138bc75d-0d04-0410-961f-82ee72b054a4>
Date: Mon Oct 10 20:40:59 2016 +0000
PR tree-optimization/71947
* tree-ssa-dom.c (cprop_into_stmt): Avoid replacing A with B,
then
B with A within a single statement.
PR tree-optimization/71947
* gcc.dg/tree-ssa/pr71947-1.c: New test.
* gcc.dg/tree-ssa/pr71947-2.c: New test.
* gcc.dg/tree-ssa/pr71947-3.c: New test.
* gcc.dg/tree-ssa/pr71947-4.c: New test.
* gcc.dg/tree-ssa/pr71947-5.c: New test.
* gcc.dg/tree-ssa/pr71947-6.c: New test.
git-svn-id: svn+ssh://gcc.gnu.org/svn/gcc/trunk@240947
138bc75d-0d04-0410-961f-82ee72b054a4
diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index 1738bc7..16e25bf 100644
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1,3 +1,9 @@
+2016-10-10 Jeff Law <l...@redhat.com>
+
+ PR tree-optimization/71947
+ * tree-ssa-dom.c (cprop_into_stmt): Avoid replacing A with B, then
+ B with A within a single statement.
+
2016-10-10 Bill Schmidt <wschm...@linux.vnet.ibm.com>
PR tree-optimization/77824
diff --git a/gcc/testsuite/ChangeLog b/gcc/testsuite/ChangeLog
index 04966cf..e31bcc6 100644
--- a/gcc/testsuite/ChangeLog
+++ b/gcc/testsuite/ChangeLog
@@ -1,3 +1,13 @@
+2016-10-10 Jeff Law <l...@redhat.com>
+
+ PR tree-optimization/71947
+ * gcc.dg/tree-ssa/pr71947-1.c: New test.
+ * gcc.dg/tree-ssa/pr71947-2.c: New test.
+ * gcc.dg/tree-ssa/pr71947-3.c: New test.
+ * gcc.dg/tree-ssa/pr71947-4.c: New test.
+ * gcc.dg/tree-ssa/pr71947-5.c: New test.
+ * gcc.dg/tree-ssa/pr71947-6.c: New test.
+
2016-10-10 Thomas Koenig <tkoe...@gcc.gnu.org>
PR fortran/77915
diff --git a/gcc/testsuite/gcc.dg/tree-ssa/pr71947-1.c
b/gcc/testsuite/gcc.dg/tree-ssa/pr71947-1.c
new file mode 100644
index 0000000..b033495
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/tree-ssa/pr71947-1.c
@@ -0,0 +1,19 @@
+/* { dg-do compile } */
+/* { dg-options "-O2 -fno-tree-vrp -fdump-tree-dom-details" } */
+
+
+int f(int x, int y)
+{
+ int ret;
+
+ if (x == y)
+ ret = x ^ y;
+ else
+ ret = 1;
+
+ return ret;
+}
+
+/* { dg-final { scan-tree-dump "Folded to: ret_\[0-9\]+ = 0;" "dom2"
} } */
+
+
diff --git a/gcc/testsuite/gcc.dg/tree-ssa/pr71947-2.c
b/gcc/testsuite/gcc.dg/tree-ssa/pr71947-2.c
new file mode 100644
index 0000000..de8f88b
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/tree-ssa/pr71947-2.c
@@ -0,0 +1,16 @@
+/* { dg-do compile } */
+/* { dg-options "-O2 -fno-tree-vrp -fdump-tree-dom-details" } */
+
+
+int f(int x, int y)
+{
+ int ret;
+ if (x == y)
+ ret = x - y;
+ else
+ ret = 1;
+
+ return ret;
+}
+
+/* { dg-final { scan-tree-dump "Folded to: ret_\[0-9\]+ = 0;" "dom2"
} } */
diff --git a/gcc/testsuite/gcc.dg/tree-ssa/pr71947-3.c
b/gcc/testsuite/gcc.dg/tree-ssa/pr71947-3.c
new file mode 100644
index 0000000..e79847f
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/tree-ssa/pr71947-3.c
@@ -0,0 +1,12 @@
+/* { dg-do compile } */
+/* { dg-options "-O2 -fno-tree-vrp -fdump-tree-dom-details" } */
+
+int f(int x, int y)
+{
+ int ret = 10;
+ if (x == y)
+ ret = x - y;
+ return ret;
+}
+
+/* { dg-final { scan-tree-dump "Folded to: ret_\[0-9\]+ = 0;" "dom2"
} } */
diff --git a/gcc/testsuite/gcc.dg/tree-ssa/pr71947-4.c
b/gcc/testsuite/gcc.dg/tree-ssa/pr71947-4.c
new file mode 100644
index 0000000..a881f0d
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/tree-ssa/pr71947-4.c
@@ -0,0 +1,22 @@
+/* { dg-do compile } */
+/* { dg-options "-O2 -fno-tree-vrp -fdump-tree-dom-details" } */
+
+
+
+static inline long load(long *p)
+{
+ long ret;
+ asm ("movq %1,%0\n\t" : "=r" (ret) : "m" (*p));
Hi, shouldn't this test be restricted to target machines that understand
MOVQ?
Same issue for gcc.dg/tree-ssa/pr71947-5.c below.
It shouldn't actually matter -- we compile to assembly, but never try to
assemble the resulting code. One could replace the movq with anything.
The test actually verifies simplification of the return statement in
"foo". In one test it should be 2*ret and the other 0. In neither case
should it load *mem again.
jeff
