On 08/03/16 17:38, Jeff Law wrote:
> cse.c changes look good, but I'd really like to see a testcase for each
> issue in the dejagnu framework. Extra points if you tried to build a
> unit test using David M's framework, but that isn't required.
>
> The testcase from 70903 ought to be trivial to add to the dejagnu suite.
> 71779 might be more difficult, but if you could take a stab, it'd be
> appreciated.
>
Yes, sure. I had assumed that the pr70903 test case is using some
target-specific vector types, but now I see that it even works as-is in
the gcc.c-torture/execute directory.
So I've added the test case to the cse patch. And quickly verified that
it works on x86_64-linux-gnu.
The pr71779 test case will be pretty difficult to reduce, because it
depends on combine to do the incorrect transformation and lra to spill
the subreg, and on the stack content at runtime to be non-zero.
But technically it *is* already in the isl-test suite, so if isl is
in-tree, it is always executed by make check or make check-isl.
It is just that gmp/mpfr/mpc and isl test results are not included by
contrib/test_summary, but that should be fixable. What do you think?
Actually that should not be too difficult, as there are test-suite.log
files that we could just added to the test_summary output as-is, for
instance:
cat isl/test-suite.log
==================================
isl 0.16.1: ./test-suite.log
==================================
# TOTAL: 5
# PASS: 5
# SKIP: 0
# XFAIL: 0
# FAIL: 0
# XPASS: 0
# ERROR: 0
.. contents:: :depth: 2
Are the patches OK now?
Thanks
Bernd.
2016-08-01 Bernd Edlinger <bernd.edlin...@hotmail.de>
PR rtl-optimization/70903
* cse.c (cse_insn): If DEST is a paradoxical SUBREG, don't record DEST.
testsuite:
2016-08-01 Bernd Edlinger <bernd.edlin...@hotmail.de>
PR rtl-optimization/70903
* gcc.c-torture/execute/pr70903.c: New test.
Index: gcc/cse.c
===================================================================
--- gcc/cse.c (revision 238915)
+++ gcc/cse.c (working copy)
@@ -5898,15 +5898,7 @@ cse_insn (rtx_insn *insn)
|| GET_MODE (dest) == BLKmode
/* If we didn't put a REG_EQUAL value or a source into the hash
table, there is no point is recording DEST. */
- || sets[i].src_elt == 0
- /* If DEST is a paradoxical SUBREG and SRC is a ZERO_EXTEND
- or SIGN_EXTEND, don't record DEST since it can cause
- some tracking to be wrong.
-
- ??? Think about this more later. */
- || (paradoxical_subreg_p (dest)
- && (GET_CODE (sets[i].src) == SIGN_EXTEND
- || GET_CODE (sets[i].src) == ZERO_EXTEND)))
+ || sets[i].src_elt == 0)
continue;
/* STRICT_LOW_PART isn't part of the value BEING set,
@@ -5925,6 +5917,11 @@ cse_insn (rtx_insn *insn)
sets[i].dest_hash = HASH (dest, GET_MODE (dest));
}
+ /* If DEST is a paradoxical SUBREG, don't record DEST since the bits
+ outside the mode of GET_MODE (SUBREG_REG (dest)) are undefined. */
+ if (paradoxical_subreg_p (dest))
+ continue;
+
elt = insert (dest, sets[i].src_elt,
sets[i].dest_hash, GET_MODE (dest));
Index: gcc/testsuite/gcc.c-torture/execute/pr70903.c
===================================================================
--- gcc/testsuite/gcc.c-torture/execute/pr70903.c (revision 0)
+++ gcc/testsuite/gcc.c-torture/execute/pr70903.c (working copy)
@@ -0,0 +1,19 @@
+typedef unsigned char V8 __attribute__ ((vector_size (32)));
+typedef unsigned int V32 __attribute__ ((vector_size (32)));
+typedef unsigned long long V64 __attribute__ ((vector_size (32)));
+
+static V32 __attribute__ ((noinline, noclone))
+foo (V64 x)
+{
+ V64 y = (V64)(V8){((V8)(V64){65535, x[0]})[1]};
+ return (V32){y[0], 255};
+}
+
+int main ()
+{
+ V32 x = foo ((V64){});
+// __builtin_printf ("%08x %08x %08x %08x %08x %08x %08x %08x\n", x[0], x[1], x[2], x[3], x[4], x[5], x[6], x[7]);
+ if (x[1] != 255)
+ __builtin_abort();
+ return 0;
+}
2016-08-01 Bernd Edlinger <bernd.edlin...@hotmail.de>
PR rtl-optimization/71779
* emit-rtl.c (set_reg_attrs_from_value): Only propagate REG_POINTER,
if the value was sign-extended according to POINTERS_EXTEND_UNSIGNED
or if it was truncated.
Index: gcc/emit-rtl.c
===================================================================
--- gcc/emit-rtl.c (revision 238915)
+++ gcc/emit-rtl.c (working copy)
@@ -1156,7 +1156,11 @@ set_reg_attrs_from_value (rtx reg, rtx x)
{
#if defined(POINTERS_EXTEND_UNSIGNED)
if (((GET_CODE (x) == SIGN_EXTEND && POINTERS_EXTEND_UNSIGNED)
- || (GET_CODE (x) != SIGN_EXTEND && ! POINTERS_EXTEND_UNSIGNED))
+ || (GET_CODE (x) == ZERO_EXTEND && ! POINTERS_EXTEND_UNSIGNED)
+ || (paradoxical_subreg_p (x)
+ && ! (SUBREG_PROMOTED_VAR_P (x)
+ && SUBREG_CHECK_PROMOTED_SIGN (x,
+ POINTERS_EXTEND_UNSIGNED))))
&& !targetm.have_ptr_extend ())
can_be_reg_pointer = false;
#endif
