This PR is a nasty wrong code bug due to my fluffing a test for a
backedge in gimple-ssa-backprop.c. Backedges are supposed to be
from definitions in the statement we're visiting to uses in statements
that we haven't visited yet. However, the check failed to account for
PHIs in the current block that had already been processed, so if two
PHIs in the same block referenced each other, we'd treat both
references as backedges.
In more detail:
The first phase of the pass goes through all statements in an arbitrary
order, making optimistic assumptions about any statements that haven't
been visited yet. The second phase then calculates a correct
(supposedly maximal) fixed point.
Although the first phase order is arbitrary in principle, we still use
the CFG rpo to cut down on the backedges. This means that the only
thing that's truly arbitrary is the order that we process the PHIs
in a block. Any order should be OK and should eventually give the
same results. But we have to follow whatever order we pick,
and the pass wasn't doing that.
Tested on aarch64-linux-gnu (with and without SVE), aarch64_be-elf
and x86_64-linux-gnu. OK for trunk and all release branches?
Sorry for what in hindsight was a really stupid bug.
Richard
2018-06-01 Richard Sandiford <richard.sandif...@linaro.org>
gcc/
PR tree-optimization/85989
* gimple-ssa-backprop.c (backprop::m_visited_phis): New member
variable.
(backprop::intersect_uses): Check it when deciding whether this
is a backedge reference.
(backprop::process_block): Add each phi to m_visited_phis
after visiting it, then clear it at the end.
gcc/testsuite/
PR tree-optimization/85989
* gcc.dg/torture/pr85989.c: New test.
Index: gcc/gimple-ssa-backprop.c
===================================================================
--- gcc/gimple-ssa-backprop.c 2018-05-18 09:26:37.726714678 +0100
+++ gcc/gimple-ssa-backprop.c 2018-06-01 13:36:04.223449400 +0100
@@ -260,6 +260,11 @@ dump_usage_info (FILE *file, tree var, u
post-order walk. */
auto_sbitmap m_visited_blocks;
+ /* A bitmap of phis that we have finished processing in the initial
+ post-order walk, excluding those from blocks mentioned in
+ M_VISITED_BLOCKS. */
+ auto_bitmap m_visited_phis;
+
/* A worklist of SSA names whose definitions need to be reconsidered. */
auto_vec <tree, 64> m_worklist;
@@ -496,8 +501,11 @@ backprop::intersect_uses (tree var, usag
{
if (is_gimple_debug (stmt))
continue;
- if (is_a <gphi *> (stmt)
- && !bitmap_bit_p (m_visited_blocks, gimple_bb (stmt)->index))
+ gphi *phi = dyn_cast <gphi *> (stmt);
+ if (phi
+ && !bitmap_bit_p (m_visited_blocks, gimple_bb (phi)->index)
+ && !bitmap_bit_p (m_visited_phis,
+ SSA_NAME_VERSION (gimple_phi_result (phi))))
{
/* Skip unprocessed phis. */
if (dump_file && (dump_flags & TDF_DETAILS))
@@ -505,7 +513,7 @@ backprop::intersect_uses (tree var, usag
fprintf (dump_file, "[BACKEDGE] ");
print_generic_expr (dump_file, var);
fprintf (dump_file, " in ");
- print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
+ print_gimple_stmt (dump_file, phi, 0, TDF_SLIM);
}
}
else
@@ -629,7 +637,12 @@ backprop::process_block (basic_block bb)
}
for (gphi_iterator gpi = gsi_start_phis (bb); !gsi_end_p (gpi);
gsi_next (&gpi))
- process_var (gimple_phi_result (gpi.phi ()));
+ {
+ tree result = gimple_phi_result (gpi.phi ());
+ process_var (result);
+ bitmap_set_bit (m_visited_phis, SSA_NAME_VERSION (result));
+ }
+ bitmap_clear (m_visited_phis);
}
/* Delete the definition of VAR, which has no uses. */
Index: gcc/testsuite/gcc.dg/torture/pr85989.c
===================================================================
--- /dev/null 2018-04-20 16:19:46.369131350 +0100
+++ gcc/testsuite/gcc.dg/torture/pr85989.c 2018-06-01 13:36:04.223449400
+0100
@@ -0,0 +1,31 @@
+/* { dg-do run } */
+
+#define N 9
+
+void __attribute__((noipa))
+f (double x, double y, double *res)
+{
+ y = -y;
+ for (int i = 0; i < N; ++i)
+ {
+ double tmp = y;
+ y = x;
+ x = tmp;
+ res[i] = i;
+ }
+ res[N] = y * y;
+ res[N + 1] = x;
+}
+
+int
+main (void)
+{
+ double res[N + 2];
+ f (10, 20, res);
+ for (int i = 0; i < N; ++i)
+ if (res[i] != i)
+ __builtin_abort ();
+ if (res[N] != 100 || res[N + 1] != -20)
+ __builtin_abort ();
+ return 0;
+}
