It turns out RTL CSE tries to remove redundant stores but fails to
do the usual validity check what such a change is TBAA neutral to
later loads.
This now triggers with the PR93946 testcases on nios2.
Bootstrap & regtest running on x86_64-unknown-linux-gnu.
Richard.
2020-04-08 Richard Biener <rguent...@suse.de>
PR rtl-optimization/93946
* cse.c (cse_insn): Record the tabled expression in
src_related. Verify a redundant store removal is valid.
---
gcc/cse.c | 13 +++++++++++--
1 file changed, 11 insertions(+), 2 deletions(-)
diff --git a/gcc/cse.c b/gcc/cse.c
index 3e8724b3fed..f07bbdbebad 100644
--- a/gcc/cse.c
+++ b/gcc/cse.c
@@ -5074,7 +5074,7 @@ cse_insn (rtx_insn *insn)
to prefer it. Copy it to src_related. The code below will
then give it a negative cost. */
if (GET_CODE (dest) == code && rtx_equal_p (p->exp, dest))
- src_related = dest;
+ src_related = p->exp;
}
/* Find the cheapest valid equivalent, trying all the available
@@ -5332,7 +5332,16 @@ cse_insn (rtx_insn *insn)
&& rtx_equal_p (trial, dest)
&& !side_effects_p (dest)
&& (cfun->can_delete_dead_exceptions
- || insn_nothrow_p (insn)))
+ || insn_nothrow_p (insn))
+ /* We can only remove the later store if the earlier aliases
+ at least all accesses the later one. */
+ && (!MEM_P (trial)
+ || ((MEM_ALIAS_SET (dest) == MEM_ALIAS_SET (trial)
+ || alias_set_subset_of (MEM_ALIAS_SET (dest),
+ MEM_ALIAS_SET (trial)))
+ && (!MEM_EXPR (trial)
+ || refs_same_for_tbaa_p (MEM_EXPR (trial),
+ MEM_EXPR (dest))))))
{
SET_SRC (sets[i].rtl) = trial;
noop_insn = true;
--
2.16.4
